#10: Life Science x AI
Welcome back to your weekly dose of AI news for Life Science!
This week, we have some exciting developments lined up for you:
Chain of Diagnosis (CoD) is making AI-driven medical diagnostics more transparent with its physician-like reasoning framework 💡
Apollo is expanding the reach of medical AI, bringing multilingual models to healthcare for billions 🚀
Life Science tools of the week ⚒️
Dive into these game-changing innovations and explore how they are transforming the biotech and healthcare landscapes!
Chain of Diagnosis (CoD): Pioneering Transparent AI in Healthcare with DiagnosisGPT💡
AI-driven diagnostics often face a key challenge: interpretability. The Chain of Diagnosis (CoD) addresses this by mirroring the way physicians think, offering a transparent, step-by-step diagnostic process. Unlike traditional models that give black-box results, CoD integrates disease confidence scores at each stage, ensuring doctors understand AI decisions. Powered by this framework, DiagnosisGPT handles over 9,600 diseases, significantly outperforming other AI models in real-world settings, delivering accuracy with transparency.
Key insights 👇
📌 Physician-like transparency: CoD breaks down complex diagnoses into interpretable steps, enhancing decision-making.
📌 Superior performance: DiagnosisGPT, driven by CoD, surpasses current large language models in medical benchmarks.
📌 Real-world readiness: With CoD, doctors can better trust and apply AI diagnostics in patient care, bridging the gap between machine predictions and clinical reasoning.
The Bottom Line: The CoD framework is a major step forward in making AI-driven medical diagnostics more transparent and clinically actionable. We see this as a critical advancement in the future of AI in healthcare.
Apollo: Democratizing Medical AI with Multilingual Models 🚀
Apollo aims to revolutionize medical AI by making it accessible to 6.1 billion people through multilingual large language models (LLMs). The team developed ApolloCorpora, a dataset in six major languages—English, Chinese, Hindi, Spanish, French, and Arabic. Apollo’s models, from 0.5B to 7B parameters, outperform other LLMs of similar size, with the Apollo-7B model excelling in various language benchmarks. These models can even enhance larger LLMs via proxy-tuning, making them adaptable without additional fine-tuning.
Key Insights 👇:
📌 Multilingual support: Apollo improves healthcare accessibility in under-resourced regions, with real-time, scalable models for local settings.
📌 Efficient and adaptable: Proxy-tuning enables Apollo models to enhance larger LLMs, democratizing medical AI access worldwide.
Next Steps: Apollo’s lightweight models can be used for medical inference and research, bringing scalable AI to clinicians and researchers with limited resources.
Life Science tools of the week 🛠️
1/ DFMDock - Protein Docking
Developed at Johns Hopkins University, DFMDock is a groundbreaking diffusion model designed to unify the separate processes of protein docking, delivering both sampling and ranking in a single framework!
📌 DFMDock predicts forces and energies simultaneously, eliminating the need for separate ranking models.
📌 It has a 44% sampling success rate compared to DiffDock-PP’s 8%, and a 16% Top-1 ranking success rate, a huge leap from 0%.
📌 DFMDock captures physical energy landscapes, outperforming models in docking tasks, especially on the Docking Benchmark 5.5 dataset.
🔗 Code
2/ Harrison.rad.1 - Radiology
Harrison.ai has unveiled Harrison.rad.1, the most advanced multimodal large language model (LLM) for radiology, designed to revolutionize AI applications in medical imaging.
📌 Harrison.rad.1 is trained on millions of DICOM images and radiology reports, specifically tailored to X-ray modalities.
📌 It processes multiple X-ray views, detects and characterizes findings, and generates comprehensive reports based on patient context.
📌 Outperforms competitors on the VQA-Rad and RadBench benchmarks, with accuracy rates up to 82% on radiological tasks.
🔗 Join the waitlist for early access!
3/ GP-GPT: Revolutionizing Gene-Phenotype Mapping
The bioinformatics landscape has a new powerhouse! GP-GPT, developed by researchers at the University of Texas at Arlington and the University of Georgia, is designed for genetic-phenotype knowledge representation and genomics analysis.
📌 Trained on over 3 million terms from large-scale datasets like OMIM, DisGeNET, and dbGaP
📌 Outperforms state-of-the-art models like GPT-4 and Llama3 across genomics tasks
📌 Excels in gene-phenotype mapping, offering superior accuracy and speed for genomics research
🔗 Paper
BITE-SIZED COOKIES FOR THE WEEK 🍪
The global effort to map the human brain has released its first data through the BICAN Rapid Release Inventory. This dataset offers early access to single-cell data that will accelerate groundbreaking brain research.
Learn about more AI news in pharma and biotech: Big pharma, biotech relations ‘won't necessarily be symbiotic’ in future AI landscape: S&P.
Chinese drugmakers WuXi AppTec and WuXi Biologics are considering divesting operations, including WuXi AppTec's Philadelphia cell and gene therapy unit, amid scrutiny from the U.S. BIOSECURE Act and a 27% revenue drop from European customers.
Researchers have developed a machine learning model that can predict Alzheimer's disease up to seven years before symptoms appear by analyzing electronic health records, highlighting the potential of artificial intelligence in early diagnosis and treatment of complex diseases.
The Cancer AI Alliance (CAIA), formed by four leading cancer centers and supported by major tech companies, aims to harness AI to analyze vast cancer data for new treatment insights. With over $40 million in funding, CAIA plans to begin producing findings by the end of 2025.
Don’t forget to take our short survey to help us understand how you currently use AI in your day-to-day!