The most valuable dataset in healthcare is the one no one can query.

For: Real-world evidence teams, pharma clinical operations, health data scientists, and anyone building regulatory submissions from observational data.

• Savana transforms unstructured clinical text into validated, structured databases using clinical NLP, operating across 300+ sites in 14 countries.

• Regulators will not accept outputs from large language models. Savana’s discriminative AI provides the reproducibility that submissions demand.

• We spoke with founder Ignacio Medrano about why free-text clinical data remains the most underleveraged layer in healthcare.

Kiin Pioneer Programme

We built a platform that helps researchers speed up their entire science, from literature review and biomarker discovery to bioinformatics and computational chemistry. If your workflow involves pulling findings from five different places before you can actually act on any of them, this is for that.

The Pioneer Programme gives academic labs and non-profits one year of free access, plus support from our science team. No cost, no data transfer, all IP stays with your institution. Applications close August, cohort starts September.

Read more about the programme

Apply now

This week we spoke with Ignacio Medrano, neurologist-turned-CEO and founder of Savana.

The problem: hospitals produce millions of words every day that never become data

Every hospital generates enormous volumes of clinical text. Radiology reports, discharge summaries, histopathology findings, physician notes and inside those documents sits rich information about what happens to patients and what doctors are thinking. Yet none of it is structured, queryable, or feeding into models.

Real-world evidence has always mattered for regulatory submissions. The bottleneck was never demand, it was collection: patient by patient, variable by variable, building registries manually for years. A single retrospective study might take 18 months of chart review before producing its first result.

“There’s no reason for humans to collect data manually anymore,” Ignacio Medrano told us. “We created a new way of doing clinical research where humans don’t have to collect data. They can dedicate their time to more interesting things, like thinking or analysing.”

The timing matters here. The European Health Data Space regulation is pushing hospitals toward structured data sharing. Simultaneously, AI scribes mean even more unstructured text will be generated in the coming years. The raw material is growing faster than anyone’s ability to manually process it.

The approach: validated NLP across 300 sites and six languages

Savana operates through its Smart Health Alliance, a network of over 300 sites across 14 countries. The platform has three layers: a hospital-facing suite for structuring local data, an interoperability layer for cross-border sharing and governance, and Next Generation Registries where pharma companies access continuously updated, AI-generated databases across multiple sites and countries.

The registries are the sharpest departure from traditional study design. Unlike static observational studies, they stream. Researchers can add new variables a year into a study and the system retrieves them retrospectively across five years of raw data, then continues capturing them prospectively. You no longer need to define every variable at protocol stage.

Amazingly, the entire system operates multilingually: Spanish, English, German, French, Portuguese, and Italian.

Why it’s different: validation that regulators will actually accept

Many NLP tools can extract information from text. The question is whether a regulator will trust the output.

“Most linguistic approaches don’t offer scientific robustness,” Ignacio Medrano explained. “We have a validation methodology that guarantees that when we transform a fragment of text into a variable, it is accurate and reliable. If we do it several times, we get the same result.”

This distinction matters more now than it did two years ago. The FDA and EMA do not accept outputs from large language models as evidence. LLMs are stochastic: the same input can produce different outputs. Savana’s discriminative AI approach provides reproducibility and auditability. So principal investigators at participating sites also review extractions through a guided interface, keeping quality anchored to clinical expertise.

For pharma teams evaluating real-world evidence platforms, this is the core differentiator. Plenty of vendors can structure text. The question is whether that output can appear in a regulatory submission without manual validation layered on top.

Real-world impact

Three weeks after COVID patients began being treated, Savana built a complete database across a European region of 2 million people, with all variables analysed and a paper submitted. Traditional registries would take 12-18 months to reach that point. The analysis identified that patients with mild symptoms would become severe or die three weeks later, enabling primary care to intervene early.

In oncology, working with Pfizer, Savana built a thrombosis prediction model for anticoagulation decisions in patients with solid tumours. It was validated against the Spanish oncology society’s database and is now referenced in their clinical guidelines. That progression, from unstructured notes to clinical guideline, is the full value chain.

Published collaborations span BMS, Johnson & Johnson, and Gilead across use cases from market access to external control arms.

The future

The more interesting direction is multimodal integration. Clinical text captures what happens to patients, but genomics, proteomics, and radiomics each add layers that text alone cannot provide.

“Doctors are inherently multimodal,” Ignacio Medrano said. “When you get into the practice of a doctor, they work with your data across all these layers. AI needs to be multimodal too.”

As AI scribes become ubiquitous, Savana’s role grows rather than shrinks. Scribes produce unstructured output. Transforming that into reliable databases still requires validated, discriminative AI. The generative layer creates the text. The validated extraction layer makes it usable as evidence. Different problems, different architectures.

Kiin’s view

Savana occupies a pretty defensible position. The regulatory gap between “we extracted this with GPT-4” and “we extracted this with a validated, reproducible pipeline” is not closing any time soon. The moat is not the NLP itself, it’s the validation methodology, the 300-site network, and the multilingual coverage that took a decade to build. Competitors can build extractors but replicating the clinical validation infrastructure across 14 countries is a different proposition entirely.

Ignacio on LinkedIn

Savana Website

Savana on LinkedIn

Thanks for reading Kiin Bio Weekly!

Share Kiin Bio Weekly

💬 Get involved

We’re always looking to grow our community. If you’d like to get involved, contribute ideas or share something you’re building, fill out this form or reach out to me directly.

Subscribe now to stay at the forefront of AI in Life Science and keep up with this upcoming season of deep dives.

Connect With Us

Have questions on this or suggestions for our next deep dive? We’d love to hear from you!

📧 Email Us | 📲 Follow on LinkedIn | 🌐 Visit Our Website

• Pan-human Azimuth is what happens when you take Azimuth and make it work across the entire human body at once. 27 million cells, 23 tissues, one unified hierarchy. From Satija Lab at the New York Genome Center.

• AAMFM builds on ESM3 to design antibody CDR loops conditioned on the actual antigen structure. The key move is using direct preference optimisation with structural priors to steer designs toward binding.

• The Meijers and Marks labs built a minimal synthetic antibody library from scratch, screened it against ten cell surface targets, then used logistic regression to rescue binders that the experimental selection missed. The dataset is public and ML-ready.

Kiin Pioneer Programme

We built a platform that helps researchers speed up their entire science, from literature review and biomarker discovery to bioinformatics and computational chemistry. If your workflow involves pulling findings from five different places before you can actually act on any of them, this is for that.

The Pioneer Programme gives academic labs and non-profits one year of free access, plus support from our science team. No cost, no data transfer, all IP stays with your institution. Applications close August, cohort starts September.

Read more about the programme

Apply now

Pan-human Azimuth: Organism-scale annotation with a unified cell type hierarchy

Where This Fits

If you work with single-cell data you have probably used Azimuth at some point, or one of the tissue-specific references it relies on. The problem is that those references are built per-organ. A lung reference does not talk to a kidney reference. The labels are inconsistent across tissues, the hierarchies are different, and if you want to compare immune populations across organs you are stuck reconciling vocabularies by hand. scGPT and SCimilarity attempt organism-wide annotation through foundation models, but recent benchmarks show they produce noisier labels with fragmented cell type assignments. Pan-human Azimuth takes the supervised reference-mapping approach and scales it to the whole body.

What It Is

Sourav Sarkar, Zhuoyan Li, Rahul Satija, and colleagues at the New York Genome Center and collaborating institutions built a supervised neural network classifier trained on 27 million cells from 23 tissues. The training data was assembled from Azimuth references, DISCO, CELLxGENE, GTEx, and HuBMAP, then systematically re-annotated onto a single hierarchical cell type tree built on the DISCO typology with Cell Ontology mappings. Every cell gets up to eight levels of hierarchical annotation with calibrated confidence scores.

Applied to the full Tabula Sapiens v2 atlas (1.1 million cells, 28 tissues), it returned a median confidence score of 0.95 at the finest level. It annotated 85.9 million cells from scBaseCamp in 13.5 hours on a single A100. It also extends to spatial transcriptomics: on Visium HD kidney cortex data it recovered canonical cortical organisation and distinguished healthy from sclerosed glomeruli in agreement with expert pathology.

Why This Is Cool

The interesting finding is the fibroblast work. Because Pan-human Azimuth applies a shared fibroblast hierarchy across all tissues, they could show for the first time that fibroblast state composition reproducibly encodes tissue of origin. Tissue-specialised fibroblast states like CCL11+ in the GI tract and G0S2+PPP1R14A+ in lung are highly specific, and these patterns replicate across independent datasets. That kind of cross-tissue comparison was simply impossible when each organ had its own annotation vocabulary. Available as panhumanpy (Python) and through cloud, R, and Python interfaces at satijalab.org/pan_human_azimuth.

Read the paper.

Try the code.

AAMFM: An antigen-specific antibody foundation model for functional antibody design

Where This Fits

Computational antibody design has been moving toward foundation models. Tools like DiffAb and dyMEAN generate CDR loops given a target structure, and general protein language models like ESM-2 provide useful representations, but they do not condition on the antigen at the point of generation. The gap is that antibody design is inherently a two-body problem: you need to model the antibody-antigen interface jointly, not just produce plausible antibody sequences in isolation. AAMFM is a multimodal foundation model that attempts to close this gap by conditioning CDR design directly on antigen structure and epitope context.

What It Is

Xiaoliang Shi, Zichen Wang, Runze Ma, Zhongyue Zhang, and Shuangjia Zheng built AAMFM as a three-stage system on top of ESM3. First, continual pretraining on antibody structures. Second, supervised fine-tuning for CDR design conditioned on the antigen via a GearNet adapter that injects antigen geometric features. Third, alignment with Calibrated Direct Preference Optimisation (Cal-DPO) using structural priors as preference signals so that the model learns to favour designs with better predicted binding geometry.

The evaluation uses RMSD, amino acid recovery, paratope hit rate, and pseudo log-likelihood (via AntiBERTy). The authors report state-of-the-art performance on functional antibody design benchmarks, though the preprint does not include experimental validation of designed sequences.

Why This Is Cool

The Cal-DPO stage is the methodologically interesting part. DPO has been used extensively in language model alignment but applying it to protein design with structural priors as the preference signal is relatively new. The idea is that instead of just maximising sequence recovery against known antibodies, you can rank candidate designs by their predicted structural compatibility with the target and use those rankings as training signal. Whether this translates to better binders in practice is still unproven since there is no wet-lab validation here. Worth following if they publish binding data.

Read the paper.

Try the code.

High-throughput machine learning-aided antibody discovery for cell surface antigens

Where This Fits

ML-based antibody design gets a lot of attention, but most computational approaches still depend on training data that does not exist for many targets. You need paired antibody-antigen binding data at scale, and for most antigens that data is sparse or proprietary. This paper from the Meijers and Marks labs takes a different angle: instead of training better generative models on limited data, build a synthetic library that is explicitly designed to be ML-compatible from the start, screen it experimentally against many targets at once, and release the whole dataset publicly.

What It Is

Deepash Kothiwal, Aaron Kollasch, Timothy Springer, Debora Marks, Rob Meijers, and a large team across the Institute for Protein Innovation, Harvard, and the Broad Institute built a minimal synthetic Fab yeast display library. The library encodes diversity in the CDRH3 loop within a compact antigen recognition module (ARM) of fewer than 100 nucleotides, using position-specific amino acid frequencies derived from 9.5 million CDRH3 sequences in the Observed Antibody Space database. Estimated diversity: roughly one billion unique clones.

They screened this against ten cell surface glycoproteins in parallel, including PD-L1, PD-L2, TIGIT, ROBO1, ROBO2, DKK1, LOX1, DCC, IL-23R, and syncytin-2. The campaign yielded 424 antibodies with sufficient production, 301 passing aggregation and polyreactivity filters, 103 with K_D below 10 nM by SPR, and 118 with EC50 below 25 nM by cell-based assay. Then they applied logistic regression to the deep sequencing data from PD-L2 and ROBO2 selections and identified additional low-frequency binders that the experimental FACS sorting had missed.

Why This Is Cool

The library design is deliberately minimal so that all diversity lives in one short, sequenceable region. That makes the entire dataset immediately usable for ML: you get a compact sequence representation linked to experimental binding outcomes across multiple targets, with no ambiguity about which residues drive recognition. The public release of this data is probably more valuable than any single antibody it produced. Most groups training antibody ML models are data-limited. This gives them a clean, multi-target training set with negative examples built in. The logistic regression rescue of missed binders is also a nice proof-of-concept that even simple ML can add value on top of experimental selection when the data is structured correctly.

Read the paper.

🗓️ Events & Competitions

The best competitions, hackathons, and community challenges in AI x life sciences, curated weekly. Know something worth featuring? Reply and let us know.

More upcoming events:

BioHackathon Europe 2026 | November 9-13, Barcelona

ELIXIR’s annual international bioinformatics hackathon, running since 2018. 160+ participants, five days of collaborative coding on open bioinformatics infrastructure and tools. The call for project proposals has now closed.

Thanks for reading!

💬 Get involved

We’re always looking to grow our community. If you’d like to get involved, contribute ideas or share something you’re building, fill out this form or reach out to me directly.

Connect With Us

Have questions or suggestions? We'd love to hear from you!

📧 Email Us | 📲 Follow on LinkedIn | 🌐 Visit Our Website

Most of what we cover in the newsletter is preclinical and clinical: how drugs get discovered, how they get tested, whether the science holds up. We started wondering what happens after approval. Do approved drugs actually sell well? If not, why, and what would fix it?

• One-third of launches miss expectations and this pattern hasn’t moved in a decade. Deloitte’s 2012-2021 US launch cohort: 34% missed, 66% met or beat. Clinical approval fixes the science; it does not fix reimbursement. Whether payers will fund a drug, at what price, for which patients, and how quickly involves different decision-makers with different evidence requirements. 57% of launch failures trace to limited market access, 47% to inadequate understanding of market needs, 41% to poor differentiation (Deloitte, 2022).

• The endpoint problem is central. A drug can be approvable but commercially weak if the trial proves an endpoint (i.e. a “measurable outcome”) that regulators accept but payers, doctors, or patients do not value enough to justify price, switching, or broad coverage.

• The fix is about who’s in the room, and when. Drug discovery is a long value chain. The people who understand reimbursement, access, and real-world uptake need to influence decisions at the start: target selection, trial design, evidence planning. The best programmes run market access due diligence before they start, not after Phase 3 locks.

Do you fancy getting free access to a platform that helps speed up your scientific research?

We're running a programme that gives academic and nonprofit labs a full year of access to our platform where you can run lit reviews, target discovery, and bioinformatics in one place, among many other things. We also throw in direct support from our science team.

Read more about the programme

Apply now

For this piece, I spoke with Gerardo Martinez, Senior Director of Global Market Access at CSL, and Maria Garcia, who leads the medical committee on independent reimbursement risk assessments at MARA Rating. Their insight runs throughout.

How market access works in practice

Once a drug is approved by a regulator (FDA in the US, EMA in Europe), it still needs to clear a second gate before patients can actually get it. That gate is Health Technology Assessment (HTA).

In England, NICE (National Institute for Health and Care Excellence) evaluates whether a drug offers enough clinical benefit relative to its cost to justify NHS funding. In Germany, G-BA (Federal Joint Committee) runs a benefit assessment that determines pricing tiers. In France, HAS (Haute Autorité de Santé) scores therapeutic improvement on a five-point scale that directly sets the reimbursement level. In the US, the system is more fragmented: CMS (Centers for Medicare & Medicaid Services) makes coverage decisions for government insurance, while private payers and pharmacy benefit managers each set their own formulary rules.

These bodies ask different questions from regulators. Regulators ask: is this drug safe and effective? HTA (Health Technology Assessment) bodies ask: is it better than what we already pay for, by enough to justify the price difference? That gap is where most of the launch failures in this piece originate.

How drugs are valued (and where the model breaks)

Before getting into how launches fail, it is helpful to understand how drugs are valued before they enter the market.

Drug valuation starts as a risk-adjusted net present value (rNPV) model. You estimate future cash flows based on patient population, discount them by the probability of the drug making it through each development stage, and subtract remaining costs. Each stage transition (preclinical to Phase 1, Phase 1 to 2, and so on) gets its own probability of success, so the model penalises early-stage assets more heavily. It forces uncertainty into the maths. The problem is that the model looks quantitative while hiding qualitative judgments about payer behaviour, patient burden, and whether anyone will actually switch from the existing standard of care.

The assumptions that matter most: how many patients will be eligible and diagnosed, what price payers will accept, how quickly physicians will switch, how long patients will stay on therapy, and whether a competitor will reset the value bar before you reach peak sales. A one-point shift in any of those can dominate the model. rNPV is useful for comparing assets in a portfolio. It is dangerous when people treat its output as a prediction rather than a structured guess.

Estimates of what it costs to develop and approve a single drug range from ~$1 billion to $2.6 billion depending on methodology and how you account for failures. When this much capital is at stake, getting the commercial assumptions wrong early (as is the case for 34% of the drugs that underperform in sales) means the loss compounds for years.

When the trial answers the wrong question

Regulatory endpoints and payer endpoints overlap, but they are not the same thing. The FDA will approve a drug based on a surrogate biomarker if the evidence is strong enough. Payers, HTA bodies, and physicians often need something more: proof that the drug changes outcomes patients actually care about, measured against a comparator that reflects real-world practice rather than placebo.

This is exactly what Gerardo described from the payer side of the table. “Payers are not only asking, ‘Does it work?’” he told me. “They are asking, ‘Does it work better than what we already fund? In which patients? How meaningful is the benefit? Is it worth the opportunity cost?’” The strongest access cases don’t avoid uncertainty. They anticipate it, make it visible early, and address it directly.

MARA’s analysis of that 2025 cohort identified exactly where the evidence falls short. The most common gaps were not in core clinical efficacy. They were in resource use and cost implications (flagged in 42 of 46 drugs), health-related quality of life evidence (37 of 46), unresolved uncertainty at launch (26 of 46), and comparator selection (20 of 46). These are the domains payers use to judge comparative value, implementation risk, and affordability. None of them are guaranteed by FDA approval.

The cleanest example of what happens when they don’t is Aduhelm (aducanumab), a drug for Alzheimer’s disease from Biogen. FDA granted accelerated approval based on amyloid plaque reduction, a biomarker CMS (Centers for Medicare & Medicaid Services) responded by limiting Medicare coverage to patients enrolled in clinical trials. Physicians were sceptical, uptake was negligible, and Biogen discontinued the drug in 2024 after investing several billion dollars in development. The biology worked, the regulatory bar was met, and the commercial outcome was still a catastrophe.

The useful contrast is Leqembi (lecanemab), in the same disease, same drug class. The difference: the confirmatory trial measured clinical decline on CDR-SB, a cognition and function endpoint. FDA converted it to traditional approval. CMS broadened coverage. The Alzheimer’s market did not reject anti-amyloid drugs categorically. It needed evidence that connected the mechanism to something patients and families could observe.

The same pattern shows up elsewhere. Avastin, Makena, Iressa: in each case, the endpoint problem was visible long before the commercial failure arrived.

There is also the rarer, more catastrophic version: post-market safety. Phase 3 trials are typically too small and too short to catch every rare or delayed adverse event. Once a drug reaches real-world populations, broader exposure can turn “manageable safety uncertainty” into boxed warnings or full market withdrawal. A 2025 JAMA Internal Medicine analysis of 560 drugs approved 2001-2019 found 23.2% experienced a post-market safety action, meaning that 13 drugs not only caused patients issues but also caused financial loss to pharma companies, e.g. through slower prescriptions or adoption.

When it works: the GLP-1 story

GLP-1 drugs are the opposite case. Wegovy and Zepbound did not rely on an abstract biomarker. The primary endpoint was body weight reduction, something patients can see, physicians can track, and media can amplify. The effect size was large enough to reset expectations for the entire obesity category. Then the SELECT trial added cardiovascular outcome data, moving the class from “weight loss drug” toward “cardiovascular risk reduction in patients with obesity.”

The trials also got the population right. The label covers adults with obesity (BMI 30+) or overweight (BMI 27+) with at least one weight-related condition. That is enormous, well-defined enough for payers to recognise as medically necessary, and broad enough to generate demand at scale. WHO estimates 2.5 billion adults are overweight globally and 890 million have obesity.

The result: Ozempic and Mounjaro alone generated $71 billion in 2025 revenue. For context, that is roughly 2.5x the combined revenue of OpenAI and Anthropic in the same year. Only ~2% of the 800 million eligible patients currently have access. Worth noting: most GLP-1 prescriptions are paid out-of-pocket or through private insurance rather than through HTA-assessed public reimbursement, so the market access dynamics here are different from most drug classes. Demand so massively exceeded expectations that supply became the bottleneck, not coverage.

What made it work: the endpoint was clinically measurable, commercially legible, and expandable into harder outcomes. Start with something everyone understands, then add evidence that justifies broader coverage. Few drug classes have executed that sequence this well.

Who gets it right (and why)

Deloitte’s data shows consistent patterns in what predicts launch success. Drugs granted priority review (meaning the FDA considered them a significant improvement over existing options) met or beat forecasts 74% of the time, versus 47% for standard review. Specialty drugs (typically higher-cost treatments administered under physician supervision or requiring special handling) outperformed at 72% versus 48%. Orphan drugs (for rare diseases affecting fewer than 200,000 patients) hit 73% versus 57%. First-in-class drugs and products reimbursed under the medical benefit (paid directly by insurers rather than through pharmacy channels) also outperform.

This does not mean “rare and expensive always wins.” It means launches go better when the unmet need is visible, the patient population is defined, the specialist community is concentrated, and the product gives payers a coherent reason to change behaviour. The more diffuse the population and the more crowded the market, the harder the launch. General medicine is especially brutal: 58% of general medicine launches missed expectations in the 2012-2021 cohort.

Maria Garcia, whose team at MARA Rating assesses assets calibrated to real-world HTA decisions, described the pattern she sees at scale: “Some companies repeatedly produce evidence packages with strong reimbursement readiness. Others show recurring gaps regardless of asset or indication. That’s not a coincidence. It’s organisational discipline, embedded at programme design level.” That distinction is financially material. It shows up in launch delays, managed entry terms, and access rates that never reach modelled uptake.

What the disciplined ones do differently is structurally simple, she said: “They integrate market access into trial design, not after it. They run HTA scientific advice before Phase 3 lock. They don’t confuse regulatory approval with reimbursement readiness. Those are two different standards, assessed by two different bodies, under two different evidentiary logics.”

Company size matters too, in a counterintuitive direction. In orphan drug launches, 92% of small company launches and 79% of midsize launches met or beat expectations, compared to 53% for large pharma. Smaller companies have faster and easier communication thanks to their size and have normally built the whole organisation around one disease area and one patient journey. Large companies have more muscle, but also more silos and more inherited constraints from acquired assets where the trial design was already locked.

Modality choice at the preclinical stage also constrains access strategy years later. The commercial model for an oral small molecule in general medicine is fundamentally different from a cell therapy in rare disease, and the evidence package needs to reflect that from the start. I asked Gerardo whether rare disease is genuinely easier because there are often no existing treatments. His answer was blunt: “Rare disease is not easier. It is different. The absence of treatment can help the narrative. It can also make the evidence problem harder.”

