AI in Life Science: Weekly Insights
Weekly Insights | July 3, 2025
In this issue:
Welcome back to your weekly dose of AI news for Life Science!
This week, we have some exciting new models lined up for you:
Dive into these game-changing innovations and explore how they are transforming the biotech and healthcare landscapes!
What’s slowing you down? Let us tackle it.
AlphaGenome: DeepMind’s AI for Decoding the Genome 🧬
DeepMind’s new model, AlphaGenome, advances our ability to interpret the human genome by predicting the effects of genetic variation with unprecedented detail. Building on DeepMind’s breakthroughs in protein structure and function prediction, AlphaGenome expands this capability to the most challenging parts of the genome: the non-coding regions. It delivers detailed, base-pair-resolution predictions of how DNA variants affect gene regulation, splicing, chromatin state, and 3D genome contacts. This makes it possible to decode how genetic variation influences human health and disease, and helps researchers prioritise which variants to study in the lab or target with precision therapies.
🔨Applications
Predict the regulatory effects of millions of genetic variants, including in regions of the genome where functional impact has been hardest to infer.
Support researchers working on rare diseases, complex traits, and variant prioritisation for experimental validation.
Serve as an open benchmark for improving computational methods in human genetics.
📌 Key Insights
Massive Long-Context Coverage - AlphaGenome predicts variant effects over 1 million base pairs at single-nucleotide resolution, integrating gene expression, splicing, chromatin accessibility, and 3D genome structure in a single model which is a first at this scale.
Superior Predictive Performance - AlphaGenome outperforms leading benchmarks on 24 of 26 tests, including a 25% gain for predicting the sign of eQTL effects and up to 59% improvement for classifying splicing QTLs. It also improves 3D chromatin contact prediction by 42% compared to Orca.
Unified Variant Interpretation - By modelling multiple genomic modalities together, AlphaGenome helps researchers decode how non-coding and distal variants shape gene regulation, bridging gaps left by narrower, single-task tools.
STATE: Simulating Cellular Responses to Drug and Genetic Interventions 🧪
Therapeutic discovery relies on accurately predicting the impact of cellular perturbations. While there are many computational models for this very task- which offer great potential over experimental methods, they’re often weak in generalising effects from experimentally observed cellular contexts to unobserved ones. Introducing STATE, an AI model that can be used to find drug/genetic interventions to shift a diseased cell's state to a healthy one, all while accounting for cellular heterogeneity. STATE includes State Transition (ST), a transformer that predicts how cells change due to a treatment, and State Embedding (SE), a state-of-the-art 600M parameter foundation model for encoding single cell state.
🔨Applications
Precision Responses - Allows detection of cell type-specific differential expressions, including responses of stem cells, cancer cells, and immune cells to drugs, cytokines, and genetic perturbations.
Larger Exploration - Enables scientists to run millions of in silico perturbations to narrow down their hypotheses, a step closer towards solving the 2^20k combinations problem.
Development of New Tools - Serves as a backbone for developing new models for various tasks, such as repurposing existing drugs, predicting patient-specific treatment response or understanding potential adverse reactions in cell types beyond the intended drug target.
📌 Key Insights
Largest Perturbation Data - STATE is trained on observational data from nearly 170 million cells and single-cell perturbation data from more than 100 million cells across 70 cell lines, more than any model to date.
Superior Performance - STATE demonstrated a 50% improvement in distinguishing perturbation effects and achieved 2x the accuracy in identifying true differentially expressed genes.
Associated Tools - The authors also released CELL-LOAD, an efficient data loading library for single-cell perturbation data, and CELL-EVAL, a comprehensive evaluation framework for virtual cell modelling that advances beyond conventional metrics.
Chai-2: A New Benchmark for De Novo Antibody Design 🧬
Designing functional antibodies entirely from scratch has remained one of protein science’s toughest challenges. Chai-2, developed by Chai Discovery, sets a new benchmark, achieving up to 16% hit rates for fully de novo antibody designs, representing a more than 100-fold improvement over previous computational methods. In a single round of lab validation, Chai-2 generated binders for 50% of tested targets, all without relying on pre-existing antibody scaffolds for those antigens.
🔨Applications
Reduce the scale and cost of antibody discovery by designing ≤ 20 candidates per target and moving directly to wet-lab testing.
Rapidly generate miniproteins, VHHs, or scFvs for novel targets that lack existing binders in the Protein Data Bank.
Accelerate the design-validation loop, completing the full workflow from AI generation to confirmed hits in under two weeks.
📌Key Insights
Reliable de novo design at practical scale - Chai-2 achieves up to 16% hit rates for fully de novo antibodies, a leap of over 100× compared to older methods. It delivers binders for 50% of novel targets tested, often with only 20 designs per target.
Novelty and diversity confirmed - Experimental results show the designed antibodies are structurally and sequence-wise distinct from known binders in SAbDab. Many targets yield multiple unique structural clusters, demonstrating Chai-2’s capacity to explore diverse solutions.
Fast lab validation cycle - Designs move from in silico output to wet-lab confirmation in about two weeks using standard 24-well plate assays. Binding affinities often reach low nanomolar or picomolar range, with favourable developability profiles.
Did you find this newsletter insightful? Share it with a colleague!
Subscribe Now to stay at the forefront of AI in Life Science.
Connect With Us
Have questions or suggestions? We'd love to hear from you!
📧 Email Us | 📲 Follow on LinkedIn | 🌐 Visit Our Website