Maria went deeper on the friction points that catch teams off guard. In rare diseases, HTA bodies are built to answer “compared to what?” and that anchor often doesn’t exist. In gene therapies, durability uncertainty is the real wall: you’re asking a payer to fund twenty years of projected benefit on two years of follow-up. Then there are budget impact optics. “A therapy treating three hundred patients can price at seven figures and still cost the system less than a common chronic disease,” she pointed out. “The optics dominate the economics, and companies consistently underestimate it.”

How do we fix this?

A growing number of companies are building tools to close the gap between clinical development and market access.

MARA Rating takes a different approach from most: independent, standardised reimbursement risk scores. They rate assets from A++ to C across domains calibrated to historical HTA outcomes, giving portfolio teams an external benchmark for access risk the same way credit ratings benchmark financial risk. Clinical, regulatory, and manufacturing risks are routinely benchmarked in portfolio governance. Reimbursement risk often is not.

On the analytics side, Datavant builds real-world evidence platforms that let teams model whether a single study design can serve both regulatory and HTA purposes. Panalgo (now part of Norstella) offers self-service RWE analytics so teams can rapidly test whether their chosen endpoint will matter to payers during trial design rather than after it. IQVIA‘s Launch Excellence suite specifically targets the forecast-miss problem with payer landscape intelligence and formulary tracking at scale.

On the strategy and intelligence side, Lumanity models different trial design scenarios against likely HTA outcomes across markets, explicitly positioning around the clinical-to-commercial alignment problem. Certara‘s Evidence & Access division links pharmacometrics directly to HTA submission requirements, so clinical pharmacology decisions get pressure-tested against payer needs early. Atheneum runs an AI-matched expert network connecting pharma teams to payer and HTA advisors early in development. Sometimes the fix is not better analytics; it is getting clinical teams in front of the right people before Phase 3 locks.

The alignment problem

Market access teams typically get involved too late. In most pharma companies, market access sits under commercial, grouped with sales rather than at the executive table. So when clinical teams are locking Phase 2/3 designs, market access isn’t in the room. Maria Garcia was direct about this: “Commercial is not the same as market access. It’s a very different understanding of the market.” Commercial teams handle sales strategy and conference presence. Market access determines whether anyone can actually pay for what you’re promoting. Because they share an org chart, they get conflated, and market access professionals end up without authority to redirect development even when the reimbursement outlook is clearly poor.

Gerardo framed this as the central misconception: “The biggest misconception is that market access starts after the product is developed. It is not just a launch function. It is a strategic bridge between innovation and real-world adoption.” The companies that get this right expose asset teams to market access constraints while there is still time to change the endpoint, comparator, population, or evidence plan. The ones that get it wrong throw a finished clinical package over the wall and wonder why the launch underperforms

The problem goes deeper than timing. Before starting any programme, teams should be running market access due diligence to assess whether a programme is even worth progressing. What matters is the speed of the feedback loop between what the science can deliver and what the market will accept.

You cannot change organisational structure overnight. What you can do is build systems that translate and pass information between teams, surfacing what matters at each stage. AI is well-suited to this. Gerardo was measured: “AI can help us move faster. It can support evidence mapping, landscape scanning, analog analysis, payer objection tracking. It can help us see patterns earlier and pressure-test assumptions with more speed.” The hype is the idea that AI replaces market access judgment. It does not. “Market access is interpretation, prioritisation, and judgment under uncertainty. AI gets us to a sharper first draft. The real value still comes from human judgment.”

The strongest use case is giving preclinical and clinical teams a live view of market access constraints. Which endpoints convinced payers in which countries? What does the standard of care look like by market, by patient subgroup? These questions are answerable from HTA decisions, trial registries, claims data, and payer policies, but nobody has the time to synthesise them manually for every asset in a portfolio. That is the gap a platform can fill: not replacing judgment, but making sure the right information reaches the right team before the design is locked.

Data sources: Deloitte launch cohort analyses (2020, 2023, market access framework), BIO/Informa/QLS clinical success rates 2011-2020, Wouters et al. JAMA 2020, MARA Rating 2025 NME Analysis. FDA regulatory documents, WHO/OECD pharmaceutical pricing policy publications. Full source notes available on request.

💬 Want to be featured in Kiin Bio Weekly?

Each issue we speak directly with researchers, scientists, and builders working at the frontier of AI in life sciences. If you're working on something in this space and think it would resonate with our community, I'd love to hear from you. Fill out this form or reach out to me directly.

Found this useful? Forward it to a colleague or friend you think would be interested, it's the best way to help the newsletter grow.

Share Kiin Bio Weekly

Subscribe now to stay at the forefront of AI in Life Science. Every week: primers, deep dives, and direct conversations with the people building the field.

Subscribe now

Connect With Us

Have questions on this or suggestions for our next deep dive? We’d love to hear from you!

📧 Email Us | 📲 Follow on LinkedIn | 🌐 Visit Our Website

• scVision throws out the tokenisation paradigm for single-cell models entirely and represents cells as images instead. The results are surprisingly strong.

• conDitar-dev does something rare in generative drug design: it goes to the bench. Real synthesis, real binding assays, real selectivity data.

• AllTheBacteria is less a paper and more a piece of infrastructure. 2.4 million bacterial genomes, uniformly processed, with a proof-of-concept antibiotic discovered from querying it.

Kiin Pioneer Programme

We built a platform that helps researchers speed up their entire science, from literature review and biomarker discovery to bioinformatics and computational chemistry. If your workflow involves pulling findings from five different places before you can actually act on any of them, this is for that.

The Pioneer Programme gives academic labs and non-profits one year of free access, plus support from our science team. No cost, no data transfer, all IP stays with your institution. Applications close August, cohort starts September.

Read more about the programme

Apply now

scVision: A vision foundation model for single-cell biology

Where This Fits

Everyone building single-cell foundation models has been pretty much following the same NLP playbook since 2023: treat genes as tokens, cells as sentences and train a transformer. scGPT, Geneformer, scFoundation all do this. It works reasonably well for cell-type annotation, but the tokenisation step is where information is lost. You have to bin continuous expression values into discrete categories, and then you throw away co-expression structure entirely. There have been several independent evaluations over the past year questioning whether these models actually beat simpler baselines once you control for dataset size. scVision does something different: it asks whether vision, not language, is the right modality for this data.

What It Is

Ridvan Yesiloglu, Md Tauhidul Islam, and colleagues at Stanford represent each cell’s transcriptome as a 2D image. Genes get placed on a 104x104 grid using Gromov-Wasserstein optimal transport so that co-regulated genes end up as spatial neighbours. This layout is shared across all cells, which means attention patterns in the model map directly onto gene programmes without any post-hoc interpretability analysis.

The model is an 86M-parameter Vision Transformer pretrained with masked image modelling on 72 million human cells. In zero-shot cell-type annotation across six held-out tissues it ranked first on every atlas, balanced accuracy 0.47 to 0.83 depending on the tissue. The few-shot results are arguably more impressive: one labelled cell per type with scVision matched the performance of 50 labelled cells with the competing models.

Why This Is Cool

The benchmark numbers are strong but annotation is the easiest downstream task, so I would not get too excited about those alone. What is more interesting is what the representation enables. Because genes are spatially arranged by co-expression, you can do things like mask a neighbourhood of genes and ask what the model predicts, or run perturbation experiments on patches of co-regulated genes, or read attention maps as pathway activations. None of those operations have analogues in the token-based models. Whether this actually helps with harder tasks like perturbation prediction or drug response is still an open question. Worth watching.

Read the paper.

Try the demo.

conDitar-dev: Generating developable 3D molecules via pocket-conditioned diffusion

Where This Fits

There are now a lot of diffusion-based methods for structure-based drug design: TargetDiff, Pocket2Mol, DiffSBDD, and others. They generate molecules conditioned on a protein binding pocket and they generally optimise for binding affinity. The well-known problem is that affinity alone does not make a drug. The molecules these tools produce tend to have poor solubility, metabolic instability, or toxicity flags, and medicinal chemists end up spending months fixing properties by hand. conDitar-dev tries to fix this at the source by incorporating developability constraints directly into the generation process rather than treating it as a downstream optimisation.

What It Is

Ruoxi Gao, Xia Ning, and collaborators at Ohio State, the University of Minnesota, Google, and Sanofi built a three-module framework. A pretrained pocket encoder (msPRL) learns multi-scale binding site representations. A diffusion model (conDitar) generates 3D ligands conditioned on those. A plug-and-play optimiser (paOPT) steers the diffusion trajectory toward better ADMET properties at generation time, no retraining required.

On their new human-targets-only benchmark (CDH, 84 disease-relevant proteins), conDitar achieves average Vina D scores of -8.85 kcal/mol, 7.4% better than the next baseline. Adding paOPT improves ADMET properties by up to 73% while maintaining comparable affinity. The CDH benchmark itself is a useful contribution since CrossDocked2020 mixes in non-human targets which confounds therapeutic relevance.

The part that stands out: they synthesised compounds and tested them. Two conDitar-dev molecules for PD-L1 showed SPR-derived K_D values of 3.49 and 3.75 micromolar. For CSF1R, hit expansion on conDitar-dev designs identified selective inhibitors with IC50 as low as 200 nM and promiscuity hit rates under 2.5%.

Why This Is Cool

Most generative drug design papers stop at computed docking scores. This one went to the bench, which immediately puts it in different territory. Whether micromolar PD-L1 binders are interesting enough to advance is debatable, but the workflow producing testable, druglike matter on the first pass is the point. The paOPT module being training-free and pluggable is also notable since ADMET requirements change between programmes and you do not want to retrain your generative model every time the target product profile shifts.

Read the paper.

Try the code.

AllTheBacteria: A community resource for bacterial genomics and antimicrobial discovery

Where This Fits

Human genetics has gnomAD: hundreds of thousands of genomes, uniformly processed, searchable, freely available. Microbiology has not had an equivalent. The raw sequencing data is there in INSDC archives but it is inconsistently assembled, inconsistently annotated, and largely inaccessible unless your group can build and maintain its own processing pipeline. AllTheBacteria does for bacteria what gnomAD did for human genetics. 2.4 million genomes from 11,273 species, all processed the same way, all queryable.

What It Is

Martin Hunt, Marcelo Torres and a large international consortium took all public bacterial and archaeal short-read whole-genome sequencing data from INSDC and ran it through uniform assembly, quality control, and annotation. Every genome gets standardised taxonomy, gene calls, antimicrobial resistance predictions, antiphage-defence annotations, protein structure predictions, and AI-ready sequence tables.

To demonstrate what you can do with this, the team ran their deep learning model APEX 1.1 against AllTheBacteria proteomes and identified 1,867 candidate encrypted antimicrobial peptides. They synthesised 24 of them, tested against 20 clinically relevant pathogens including resistant strains, and found multiple hits with low-micromolar activity. A lead molecule, ATB20, reduced Acinetobacter baumannii burden in a mouse skin abscess model with efficacy comparable to polymyxin B.

Why This Is Cool

This is infrastructure. The antimicrobial peptide discovery is the validation experiment showing what becomes possible when 2.4 million annotated genomes are openly available and searchable. The barrier for training ML models on bacterial sequences, running comparative genomics at scale, or doing real-time resistance surveillance used to be whether your group could maintain its own pipeline. That barrier is now gone. The fact that a relatively straightforward query immediately turned up a validated antibiotic active against a WHO priority pathogen tells you how much untouched biology is sitting in public archives.

Read the paper.

Try the code.

🗓️ Events & Competitions

The best competitions, hackathons, and community challenges in AI x life sciences, curated weekly. Know something worth featuring? Reply and let us know.

More upcoming events:

BioHackathon Europe 2026 | November 9-13, Barcelona

ELIXIR’s annual international bioinformatics hackathon, running since 2018. 160+ participants, five days of collaborative coding on open bioinformatics infrastructure and tools. The call for project proposals has now closed.

Thanks for reading!

💬 Get involved

We’re always looking to grow our community. If you’d like to get involved, contribute ideas or share something you’re building, fill out this form or reach out to me directly.

Connect With Us

Have questions or suggestions? We'd love to hear from you!

📧 Email Us | 📲 Follow on LinkedIn | 🌐 Visit Our Website

Why the molecular format you feed a language model matters as much as the model itself, and what a five-day experiment over New Year’s proved about it.

For computational chemists, drug discovery scientists, and anyone building or evaluating agentic AI systems that touch molecular structures.

• Language models can now reason about chemistry, yet they still routinely fail at writing molecules correctly. The bottleneck is not intelligence. It is representation: SMILES and IUPAC were designed for humans and databases, not for models that generate text token by token.

• Across a benchmark of 78,000 questions, MolJSON consistently outperformed existing formats. On the constrained generation task, GPT-5 achieved 95.3% accuracy outputting MolJSON (a simple atom-and-bond list in JSON), compared with 64.0% for SMILES and 76.3% for IUPAC. The format appears in zero training corpora. The advantage is purely structural.

• This piece covers where existing representations break down, how MolJSON works, why the numbers are not close, and what it means for anyone building tools that ask language models to propose or modify molecules

Kiin Pioneer Programme

We built a platform that helps researchers speed up their entire science, from literature review and biomarker discovery to bioinformatics and computational chemistry. If your workflow involves pulling findings from five different places before you can actually act on any of them, this is for that.

The Pioneer Programme gives academic labs and non-profits one year of free access, plus support from our science team. No cost, no data transfer, all IP stays with your institution. Applications close August, cohort starts September.

Read more about the programme

Apply now

This week we spoke with Nicholas Runcie, a DPhil student in statistics at Oxford, about a format for molecular representation that started as a five-day sprint over New Year’s and ended up reshaping how his group thinks about language models in chemistry.

The problem: models that can reason but cannot write

In January 2025, Nicholas Runcie set out to prove that language models could not do chemistry. He built ChemIQ, a benchmark of 816 questions spanning everything from counting carbon atoms to solving NMR spectra. The results surprised him. When DeepSeek R1 and OpenAI’s o3-mini arrived with chain-of-thought reasoning, they scored 50-57% on tasks where previous models had achieved 3-7%. On IUPAC naming, a task where earlier models scored near zero, reasoning models reached 29-44%. They were not just pattern-matching. They were building internal representations of molecular graphs and reasoning about them step by step.

The problem was reliability. A model might reason beautifully about a molecule’s structure, then botch the output because SMILES notation requires you to linearise a graph into a single traversal path with nested parentheses and ring-closure digits. Get one bracket wrong and the molecule is invalid. IUPAC names are just as problematic: they demand perfect recall of nomenclature rules, locant numbering, and substituent priority. These formats were designed for chemists reading printed pages and for databases storing canonical strings. They were never designed for sequence models generating text token by token.

Any agentic system that needs to propose, modify, or communicate molecules needs its language model to write structures reliably. If the representation introduces systematic errors, the pipeline breaks downstream.

Figure 1: Six ways to say ‘acetic acid’. Only MolJSON explicitly lists atoms and bonds without requiring the model to linearise the molecular graph or recall nomenclature rules.

The idea: just write out the atoms and the bonds

Instead of encoding a molecule as a traversal (SMILES) or a name (IUPAC), write it as a JSON object with two arrays: one listing atoms with identifiers and elements, another listing bonds with source, target, and order. No traversal path to maintain. No nomenclature rules to recall. No ring-closure digits to track.

Want to add a methyl group to a benzene ring? Add one entry to the atoms array and one to the bonds array. In SMILES, you would rewrite the entire string. In IUPAC, you would recalculate every locant and potentially rename the parent chain.

{

  “atoms”: [

    {”id”: “C1”, “element”: “C”},

    {”id”: “C2”, “element”: “C”},

    {”id”: “O1”, “element”: “O”},

    {”id”: “O2”, “element”: “O”}

  ],

  “bonds”: [

    {”source”: “C1”, “target”: “C2”, “order”: 1},

    {”source”: “C2”, “target”: “O1”, “order”: 2},

    {”source”: “C2”, “target”: “O2”, “order”: 1}

  ]

}

Acetic acid (CH3COOH) in MolJSON. Two carbons, two oxygens, three bonds. The entire molecule is a readable list with no special syntax to memorise.

The format also exploits something about how modern language models are trained. Structured JSON output is a first-class capability in every major commercial model because it underpins function calling and agent tool use. Models have been explicitly optimised to produce valid JSON conforming to a schema. MolJSON rides that existing capability rather than asking models to master a chemistry-specific syntax they encounter rarely in training data.

Runcie had the idea in January 2025 but shelved it. SMILES and IUPAC were good enough at the time. The format only came to life because OpenAI gave him a grant of API credits with a New Year’s Eve deadline. With five days and $4,000 to spend, he designed and ran every experiment without iteration. The results in the paper are exactly what he obtained on his first attempt.

Why it’s different: the numbers are not close

Runcie’s benchmark tested three tasks across 78,045 questions using GPT-5-nano, GPT-5-mini, GPT-5, and Claude Haiku 4.5. The molecules were sampled from PubChem deposits made after October 2025 to reduce contamination from training data. Three findings stand out.

Translation. When asked to convert a molecule from one representation to another, GPT-5 achieved 71.0% accuracy translating IUPAC names to MolJSON, compared with 43.7% translating the same IUPAC names to SMILES. The gap held across all model sizes and both directions.

Shortest-path reasoning. Given a molecule and asked to count bonds between two halogen atoms, GPT-5 answered 98.5% of questions correctly when the molecule was presented in MolJSON, versus 92.2% for SMILES and 82.7% for IUPAC. MolJSON also used roughly 1.8 times fewer reasoning tokens than SMILES on this task, meaning the model spent less compute parsing the structure before it could reason about it.

Constrained generation. When asked to generate a molecule satisfying a set of structural constraints (specific atoms, ring counts, shortest-path requirements), GPT-5 achieved 95.3% accuracy outputting MolJSON, versus 64.0% for SMILES and 76.3% for IUPAC. For molecules with a fused ring system, GPT-5 hit 92% accuracy in MolJSON, compared with 42% for SMILES.

The pattern is consistent: MolJSON outperforms formats that language models have been heavily exposed to in training. SMILES strings are ubiquitous in PubChem, ChEMBL, and ZINC. IUPAC names appear in millions of papers. MolJSON appears in none of their training corpora. The advantage comes from structural alignment between the format and how models process information, not from data familiarity.

SMILES and IUPAC accuracy also degrades with molecular size and ring complexity. Fused ring systems cause particular problems for IUPAC. MolJSON remains robust because it represents the graph directly rather than encoding it through a traversal or naming convention.

The future: from format to fully autonomous chemistry

Runcie is clear that MolJSON as published is not a finished standard. He has already heavily modified the format for his current work, keeping the core principle (explicit atom and bond lists) while optimising for specific use cases. He has moved from JSON to YAML for compatibility with Codex agents, and he is currently exploring strategies for encoding stereochemistry and other molecular features

The more interesting signal is where this is heading. Runcie is building agents for autonomous chemistry research, and MolJSON is a core part of that framework. Reliable molecular I/O is what lets an agent propose a structure, check it, modify it, and hand it to the next step without a human correcting format errors along the way.

This fits within Runcie’s broader research programme: developing systems capable of fully automated chemistry research. If a computer were to design a drug, how would it do it? His hypothesis is that language models combining broad knowledge with structured reasoning may solve chemistry problems humans have not yet imagined. MolJSON is one piece of that infrastructure, ensuring the model can reliably read and write molecular structures so that higher-level reasoning is not bottlenecked by format errors.

Kiin’s view.

The most important thing about this paper is not MolJSON itself. It is the demonstration that molecular representation is a first-order design decision when building LLM-based chemistry systems, and that the defaults everyone reaches for (SMILES, IUPAC) are measurably suboptimal. The field has spent years fine-tuning models and building agentic scaffolding while leaving the input/output format as an afterthought. This paper puts a number on what that costs: a 30-percentage-point accuracy gap on constrained generation, same model, same molecules.

These are also first-attempt results with no prompt engineering or iterative refinement. Someone thinking harder about schema design or hybrid representations could likely push them further. That is Runcie’s explicit point: this is a starting direction, not a destination. For anyone building tools that ask language models to propose or modify molecules, the takeaway is immediate. Test your representation. The gap between “works sometimes” and “works reliably” may be a schema change, not a model upgrade.

GitHub: github.com/oxpig/MolJSON

Thanks for reading Kiin Bio Weekly!

Share Kiin Bio Weekly

💬 Get involved

We’re always looking to grow our community. If you’d like to get involved, contribute ideas or share something you’re building, fill out this form or reach out to me directly.

Subscribe now to stay at the forefront of AI in Life Science and keep up with this upcoming season of deep dives.

Connect With Us

Have questions on this or suggestions for our next deep dive? We’d love to hear from you!

📧 Email Us | 📲 Follow on LinkedIn | 🌐 Visit Our Website

• Vilya-1 predicts macrocycle conformations across diverse chemistries with a single all-atom model, beating physics-based methods and existing deep learning approaches on geometric accuracy. Macrocycles are notoriously hard to model, and this is the first foundation model built specifically for them.

• TWIN is an implicit solvent ML potential trained on ab initio data that runs two orders of magnitude faster than explicit solvent while approaching DFT accuracy. If this holds up broadly, it removes one of the biggest bottlenecks in drug and peptide simulation.

• UCE is now published in Nature after a 2023 preprint. A single-cell foundation model that embeds any cell type without retraining, trained on the Tabula Sapiens atlas. The zero-shot transfer results across tissues and species are what earned it the Nature publication.

Kiin Pioneer Programme

We built a platform that helps researchers speed up their entire science, from literature review and biomarker discovery to bioinformatics and computational chemistry. If your workflow involves pulling findings from five different places before you can actually act on any of them, this is for that.

The Pioneer Programme gives academic labs and non-profits one year of free access, plus support from our science team. No cost, no data transfer, all IP stays with your institution. Applications close August, cohort starts September.

Read more about the programme

Apply now

Vilya-1: Macrocycle structure prediction foundation model

🧪 Where This Fits

Macrocyclic peptides occupy a sweet spot in drug design: large enough to hit protein-protein interactions that small molecules cannot reach, small enough to potentially cross cell membranes. The problem is that their conformational behaviour is extremely difficult to predict. They are too large for small-molecule force fields to handle well, too flexible for protein structure prediction methods, and too chemically diverse (cyclic, branched, non-natural amino acids, N-methylation) for any single modelling approach to cover.

Current options are either expensive physics-based sampling (molecular dynamics, metadynamics) that takes days per compound, or deep learning methods built for linear peptides that do not generalise to the cyclic, heavily modified structures that matter therapeutically. Tools like AlphaFold and OpenFold were not designed for this chemical space. Vilya-1 from Vilya Research is the first foundation model purpose-built for macrocycle structure prediction, using a unified all-atom representation that handles diverse topologies and chemical modifications in a single model.

🔍 What It Is

• Predicting macrocycle 3D structure is hard because they span a chemical space between small molecules and proteins that existing methods handle poorly. Sturmfels, Salem, Hiranuma et al. from Vilya Research present Vilya-1, a deep learning foundation model trained on heterogeneous structural datasets across diverse macrocycle topologies and chemistries.

• Uses a uniform all-atom representation that accommodates cyclic peptides, non-natural amino acids, N-methylated residues, and other modifications without needing separate model architectures for each. Trained on a mix of experimental and computational structural data.

• Improved geometric accuracy compared to physics-based methods and existing deep learning alternatives. Coverage extends to small molecules. Also supports predicting developability properties like membrane permeability, and generative design of novel macrocycles with specified properties.

💡 Why This Is Cool

Macrocycles are having a moment in pharma. Peptide drugs like semaglutide proved the market, and companies like Bicycle Therapeutics are building pipelines around constrained cyclic peptides. The structural prediction gap has been a real bottleneck: you can design a macrocycle computationally, but confirming it folds the way you expect requires either NMR (expensive, slow) or crystal structures (often impossible for flexible molecules). A model that gives you reliable conformational predictions for chemically diverse macrocycles would accelerate the design-make-test cycle considerably. The extension to permeability prediction is smart, since oral bioavailability is the other major hurdle for this drug class. No public code yet, which limits immediate evaluation by the community.

📃 Read the paper.

No public code repository available at time of writing.

TWIN: Transferable implicit solvent machine learning potential approaching ab initio accuracy

🧪 Where This Fits

Molecular dynamics simulations of drugs and proteins in solution face a fundamental trade-off: explicit solvent (modelling every water molecule) is accurate but extremely expensive computationally, while implicit solvent models (treating water as a continuum) are fast but sacrifice accuracy. For drug design, this means you either spend days simulating a single compound in explicit water, or you get fast results from implicit models that often disagree with experiment.

Previous implicit solvent approaches (GB/SA, COSMO) are based on empirical physics approximations that break down for many drug-like molecules. Recent ML potentials (ANI, MACE) have improved accuracy for gas-phase simulations, but extending them to solvation has been limited. TWIN from TU Munich takes an equivariant graph neural network, trains it exclusively on ab initio and experimental solvation data (no empirical force field data), and produces an implicit solvent potential that transfers across drugs, peptides, and proteins.

🔍 What It Is

• Simulating molecules in water is either accurate and slow (explicit solvent) or fast and unreliable (implicit solvent). Eckwert and Zavadlav from TU Munich present TWIN (Transferable Water Implicit Network), an ML potential that combines the speed of implicit solvent with accuracy approaching DFT-level explicit solvent calculations.

• Built on an equivariant graph neural network trained exclusively on ab initio quantum mechanical data and experimental measurements. Avoids any dependence on empirical force field parameters, which is what limits the transferability of traditional implicit solvent models.

• Two orders of magnitude faster timestep evaluation than explicit solvent approaches. Outperforms prior ML-based implicit solvent models on crystallographic and NMR benchmarks. Transfers across drug molecules, peptides, and proteins without retraining.

💡 Why This Is Cool

Solvent is the hidden cost of computational drug design. Every binding free energy calculation, every conformational sampling run, every MD simulation spends most of its compute on water molecules rather than the drug you care about. If TWIN delivers on the promise of DFT-level solvation accuracy at implicit-solvent speed, and transfers reliably across chemical space, it could make free energy perturbation and metadynamics calculations accessible at scale rather than as expensive one-off studies. The key caveat: “approaching ab initio accuracy” on benchmarks does not always translate to “accurate enough for rank-ordering drug candidates.” The real test is whether medicinal chemists can trust it for prospective predictions. Two orders of magnitude speedup is the kind of change that alters what questions you can afford to ask.

📃 Read the paper.

No public code repository available at time of writing.

UCE: Universal cell embedding provides a foundation model for cell biology

🧪 Where This Fits

Single-cell foundation models have proliferated over the past two years: scGPT, Geneformer, scFoundation, and others all aim to learn general representations of cells from large-scale scRNA-seq data. The shared limitation is that most require fine-tuning on task-specific labelled data, and they struggle to generalise to cell types or tissues not seen during training. They also typically use gene-token vocabularies tied to a specific species, making cross-species transfer difficult.

UCE from Jure Leskovec and Stephen Quake’s groups at Stanford takes a different approach: it embeds cells using protein-level information (via ESM embeddings of gene products) rather than gene-name tokens, which makes the vocabulary species-agnostic. Trained on 36 million cells from the Tabula Sapiens atlas across multiple human tissues. The preprint appeared in late 2023, and the Nature publication confirms the approach held up through peer review with additional validation.

🔍 What It Is

• Single-cell foundation models require fine-tuning for new tasks and do not transfer well across species or unseen cell types. Rosen, Roohani, Agrawal et al. from Stanford present UCE (Universal Cell Embedding), a foundation model that generates cell embeddings in a shared space without any task-specific retraining.

• Uses ESM protein language model embeddings to represent genes (rather than arbitrary gene tokens), making the representation biologically grounded and species-transferable. Trained on 36 million cells across diverse human tissues from the Tabula Sapiens consortium.

• Zero-shot cell type classification across tissues and species without fine-tuning. Embeddings capture biological relationships: similar cell types cluster together even across different organs and organisms. Published in Nature after extended peer review.

💡 Why This Is Cool

The protein-embedding approach is what makes this work where others have not. When you represent genes as ESM embeddings of their protein products, you get biological similarity for free: genes with similar functions have similar representations regardless of what organism they come from. That is why UCE can transfer across species without retraining, while models using gene-name tokens cannot. The Nature publication after a 2023 preprint suggests the reviewers were convinced this is not just a benchmark artefact. For single-cell researchers, the practical question is whether UCE embeddings are useful enough to replace the fine-tuning workflows they already have with scGPT or Geneformer. The 33-layer model and code are public, so that comparison is straightforward to make.

📃 Read the paper.

💻 Try the code.

🗓️ Events & Competitions

The best competitions, hackathons, and community challenges in AI x life sciences, curated weekly. Know something worth featuring? Reply and let us know.

More upcoming events:

BioHackathon Europe 2026 | November 9-13, Barcelona

ELIXIR’s annual international bioinformatics hackathon, running since 2018. 160+ participants, five days of collaborative coding on open bioinformatics infrastructure and tools. The call for project proposals has now closed.

Thanks for reading!

💬 Get involved

We’re always looking to grow our community. If you’d like to get involved, contribute ideas or share something you’re building, fill out this form or reach out to me directly.

Connect With Us

Have questions or suggestions? We'd love to hear from you!

📧 Email Us | 📲 Follow on LinkedIn | 🌐 Visit Our Website

Open scientific data has an access problem, and a former Terra engineer is building the fix

For computational biologists, genomics researchers, and any scientist working with large open-source datasets they can’t actually use.

• Open data mandates have flooded repositories with petabytes of publicly available scientific data. In practice, most academic researchers can’t touch it. Downloading costs thousands in egress fees, requires cloud infrastructure they don’t have, and takes weeks of setup before a single analysis runs.

• Scigantic puts compute where the data already lives. Nearly 1 exabyte of open-source data appears as local files in a notebook, with zero transfer fees and built-in fine-tuning for foundation models like ESM and DNABERT. It was built by someone who spent years on Terra at the Broad Institute and saw exactly where enterprise genomics platforms fail the people who need them most.

• This piece covers what the platform does, why the person behind it is unusually well-positioned to build it, and what it means for the growing gap between “data exists” and “researchers can work with it.”

Kiin Pioneer Programme

We built a platform that helps researchers speed up their entire science, from literature review and biomarker discovery to bioinformatics and computational chemistry. If your workflow involves pulling findings from five different places before you can actually act on any of them, this is for that.

The Pioneer Programme gives academic labs and non-profits one year of free access, plus support from our science team. No cost, no data transfer, all IP stays with your institution. Applications close August, cohort starts September.

Read more about the programme

Apply now

This week we spoke with Aaron Kanzer, founder of Scigantic, a platform that lets scientists work with massive open-source datasets without ever downloading them. Aaron is a solo founder running the company out of Boston, fully bootstrapped, with no outside funding. Before Scigantic he was a senior engineer at the Broad Institute working on Terra (the genomics platform serving 65,000 users across 80 petabytes of data), a founding engineer on MIT’s LINC project mapping neural circuits, and most recently scaled ML operations at Generate:Biomedicines. He built Scigantic because he kept watching the same problem go unsolved from the inside.

The problem: open data that isn’t actually open

There is a quiet fiction in science right now. Funders mandate open data sharing. Repositories grow by petabytes each year. Papers cite datasets that are technically available to anyone. The word “open” appears in every grant report.

The reality for a postdoc with a laptop and no cloud budget is different. “Open” means the data exists on a server somewhere. To actually use it, you need to download terabytes over days or weeks, pay egress fees that can run into thousands of dollars per transfer, set up cloud infrastructure you were never trained to manage, and hope your local machine has the storage to hold it all. Most don’t.

The origin was personal. At MIT’s McGovern Institute, Aaron was working on the LINC project, mapping neural circuits through high-resolution brain imaging. “I noticed that these amazing datasets were being generated within the LINC project,” he told us. “Incredible high-resolution images of the brain with a ‘wow’ factor. However, the data engineering necessary to access these images was so technical and challenging.” The data existed. It was extraordinary. And almost nobody outside a small group of engineers could actually get to it. “How might we improve this process for anyone in the scientific community?” That question became Scigantic.

The problem compounds with AI models. Tools like ESM and OpenFold are open-source, technically free to use. In practice, fine-tuning them on your own data requires weeks of infrastructure setup, or you go back to the original team and ask them to run it for you. The models are open. The ability to actually use them is not.

As Kanzer puts it: “Open-source models are great as a base case for scientific discovery. If you have a more targeted hypothesis, or a specific therapeutic you are developing, fine-tuning becomes integral. Setting up fine-tuning though still has a steep data engineering learning curve before you can conduct any science.” The bottleneck is not the model. It is everything between the model and actually using it on your own data.

For context: AWS charges between $0.09 and $0.12 per gigabyte for data transfer out of S3. A single copy of the AlphaFold database is over 20 terabytes. That is roughly $2,000 just to download one dataset, before you’ve done a single calculation. Multiply that across the dozens of datasets a real project might touch, and “open” starts to feel like a word with a hidden paywall attached.

The approach: compute moves to the data

Scigantic inverts the model. Instead of moving data to the researcher, it moves the researcher’s compute to where the data already lives. You open a JupyterHub notebook. The datasets you need appear as local files on your machine. You write code as if everything is sitting on your hard drive. Nothing ever downloads.

The mechanism is elegant: FUSE mounts make cloud-hosted data appear as a local filesystem. Every file read translates to a range-GET against the object store, pulling only the bytes you actually need. Your analysis runs in the same cloud region as the data. When you’re done, only the results leave: a trained model, a table, a figure. Typically megabytes, not terabytes.

Kanzer’s analogy is a library. “Rather than taking the entire book off the shelf, instead hand me the specific page I’m looking for. No one is reading the entire book, so why move the entire book?” That is what FUSE mounts achieve under the hood. The data stays in the cloud. Your notebook reads only the bytes it needs, as if they were sitting on a local drive. “FUSE is great because it elegantly tricks the agent into operating as if the data is a local directory,” Kanzer says. “It’s a more clever mousetrap when dealing with large-scale datasets.”

For fine-tuning, the workflow is similarly stripped back. Upload a CSV of sequences and labels, pick a foundation model, train. LoRA on frozen backbones means the whole thing fits on 24-48 GB cards. No training loops to write. No dependency management. No fighting with CUDA versions at 2am.

The platform currently links nearly 1 exabyte of open-source data spanning structural biology, genomics, neuroscience, and climate science. Datasets include UniProt, the full TCGA cancer genomics archive, CZ CELLxGENE Census, Allen Brain Atlas, and dozens more.

Why it’s different: built by the person who saw the limits from the inside

What makes this more than another cloud notebook is who built it and why. Aaron spent time working on Terra at the Broad Institute, the largest open-source genomics platform in the world. He knows exactly what works about that system and exactly where it breaks down for the people who need it most. Terra is powerful, complex, enterprise-focused, and genomics-specific. If you’re a neuroscientist or a climate researcher, it doesn’t serve you. If you’re a postdoc without a bioinformatics team behind you, the learning curve alone can take weeks.

“Terra is an amazing tool for researchers, but I felt it was tightly coupled to specific genomic hypotheses,” Kanzer says. “Exploratory data analysis and fine-tuning require flexibility, so Scigantic is trying to abstract data engineering and ultimately let the researcher decide.” The distinction matters. Terra serves genomics well. If you are a neuroscientist, a climate scientist, or simply someone who wants to explore data before committing to a hypothesis, it was not built for you. “Data-scaling issues are not isolated to genomics or life sciences,” Kanzer adds. “They occur in all forms of science.”

One example of what “differently” looks like in practice: comparing ESM predictions against OpenFold outputs. Today, that means two separate environments, two sets of dependencies, two data pipelines. On Scigantic, it’s one notebook session. Both models, same data, side by side.

The platform also includes an AI assistant built into JupyterLab that has context on every mounted dataset’s schema. It can point you toward relevant columns, suggest analyses, or nudge you back if you’re heading in an unproductive direction. It is not a chatbot bolted onto a product. It knows what data you’re looking at.

Who it’s for

Scigantic is built for academics first. Postdocs, PhD students, researchers at institutions that don’t have dedicated cloud engineering teams (which is most of them). The pricing reflects that: a free tier exists, and the researcher plan is $19 per month.

The logic is deliberate, not a limitation. “The most impactful scientific breakthroughs originate in academia before being commercialised,” Kanzer says. “The Human Genome Project. The Protein Data Bank. Scigantic believes that prioritising academics first will lead to a much larger impact later on.” Today’s postdoc running analyses on the free tier is tomorrow’s head of computational biology choosing infrastructure for an entire department.

Early feedback suggests the platform points researchers in productive directions faster than existing tools. The AI assistant in particular seems to help people who know what question they want to ask, but don’t know which dataset or which column holds the answer.

Early signs suggest it is working. One MIT postdoc put it directly: “Scigantic’s ability to quickly navigate large S3 buckets is incredibly valuable. Gemini, Claude, Codex, etc. try to download large datasets in real-time, often never getting a solid answer to my question. I’m thoroughly impressed how fast it gets me onboarded to the data.” The bar here is not perfection. It is whether the platform gets a researcher to a productive starting point faster than the alternatives. By that measure, it seems to be landing.

The future

Aaron has started conversations with multiple renown research institutions. He’s exploring what a partnership with Kiin could look like. The vision is to become the default access layer between open scientific data and the researchers who need it, across every domain, not just genomics.

“We see multiple partnerships forming with Scigantic in the next year, further validating our hypothesis that navigating unstructured, large-scale data is the bottleneck,” Kanzer says. NASA conversations are already underway. The thesis is simple: if the bottleneck is access, and access keeps getting harder as datasets grow, then the platform that solves it once becomes the default layer everyone builds on.

The fact that this is bootstrapped and solo matters. There’s no board pushing toward enterprise sales or premature monetisation. The roadmap can stay focused on what academics actually need, for now. That is a strategic choice, not a limitation.

Kiin’s view

The “why now” here is straightforward. Open data mandates from NIH, Wellcome, and the EU are flooding repositories with petabytes of new data every year. Foundation models for biology are proliferating faster than any lab can keep up. The gap between “data exists” and “I can work with it” is widening, not closing. Someone was going to build this access layer. The fact that it’s being built by someone who spent years inside Terra, who understands both the infrastructure and the user pain at a level most founders don’t, is what makes it credible.

The risk is execution at scale with a single person. The opportunity is that the product is simple enough in concept (notebook + mounted data + fine-tuning) that it doesn’t need a 50-person engineering team to work. It needs to work reliably for the 10,000 postdocs who currently can’t access the data their own field generated. If it does that, the enterprise customers will follow.

Thanks for reading Kiin Bio Weekly!

Share Kiin Bio Weekly

💬 Get involved

We’re always looking to grow our community. If you’d like to get involved, contribute ideas or share something you’re building, fill out this form or reach out to me directly.

Subscribe now to stay at the forefront of AI in Life Science and keep up with this upcoming season of deep dives.

Connect With Us

Have questions on this or suggestions for our next deep dive? We’d love to hear from you!

📧 Email Us | 📲 Follow on LinkedIn | 🌐 Visit Our Website

• C3P uses proteins as a training signal to learn promoter representations, which lets you find co-regulated genes across bacterial genomes without any experimental data. Simple idea, works well.

• BioMatrix is another biological foundation model, but this one actually covers the full matrix: molecule sequences, molecule structures, protein sequences, protein structures, and natural language, all in one model. State-of-the-art or competitive on 77 of 80 benchmarks.

• JEDEL automates DNA-encoded library design from a pharmacophore, and everything it generates is synthesisable from purchasable building blocks. That constraint is what makes it useful rather than academic.

Kiin Pioneer Programme

We built a platform that helps researchers speed up their entire science, from literature review and biomarker discovery to bioinformatics and computational chemistry. If your workflow involves pulling findings from five different places before you can actually act on any of them, this is for that.

The Pioneer Programme gives academic labs and non-profits one year of free access, plus support from our science team. No cost, no data transfer, all IP stays with your institution. Applications close August, cohort starts September.

Read more about the programme

Apply now

C3P: Contrastive promoter-protein pretraining for bacterial gene regulation

🧪 Where This Fits

Genome language models (like DNABERT, Nucleotide Transformer, Evo) have gotten good at learning sequence representations, but they still struggle with regulatory DNA. Promoters are short, non-coding, and their function depends on context that pure sequence models have trouble capturing. The standard approach is to pretrain on raw DNA and hope the model picks up regulatory grammar along the way. It mostly does not.

C3P takes a different angle entirely. Rather than trying to learn promoter function from DNA sequence alone, it uses the protein that a promoter regulates as a supervisory signal. The logic: promoters that regulate similar functions should have similar representations, and protein language models already capture functional similarity well. So you can transfer that knowledge to the promoter side through contrastive learning. It is a simple idea, and arguably obvious in hindsight, but nobody has done it at this scale before.

🔍 What It Is

• Genome language models learn poor representations of regulatory DNA, limiting their use for predicting gene regulation. Dufault, Xu, and Moses from the University of Toronto present C3P, a contrastive learning framework that pairs bacterial promoters with their downstream proteins to learn promoter representations.

• Trained on 88 million promoter-protein pairs using a CLIP-style objective. The promoter encoder learns to align with embeddings from a frozen protein language model, so promoters regulating functionally similar proteins end up with similar representations.

• Multi-fold improvement over leading genome language models on regulatory annotation prediction. Enables zero-shot co-regulated gene retrieval: given a promoter, find other promoters driving similar functions, across genomes, with no experimental data required.

💡 Why This Is Cool

The clever bit here is recognising that we already have a strong signal for regulatory function, it just lives in protein space rather than DNA space. Protein language models have encoded functional relationships that took decades of biochemistry to establish. C3P bridges that knowledge back to the regulatory side. This matters for microbiology because most bacterial genomes have no experimental regulatory data at all. If this approach generalises beyond the training distribution (which the zero-shot retrieval results suggest it might), it opens up regulatory annotation for millions of uncharacterised organisms. The limitation is that it only works where you have a clear promoter-protein pair, which excludes non-coding RNAs and complex eukaryotic regulation.

📃 Read the paper.

💻 Try the code.

BioMatrix: A comprehensive biological foundation model spanning sequences, structures, and language

🧪 Where This Fits

The biological foundation model space has been fragmented. You have protein language models (ESM, ProtTrans), molecule models (MolBERT, ChemBERTa), structure predictors (AlphaFold, ESMFold), and various attempts to combine two of these modalities. What nobody has done convincingly is put all five modalities (molecule sequence, molecule structure, protein sequence, protein structure, and natural language) into one architecture that can both read and generate all of them. Previous multi-modal attempts (BioMedGPT, Galactica) either used modality-specific encoders bolted onto a language model, or covered text plus one other modality.

BioMatrix from Tsinghua and collaborators argues you do not need specialised encoders at all. You tokenise everything into a shared vocabulary, train with standard next-token prediction, and let the model sort out the relationships. The scale helps: 304 billion tokens across all modalities, built on Qwen3 at 1.7B and 4B parameters.

🔍 What It Is

• Existing biological AI models are specialised to one or two data types and cannot transfer knowledge between molecules and proteins natively. Pei et al. from Tsinghua present BioMatrix, a multimodal foundation model that unifies molecule sequences, molecule structures, protein sequences, protein structures, and natural language in a single architecture.

• All modalities are mapped to a shared token space using learned structure tokenisers (VQ-VAE for molecular and protein 3D structures). The model is then trained with a standard next-token prediction objective, with no external encoders or modality-specific output heads. Built on Qwen3, trained on 304.4 billion tokens.

• State-of-the-art or competitive on 77 out of 80 downstream tasks across property prediction, molecule generation, protein folding, captioning, and binding affinity prediction. Outperforms modality-specific specialist models on several benchmarks.

💡 Why This Is Cool

The interesting question is not “does a bigger model do well on benchmarks” (it does, unsurprisingly) but whether unifying modalities in one model creates emergent cross-modal capabilities that specialist models cannot replicate. The paper does not fully answer this yet, and most of the 80 tasks are single-modality evaluations that a specialist could handle. The real test will be tasks that require reasoning across modalities simultaneously: “given this protein structure and this molecule, predict binding and explain why.” If BioMatrix can do that better than a pipeline of specialists, the unified approach is vindicated. If it just matches specialists on their own turf, the value proposition is convenience rather than capability. The Apache 2.0 license and public weights make it easy to test either way.

📃 Read the paper.

💻 Try the code.

JEDEL: Zero-shot DNA-encoded library design for early-stage drug discovery

🧪 Where This Fits

DNA-encoded libraries (DELs) are combinatorial chemistry at industrial scale: you attach DNA barcodes to building blocks, combine them through validated reactions, and screen the resulting millions of compounds against a target. The design problem is choosing which building blocks and reactions to include. Traditional DEL design relies on chemical diversity heuristics or brute-force enumeration, neither of which accounts for whether the resulting library will actually produce binders for your specific target.

Generative drug design models (like those from Recursion, Insilico, etc.) can propose target-specific molecules, but they produce virtual structures that then require separate synthesis planning, which often fails. JEDEL closes this gap by designing libraries that are target-aware and synthesisable by construction: every output is built from purchasable building blocks through validated reactions.

🔍 What It Is

• Current DEL design ignores target information, and generative models propose molecules that are often unsynthesisable. Jocys et al. from the University of Surrey present JEDEL, a framework that converts 3D pharmacophore information from known active ligands into synthesis instructions for target-focused DNA-encoded libraries.

• Takes pharmacophore representations as input, searches a space of purchasable building blocks and validated combinatorial reactions, and assembles libraries that are synthesisable by definition. Requires no target-specific retraining, operating in a zero-shot manner across different protein targets.

• Tested across 18 protein targets, JEDEL outperformed random and diversity-based baselines on predicted binding affinity, pharmacophore recovery, and sample efficiency. The constraint to purchasable reagents means every designed library can be made in the lab immediately.

💡 Why This Is Cool

DELs are already a workhorse in pharma (Novartis, GSK, and X-Chem all run them at scale), so improvements here have a clear commercial path. What JEDEL does is shift library design from “maximise chemical diversity and hope something binds” to “build a library biased toward your target’s pharmacophore.” The zero-shot aspect means you do not need to retrain for each new campaign. The synthesisability constraint is the real differentiator from generative models: the gap between “here is a molecule that might bind” and “here is a library you can make on Monday” is where most computational drug design papers lose their translational value. The limitation is that pharmacophore inputs require existing active compounds, so this is a hit expansion tool rather than a de novo discovery method. For teams already running DEL campaigns, this looks immediately applicable.

📃 Read the paper.

No public code repository available at time of writing.

🗓️ Events & Competitions

The best competitions, hackathons, and community challenges in AI x life sciences, curated weekly. Know something worth featuring? Reply and let us know.

More upcoming events:

BioHackathon Europe 2026 | November 9-13, Barcelona

ELIXIR’s annual international bioinformatics hackathon, running since 2018. 160+ participants, five days of collaborative coding on open bioinformatics infrastructure and tools. The call for project proposals has now closed.

Thanks for reading!

💬 Get involved

We’re always looking to grow our community. If you’d like to get involved, contribute ideas or share something you’re building, fill out this form or reach out to me directly.

Connect With Us

Have questions or suggestions? We'd love to hear from you!

📧 Email Us | 📲 Follow on LinkedIn | 🌐 Visit Our Website

• For computational chemists, drug discovery scientists, and anyone running or interpreting docking campaigns.

• Docking scores do not predict binding affinity. The OpenBind benchmark confirmed it: molecular weight alone outperformed Gnina, Boltz-2, and other state-of-the-art models on experimental binding data.

• The real value of docking is pose prediction, not ranking compounds. Most teams using it well have stopped asking it to estimate affinity altogether.

• This piece covers where docking fails, where it works (template-guided lead optimisation), and why pooling imperfect methods beats waiting for a perfect one.

Freebie alert: We know how hard science is. That’s why we built the Pioneer Programme.

We’re selecting academic and nonprofit teams to get one year of free access to our drug discovery platform, with support from our science team. If you spend more time pulling together findings from different sources than actually acting on them, it’s worth applying.

No cost, no data transfer, all IP stays with your institution. Applications close August, cohort starts September.

Read more about the programme

Apply now

For this piece, I spoke with Rachael Skyner, our cheminformatician at Kiin Bio, about the mechanics of molecular docking, why scoring remains unsolved, and how template-guided approaches are changing what is possible in lead optimisation. Her insight runs throughout.

“The search problem is not the hard part. It’s ranking them and picking the right pose.”

Why this matters now

Thanks to AlphaFold and its successors, drug discovery teams now have predicted structures for essentially every human protein. Generative chemistry tools are producing novel molecules faster than medicinal chemists can evaluate them. Having a structure to dock into is no longer the problem. Trusting what the docking tells you is. The OpenBind benchmark, published earlier this year, put hard numbers on this trust deficit for the first time, and the results were not encouraging.

Docking can tell you roughly where a molecule might land. It cannot reliably tell you how tightly it will stick. Most practitioners have adapted by splitting the problem in two: get the pose right first, worry about affinity later. That distinction shapes everything about how docking is actually used today.

The scoring problem nobody has solved

Traditional docking scores are supposed to estimate binding affinity. They do not. The physics approximations used in scoring are rough even for a static system, and on top of that, docking ignores the fact that binding is actually a dynamic process. Binding in the body involves solvent reorganisation, protein conformational changes, entropy costs, and timescale-dependent interactions that no static pose can capture.

“It’s an oversimplification of what’s actually going on with binding,” says Skyner. “It’s not just binding that contributes to affinity in the end. There are lots of other things that can influence how strongly a molecule binds. You’re treating it as a static problem when in reality it’s very dynamic.”

Increasingly, people are separating pose prediction from affinity estimation altogether. Modern scoring methods from tools like Gnina use convolutional neural networks trained on the PDBbind dataset. Rather than predicting binding affinity directly, the CNN score predicts confidence in whether the pose is correct, which is a more tractable and more useful question to ask.

Protein flexibility and the induced fit problem

Traditional rigid docking assumes the protein does not move, which is obviously wrong. Proteins reshape themselves around ligands in a process called induced fit, and this is why rigid receptor docking often fails for novel chemotypes.

Physics-based approaches like Rosetta and Schrodinger’s IFDMD allow limited residue rotation or flipping at the binding site. Machine learning methods like DiffDock take a generative approach, producing the molecule directly into the protein. Co-folding methods fold the protein and ligand simultaneously, addressing flexibility implicitly by allowing both to move at the same time during prediction.

The problem is that induced fit methods produce multiple protein conformations alongside multiple ligand conformations, and the analysis becomes exponentially more complex. “From an analysis point of view, it’s really difficult to deal with those structures,” says Skyner. “You have to start doing more complicated things like clustering together both the ligand and the protein conformation to see which ones come up most often.”

For high-throughput virtual screening where you want to process hundreds of thousands of molecules, this complexity is impractical. Induced fit docking belongs in the later stages of a project, when you already have a binder and want hypotheses about its binding mode.

Traditional rigid docking assumes the protein does not move. Proteins move constantly. They reshape themselves around ligands in a process called induced fit, and this is why rigid receptor docking often fails.

The virtual screening funnel

In practice, docking is rarely used alone. It sits inside a virtual screening funnel: start with a compound library (thousands to millions of molecules), run fast docking as a heuristic filter (anything scoring worse than minus eight kilocalories per mole gets discarded), then rescore the surviving hits with more expensive methods.

Rescoring typically involves energy minimisation with solvent. Methods like MMGBSA and MMPBSA incorporate water into the calculation, which basic docking ignores entirely. “When you start thinking about the solvent as well, you start getting much better correlation with experimental affinities,” says Skyner. These methods are computationally expensive, though still feasible for the few hundred molecules that survive initial filtering.

ML-based rescoring adds another option: build a local model of the docking score and then only dock the subset your model flags as likely to score well. This can cut the number of molecules you need to physically dock in half.

Where docking actually works: template-guided lead optimisation

Docking works best in template-guided approaches during lead optimisation, not in blind global searches or massive virtual screens. This is where Skyner’s work focuses.

At this stage, you already have a molecule with known binding evidence. You are working through a congeneric series: molecules that share a common core (the maximum common substructure, or MCS) with small modifications around the periphery. A chlorine added here, a fluorine swapped there, an atom changed in a ring. You already know the molecule binds. What you want to understand is whether the modification changes how it sits in the pocket.

Template-guided docking uses the known crystal structure of a reference compound to anchor the search. Rather than exploring the entire binding pocket from scratch, you start from the position of the known binder and run a local minimisation. Skyner has developed an MCS-guided docking approach that generates low-energy conformers of the ligand, rigidly aligns them to the reference structure via the MCS, and performs local optimisation from that aligned position using AutoDock Vina.

“If you give it the right starting position, it’s got a much better chance of getting the pose correct,” says Skyner. “People might do this by default, but it really, really helps in these scenarios.”

Ensembles, pooling, and what comes next

No single docking method dominates across all targets. In Skyner’s evaluation of 22 targets (1,888 structures), pooling results from multiple methods and rescoring with Gnina’s CNN score consistently outperformed any individual approach. The pooled “oracle” (checking whether any method generated the correct pose) found it 90% of the time using RMSD <2 angstroms as the heuristic for a correct binding pose, and 54% by the more stringent SuCOS metric, which evaluates shape and pharmacophore overlap rather than just atomic positions.

The practical recommendation: combine Vina (fast, global search) with MCS-guided Vina (template-aware), pool the poses, and rescore with the CNN score. Two methods captured nearly all the benefit of three, with substantially less compute.

The preparation of proteins and ligands is still the most error-prone step in any docking workflow. Get the protonation states wrong, miss a stereoisomer, assign incorrect charges, and the docking algorithm has no chance of finding the right answer. Errors compound downstream into molecular dynamics and other physics-based follow-up methods. “If you get it wrong from the beginning, it really affects everything else,” says Skyner.

Docking will not predict your binding affinities, and it will not reliably rank compounds within a series. The industry spent years asking it to do both, and the benchmarks now confirm what practitioners quietly knew. The gains are coming from asking docking the right question in the right context: use it for pose prediction rather than affinity estimation, use template-guided exploration rather than blind search, pool results from multiple methods rather than betting on one. The teams actually getting value out of docking today have mostly just stopped asking it to be something it was never designed to be.

💬 Want to be featured in Kiin Bio Weekly?

Each issue we speak directly with researchers, scientists, and builders working at the frontier of AI in life sciences. If you're working on something in this space and think it would resonate with our community, I'd love to hear from you. Fill out this form or reach out to me directly.

Found this useful? Forward it to a colleague in computational biology, it's the best way to help the newsletter grow.

Share Kiin Bio Weekly

Subscribe now to stay at the forefront of AI in Life Science. Every week: primers, deep dives, and direct conversations with the people building the field.

Subscribe now

Connect With Us

Have questions on this or suggestions for our next deep dive? We’d love to hear from you!

📧 Email Us | 📲 Follow on LinkedIn | 🌐 Visit Our Website

• Google’s AMIE system matched or beat primary care physicians on clinical management reasoning across 100 multi-visit scenarios. Published in Nature, and the study design is more rigorous than most in this space.

• CANVAS turns standard H&E slides into virtual spatial proteomics maps, predicting tumour microenvironment neighbourhoods from cheap histology. Validated across 5,000 patients and 9 cancer types.

• AMPGAN v3 is the first generative model for antimicrobial peptides that handles non-canonical amino acids and chemical modifications. Two of five generated candidates showed real antimicrobial activity. They also built an agentic pipeline around it, which is where this gets interesting.

Kiin Pioneer Programme

We built a platform that helps researchers speed up their entire science, from literature review and biomarker discovery to bioinformatics and computational chemistry. If your workflow involves pulling findings from five different places before you can actually act on any of them, this is for that.

The Pioneer Programme gives academic labs and non-profits one year of free access, plus support from our science team. No cost, no data transfer, all IP stays with your institution. Applications close August, cohort starts September.

Read more about the programme

Apply now

AMIE: Conversational AI for disease management

🧪 Where This Fits

Most AI-in-medicine papers test whether a model can diagnose from a static vignette. That is a solved problem at this point, or at least a well-explored one. What has not been tested seriously is whether an AI system can manage a patient over time: adjust treatment plans across multiple visits, respond to new lab results, and prescribe medications safely. That is what primary care actually involves, and it is where AMIE (Articulate Medical Intelligence Explorer) from Google DeepMind now enters.

Previous AMIE work showed the system could match physicians on diagnostic conversations. This paper extends it to management reasoning, which is harder because it involves sequential decisions, guideline interpretation, and medication safety. The comparison set is interesting: 21 board-certified primary care physicians across 100 multi-visit case scenarios grounded in UK NICE and BMJ Best Practice guidelines.

🔍 What It Is

• Management reasoning for longitudinal care is largely untested for AI. Liévin et al. from Google DeepMind present an agentic version of AMIE optimised for multi-visit clinical management and medication reasoning.

• Uses Gemini’s long-context window to combine in-context retrieval of clinical guidelines with structured reasoning. A multi-agent architecture handles dialogue, management planning, and medication lookup from drug formularies (OpenFDA, BNF).

• In a blinded virtual OSCE study, AMIE scored significantly higher than PCPs on treatment preciseness (96% vs. 62%, p<0.001) and guideline alignment (93% vs. 75% by visit 3). On the RxQA medication benchmark, AMIE outperformed PCPs on harder questions (57.9% vs. 47.8%, p<0.001). Specialist physicians and patient actors preferred AMIE 47% of the time vs. 7% for PCPs.

💡 Why This Is Cool

The honest reaction here is: this is impressive and somewhat uncomfortable. AMIE is not just pattern-matching against guidelines, it is reasoning about how to adjust a plan given what happened at the last visit. The study design (blinded OSCE, specialist evaluators, real guidelines) is more credible than most in this space. The medication reasoning results are particularly notable because prescribing errors are a leading cause of preventable harm. That said, this is still a virtual scenario, not a real clinic with real patients who do unexpected things. The gap between “performs well in structured evaluation” and “can safely manage my mum’s hypertension” remains large. What this does establish is that the technical capability exists. The regulatory and deployment questions are now the binding constraint, not the model performance.

📃 Read the paper.

💻 Try the code.

CANVAS: Virtual spatial tumor profiling from histopathology

🧪 Where This Fits

Spatial proteomics (CODEX, MIBI, etc.) can map the tumour microenvironment at single-cell resolution, but it costs thousands per sample and requires specialised equipment. Standard H&E histopathology costs almost nothing and is already collected for every cancer patient. The question is whether you can infer the spatial biology from the cheap stain. Previous attempts have tried to predict individual protein expression from H&E, but that approach is fragile: sensitive to staining variation, limited to a handful of markers, and accuracy drops quickly. CANVAS takes a different approach, predicting cellular neighbourhood patterns rather than individual proteins, which is a more robust prediction target.

This connects to a broader trend in computational pathology where foundation models (UNI, Virchow, CONCH) have made feature extraction from H&E much more powerful, and the question is now what downstream tasks those features can support. CANVAS uses these pretrained features to bridge modalities rather than training from scratch.

🔍 What It Is

• Spatial proteomics reveals the tumour microenvironment in detail but costs too much to scale. Li et al. from Stanford present CANVAS, an AI platform that predicts spatial cellular neighbourhood structures directly from standard H&E slides, trained on an atlas of 18 million cells profiled by 41-plex CODEX imaging.

• CANVAS defines 10 reproducible cellular neighbourhoods from CODEX data across 457 lung cancer patients, then trains a pathology foundation model to predict these neighbourhood patterns from co-registered H&E images. It operates at the ecological niche level rather than individual cell types.

• Applied to over 5,000 patients across 9 cancer types, CANVAS-derived spatial features predicted immunotherapy response with AUCs above 0.75 at 6, 12, and 24 months. The spatial signature stratified patients by progression-free survival (HR = 2.42, p<0.001) and outperformed established biomarkers including TMB, PD-L1 expression, and TLS. Validated externally on a Cancer Moonshot Biobank cohort.

💡 Why This Is Cool

This matters for a specific reason: immunotherapy response prediction is a clinical problem where existing biomarkers (PD-L1, TMB) work poorly. About 20-30% of patients respond to checkpoint inhibitors, and we are bad at predicting who they will be beforehand. CANVAS proposes that the spatial organisation of the tumour microenvironment, inferred from a slide that already exists in every pathology lab, is more informative than the molecular markers we have been relying on. If the external validation holds up across broader cohorts (the Moonshot cohort is small at n=40), this could actually change who gets prescribed immunotherapy. The non-commercial license limits immediate industry adoption, but for academic cancer centres this is usable now.

📃 Read the paper.

💻 Try the code.

AMPGAN v3: Agentic discovery of non-canonical antimicrobial peptides

🧪 Where This Fits

Antimicrobial resistance causes over a million deaths annually, and no new antibiotic class has been commercialised since 2000. Antimicrobial peptides (AMPs) are attractive because they disrupt membranes through physical interactions, making resistance harder to develop. Generative models for AMP design exist (PepGAN, HydrAMP, AMP-Designer), but they all share two limitations: they only work with natural L-amino acids, and they require manual filtering of outputs. Real therapeutic peptides need D-amino acids and terminal modifications to survive in the body. AMPGAN v3 is the first generative model that handles these non-canonical chemistries, and PepCraft wraps it in a multi-agent pipeline that automates the filtering.

The field has been generating lots of candidate peptides computationally, but the translation gap to actual antimicrobials has been wide. Most papers stop at predicted activity scores. This one synthesises candidates and tests them against real bacteria, which is the bar that matters.

🔍 What It Is

• Generative models for antimicrobial peptides are limited to natural amino acids and produce outputs that need extensive manual curation. Jung et al. from the University of Vermont and Purdue present AMPGAN v3, a conditional GAN that generates antimicrobial peptides with D-amino acids and terminal modifications, paired with PepCraft, a multi-agent framework for automated AMP discovery.

• AMPGAN v3 separates adversarial training across two discriminators: one for sequence realism, one for antimicrobial activity prediction. This fixes the training instability that plagued earlier versions (only ~10% of AMPGAN v2 runs produced usable models). PepCraft uses a Planning Agent to coordinate specialised executors for generation, physicochemical filtering, and database verification.

• Two of five synthesised candidates showed clear antimicrobial activity against Gram-positive strains, with the best reaching MIC of 8 μg/mL against B. subtilis. The candidates spanned three structural classes (alpha-helical, beta-hairpin, random coil) and incorporated D-amino acids and amidation, which previous generative methods cannot produce. PepCraft’s prioritisation recommendations aligned with the wet-lab results.

💡 Why This Is Cool

The wet-lab hit rate (2/5) is respectable for a generative model, and the chemical space expansion is the real contribution. Every other AMP generator is restricted to natural amino acids, which means their outputs degrade rapidly in serum. Expanding the vocabulary to include D-amino acids and terminal caps makes the generated peptides actually viable as therapeutics rather than just interesting sequences. The agentic pipeline (PepCraft) is early-stage and exploratory, but it points toward a pattern we will see more of: generative models wrapped in verification agents that can filter, cross-reference, and prioritise without human intervention. This was accepted at the ICML 2026 GenBio workshop, not a top venue, and the validation is limited to Gram-positive bacteria. Worth watching, not yet proven at scale.

📃 Read the paper.

💻 Try the code.

🗓️ Events & Competitions

The best competitions, hackathons, and community challenges in AI x life sciences, curated weekly. Know something worth featuring? Reply and let us know.

More upcoming events:

Creative Disruption Forum: Modern Drug Discovery | June 18, NIAB Cambridge

A full-day forum for biotech and R&D leaders exploring how technology is changing small molecule drug discovery. Keynote interviews with industry thought leaders followed by workshops under Chatham House Rules, limited to 60 attendees. Part of Cambridge Wide Open Week. Organised by Graham Combe and Prof Tony Sedgwick. £60 for biotech companies.

London Protein Design Day | June 23, Imperial College London

The first edition of a one-day symposium bringing together London’s protein design community and beyond. Programme spans AI-driven design, molecular dynamics, and bioinformatics, with applications across enzymes, antibodies, and materials. Organised by Pietro Sormanni, Rebecca Birolo, and Jakub Lála. In person only.

BioHackathon Europe 2026 | November 9-13, Barcelona

ELIXIR’s annual international bioinformatics hackathon, running since 2018. 160+ participants, five days of collaborative coding on open bioinformatics infrastructure and tools. The call for project proposals has now closed.

Thanks for reading!

💬 Get involved

We’re always looking to grow our community. If you’d like to get involved, contribute ideas or share something you’re building, fill out this form or reach out to me directly.

Connect With Us

Have questions or suggestions? We'd love to hear from you!

📧 Email Us | 📲 Follow on LinkedIn | 🌐 Visit Our Website

• mRNAutilus generates entire therapeutic mRNA sequences from scratch, and the wet-lab numbers are hard to argue with: 400x over wild-type expression, beating commercial Spike constructs.

• EpiFormer brings geometric deep learning to epitope prediction with a 40% F1 boost. A nice complement to last week’s ESM binder design coverage, now from the antigen side.

• Folddisco indexes 53 million protein structures and finds structural motifs in seconds. The Steinegger lab keeps quietly building infrastructure that makes everyone else’s work faster.

Kiin Pioneer Programme

We built a platform that helps researchers speed up their entire science, from literature review and biomarker discovery to bioinformatics and computational chemistry. If your workflow involves pulling findings from five different places before you can actually act on any of them, this is for that.

The Pioneer Programme gives academic labs and non-profits one year of free access, plus support from our science team. No cost, no data transfer, all IP stays with your institution. Applications close August, cohort starts September.

Read more about the programme

Apply now

mRNAutilus: Multi-Objective-Guided Discrete Generation of mRNA with Optimized Therapeutic Properties

🔬 Designing full-length therapeutic mRNAs means optimising stability, translation efficiency, and codon usage simultaneously. Current methods tackle these objectives piecemeal, stitching together separately optimised UTRs and coding regions, which leaves performance on the table.

Patel et al. from the Chatterjee lab at Duke present mRNAutilus, a generative framework that designs complete mRNA transcripts optimised across multiple properties at once.

🧬 The system trains a masked discrete diffusion model on millions of full-length mRNAs, then steers generation with Monte Carlo tree guidance to hit multiple objectives without retraining. It operates on whole transcripts rather than modular components.

⚡ Zero-shot mRNAutilus designs encoding firefly luciferase achieved over 400-fold higher expression than wild-type, outperforming commercial baselines. For SARS-CoV-2 Spike, designs matched or surpassed both clinically used constructs and lab-optimised sequences. The framework also generalised to prime editing guides and targeted protein degradation, which suggests this is not a one-trick benchmark result.

🧪 Where This Fits

mRNA therapeutics have had a sequencing problem disguised as a design problem. Post-COVID, the bottleneck is no longer “can we make mRNA drugs?” but “can we make them well enough to compete on potency and manufacturing cost?” Tools like LinearDesign (Zhang lab, 2023) optimise coding sequences for stability, and UTR-focused approaches pick regulatory elements, but these treat the transcript as a set of independent modules. mRNAutilus is the first framework I have seen that treats the entire transcript as a single generative object, which matters because interactions between UTRs and coding regions affect folding and translation in ways modular approaches miss.

The wet-lab validation is what separates this from yet another generative model paper. Beating commercial constructs for Spike expression is a meaningful bar, not an in silico benchmark. The generalisation to prime editing and degraders suggests the architecture is flexible enough to not be overfit to reporter assays. The timing makes sense too: masked diffusion models have matured enough (thanks to protein and genomics applications) that applying them to mRNA sequences is a natural next step, and the Monte Carlo tree guidance borrows from AlphaGo-era decision strategies to handle multi-objective trade-offs without expensive retraining.

For readers working in mRNA therapeutics: this is worth watching closely. The code is not yet public, which limits immediate adoption, but the approach could compress the design-test cycle considerably once available.

💡 Why This Is Cool

The shift from “optimise one property” to “generate the whole thing optimised” matters more than it sounds. The history of biologics design is littered with tools that optimised one metric while inadvertently breaking another. If multi-objective generation holds up across more constructs and delivery contexts, it moves mRNA design closer to what protein design achieved with diffusion models over the past two years. The open question is whether the approach scales to longer, more complex transcripts and novel target classes beyond the well-studied ones shown here.

📃 Read the paper.

EpiFormer: Learning Antigen-Antibody Interactions for Epitope Prediction via Geometric Deep Learning

🔬 Predicting which surface residues an antibody will target on an antigen remains stubbornly difficult. Most existing methods treat the antigen in isolation, ignoring the antibody entirely or bolting antibody information on as a late-stage afterthought.

Ahmed et al. from Georgia State University introduce EpiFormer, a geometric deep learning framework that models antigen-antibody interactions through interleaved cross-attention within GNN encoding layers.

🧬 Rather than encoding antigen and antibody separately then combining representations at the end, EpiFormer threads cross-attention between the two structures at every encoding layer. This allows bidirectional information flow throughout the representation, so the model learns how antibody geometry constrains which epitope residues are accessible.

⚡ On standard benchmarks, EpiFormer achieves over 40% improvement in F1 score compared to previous best methods. That is a substantial jump for a prediction task where incremental gains of 2-5% have been typical. The model operates on 3D structural inputs from antibody-antigen complexes.

🧪 Where This Fits

Epitope prediction sits upstream of antibody engineering: if you know where an antibody binds, you can design better binders and prioritise vaccine targets. Previous approaches like DiscoTope and ElliPro use geometry and surface properties of the antigen alone, which is a bit like predicting where a key fits without looking at the lock. More recent methods (PECAN, AsEP) incorporated paratope information but typically as a separate encoding step with late fusion.

EpiFormer’s contribution is architectural rather than data-driven. The interleaved cross-attention ensures antibody context informs antigen representations from the start rather than being concatenated at the decision layer. This connects naturally to last week’s coverage of ESM-based binder design: that work generates antibodies given a target, while EpiFormer predicts where on the target those antibodies will land. Together they cover both directions of the same binding prediction problem.

The 40% F1 improvement is striking, though it warrants some caution. Epitope prediction benchmarks are notoriously sensitive to train/test splitting, and structural epitope datasets remain small (a few thousand complexes in SAbDab). Whether this holds on truly novel antigen folds or just reflects better exploitation of known structural patterns is an open question. The code is available, which helps.

💡 Why This Is Cool

The “interleave information early rather than fuse late” lesson keeps appearing across structural biology. AlphaFold did it for MSA and structure tracks. ESM3 does it for sequence, structure, and function. EpiFormer applies the same intuition to a paired prediction problem. The field has been underestimating how much cross-modal information gets lost in late-fusion architectures, and each new result in this direction makes that clearer. For antibody discovery teams, this is immediately useful if it generalises beyond the benchmark setting.

📃 Read the paper.

💻 Try the code.

Structural Motif Search Across the Protein Universe with Folddisco

🔬 Finding recurring 3D structural motifs (zinc fingers, catalytic triads, protein-protein interaction surfaces) across millions of predicted structures is computationally prohibitive. Existing methods either cannot handle discontinuous motifs or choke on databases beyond a few hundred thousand structures.

Kim et al. from the Steinegger lab at Seoul National University present Folddisco, a structural motif search tool that indexes 53 million AFDB50 structures in a 1.45 TB index and returns query results in seconds.

🧬 Folddisco encodes proximal residue pairs into geometric feature sets (distances, angles, and side-chain orientation via torsion angles), stores them in a position-independent inverted index, and ranks hits using an IDF-based coverage score that rewards rare features. This handles both short continuous motifs and long discontinuous ones.

⚡ Indexing is 11x faster to build and 4x more storage-efficient than previous state-of-the-art. Query speed is 20-fold faster than pyScoMotif on the full pipeline. On the zinc finger benchmark against the human proteome, Folddisco outperformed both RCSB and pyScoMotif on recall while maintaining higher precision. It also successfully distinguished active from inactive GPCR conformational states using activation motifs.

🧪 Where This Fits

This is infrastructure work, and the most consequential kind. The AlphaFold database gave us 200+ million predicted structures, but searching them structurally has lagged far behind searching them by sequence (where tools like Foldseek, also from the Steinegger lab, already operate at scale). Folddisco fills the motif-search gap: given a 3D pattern of interest, find everywhere it occurs across all known and predicted protein structures.

The practical value becomes clear with the GPCR example. Being able to query “show me all structures with this activation motif” across both experimental PDB structures and AlphaFold predictions means you can study conformational states at proteome scale. That was previously manual curation work. Similarly, the zinc finger detection in uncharacterised metagenomic proteins (from ESM30) demonstrates functional annotation where sequence-based methods fail entirely.

Folddisco’s main limitation is its 20-angstrom connectivity constraint, which means it cannot detect spatially distant functional sites like remote allosteric pockets. The IDF scoring also struggles with very short motifs. These are known trade-offs for the speed gains.

💡 Why This Is Cool

The Steinegger lab has been building the search infrastructure for the structure-prediction era piece by piece: MMseqs2 for sequences, Foldseek for structure alignment, and now Folddisco for motif search. Each tool makes the previous one more useful. What matters here is not the individual benchmarks but the fact that motif search at 53-million-structure scale is now a webserver query rather than a compute cluster job. That means anyone with a structural intuition and a browser can generate hypotheses that previously required a compute cluster and custom code. The webserver is live at search.foldseek.com/folddisco.

📃 Read the paper. 💻 Try the code.

🗓️ Events & Competitions

The best competitions, hackathons, and community challenges in AI x life sciences, curated weekly. Know something worth featuring? Reply and let us know.

More upcoming events:

Creative Disruption Forum: Modern Drug Discovery | June 18, NIAB Cambridge

A full-day forum for biotech and R&D leaders exploring how technology is changing small molecule drug discovery. Keynote interviews with industry thought leaders followed by workshops under Chatham House Rules, limited to 60 attendees. Part of Cambridge Wide Open Week. Organised by Graham Combe and Prof Tony Sedgwick. £60 for biotech companies.

London Protein Design Day | June 23, Imperial College London

The first edition of a one-day symposium bringing together London’s protein design community and beyond. Programme spans AI-driven design, molecular dynamics, and bioinformatics, with applications across enzymes, antibodies, and materials. Organised by Pietro Sormanni, Rebecca Birolo, and Jakub Lála. Abstract deadline for poster/oral presentations is this Saturday (May 17). In person only.

BioHackathon Europe 2026 | November 9-13, Barcelona

ELIXIR’s annual international bioinformatics hackathon, running since 2018. 160+ participants, five days of collaborative coding on open bioinformatics infrastructure and tools. The call for project proposals opens March 16 and closes April 15 - so if you want to lead a project, that’s your window.

Thanks for reading!

💬 Get involved

We’re always looking to grow our community. If you’d like to get involved, contribute ideas or share something you’re building, fill out this form or reach out to me directly.

Connect With Us

Have questions or suggestions? We'd love to hear from you!

📧 Email Us | 📲 Follow on LinkedIn | 🌐 Visit Our Website

Who this piece is for: Computational biology professionals and hiring managers navigating the UK market in 2026.

What this covers: Where the roles are, what they pay, and what actually gets people hired, based on recruiter data and 750+ live listings.

The takeaway: The market rewards specialists who can ship, not generalists who can apply. Infrastructure roles are where demand is highest, entry-level is brutally oversaturated, and your visibility matters more than your credentials.

Freebie alert: We know how hard science is. That’s why we built the Pioneer Programme.

We’re selecting academic and nonprofit teams to get one year of free access to our drug discovery platform, with support from our science team. If you spend more time pulling together findings from different sources than actually acting on them, it’s worth applying.

No cost, no data transfer, all IP stays with your institution. Applications close August, cohort starts September.

Read more about the programme

Apply now

The computational biology job market in 2026 has split in two. We wanted to dive deeper to understand why, and try to think about what might happen next.

Traditional pharma has been making considerable cuts for over a year. Bayer alone cut roughly 14,500 roles between 2022 and 2026, with the steepest reductions in 2024 and 2025. BMS and Teva followed with thousands more. AI-native biotech, meanwhile, is hiring faster than the talent pool can seem to keep up. As a quick example, Isomorphic Labs has 21 open ML drug discovery roles in London alone.

We looked into the market dynamics, skill levels, salary benchmarks, and hiring patterns across 2,000+ UK listings on LinkedIn, and spoke to Joe Phillips, Principal Consultant BioAI at Cubiq Recruitment, a specialist recruiter in bio-AI and computational life sciences. Everything we found is in here: where the roles are, what they pay, who’s getting hired, and what we think happens next.

“There’s an enormous amount of strong academic talent coming through, but the number of genuinely junior opportunities is tiny compared to demand. The candidates who stand out usually have something beyond their degree alone to show.”

📊 Where the roles are

Across the UK, there are currently 2,000+ open positions in computational biology, bioinformatics, AI drug discovery, and adjacent infrastructure roles (source: LinkedIn Jobs, May 2026). These span big pharma, AI-native biotech startups, NHS trusts, CROs, and platform companies selling into life sciences.

• MLOps and platform engineering (1,000+ roles): The biggest category by far, and probably the most surprising if you haven’t been watching the infrastructure side of biotech. Why so many? McKinsey’s 2025 State of AI report found that 88% of companies now use AI in at least one business function (up from 78% the year before), but roughly two-thirds are still stuck in pilot mode. Companies built research teams over the last few years, proved that their models work, and now need people who can operationalise them at scale. Joe says this is the biggest shift he’s seen: “A lot of this work was previously absorbed by ML Engineers. Now the cost and complexity around compute, GPU utilisation, and inference has become significant enough that firms are hiring specialists.”

• Bioinformatics (355 roles): The broadest category, spanning clinical bioinformatics, genomics, spatial and single-cell analysis. Also the most accessible at entry level: 33% of bioinformatics listings are entry-level, compared to just 12% in ML drug discovery. For early career candidates, this is where the door is most open, across both big pharma and smaller biotech. Startups tend to offer faster progression and broader scope; pharma offers stability and established infrastructure.

• Computational biology (164 roles): Core comp bio and adjacent research scientist roles. Spans both pharma and biotech.

• ML in drug discovery (137 roles): Protein design, ADMET prediction, molecular modelling. Isomorphic Labs dominates with 21 positions, followed by Relation Therapeutics with 16. Newer players like CuspAI and Boltz have also been hiring heavily over the past 12 months. Small in absolute numbers, but likely one of the fastest-growing categories year on year: the AI in drug discovery market is expanding at around 25% annually (Verified Market Research). Even if the number of roles today looks modest, the trajectory and the funding flowing in suggest this will look very different by 2028. The roles that exist here tend to be senior, well compensated, and competitive.

• Clinical AI (~500 roles): The broadest umbrella, covering health informatics through to clinical data science.

• AI protein design (9 roles): Tiny, highly specialised, and almost exclusively mid-senior level.

The other trend Joe flags: forward-deployed engineers and technical client facing hires. “A lot of companies are commercialising scientific ML platforms now rather than running their own therapeutics pipelines, so they need engineers and scientists who can comfortably operate across product, research, and client conversations.” GTM hiring is picking up too as firms move beyond pure research mode.

📍 Geography: the Golden Triangle and beyond

If you’ve been paying attention to the London tech scene, the top of this list won’t surprise you.

“Kings Cross is obviously on fire at the moment,” says Joe. “Oxford and Cambridge are also hotspots, although of the two, Cambridge is always an easier sell to candidates because of the commutability from London.”

Outside the Golden Triangle, the Northern Arc (Leeds, Liverpool, Manchester, Sheffield) is showing up as a secondary cluster, backed by Northern Gritstone funding for life science and deep tech spinouts. Glasgow also appears consistently in bioinformatics listings, driven by NHS Scotland roles. For candidates willing to look beyond the south-east, the cost of living advantage is real, particularly when London salaries don’t always come with a proportional premium.

💰 What it pays

So where do these 2,000+ roles sit on salary? UK data for comp bio is notoriously thin, which makes it hard for candidates to know when an offer is fair. Joe shared benchmarks from his recruitment work:

The ML premium is hard to ignore. A senior ML engineer in biotech can earn nearly double what a senior bioinformatician earns. That comes down to scarcity of strong ML talent and the direct commercial impact these roles carry: GPU optimisation and inference efficiency directly affect a company’s burn rate.

Joe’s caveat: “This can massively depend on the size of the company and what they are working on. There are always outliers.” Community data backs this up. Biotech equity is less liquid than FAANG RSUs, so even when base salaries match, total compensation often lags tech. The trade-off is mission, ownership, and the fact that senior bio-AI roles are closing the gap faster than any other life sciences category.

🏠 Remote, hybrid, or on-site?

Many comp bio professionals came up during the pandemic era of fully remote work, so this question comes up constantly. The answer depends on what you actually do day to day.

Bioinformatics is the outlier for remote work (46%), probably because much of the work is pipeline-based and can run independently. Drug discovery and comp bio skew heavily in-person, particularly at early-stage companies where wet lab and dry lab collaboration matters daily.

“Fully remote roles are still fairly rare,” says Joe. “If anything, more companies are trying to get people together in person more, especially at earlier stages where collaboration between science and engineering matters a lot.”

In practice, remote flexibility correlates strongly with sub-field. Bioinformatics and genomics data science offer the most options; ML drug discovery is almost entirely on-site.

🛠️ Skills that matter now

The roles exist and the salaries are there, the question is really just what actually gets you through the door as the technical bar has moved. Python is the most used programming language in the world (GitHub’s 2024 Octoverse report placed it at number one for the first time, overtaking JavaScript). R remains critical for statistical genomics. Beyond the basics, what actually separates candidates falls into three tiers.

Table stakes (expected, not differentiating):

• Python, R, Bash/Unix

• Git, Docker, AWS or GCP

• PyTorch, scikit-learn

• Nextflow or Snakemake

• Basic ML (regression, classification, clustering)

Differentiators (what gets you to the top of the pile):

• GPU optimisation and distributed systems (the single most in-demand infrastructure skill right now)

• Protein language models (e.g. ESM-2, ProtTrans)

• JAX proficiency (driven by the DeepMind/Isomorphic ecosystem)

• Production ML deployment (Kubernetes, inference optimisation)

• Cross-functional communication: being able to sit across product, research, and client conversations

Emerging (bet on these for 2027):

• Diffusion models for molecular generation (e.g. DiffDock, FrameFlow)

• Geometric deep learning and equivariant neural networks

• LLMs for biomedical data (RAG architectures, agentic AI for research)

• Foundation models for genomics (single-cell, spatial transcriptomics)

PyTorch has decisively won over TensorFlow in bio-AI research. JAX has gone from niche to essential for structural biology. Perl has disappeared from modern curricula entirely. The field moves fast enough that what was cutting-edge in 2023 (basic AlphaFold usage) is now baseline knowledge.

What we think: The agentic AI category is the one to watch. Right now, “agentic” is mostly a buzzword on job listings. Within 18 months it will be a real job requirement, because the companies that figure out how to automate their literature review, hypothesis generation, and experimental design pipelines will move significantly faster than those relying on manual researcher effort. If you’re picking a side project to build in public, an agentic research workflow is probably the highest-signal thing you could show a hiring manager right now.

🎯 How to stand out as a company and an employee (from someone who sees 1,000 CVs)

Knowing the right skills is one thing, betting noticed in a pile of 300 applicants is another. The entry-level bottleneck is worth spelling out from both sides. Candidates are applying into a tiny number of junior roles: only 11-12% of ML drug discovery and comp bio positions are entry-level, yet the pipeline of qualified graduates is enormous. Joe puts numbers to it: a senior role might attract 50 applicants, of which maybe one is genuinely relevant. A junior role pulls 300-350, of which 7-10 are a real fit. The ratio of applicants to roles is 6-7x higher at entry level, and even then, most applications miss the mark. Companies, meanwhile, are drowning in applications and still can’t find the right people. The volume of inbound is high; the signal-to-noise ratio is low.

Joe’s advice on what separates the top candidates:

1. Show evidence of building, not just researching. “Founders and hiring teams often love seeing evidence that someone can actually build and ship things, not just research them theoretically. Particularly in bio, there’s a real appreciation for candidates who can move between experimentation and execution.”

2. Be visible where recruiters actually look. This goes beyond LinkedIn. “A huge chunk of our time is tracking publications and collaborations to map who’s working on what, which labs are producing interesting work. There’s also GitHub and Hugging Face, where we’re digging through models, software tooling and tech stacks.” Active repositories, conference contributions, hackathon results, and community involvement all increase visibility.

3. Go deep in interviews, not broad. “A lot of candidates stay very high-level when explaining their work, but hiring teams usually want to go much deeper than people expect. They want candidates to talk through how decisions were made, why certain approaches worked.” Approach interviews like an external consultant: figure out exactly what you are there to improve.

4. Bridge the academia-industry gap deliberately. “Where people sometimes struggle is around product and commercial awareness. Industry teams aren’t just thinking about whether something works academically. They’re thinking about usability, timelines, scalability, deployment and business impact.” The candidates who transition best have already sought out internships, collaborations, or commercial projects while in academia.

5. Network into the hidden job market. “The people most in-demand tend to have clear signals that they’re genuinely interested in their area outside their day job. They’re naturally around others in the space a lot of the time, so are closer to that hidden job market that’s often rife with word-of-mouth opportunities.”

What we think: The pipeline problem in comp bio is structural, not cyclical. Universities produce graduates faster than the industry can absorb at junior level, while senior roles go unfilled for months. This won’t close by itself. The people who break through are the ones who’ve already demonstrated they can operate above their experience level. A polished GitHub with one well documented, production quality project is worth more than five papers in middling journals. Hiring managers are pattern-matching for “can this person ship something on day one,” and the evidence needs to be visible before the interview.

🏢 For companies: how to compete for talent

Now the contrary. If you’re a hiring manager or founder reading this, you already know the challenge. You’re getting hundreds of applications per role and still can’t find the right people. Startups are competing with Isomorphic Labs (£1.6 billion raise in 2025), Google DeepMind, and Boltz for the same small talent pool. Joe’s take on what works:

“A lot of candidates want more ownership, closer access to founders, more influence over direction of a product, and thrive on the ability to actually see their work shape a product directly. The larger players cannot offer this at the same level.”

Smaller companies can lean into that. What the ones that hire well do differently:

• Clarity of mission: Strong conviction about what they are building and why it matters. “If people believe in the founding team and what they’re standing for, it counts for a lot.”

• Process efficiency: “Slow feedback, too many stages, or technical tasks that take hours of a candidate’s time can quickly put people off.” Even unsuccessful candidates should leave with a good impression.

• Communication throughout: “Companies sometimes assume that if there’s no update, there’s no reason to contact the candidate. From the candidate’s side, silence feels like being ghosted.” Even an update saying they are still in consideration keeps people engaged.

• Clear expectations: “So often role details change mid-search, sometimes several times, which sends a mixed message to market and damages perception to the target talent pools.”

What we think: The talent competition in bio-AI is asymmetric in a way that favours startups, if they play it right. The big players offer prestige and salary. They cannot offer speed, ownership, or the feeling of shaping something from scratch. The startups that lose candidates to Isomorphic or DeepMind are usually the ones that ran a slow, unclear process, not the ones that lost on compensation alone. Your hiring process is your first product demo. If it’s confusing or inconsistent, strong candidates will read that as a signal about what working there is actually like.

⚡ The market in motion

Two things are true at the same time in computational biology right now. Traditional pharma is contracting (patent cliffs, layoffs, restructuring), while AI-native biotech is expanding fast. Isomorphic Labs raised £1.6 billion with no molecules in clinical trials. UK seed investment leapt 19% in 2025. Two UK biotech companies hit unicorn status (Verdiva Bio and Isomorphic Labs). The computational biology market overall is growing at 13% annually toward $22 billion by 2034.

For candidates, the opportunity is real, but “learn Python and apply broadly” doesn’t work anymore. The market rewards specialists who can ship production systems and communicate across disciplines. Infrastructure roles (MLOps, GPU optimisation, deployment) are where the most acute demand is. The salary premium for ML over traditional bioinformatics is widening. Remote work exists, though it’s not the default.

The people who do best in this market aren’t necessarily the most credentialed. They tend to be the ones who are visible, who adapt quickly, and who build things in the open.

💬 Want to be featured in Kiin Bio Weekly?

Each issue we speak directly with researchers, scientists, and builders working at the frontier of AI in life sciences. If you're working on something in this space and think it would resonate with our community, I'd love to hear from you. Fill out this form or reach out to me directly.

Found this useful? Forward it to a colleague in computational biology, it's the best way to help the newsletter grow.

Share Kiin Bio Weekly

Subscribe now to stay at the forefront of AI in Life Science. Every week: primers, deep dives, and direct conversations with the people building the field.

Subscribe now

Connect With Us

Have questions on this or suggestions for our next deep dive? We’d love to hear from you!

📧 Email Us | 📲 Follow on LinkedIn | 🌐 Visit Our Website

Three papers this week that all ask some version of the same question: are we actually measuring the right thing? SwitchCraft pushes protein design past the single-structure assumption that underpins most current methods. TadA-Bench asks whether protein language models can predict the future of evolution or only interpolate its past. Chem-PerturBridge harmonises the messy pile of perturbation transcriptomics data and reveals just how little of it agrees at the gene level. The connecting thread: the foundation models exist, but the design paradigms and benchmarks haven’t caught up.

We just opened up our Kiin Pioneer Programme: free access to our platform for academic and nonprofit research teams for a year.

The short version: we’ve built a place where scientists can collaborate on their drug discovery work without everything living in disconnected tools and someone’s local files. Literature reviews, target discovery, bioinformatics, all in one place. When one person finds something interesting and someone else has relevant data, the platform catches that and suggests what to look at next. Everything’s tracked, so six months later you actually know what was done and why.

We’re looking for teams who are trying to move faster on questions like: which targets should we prioritise? How do we make sense of conflicting evidence? What’s actually worth testing next?

No cost, no data transfer, all IP stays with your institution. Applications close in August, first cohort starts in September.

Read more about the programme

Apply now

SwitchCraft: A Programmatic Framework for Designing State-Switching Proteins

🔬 Protein design tools optimise for one structure. Real proteins switch between conformations to do their jobs: allosteric enzymes toggle between active and inactive states, biosensors change shape upon binding. Designing proteins that intentionally switch between defined states has been a manual, low-throughput exercise.

Jing, Bafna, and colleagues from MIT’s Berger lab built SwitchCraft, a framework that designs proteins with specified multi-state behaviour by backpropagating through differentiable structure prediction models.

🧬 The core idea: treat Boltz-1 (an open-source structure prediction model) as a differentiable loss function. Define the desired structural states, then optimise the sequence so that it folds into all of them under appropriate conditions. The framework is programmatic. You compose design objectives from modular building blocks rather than training a new model per task.

⚡ They demonstrate allosteric regulation of protein motifs, discrimination between bound ligand identities, and fluorescent biosensor design. The biosensor results are particularly telling: designed sequences show distinct fluorescence states depending on which ligand is bound. That’s functional multi-state behaviour that hasn’t been accessible through standard single-state design pipelines.

🧪 Where This Fits

This sits at the leading edge of computational protein design, downstream of structure prediction but upstream of experimental characterisation. The reason it exists now is simple: differentiable structure prediction. Until models like Boltz-1, ESMFold, and AlphaFold became fast and accurate enough to use as gradient-providing modules, you couldn’t backpropagate through structure. RFdiffusion and ProteinMPNN design single structures beautifully, but they don’t handle the multi-state problem. SwitchCraft does something conceptually different: it treats structure prediction as a subroutine rather than the end goal. The limitation is that experimental validation here is still computational, relying on Boltz-1’s predictions of the designed states. Whether these sequences actually fold into multiple states in a wet lab remains open, though the biosensor designs are testable. If you work on biosensors, allosteric switches, or molecular logic gates, this is worth trying now.

💡 Why This Is Cool

This is what it looks like when structure prediction becomes infrastructure rather than the main event. The field spent five years building accurate folding models. Now those models are components in design loops. Multi-state protein design has been a goal since the Kuhlman lab’s early work on conformational switches, but the computational tools never matched the ambition. SwitchCraft doesn’t solve the full problem (experimental validation is still the bottleneck), but it makes the design step tractable in a way it simply wasn’t before.

📄 Read the paper.

💻 Try the code.

TadA-Bench: A Million-Variant Benchmark for Future-Round Discovery Toward Agentic Protein Engineering

🔬 Protein fitness prediction benchmarks typically test whether models can fill in gaps within a mutational landscape. That’s interpolation. What protein engineers actually need is extrapolation: given rounds 1 through 10 of directed evolution, can you predict what we’ll find useful in round 11? No existing benchmark properly tests this.

Gao and colleagues from Shanghai Jiao Tong University built TadA-Bench from 31 rounds of real TadA (tRNA adenosine deaminase) directed evolution, totalling roughly one million variants.

🧬 The benchmark enforces chronological evaluation: models train on earlier rounds and must predict which variants from later rounds will be experimentally validated. It provides aligned DNA, RNA, and protein sequences, and uses a Seq2Graph method to create comparable activity measurements across rounds that originally used different assay conditions.

⚡ The headline finding is sobering. Current protein language models (including ESM-2 and various fine-tuned variants) struggle with temporal prediction even when they perform well on standard interpolation benchmarks. The gap between interpolation and extrapolation performance is substantial. Good performance on DMS datasets doesn’t mean your model can guide the next round of experiments. Many practitioners suspected this; now there’s a number attached to it.

🧪 Where This Fits

This is a benchmark, not a tool, and its value depends on adoption. It fills a gap the field has known about for years. ProteinGym and similar resources test interpolation well, but nobody had assembled a large-scale temporal benchmark from real directed evolution campaigns. The 31 rounds of TadA evolution provide unusually rich temporal structure (most published datasets have 3-5 rounds at most). The “agentic protein engineering” framing is timely: as more groups wire LLMs into experimental design loops, you need evaluation frameworks that test whether the model’s suggestions actually lead somewhere productive. The dataset is on Hugging Face and the code is open on GitHub, which lowers the barrier to adoption. The main caveat is generalisability: TadA is one enzyme. Performance on this benchmark won’t guarantee performance on your protein of interest. Still, it’s the best temporal test we have.

💡 Why This Is Cool

The protein ML field has a measurement problem. Papers report NDCG on held-out DMS positions and claim their model “guides protein engineering.” This benchmark calls that bluff. Can your model predict the future, or only reconstruct the past? The distinction matters enormously for anyone running a real directed evolution campaign. If current models fail at temporal prediction (and they largely do), that’s uncomfortable but useful information. It tells you where the actual research gap is, which is more valuable than another leaderboard.

📄 Read the paper.

💻 Code and data.

🤗 Access the dataset.

Chem-PerturBridge: A Harmonized Compendium of Small Molecule Perturbation Transcriptomic Effects

🔬 The field has generated enormous amounts of small molecule perturbation transcriptomics data (L1000, sci-Plex, Tahoe, and many smaller datasets). The problem: nobody knows how well they agree with each other, and combining them for model training requires harmonisation that hasn’t been done systematically.

Szałata and colleagues from the Theis lab at Helmholtz Munich built Chem-PerturBridge, standardising 37,000+ compounds across 1.25 million samples from eight assay types with consistent metadata.

🧬 The compendium spans nine datasets including sci-Plex3, Tahoe, L1000, OP3, DILImap, and others. The harmonisation pipeline normalises cell line annotations, compound identifiers, and gene nomenclature, making cross-dataset comparison possible for the first time at this scale.

⚡ The uncomfortable finding: gene-level agreement between datasets is poor. When the same compound is tested in the same cell line across different platforms, correlation at the individual gene level is low. Directional consistency (is the gene up or down?) is better, and turns out to be sufficient for improving compound representation learning. Models trained on the harmonised directional data outperform those trained on any single dataset alone

🧪 Where This Fits

This sits upstream of almost everything in computational chemical biology: virtual screening, target identification, mechanism-of-action prediction, and toxicity forecasting all depend on perturbation data quality. The honest finding here is that the perturbation transcriptomics field has a reproducibility problem at the gene level, even between high-quality experiments. That’s not new as a suspicion, but quantifying it across eight assay types is valuable. The positive takeaway is that directional signals persist, and they’re enough to learn useful compound representations. For practitioners, this means you should probably stop training on raw gene-level expression values from a single perturbation study and start using directional features from multiple sources. The resource is open (MIT licence on code, upstream licences on data), which is appropriate for something positioned as community infrastructure.

💡 Why This Is Cool

The field has been treating perturbation transcriptomics datasets as interchangeable training data without checking whether they actually say the same thing. They don’t, at least not at the resolution most people assume. This paper does the unglamorous work of quantifying that disagreement and showing what signal does survive. The practical consequence: if you’re building models on LINCS L1000 data alone, you’re probably leaving performance on the table. Directional agreement across platforms is a more robust training signal than absolute expression values from one platform. This is infrastructure work rather than a methods advance, but it’s the kind that makes everything downstream more trustworthy.

📄 Read the paper.

💻Try the code.

🗓️ Events & Competitions

The best competitions, hackathons, and community challenges in AI x life sciences, curated weekly. Know something worth featuring? Reply and let us know.

More upcoming events:

Creative Disruption Forum: Modern Drug Discovery | June 18, NIAB Cambridge

A full-day forum for biotech and R&D leaders exploring how technology is changing small molecule drug discovery. Keynote interviews with industry thought leaders followed by workshops under Chatham House Rules, limited to 60 attendees. Part of Cambridge Wide Open Week. Organised by Graham Combe and Prof Tony Sedgwick. £60 for biotech companies.

London Protein Design Day | June 23, Imperial College London

The first edition of a one-day symposium bringing together London’s protein design community and beyond. Programme spans AI-driven design, molecular dynamics, and bioinformatics, with applications across enzymes, antibodies, and materials. Organised by Pietro Sormanni, Rebecca Birolo, and Jakub Lála. Abstract deadline for poster/oral presentations is this Saturday (May 17). In person only.

BioHackathon Europe 2026 | November 9-13, Barcelona

ELIXIR’s annual international bioinformatics hackathon, running since 2018. 160+ participants, five days of collaborative coding on open bioinformatics infrastructure and tools. The call for project proposals opens March 16 and closes April 15 - so if you want to lead a project, that’s your window.

Thanks for reading!

💬 Get involved

We’re always looking to grow our community. If you’d like to get involved, contribute ideas or share something you’re building, fill out this form or reach out to me directly.

Connect With Us

Have questions or suggestions? We'd love to hear from you!

📧 Email Us | 📲 Follow on LinkedIn | 🌐 Visit Our Website

I keep coming back to this question of whether the protein ML field has a modelling problem or a data problem. This week's papers make the case from both sides. Biohub's ESM release argues that scale and architecture can solve binder design from scratch. OpenADMET argues the opposite: that no model will work until the training data stops being inconsistent rubbish. SynFit sits somewhere in between, showing what you can get when you have good multi-property data and a framework that knows how to use it.

We just launched our Kiin Pioneer Programme, giving academic and nonprofit research teams one year of free access to our drug discovery platform!

KiinOS is a platform where scientists can run literature reviews, target discovery, and bioinformatics in one place. It keeps a record of what’s been done, by who, and what came out of it. So if one person finds a promising target and someone else has relevant data, the platform connects those results and suggests what to pursue next.

We’re looking for teams asking: which targets should we prioritise? How do we interpret conflicting evidence? Which hypotheses are worth testing next?

There’s no cost, no data transfer, and all IP stays with your institution. Applications close in August, with the first cohort starting September.

Read more about the programme

Apply now

ESM: A World Model of Protein Biology

🔬 We can predict how proteins fold, but designing new ones that actually bind specific targets, particularly antibodies, still requires months of expensive lab screening just to find starting candidates.

Biohub has released ESM: a protein language model (ESMC, trained on 2.8 billion sequences), a structure prediction and design model (ESMFold2), and a map of 6.8 billion sequences (ESM Atlas). All MIT-licensed.

🧬 ESMFold2 learns protein representations from evolutionary data, then searches that learned space for proteins predicted to bind a given target. It scores candidates using its own confidence estimates, so the entire design loop is computational. Structure prediction runs from a single sequence without needing alignment databases.

⚡ Designed binders for five cancer/immunology targets were validated in the lab: minibinder success rates of 70%, scFv antibody success rates of 21%. A PD-L1 binder hit 4.3 nM affinity and blocked immune checkpoint suppression in cells. Cryo-EM confirmed an EGFR binder matched the prediction at 1.2 Å RMSD.

🧪 Where This Fits

This positions ESM as a competitor to both AlphaFold 3 (for structure prediction) and RFdiffusion/ProteinMPNN (for binder design). The antibody-antigen prediction results are interesting because this is where AlphaFold has historically struggled most. The binder design piece is where Biohub is making their boldest claim: that the earliest stage of therapeutic protein discovery can happen computationally in days rather than months. Worth noting that the 21% scFv success rate, while a step up from near-zero for most computational methods, still means roughly 80% of designs fail in the lab. These are first-round hits. David Baker’s group (RFdiffusion) and Generate Biomedicines (Chroma) are the obvious comparisons for generative protein design. ESMFold2 is differentiated by working from a language model backbone rather than a diffusion architecture, which changes how it scales with compute at inference time.

💡 Why This Is Cool

The sparse autoencoder analysis is what I find most thought-provoking. ESMC independently recovered biological concepts like the nucleophilic elbow motif across 75 of 99 relevant enzymes, despite never being told what one is. That’s a model learning the grammar of protein biology from sequence alone. Whether the binder design numbers hold up across a broader and more difficult target set remains to be seen. The cryo-EM validation is reassuring, but five targets is still five targets.

📄 Read their press release.

💻 Try the tool.

SynFit: Synergistic Contrastive Learning for Multi-Objective Protein Fitness Prediction and Optimisation

🔬 Protein engineering almost always requires optimising multiple properties at once, but current ML fitness predictors handle each property independently. Train separate models for yield and selectivity and you’ll get variants that excel at one while tanking the other.

Georgia Tech and UC Santa Barbara developed SynFit, a multi-objective framework that fine-tunes protein language models on experimental fitness data across multiple assays simultaneously.

🧬 SynFit combines a shared ESM2 encoder with property-specific prediction heads, using contrastive learning to capture cross-property relationships from deep mutational scanning data. Predictions are integrated via Pareto sorting to find variants that improve everything at once.

⚡ On Pareto front analysis across 20 proteins, SynFit hits the optimal front 70% of the time versus 60% for ProteinNPT and 55% for ConFit. The wet-lab result is more convincing: 83 out of 100 designed hextuple mutants for a biocatalytic borylation enzyme showed simultaneously improved yield and enantioselectivity, with multiple variants beating everything in the training data.

🧪 Where This Fits

This sits squarely in the protein engineering workflow after you have initial DMS data and want to explore combinatorial sequence space. ConFit, from the same group, is the direct predecessor, and SynFit extends it with the multi-objective training component. ProteinNPT (Notin et al., ICML 2023) tackles a related problem through non-parametric transformers but doesn’t explicitly model cross-property correlations. The wet-lab validation on biocatalytic borylation is well chosen because it’s a genuinely new-to-nature reaction where directed evolution data is sparse, making the Pareto optimisation problem harder. The limitation is that you still need initial multi-property DMS data for your protein of interest, which isn’t always available. The KRAS case study (identifying shared functional residues across six binding partners) is a nice mechanistic demonstration but relies on an unusually well-characterised system.

💡 Why This Is Cool

The 83/100 result is more impressive than any of the benchmark numbers. Getting computationally designed hextuple mutants to simultaneously beat the training set on both yield and enantioselectivity, in a single round without iterative screening, is a practical result that enzyme engineers will care about. It suggests ML-guided combinatorial design can start to compress the “design-build-test” cycle for multi-objective problems. The architecture is straightforward enough that adoption shouldn’t be difficult once the code is released.

📄 Read the paper.

💻Try the code.

Mapping the Avoid-ome: A Systematic Open-Science Approach to Predictive ADMET

🔬 Around 30% of clinical drug failures trace back to ADMET problems. The ~100 proteins responsible (CYPs, hERG, transporters, nuclear receptors) are well known, but existing ML models train on data cobbled from dozens of labs using different protocols. A recent analysis found almost no correlation between IC50 values for the same compound measured by different groups.

Fraser (UCSF), Edgar (Octant), Chodera (MSKCC), and Walters (OMSF) have launched OpenADMET, an ARPA-H and Gates Foundation-funded consortium generating systematic, internally consistent ADMET datasets and releasing everything publicly.

🧬 The consortium runs assays across the full “Avoid-ome” panel at industrial scale: 30,000 compounds per run in 1536-well plates, under $0.40 per compound. Active learning selects informative compounds for expansion, and structural biology (100+ PXR crystal structures so far) resolves binding modes.

⚡ This is a programme-level perspective paper rather than a single dataset release. The first community challenge has run. They’re screening tens of thousands of compounds weekly. I appreciate the honesty that this is long-term infrastructure. They’re not claiming to have solved ADMET prediction; they’re arguing nobody will until the data problem is addressed.

🧪 Where This Fits

This is an upstream data-generation effort, not a prediction tool. It sits before everything else in the ADMET pipeline: the datasets it produces will train the next generation of models. TDC (Therapeutics Data Commons) and ChEMBL aggregate existing literature data. OpenADMET’s bet is that literature data is too inconsistent to train reliable models, and that generating internally consistent measurements from scratch, with structural validation, justifies the cost. Tools like ADMET-AI (Swanson et al., 2024) would be downstream consumers of these datasets. The federated learning alternative (training behind pharma company firewalls) is explicitly discussed and dismissed: it can’t generalise beyond local chemical space and doesn’t produce the structural understanding needed for true mechanistic models.

💡 Why This Is Cool

The framing is what matters here. By defining ADMET as a finite structural biology problem (map the interactions with roughly 100 proteins and you’ve covered most failure modes), they turn an open-ended prediction challenge into a bounded experimental campaign. The question is whether $0.40-per-compound assays and active learning can produce models that generalise to novel chemical matter outside the training distribution. The open-science commitment, with Gates Foundation backing, makes this more credible than most “we’ll share data eventually” promises from pharma-adjacent initiatives.

📄 Read the paper

💻 Learn more

🗓️ Events & Competitions

The best competitions, hackathons, and community challenges in AI x life sciences, curated weekly. Know something worth featuring? Reply and let us know.

More upcoming events:

Creative Disruption Forum: Modern Drug Discovery | June 18, NIAB Cambridge

A full-day forum for biotech and R&D leaders exploring how technology is changing small molecule drug discovery. Keynote interviews with industry thought leaders followed by workshops under Chatham House Rules, limited to 60 attendees. Part of Cambridge Wide Open Week. Organised by Graham Combe and Prof Tony Sedgwick. £60 for biotech companies.

London Protein Design Day | June 23, Imperial College London

The first edition of a one-day symposium bringing together London’s protein design community and beyond. Programme spans AI-driven design, molecular dynamics, and bioinformatics, with applications across enzymes, antibodies, and materials. Organised by Pietro Sormanni, Rebecca Birolo, and Jakub Lála. Abstract deadline for poster/oral presentations is this Saturday (May 17). In person only.

BioHackathon Europe 2026 | November 9-13, Barcelona

ELIXIR’s annual international bioinformatics hackathon, running since 2018. 160+ participants, five days of collaborative coding on open bioinformatics infrastructure and tools. The call for project proposals opens March 16 and closes April 15 - so if you want to lead a project, that’s your window.

Thanks for reading!

💬 Get involved

We’re always looking to grow our community. If you’d like to get involved, contribute ideas or share something you’re building, fill out this form or reach out to me directly.

Connect With Us

Have questions or suggestions? We'd love to hear from you!

📧 Email Us | 📲 Follow on LinkedIn | 🌐 Visit Our Website

This week we’re looking at clinical AI, specifically the type that nobody’s talking about. I got connected to Ignacio through his work at Savana, where they’ve been extracting real-world evidence from clinical records for years. What struck me the most was how much of the conversation around AI in medicine misses the point: everyone’s focused on ChatGPT and scribes, while the predictive models that actually enable personalised medicine are sitting there with regulatory approval and population-level evidence, largely unused.

We got on a call, and the result is this primer.

Freebie alert: We know how hard science is. That’s why we built the Pioneer Programme.

We’re selecting academic and nonprofit research teams to get one year of free access to our drug discovery platform plus hands-on support from our science team. If your research bottleneck isn’t data but connecting the findings you already have, this is for you.

We’re looking for teams asking: which targets should we prioritise? How do we interpret conflicting evidence across datasets? Which hypotheses are worth testing next? Where are the strongest translational opportunities?

No cost. No data transfer. All IP stays with your institution. Applications close August, cohort starts September.

Read more about the programme

Apply now

Healthcare systems across Europe are breaking. Waiting lists stretch to 14 months. The UK Health Secretary declared the NHS “broken.” Spain is close behind. Populations are ageing, medicines are expensive, and the workforce cannot scale. Into this crisis arrives artificial intelligence, not as a future promise, but as a present reality. The FDA has approved close to 1,500 AI use cases across clinical specialties. Meta-analyses now demonstrate cost-effectiveness in diabetes, colon cancer, and mammography screening. The question is no longer whether AI works in medicine. It is why adoption remains so slow.

Yet most clinicians, when asked about AI, will talk about ChatGPT and medical scribes. They are looking at the louder revolution while the quieter, more consequential one unfolds beneath it.

We spoke to Ignacio H. Medrano, neurologist-turned-CEO of Savana, about the real state of clinical AI: what is proven, what is hype, and what clinicians are getting wrong about the pace of change.

Ignacio H. Medrano, CEO of Savana

“It’s unethical not to use AI these days in certain areas, because it’s proven, and it’s proven at scale across populations.”

🔀 Two types of AI, one common misunderstanding

The most pervasive misconception among clinicians today is that generative AI (LLMs, chatbots, multi-modal models) is the only AI that matters. It is visible, fast-moving, and immediately useful: scribes that eliminate documentation burden, literature tools like Open Evidence replacing PubMed searches, agents managing waiting lists. These applications are exploding because they save time and, critically, do not require clinical validation. They handle documentation, not decisions.

But the deeper disruption is discriminative AI: predictive, classification-based models that have existed for over a decade. This is the AI that enables precision medicine. Granular predictions for individual patients about immunotherapy response, relapse probability in multiple sclerosis, optimal drug sequencing in haematologic cancers. It takes statistics to a level where you can determine the actual probability of a specific outcome for a specific patient.

Discriminative AI started earlier but arrives later in practice. Every algorithm requires validation, external replication, meta-analysis, and integration into clinical guidelines. That pipeline is slow. But it is the pipeline that delivers personalised medicine, and it is now producing results.

“A big misconception is forgetting that discriminative, predictive AI is the real silent disruption,” says Medrano. “The other is underestimating the speed at which agentic AI is arriving. People think this takes 20 years. It’s happening now, like thunder.”

✅ Separating signal from noise: what makes clinical AI “real”

With close to 1,500 FDA-approved AI applications and an exponential curve of machine learning publications on PubMed, distinguishing proven tools from hype requires a framework.

Medrano uses three levels:

1. Regulatory approval: FDA or EMA clearance confirms correct dataset construction, generalisation to new cohorts, and absence of bias. This is the minimum threshold.

2. Population-level evidence: Publications demonstrating that algorithms work in general populations and are cost-effective. Meta-analyses now exist for AI in diabetes management, colon cancer screening, and mammography. A National Library of Medicine study proved chatbots reduce severe mental health events. A cardiology trial demonstrated that AI applied to ECGs reduces cardiovascular mortality.

3. Real-world deployment: Hospitals actually running these systems in clinical workflows. Hospital Clinic Barcelona has used AI to predict sepsis in intensive care for over three years. Finland deploys predictive models on GP workstations for population segmentation. In China, Ping An Good Doctor attends 100 patients daily without human involvement. In Utah, AI autonomously renews prescriptions.

The gap between levels two and three, between proven effectiveness and actual deployment, is where the real problem lives.

🚧 Three barriers, one that matters most

Technical: Clinical information remains fragmented across systems. Standards like HL7 FHIR are improving interoperability, but the problem is not fully solved. Training and validating models still requires stitching together disparate data sources.

Regulatory: Largely resolved. European regulation now recognises that anonymised data used for research does not require individual informed consent. The European Health Data Space will mandate hospitals to share data within one to two years. Countries like Switzerland, France, Germany, and the UK already permit compliant secondary use of clinical data.

Cultural: The real bottleneck. Not rejection; most managers and clinicians accept AI is inevitable. The problem is twofold. First, pure ignorance: hospital managers do not realise they could build pharmacogenomic models today that predict which patients will respond to expensive biologics, saving millions by paying only for drugs that will work. Second, politics: an algorithm that reduces the need for human labour is nearly impossible to sell when unions and the public demand 10 new nurses before any technology investment. Politicians choose nurses even when they understand AI’s value.

“It’s not belief anymore, it’s knowledge,” says Medrano. “Managers just don’t understand the power of what’s already possible with the data they have.”

📊 Real-world evidence: from luxury to necessity

Real-world evidence has always mattered. It is the difference between reading about a country and landing there. Clinical trials tell you what should happen under controlled conditions. Real-world evidence tells you what actually happens.

The barrier was always collection. Patient by patient, variable by variable, manually assembling registries over years. Exhausting, expensive, and therefore underutilised. Now, computational systems can extract this information automatically, reliably, and at scale from electronic health records. Once extraction became feasible, demand exploded. Regulators began requesting it. Pharma began requiring it.

This is where initiatives like the UK’s Foresight programme become significant: 57 million medical records feeding predictive models for 100 diseases at 20-year horizons. The Scandinavian countries (Norway and Denmark) are sharing data internationally and validating models across borders. These are not pilots. They are national-scale infrastructure decisions.

🤖 The convergence: agentic AI meets predictive models

Here is where the field is heading, and where most clinicians have not yet looked. The users of sophisticated discriminative AI models (multi-modal predictive algorithms trained on clinical text, genomics, proteomics, radiomics) will not be human doctors. They will be certified agentic AI systems.

Generative AI agents will orchestrate clinical workflows. Discriminative AI will provide the predictions those agents act on. The agent decides what to ask; the predictive model provides the answer. The human clinician supervises, validates, and handles what requires physical presence.

A meta-analysis of 15 studies already shows that in 13 out of 15, AI chatbots were rated as more empathic than human clinicians. Not even communication, the last presumed advantage of human clinicians, remains unchallenged.

“If doctors keep doing the same thing, they’ll become pointless,” says Medrano. “People will turn to their phones, where an agentic AI healthcare service for 10 euros will give them advice, pull their data, run algorithms. And then that agent will hire humans to perform the physical tasks it cannot.”

⚡ Where this leaves us

The infrastructure is being built. Data-sharing mandates are arriving. Validation evidence is accumulating. The two types of AI, generative and discriminative, are converging toward agentic systems that will reshape how healthcare is delivered.

The bottleneck is not technology or regulation. It is the speed at which institutions recognise what is already possible, and act before the system breaks entirely, or before patients simply route around it.

Big thanks Ignacio for meeting with us and sharing his insights for this primer!

💬 Want to be featured in Kiin Bio Weekly?

Each issue we speak directly with researchers, scientists, and builders working at the frontier of AI in life sciences. If you're working on something in this space and think it would resonate with our community, I'd love to hear from you. Fill out this form or reach out to me directly.

Found this useful? Forward it to a colleague in drug discovery or computational chemistry, it's the best way to help the newsletter grow.

Share Kiin Bio Weekly

Subscribe now to stay at the forefront of AI in Life Science. Every week: primers, deep dives, and direct conversations with the people building the field.

Subscribe now

Connect With Us

Have questions on this or suggestions for our next deep dive? We’d love to hear from you!

📧 Email Us | 📲 Follow on LinkedIn | 🌐 Visit Our Website

We know how hard science is. That’s why we built the Pioneer Programme.

We’re selecting academic and nonprofit research teams to get one year of free access to our drug discovery platform plus hands-on support from our science team. If your research bottleneck isn’t data but connecting the findings you already have, this is for you.

We’re looking for teams asking: which targets should we prioritise? How do we interpret conflicting evidence across datasets? Which hypotheses are worth testing next? Where are the strongest translational opportunities?

No cost. No data transfer. All IP stays with your institution. Applications close August, cohort starts September.

Read more about the programme

Apply now

Proteo-R1: Reasoning Foundation Models for De Novo Protein Design

🔬 Current deep learning methods for protein design generate molecular structures without explicitly reasoning about which residues matter or why. All residues are treated uniformly during generation, and design intent is implicitly buried in diffusion parameters. This makes it hard to interpret why a model made a particular choice, or to reuse that logic on a different target.

Researchers from Stanford, RIKEN, and collaborating institutions introduce Proteo-R1, a framework that separates molecular understanding from geometric generation. A multimodal large language model first reasons about binding interactions, then passes residue-level constraints to a diffusion model that generates the structure.

🧬 The understanding expert (a multimodal LLM) analyses protein sequences, AF3-style structural representations, and textual context to identify key interaction residues and predict their amino acid identities. These sparse, residue-level decisions are passed as hard constraints to the generation expert, an AlphaFold3-style diffusion model that performs conditional co-design while respecting the fixed interaction anchors. Training proceeds through three stages: multimodal alignment, structural reasoning mid-training, and joint reasoning-guided design on antibody-antigen complexes from SAbDab.

⚡ On simultaneous multi-CDR antibody redesign, Proteo-R1 achieves the lowest or near-lowest per-CDR RMSD in five of six regions, with interface improvement (IMP) of 56.58%. It produces the lowest steric clash rates (0.50% intra-chain, 0.14% inter-chain) and best dihedral distribution divergence among all tested methods. On CDR-H3 design via the RAbD benchmark, it obtains the best lDDT (0.9693), TM-score (0.9816), and DockQ (0.801) over DGENet, BoltzGen, and MFDesign.

🔬 Applications and Insights

1️⃣ Interpretable Antibody Engineering

The reasoning expert produces explicit, human-readable justifications for which residues it selects as interaction anchors. You can inspect and modify the design logic directly.

2️⃣ Controllable CDR Design

Researchers can edit the reasoning outputs (e.g. specifying different hotspot residues) to steer the generative model without retraining, giving fine-grained control over binding specificity.

3️⃣ Structure-Sequence Consistency

Proteo-R1 is the only method that achieves positive structure-sequence consistency (Δ = IF-AAR minus AAR) across five of six CDR regions. Its designs are structurally grounded rather than simply recovering native sequences.

4️⃣ Generative Backend Flexibility

The reasoning expert works with alternative generative frameworks. Tested with UniMoMo, it improves IMP from 65% to 67.79% and binding energy from 8.46 to 7.35 ΔG without architectural changes to either component.

💡 Why This Is Cool Proteo-R1 separates the “what should we build” question from the “how do we build it” question. That mirrors how human protein engineers actually work: identify critical interaction residues first, then optimise geometry under those constraints. The reasoning module plugs into different generative backends, and the design logic is readable and editable rather than locked inside a diffusion trajectory.

📄 Read the paper

💻 Try the code

ActivityFinder: Toward the Fully Automatic Integration of Structural and Binding Affinity Data

🔬 Building computational models that predict binding affinity requires both 3D protein-ligand structures and experimental activity measurements. These data live in separate databases (PDB for structures, ChEMBL for bioactivities) with no fully automated way to link them. Identifier-based approaches miss connections where sequences diverge, ligand representations vary, or binding-site mutations are present.

A team at the University of Hamburg developed ActivityFinder, a fully automated pipeline that links crystal structures of protein-ligand complexes directly to wet-lab bioactivity data. It requires only PDB files and a ChEMBL database dump, with no external services or continuous data connections.

🧬 ActivityFinder works in two stages. First, it builds an ActivityDB instance by parsing PDB structures into the NAOMI data format, extracting ligand representations as six different string types (InChI, InChIKey, canonical SMILES with and without stereo, InChI connection layer, InChI hydrogen layer), and creating BLAST databases from protein sequences. Second, it queries this database using sequence alignments (at 80%+ identity) and detailed chemical structure matching to cross-reference structural and bioactivity records. It tracks mutations and binding-site residues at atomic resolution.

⚡ Applied to 226,302 PDB structures and ChEMBL 35, ActivityFinder linked 20,197 PDB entries involving 13,734 PDB ligands to 17,829 unique ChEMBL ligands across 2,585 ChEMBL targets, covering over one million bioactivity data points. Compared to BioChemGraph (an identifier-based method), ActivityFinder identifies 46,287 unique structure-activity triplets versus 19,333. Of the 6,575 triplets unique to ActivityFinder, 1,082 are entirely novel: new combinations of PDB complex, ChEMBL target, and ChEMBL molecule.

🔬 Applications and Insights

1️⃣ Proprietary Data Integration

Because it requires only local PDB files and an SQL database, ActivityFinder works for pharmaceutical companies that want to link in-house crystal structures to internal bioactivity records without sending data to external services.

2️⃣ Mutation-Aware Linking

The tool explicitly tracks sequence variants and binding-site mutations at atomic resolution. Researchers can study how specific point mutations affect ligand binding affinity across related structures.

3️⃣ Training Data for ML Scoring Functions

The linked structure-activity pairs are ready-made training datasets for machine learning scoring functions. Quality annotations (confidence levels 1 to 5) let modellers filter for the precision their application requires.

4️⃣ Expanding Known Chemical Space

Of the novel triplets ActivityFinder uniquely identifies, many involve new combinations of PDB complex, ChEMBL target, and ChEMBL molecule not found by any other method. These are new data points for structure-based drug design.

💡 Why This Is Cool

Everyone building a scoring function or binding predictor eventually has to manually assemble their own structure-activity dataset. The results are rarely shared and identifier matching misses a lot. ActivityFinder automates this from scratch, works on proprietary data, and finds 2.4x more structure-activity links than the identifier-based alternative.

📄 Read the paper

💻 Try the tool: Available via the ProteinsPlus REST API

ProteinConformers: Large-Scale and Energetically Profiled Descriptions of Protein Conformational Landscapes

🔬 Understanding protein function requires capturing how structures move across their conformational space. Existing MD trajectory databases start only from native structures (near the global energy minimum), conformer generators have no standardised benchmarks, and available datasets provide limited energetic annotations. No resource maps the full spectrum from non-native to near-native states with both structural and energetic characterisation.

Researchers at the National University of Singapore (Yang Zhang’s group) present ProteinConformers, a database of 2.7 million geometry-optimised conformations across 734 proteins, paired with energy evaluations and a benchmarking framework for multi-conformation generators.

🧬 The dataset uses a multi-seed decoy sampling strategy: for each protein, hundreds of diverse starting conformations (drawn from CASP5-15 prediction submissions) are each run through full-atom molecular dynamics simulation using GROMACS 2023 with the OPLS-AA force field. Each conformation receives five energetic evaluations (RW, RWplus, EvoEF2, Rosetta, FoldX) and pairwise similarity annotations (TM-score and RMSD). The curated benchmark subset, ProteinConformers-lite, contains 381,546 MD-refined conformers across 87 CASP14/15 proteins with 1.9 million energetic annotations.

⚡ ProteinConformers spans protein lengths from 33 to 949 residues, with conformations distributed continuously from non-native to near-native states (TM-scores covering the full 0 to 1 range). Local geometric quality matches the Top2018 reference set: dihedral angle distributions show Pearson correlations of 0.97 to 0.99, and near-native conformations fall below the Top2018 average Ramachandran outlier rate of 13%. In benchmarking five generative models, BioEmu achieves the highest coverage under strict energy thresholds (5 kJ/mol), while AlphaFlow-MD scores comparably on the CGMSmah geometric plausibility metric. Total compute cost: approximately 40 million CPU hours.

🔬 Applications and Insights

1️⃣ Benchmarking Conformation Generators

ProteinConformers-lite is the first standardised evaluation framework for models like AlphaFlow, ESMFlow, and BioEmu. It measures both diversity (how much of the energy landscape a model covers) and plausibility (geometric realism) in one benchmark.

2️⃣ Allosteric Mechanism Studies

The continuous energy surfaces from non-native to native states let researchers study allosteric transitions and functionally relevant intermediates that single-seed MD simulations would miss.

3️⃣ Drug Discovery Applications

Energy-annotated conformational ensembles support ensemble docking workflows, where sampling multiple receptor states improves virtual screening hit rates compared to docking against a single crystal structure.

4️⃣ Evaluating Energetic Realism

With five energy functions evaluated per conformation, the dataset allows systematic comparison of how well different generators produce physically plausible low-energy structures versus merely diverse ones.

💡 Why This Is Cool

If you want to compare AlphaFlow against BioEmu against ESMFlow, you currently have no standard reference to test against. ProteinConformers fills that gap. Every conformation has known energy and measured structural similarity to the native state, and the web platform lets you explore without running simulations.

📄 Read the paper

💻 Explore the database

🗓️ Events & Competitions

The best competitions, hackathons, and community challenges in AI x life sciences, curated weekly. Know something worth featuring? Reply and let us know.

More upcoming events:

London Protein Design Day | June 23, Imperial College London

The first edition of a one-day symposium bringing together London’s protein design community and beyond. Programme spans AI-driven design, molecular dynamics, and bioinformatics, with applications across enzymes, antibodies, and materials. Organised by Pietro Sormanni, Rebecca Birolo, and Jakub Lála. Abstract deadline for poster/oral presentations is this Saturday (May 17). In person only.

BioHackathon Europe 2026 | November 9-13, Barcelona

ELIXIR’s annual international bioinformatics hackathon, running since 2018. 160+ participants, five days of collaborative coding on open bioinformatics infrastructure and tools. The call for project proposals opens March 16 and closes April 15 - so if you want to lead a project, that’s your window.

Thanks for reading!

💬 Get involved

We’re always looking to grow our community. If you’d like to get involved, contribute ideas or share something you’re building, fill out this form or reach out to me directly.

Connect With Us

Have questions or suggestions? We'd love to hear from you!

📧 Email Us | 📲 Follow on LinkedIn | 🌐 Visit Our Website

We know how hard science is. That’s why we built the Pioneer Programme.

We’re selecting academic and nonprofit research teams to get one year of free access to our drug discovery platform plus hands-on support from our science team. If your research bottleneck isn’t data but connecting the findings you already have, this is for you.

No cost. No data transfer. All IP stays with your institution. Applications close August, cohort starts September.

Read more about the programme

Apply now

Single-cell biology has spent the last few years borrowing ideas from language models. Several groups have adapted transformer architectures to gene expression data, including scGPT and Geneformer, both of which pretrain on massive single-cell datasets using masked modeling objectives.

Transformers. Pretraining. Big datasets. It all felt like progress. But if you ask most practitioners what they actually use day to day, the answer is still pretty modest: embeddings that work, transfer across datasets, and do not collapse when the technology changes.

That tension is what motivated scConcept, a new foundation model for single-cell transcriptomics that steps away from gene reconstruction and instead asks a simpler, more explicit question: do these genes come from the same cell?

We spoke with Mojtaba Bahrami, in Fabian Theis’ lab at Helmholtz Munich, part of the the team behind scConcept about why the field needed a reset, what contrastive learning brings to biology, and why the future may be less about bigger models and more about better representations

🔴 The Problem

If you zoom out, most single-cell foundation models share the same basic idea. Treat genes like words. Mask some of them. Ask the model to predict what’s missing. This strategy mirrors masked language modelling introduced in BERT, where models learn by predicting missing tokens in a sentence. It works well enough on paper, but something feels off once you start using the embeddings downstream.

As Mojtaba Bahrami put it when we spoke, “In language models, training and inference are basically the same task. You predict the next token. But in single-cell, no one actually cares about predicting masked genes. People care about the embedding.”

That mismatch matters. Masked gene prediction optimises the wrong thing. The model gets good at reconstructing counts, but the cell-level representation becomes an afterthought. Most approaches average learned gene embeddings and hope for the best. This may work depending on the downstream question but often it doesn’t. If the question can not be simply answered by looking at gene level information and needs understanding higher level biological processes going on in the cell, then we have a problem.

The cracks show up quickly. Simple methods like PCA or VAEs still outperform large models on tasks like cell type annotation. Benchmarking efforts such as the scIB framework have shown that classical integration methods can remain highly competitive across datasets. Spatial assays break embeddings entirely. New technologies shift the latent space so much that “foundation” starts to feel like marketing rather than reality.

💡 The Idea

Here comes scConcept. Instead of asking the model to reconstruct genes, it asks something more aligned with how the actual biology is: What is the identity of a cell given a partial set of its gene expression. In other words, can we identify the biological processes that make up the identity of a cell consistently through looking at different views of its transcriptome profile?

Technically, this is done using contrastive learning. Each cell is split into two disjoint gene subsets. The model sees both views and is trained to pull them together in embedding space, while pushing apart views from different cells.

To make this work, the team borrowed a trick from BERT: a dedicated CLS token. In language, CLS represents a sentence. In scConcept, it represents the entire cell. The loss is applied directly to that token.

“If the model can solve this task,” Mojtaba explained, “it has to develop a high-level idea of cell identity. There’s no shortcut.”

That framing turns the embedding from a side effect into the main objective. The model is no longer rewarded for local gene accuracy, but for capturing global cellular identity.

📊 Where the data work really shows

The architecture alone is only half the story. The other half lives in how the data is presented. One of the team’s biggest insights came from a failed experiment. Early versions of scConcept used the same binning strategies as other models. The result was large, technology-driven shifts in the embedding space.

“That was the moment we realised something was fundamentally wrong,” Mojtaba said. “The model was learning the technology, not the biology.”

The fix was surprisingly simple. Instead of feeding absolute expression values, scConcept uses rank encoding. Genes are ordered by expression within each cell. Only relative relationships matter.

If gene A is higher than gene B, the model sees that. The actual counts are ignored.

This turns out to be remarkably robust. Different assays may disagree on absolute numbers, but gene rankings tend to stay stable. Rank encoding strips away much of the batch effect before the model even starts learning.

Then comes gene subsetting. During training, the model constantly sees partial views of cells, including realistic gene panels from spatial technologies. This forces the embedding to stay stable even when most genes are missing.

The result is a representation that does not panic when faced with a 300-gene panel instead of a full transcriptome.

📢 Why it is different

The evaluation results reflect that design choice. scConcept consistently outperforms other foundation models on cell type annotation, cross-technology transfer, and spatial imputation, often matching or beating domain-specific tools. More importantly, it fails more gracefully.

When information is missing, performance drops for the right reasons, not because the embedding space collapses. Closely related cell types remain close. Spatial structure is preserved. Adaptation improves things further without retraining the entire model.

One subtle but important point the authors stress is that scConcept is not a batch correction method. It does not erase real biological differences. It just stops the model from confusing technology with biology.

That distinction matters if you actually want to interpret what the model is doing.

🔮 The Future

The long-term vision goes further than single cells. Right now, scConcept is trained on around 30 million cells, deliberately matching the scale of existing models. The next step is obvious: train on hundreds of millions. Initiatives like the Human Cell Atlas and repositories such as CELLxGENE now host hundreds of millions of publicly available cells, making this scale increasingly realistic.

However, the more interesting shift is conceptual. As Mojtaba put it, “Cells are the starting point. But biology doesn’t stop there. Tissues matter. Spatial context matters. Patients matter.”

Contrastive learning over partial views opens the door to representing tissue sections, neighbourhoods, even whole samples. Instead of asking whether two gene sets come from the same cell, future models might ask whether two regions come from the same tissue state.

That feels like a natural progression, especially as spatial assays become routine.

For now, scConcept is a reminder that better questions often beat bigger models. By focusing on what practitioners actually use, rather than what looks impressive on paper, it points toward a quieter but more useful future for single-cell AI, which is probably what the field needs right now.

👨‍🔬 Get in touch with Mojtaba.

📑 Read the paper.

💻 Check out the code.

Thanks for reading Kiin Bio Weekly!

💬 Get involved

We’re always looking to grow our community. If you’d like to get involved, contribute ideas or share something you’re building, fill out this form or reach out to me directly.

Share Kiin Bio Weekly

Subscribe now to stay at the forefront of AI in Life Science and keep up with this upcoming season of deep dives.

Connect With Us

Have questions on this or suggestions for our next deep dive? We’d love to hear from you!

📧 Email Us | 📲 Follow on LinkedIn | 🌐 Visit Our Website

Keeping up with AI x life science news can get exhausting.

It’s scattered across LinkedIn, X, Substack, arXiv, Slack, newsletters... and you still somehow miss the things that actually matter. Too much noise, not enough signal.

We’re building something to fix that: a smarter, more powerful way to stay on top of what’s actually relevant to you.

But we want to build it with you, not just for you. Take 2 minutes to tell us what’s missing. What you share will directly shape what we build, and you’ll be the first to benefit from it.

Share your input

🇺🇸 We’re heading to Bio-IT World in Boston, May 19-21.

Our CEO Filippo and CTO Bogdan will be there and would love to meet anyone thinking about:

• How AI is actually changing preclinical workflows (not just the hype)

• Why drug discovery is a systems problem, not just a science one

• What it takes to go from 5-year timelines to something radically faster

No pitch, just good conversation. If any of that’s on your mind, reach out - we’ll find a time to grab a coffee.

RegVelo: Gene-Regulatory-Informed Dynamics of Single Cells

🔬 RNA velocity models cellular dynamics but ignores gene regulatory interactions. Conversely, gene regulatory network inference methods neglect dynamics entirely. No existing approach jointly captures both, limiting our ability to simulate perturbations and predict how regulatory changes drive cell fate decisions.

RegVelo from Helmholtz Munich and Fabian Theis’s lab bridges this gap. It is an end-to-end deep generative framework that jointly infers transcriptome-wide splicing kinetics and gene regulatory interactions from scRNA-seq data, producing an actionable in silico cell for perturbation simulation.

🧬 RegVelo encodes unspliced and spliced RNA into a latent space, then models transcription as a regulated process governed by a GRN weight matrix. A parallel high-dimensional ODE solver couples all gene dynamics simultaneously rather than treating genes independently. Prior GRN knowledge from ATAC-seq or public databases constrains the network, while data-driven refinement learns new regulatory edges and edge weights.

⚡ On cell cycle data, RegVelo achieves a cross-boundary correctness of 0.864, velocity consistency of 0.873, and significantly outperforms scVelo and veloVI (p < 0.001). For GRN inference, it ranks first among six methods on edge prediction (median AUROC = 0.59) and achieves AUROC = 0.95 for identifying known lineage driver TFs across four hematopoietic lineages. Predictions validated by CRISPR-Cas9 knockout and single-cell Perturb-seq in zebrafish neural crest.

🔬 Applications and Insights

1️⃣ In Silico Perturbation Screening

Masking regulons and comparing velocity fields lets researchers simulate gene knockouts computationally, predicting cell fate shifts before running wet-lab experiments.

2️⃣ Cell Fate Driver Discovery

Applied to zebrafish neural crest, RegVelo identified tfec as a key early driver and elf1 as a regulator of pigment cell fate, both validated in vivo with CRISPR-Cas9.

3️⃣ Lineage-Specific GRN Recovery

In hematopoiesis, RegVelo recovered known lineage drivers (Smarca1, Pdx1, Mnx1, Hhex) across four lineages with high ranking accuracy (AUROC = 0.95).

4️⃣ Uncertainty-Aware Predictions

As a Bayesian generative model, RegVelo quantifies intrinsic and extrinsic cell state uncertainty, giving confidence estimates for both velocities and inferred regulatory edges.

💡 Why This Is Cool

This is the first framework to couple RNA velocity with gene regulatory networks in a single generative model. Rather than inferring dynamics and regulation separately and hoping they align, RegVelo learns them jointly. The result is a model that can simulate what happens when you perturb the regulatory wiring, with predictions validated from in silico all the way to in vivo knockouts. That closes the loop from computational hypothesis to experimental confirmation.

📄 Read the paper

💻 Try the code

TTM: Triplet Tumbling Microscopy Enables In Situ Quantification of Protein Complex Assembly and Dynamics

🔬 Protein-protein interactions drive nearly every cellular process, but measuring them inside living cells remains limited. FRET requires two labels and prior knowledge of interacting partners, while fluorescence anisotropy only works for small proteins below 50 kDa. There is no broadly applicable way to quantify protein complex size and binding dynamics in situ in real time.

TTM (Triplet Tumbling Microscopy) from Calico Life Sciences solves this by measuring rotational diffusion of protein complexes using only a single fluorescent tag. By leveraging long-lived triplet states in fluorescent proteins, TTM extends the measurable timescale from nanoseconds to hundreds of microseconds, covering the full range of cellular protein complexes.

🧬 TTM uses a pulsed excitation sequence: a 488 nm pulse generates triplet states aligned with the excitation polarisation, then an infrared trigger pulse (785-940 nm) reads out their orientation after a variable delay. As proteins tumble, the triplets lose alignment at a rate proportional to complex size. Rigid fluorescent protein tags (truncated mVenus and mStayGold) ensure tag motion faithfully reports target motion.

⚡ In purified protein systems, tumbling time constants scale linearly with molecular weight (r squared = 0.99). In living U2OS cells, TTM resolves complexes from 41 to 195 kDa from single-cell recordings (r squared = 0.85). It detects the approximately 10% size change from E6AP binding HPV16 E6 protein, quantifies p53 homo-oligomerisation states across nine point mutations, and tracks rapamycin-induced dimerisation dynamics in real time at approximately 3 Hz.

🔬 Applications and Insights

1️⃣ Single-Tag Interaction Detection

Unlike FRET, TTM requires only one fluorescent label and no prior knowledge of binding partners, making it applicable to uncharacterised or unexpected interactions.

2️⃣ Live-Cell Binding Dynamics

Real-time imaging at approximately 3 Hz captures the kinetics of complex formation as it happens, not just endpoint measurements.

3️⃣ Oligomerisation State Profiling

TTM distinguishes monomers, dimers, and tetramers of p53 in cells, revealing how tetramerisation domain mutations shift the equilibrium between functional states.

4️⃣ Standard Microscope Compatibility

The hardware requirements (pulsed lasers and an intensified camera) are compatible with most fluorescence microscopes, lowering the barrier to adoption across labs.

💡 Why This Is Cool

TTM fills a gap that has persisted for decades in cell biology: measuring how big a protein complex is inside a living cell, in real time, with a single label. The ability to track binding dynamics and oligomerisation states at physiological concentrations opens the door to studying protein interactions in their native context rather than in lysates or reconstituted systems. One tag, one measurement, real answers.

📄 Read the paper

Path2Space: AI-Predicted Spatial Transcriptomics Unlocks Breast Cancer Biomarkers from Pathology

🔬 Spatial transcriptomics is transforming our understanding of tumour heterogeneity, but its high cost limits large-scale biomarker discovery. Previous efforts to predict gene expression from histopathology slides have been restricted to small gene sets, precluding survival and treatment response analyses in large clinical cohorts.

Path2Space from NIH’s National Cancer Institute and Cedars-Sinai predicts the spatial expression of thousands of genes directly from routine H&E-stained histopathology slides. Trained on extensive breast cancer spatial transcriptomics data, it outperforms 21 established methods and enables scalable biomarker discovery without molecular assays.

🧬 Path2Space uses CTransPath, a digital pathology foundation model, to extract features from colour-normalised tile images around each spatial transcriptomics spot. A multilayer perceptron predicts log-transformed expression for 14,068 genes per spot. A spatial smoothing step averages predictions with neighbouring spots to mitigate technical variability. Trained on the Bassiouni et al. cohort comprising 56,567 matched image-expression spot pairs from 14 patients.

⚡ Median gene-wise PCC of 0.38 (smoothed) across 14,068 genes, with 6,629 genes exceeding PCC > 0.4. Binary classification of high versus low expression yields median AUC of 0.70, with 3,116 genes surpassing 0.75. Generalises robustly across three independent external cohorts (HEST, Martinez, HTAN). Applied to 976 TCGA breast cancer patients, Path2Space identifies three prognostic SpatioTypes and predicts chemotherapy and trastuzumab response at accuracy levels equal to or exceeding bulk sequencing biomarkers.

🔬 Applications and Insights

1️⃣ Low-Cost Spatial Biomarker Discovery

Derives spatial gene expression landscapes from routine pathology slides without expensive molecular assays, enabling large-cohort studies previously limited by cost.

2️⃣ Prognostic Breast Cancer Subtyping

Unsupervised clustering of predicted spatial transcriptomic profiles identifies three SpatioTypes with distinct biology and survival outcomes across 976 patients.

3️⃣ Treatment Response Prediction

Spatial biomarkers from H&E slides predict response to chemotherapy and trastuzumab at accuracy levels matching or exceeding those from bulk tumour sequencing.

4️⃣ Archival Tissue Applicability

Works on standard FFPE and fresh-frozen archival tissue, meaning existing hospital slide collections can be retrospectively analysed without new sample collection.

💡 Why This Is Cool

Spatial transcriptomics has been too expensive to run on the thousands of patients needed for robust biomarker discovery. Path2Space sidesteps this entirely by inferring spatial gene expression from H&E slides that hospitals already collect for every tumour. Turning routine pathology into a spatial omics readout could democratise precision oncology for any institution with a slide scanner.

📄 Read the paper

💻 Try the code

🗓️ Events & Competitions

The best competitions, hackathons, and community challenges in AI x life sciences, curated weekly. Know something worth featuring? Reply and let us know.

More upcoming events:

London Protein Design Day | June 23, Imperial College London

The first edition of a one-day symposium bringing together London’s protein design community and beyond. Programme spans AI-driven design, molecular dynamics, and bioinformatics, with applications across enzymes, antibodies, and materials. Organised by Pietro Sormanni, Rebecca Birolo, and Jakub Lála. Abstract deadline for poster/oral presentations is this Saturday (May 17). In person only.

BioHackathon Europe 2026 | November 9-13, Barcelona

ELIXIR’s annual international bioinformatics hackathon, running since 2018. 160+ participants, five days of collaborative coding on open bioinformatics infrastructure and tools. The call for project proposals opens March 16 and closes April 15 - so if you want to lead a project, that’s your window.

Thanks for reading!

💬 Get involved

We’re always looking to grow our community. If you’d like to get involved, contribute ideas or share something you’re building, fill out this form or reach out to me directly.

Connect With Us

Have questions or suggestions? We'd love to hear from you!

📧 Email Us | 📲 Follow on LinkedIn | 🌐 Visit Our Website

Keeping up with AI x life science news can get exhausting.

It’s scattered across LinkedIn, X, Substack, arXiv, Slack, newsletters... and you still somehow miss the things that actually matter. Too much noise, not enough signal.

We’re building something to fix that: a smarter, more powerful way to stay on top of what’s actually relevant to you.

But we want to build it with you, not just for you. Take 2 minutes to tell us what’s missing. What you share will directly shape what we build, and you’ll be the first to benefit from it.

Share your input

Today we’re looking at Adaptyv, a Lausanne-based startup that’s building the experimental infrastructure the protein design revolution has been missing. We spoke with Tudor, who leads community and protein engineering efforts at Adaptyv, about why designing proteins computationally was only ever half the problem, and what it takes to close the gap between prediction and proof.

“People are realising now that data is the bottleneck for what we can currently achieve in ML for protein design. They need specialised functional data and they need to generate it fast.”

— Tudor, Adaptyv

🔴 The Problem

AI protein design has exploded. Diffusion models, language models, and structure prediction tools can now generate novel protein sequences in minutes. But there’s a persistent bottleneck that sits downstream of all that computation: actually testing whether the designs work.

Before Adaptyv, a computational protein designer who wanted to validate their binders had limited options. You could work with a contract research organisation, with complex onboarding, long timelines, and months before you saw results. You could do it yourself if you were lucky enough to have lab access. Or you could be in one of a handful of major labs, like the Baker lab, that had the infrastructure to run binding affinity measurements at scale.

For everyone else, the growing wave of independent protein designers, small academic groups, and early-stage biotechs training their own generative models, experimental validation was a wall. You could design as many proteins as you wanted on a computer, but you had no efficient way to know which ones actually folded, bound their target, or did anything useful.

“It was quite a black and white situation,” Tudor explains. “You were either in one of the bigger labs or you weren’t. And if you weren’t, you were designing stuff on the computer with no idea what it does in the real world.”

💡 The Idea

Adaptyv’s answer: a cloud lab purpose-built for protein designers.

The thesis has been core to the company since its founding by Julian Englert and Daniel Nakhaee-Zadeh, both engineers from EPFL. They initially built a microfluidics platform for high-throughput antibody-antigen binding testing, but when that approach proved too experimental, they pivoted. First to BLI and SPR-based measurement approaches in late 2023, then to an official platform launch in 2024.

The model is simple. You sign up on the Foundry platform. Upload your sequences via CSV. Add details: tags, antibody formats, terminal modifications. Select your target from a catalogue sourced from multiple suppliers. Submit. Adaptyv handles everything else.

The Foundry experiment creation workflow: choose your desired assay, choose targets for binding experiments, number of replicates, upload your sequences, and get an experiment quote for your draft experiment.

Behind the interface sits what the team calls LabOS, an orchestration brain that manages scheduling, expression, measurement, curve fitting, quality control, and results delivery across their automated systems. Users get back visualisations, raw sensorgrams, fitted binding curves, and all underlying data in a single package.

The key technical innovation is miniaturisation. By running cell-free expression reactions in microlitre volumes rather than large batch cultures, Adaptyv achieves roughly a 1,000x reduction in reagent use. No recombinant E. coli. No multi-day protein expression waits. Just small, fast, cell-free reactions that produce enough protein to measure on BLI or SPR.

📊 The Data

Current turnaround is approximately two weeks: one week for target QC if the target is new, then one week for the experiments themselves. The goal is to push that down to two to three days, and ultimately to near-instant validation that matches the speed of ML training cycles.

“Ideally, you want to get the experimental validation within the same time pressure as a training run,” Tudor says. “Almost instant. If we could make it faster, we would.”

The platform isn’t just a validation service; it’s becoming a data engine. Some customers are already running active learning loops: generate a set of designs, send them to Adaptyv, get results, retrain, repeat. Others are running large-scale campaigns to map the druggability of entire target spaces.

A typical use case: a team trains a diffusion model on PDB data, generates novel binders, and validates them through Adaptyv. The experimental results feed back into the next round of model training. Each cycle produces better designs and richer data.

“Both work hand in hand,” Tudor explains. “Now people are using us primarily for validation. But ultimately we want to go in the direction of large-scale custom data generation campaigns. We want to be a protein data centre.”

🔬 Proteinbase

The data Adaptyv was generating led naturally to their second platform: Proteinbase.

The problem it addresses is fragmentation. Existing protein databases use different protocols, different measurement methods, and different standards. Data from one source is hard to compare with data from another. Teams trying to train models on aggregated functional protein data spend enormous effort just standardising datasets, and even then, results often don’t reproduce.

“People were saying that what they’re missing is a unified database for functional proteins with standardised protocols,” Tudor says. “Everyone was trying to standardise datasets from different sources and running into the same problems. At least if everything is from a single source, it’s better for training better models.”

Proteinbase hosts competition data, provides unified downloads, and is building out community features: leaderboards, badges, knowledge sharing. It’s part database, part competition platform, part community hub.

The competitions have been a major growth driver. Recent rounds attracted over 600 participants, with hit rates up to 13% and binders in the nanomolar range. Adaptyv has also been hosting hackathons internationally in San Francisco and Berlin, with more planned across Europe. A recent GEM Bio Workshop competition hosted at ICLR on a disordered target (RBX1) drew more than 180 submissions.

💊 Who It’s For

Adaptyv serves three overlapping audiences: academic groups training and validating protein design models, biotech and pharma teams running lead optimisation or target characterisation campaigns, and independent protein designers who previously had no access to experimental validation.

Pricing is transparent and visible on the platform.

The philosophy is democratisation. Protein design has historically been concentrated in a small number of elite labs. Adaptyv is opening that up, not just through the lab infrastructure but through the community.

“Protein design has been super insular,” Tudor says. “It was either you were in one of the bigger labs or you weren’t. Giving people access to experimental validation and a community to share what they’ve learned, that’s how you grow the field.”

🔮 The Future

The next 12 months are focused on speed and automation.

On the Adaptyv platform: two-day experimental results, an expanded assay array, more standardisation, and deeper investment in agent-based workflows. The API is already designed for programmatic experiment submission, and the team is building towards a future where AI agents can control every aspect of the pipeline, from sample handling to liquid handlers to results retrieval.

“We have a two-level system,” Tudor explains. “A higher-level customer API for submitting experiments and getting results, and a lower-level API for every single machine interaction. In the future, agents could control all of it.”

On Proteinbase: more visible content, international hackathon expansion across Europe, and continued growth of the competition platform. The demand is clear: organisers are reaching out to Adaptyv to add protein design tracks to their events, and recent competitions have been oversubscribed.

The longer-term vision ties both platforms together. As the field recognises that data, not compute or model architecture, is the binding constraint on progress in ML for protein design, Adaptyv is positioning itself as the infrastructure layer that generates, validates, and distributes that data at scale.

Design is only half the problem. Adaptyv is building the other half.

👨‍🔬 Get in touch with Tudor.

💻 Adaptyv Website.

🌐 Adaptyv on LinkedIn.

🧬 Explore Proteinbase.

Thanks for reading Kiin Bio Weekly!

💬 Get involved

We’re always looking to grow our community. If you’d like to get involved, contribute ideas or share something you’re building, fill out this form or reach out to me directly.

Share Kiin Bio Weekly

Subscribe now to stay at the forefront of AI in Life Science and keep up with this upcoming season of deep dives.

Connect With Us

Have questions on this or suggestions for our next deep dive? We’d love to hear from you!

📧 Email Us | 📲 Follow on LinkedIn | 🌐 Visit Our Website

Keeping up with AI x life science news can get exhausting.

It’s scattered across LinkedIn, X, Substack, arXiv, Slack, newsletters... and you still somehow miss the things that actually matter. Too much noise, not enough signal.

We’re building something to fix that: a smarter, more powerful way to stay on top of what’s actually relevant to you.

But we want to build it with you, not just for you. Take 2 minutes to tell us what’s missing. What you share will directly shape what we build, and you’ll be the first to benefit from it.

Share your input

🇺🇸 We’re heading to Bio-IT World in Boston, May 19-21.

Our CEO Filippo and CTO Bogdan will be there and would love to meet anyone thinking about:

• How AI is actually changing preclinical workflows (not just the hype)

• Why drug discovery is a systems problem, not just a science one

• What it takes to go from 5-year timelines to something radically faster

No pitch, just good conversation. If any of that’s on your mind, reach out - we’ll find a time to grab a coffee.

MolJSON: Molecular Representations for Large Language Models

🔬 LLMs are increasingly used in chemistry for tasks like reaction prediction and structure elucidation, but they need to read and write molecular structures reliably. Previous work has defaulted to SMILES strings or IUPAC names, but no one has systematically tested whether these formats are actually good for LLMs. Both impose strict serialisation rules that may not align with how language models process information.

Researchers at Oxford introduce MolJSON, a structured JSON schema that represents molecular graphs explicitly as lists of atoms and bonds. Unlike SMILES (which requires a specific graph traversal) or IUPAC (which requires rule-based nomenclature), MolJSON presents the molecular graph directly in a format compatible with LLM structured output modes.

🧬 They evaluated five molecular representations across 78,045 algorithmically generated questions spanning translation, shortest-path reasoning, and constrained generation tasks using GPT-5-nano, GPT-5-mini, GPT-5, and Claude Haiku 4.5. MolJSON consistently outperformed all existing formats as both an input and output representation.

⚡ On translation tasks, GPT-5 achieved 71.0% accuracy converting IUPAC to MolJSON versus 43.7% for IUPAC to SMILES. For constrained generation, GPT-5 reached 95.3% accuracy generating MolJSON versus 64.0% for SMILES and 76.3% for IUPAC. MolJSON was also 1.8x more token-efficient than SMILES on reasoning tasks. Performance advantages held even though models were never explicitly trained on MolJSON.

🔬 Applications and Insights

1️⃣ Better Chemistry Agents

LLM-based chemistry systems that read and write molecules can operate more reliably by switching to MolJSON, reducing errors from format parsing failures.

2️⃣ Robust Molecular Reasoning

MolJSON maintained high accuracy even on complex molecules with fused rings and high atom counts, where SMILES and IUPAC performance degraded sharply.

3️⃣ Token-Efficient Representations

Graph-based formats let models skip the internal reconstruction step needed for traversal-based representations, using fewer reasoning tokens and reducing latency.

4️⃣ Format-Agnostic Improvement

MolJSON outperformed SMILES and IUPAC despite those formats being well-represented in LLM training data, suggesting explicit graph encodings are intrinsically better aligned with how LLMs process structure.

💡 Why This Is Cool

Everyone building LLM chemistry tools has been using SMILES because it was already there. This paper shows that is leaving significant performance on the table. A simple change in molecular representation, with no model retraining, unlocks dramatically better accuracy across translation, reasoning, and generation. The fact that LLMs spontaneously generate semantically meaningful atom identifiers (like “C_acyl” or “Npip”) in MolJSON suggests these models can reason about molecular graphs more naturally when given an explicit graph format.

📄 Read the paper

💻 Try the code

PLaTITO: Protein Language Model Embeddings Improve Generalisation of Implicit Transfer Operators

🔬 Molecular dynamics simulations are essential for understanding protein behaviour, but conventional MD is computationally prohibitive for biologically relevant timescales. Generative molecular dynamics methods learn surrogate models from trajectory data, but they typically require large collections of long MD trajectories and struggle to generalise to unseen protein systems.

PLaTITO from Chalmers, Copenhagen, and DTU introduces coarse-grained transferable implicit transfer operators (TITO) for protein molecular dynamics that generalise to out-of-distribution protein systems. By conditioning on protein language model embeddings from ESM and structure embeddings from Proteina, the model learns to transfer across diverse proteins without system-specific fine-tuning.

🧬 Trained on the mdCATH dataset (4,471 domains, ~56 ms aggregate simulation time), PLaTITO learns long-time transition densities conditioned on backbone coordinates, sequence, temperature, and time step. The architecture uses a two-stage design: a conditioning network produces per-residue representations, and a velocity network generates the flow field for sampling future conformations.

⚡ PLaTITO-Big (19M parameters) outperforms BioEmu across all equilibrium sampling metrics on fast-folding proteins while requiring substantially less training data (56 ms vs. 216 ms) and compute (1,100 GPU hours vs. 9,216). It recovers non-Arrhenius temperature-dependent folding kinetics and explores cryptic binding pockets, generating trajectories with repeated folding and unfolding events at microsecond timescales.

🔬 Applications and Insights

1️⃣ Out-of-Distribution Generalisation PLaTITO transfers to unseen proteins without fine-tuning, unlike Boltzmann Emulators that require system-specific training data.

2️⃣ Data-Efficient Training Achieving state-of-the-art equilibrium sampling with 4x less MD data and 8x less compute than BioEmu demonstrates that transfer learning can dramatically reduce the data barrier.

3️⃣ Temperature-Dependent Kinetics Explicitly conditioning on temperature lets PLaTITO capture non-Arrhenius behaviour, reproducing physically meaningful folding and unfolding rates across temperature ranges.

4️⃣ Cryptic Binding Pocket Discovery PLaTITO-Big samples conformational transitions to cryptic pockets from both apo and holo states, opening applications in drug discovery for targeting hidden binding sites.

💡 Why This Is Cool

This is the first transferable molecular dynamics model that genuinely generalises across protein systems while beating dedicated Boltzmann Emulators on their own benchmarks. The key insight is that pretrained protein language models encode enough structural and evolutionary information to let a small (3-19M parameter) dynamics model transfer across diverse folds. Generating realistic microsecond-scale folding trajectories on a single GPU in seconds, rather than months of conventional simulation, changes what is computationally accessible for studying protein dynamics.

📄 Read the paper

OpenBind: A Structure-Affinity Dataset for Structure-Based AI in Drug Discovery

🔬 Structure-based AI for drug discovery is held back by a data bottleneck. Public protein-ligand datasets are sparse, unevenly distributed, and rarely link crystallographic binding modes with quantitative affinity measurements at scale. Current ML methods for docking, cofolding, and affinity prediction are difficult to evaluate fairly because most benchmarks overlap with training data.

The OpenBind consortium (Diamond Light Source, Oxford, and partners) releases a dense structure-affinity dataset: 925 crystallographic binding events from 699 compounds with affinity measurements for 601 compounds, all targeting EV-A71 2A protease, a viral target relevant to pandemic preparedness. The data are deliberately positioned in an under-represented region of protein-ligand space relative to existing public structures.

🧬 The dataset includes fragment screen hits and follow-on compounds with KD values from Creoptix WAVEsystem measurements, creating a coherent experimental series where users can study local structure-activity relationships. Reference benchmarks span conventional docking (AutoDock Vina), ML docking (GNINA, DiffDock), cofolding (AlphaFold3, Boltz, OpenFold3), and affinity prediction methods.

⚡ Redocking achieves up to 85% success (GNINA), but cross-docking into apo structures drops below 5% due to binding-site loop conformational changes. Cofolding methods reach 36% (OpenFold3-p2), but fine-tuning on fragment-screen data boosts this to 76%, approaching redocking performance. For affinity prediction, a simple molecular-weight baseline outperforms most structure-based methods, highlighting how challenging this task remains.

🔬 Applications and Insights

1️⃣ Benchmarking Beyond Training Data

The EV-A71 2A protease complexes are dissimilar to pre-2021 PDB data, providing a genuine test of whether cofolding and docking methods generalise or just memorise near-neighbours.

2️⃣ Fragment Screens as Training Data

Fine-tuning cofolding models on fragment-bound structures doubled success rates on follow-on compounds, showing that early experimental data can feed directly into AI model improvement.

3️⃣ Separating Failure Modes

The dataset cleanly distinguishes receptor-conformation failures from ligand-placement failures, letting method developers target specific weaknesses rather than debugging aggregate metrics.

4️⃣ Affinity Prediction Reality

Check Simple baselines beating structure-based methods on this dataset is a clear signal that current affinity models may be learning chemical trends rather than genuine protein-ligand interaction physics.

💡 Why This Is Cool

OpenBind is not just releasing more structures. It is building the experimental infrastructure to generate the kind of data that structure-based AI actually needs: dense, linked structure-affinity measurements within coherent chemical series, positioned where current models are weakest. The fragment fine-tuning result is particularly striking. It shows that a relatively small crystallographic screen can transform cofolding performance on a new target, pointing toward a practical workflow where early experiments directly improve computational predictions for the same campaign.

📄 Read the data

💻 Try the benchmarks

📬 Newsletter Shout-Out

This week we're shouting out Building in BioAI, a monthly newsletter from Joe:

Building in BioAI is a monthly newsletter for those operating in, or interested in, the AI-enabled biology space. That’s founders, technical leaders, and individual contributors working within areas like therapeutics, diagnostics, and tooling.

Joe’s roundup centres on observations from within the space, including analysis of how teams are structuring themselves, what’s changing in hiring, where funding is landing, what headlines mean for growth, and how BioAI companies are thinking about commercialising what they’re building.

Each edition pulls from ongoing conversations with people doing the work day-to-day, as well as his own take on what’s hit headlines that month.

Joe recruits in this space day-to-day, and so often speaks from that vantage point. He spends most of his time inside these teams, hiring for them, speaking with founders and senior talent across the market. The aim isn’t to overstate where things are going, but to give a clear picture of what’s actually happening and why it matters if you’re hiring or looking to commercialise in BioAI.

🔗 Check it out!

🗓️ Events & Competitions

The best competitions, hackathons, and community challenges in AI x life sciences, curated weekly. Know something worth featuring? Reply and let us know.

More upcoming events:

BioHackathon Europe 2026 | November 9-13, Barcelona

ELIXIR’s annual international bioinformatics hackathon, running since 2018. 160+ participants, five days of collaborative coding on open bioinformatics infrastructure and tools. The call for project proposals opens March 16 and closes April 15 - so if you want to lead a project, that’s your window.

Thanks for reading!

💬 Get involved

We’re always looking to grow our community. If you’d like to get involved, contribute ideas or share something you’re building, fill out this form or reach out to me directly.

Connect With Us

Have questions or suggestions? We'd love to hear from you!

📧 Email Us | 📲 Follow on LinkedIn | 🌐 Visit Our Website