JJ Gao

Great functional genomics work from Genentech's AI team They identified three isolated threads of AI research on functional genomics data: 1. Predicting tracks of functional data (i.e. ChIP-seq, RNA-seq) 2. Training on sequences alone with language models, and 3. Training generative models to produce regulatory sequences. So they set out to make a "unifying multimodal masking framework for functional genomics" that creates a standard training paradigm across these different types of objectives. The framework, called Nona, takes in both DNA sequences and functional data tracks into a model where one or both inputs can be masked. In ML terms, this is basically rolling "predictive, generative, and self-supervised learning within a single framework." They showed how this paradigm improves performance for both functional data prediction and sequence design/generation. It's awesome to live in a time where top AI scientists are working at Genentech and openly publishing their research. (Link to preprint in comments.)

158

4 Comments

Are we poised to redefine response assessment in solid tumors to better correlate with long-term outcomes such overall survival? I am thrilled to share our latest collaborative work with Onc.AI at #ASCO25 on serialCTRS, a novel tumor response endpoint that leverages deep learning to quantify imaging changes over time. Unlike the traditional RECIST method which relies on a few manually selected target lesions, serialCTRS offers a more nuanced and comprehensive assessment that not only considers tumor lesions but also evaluates body composition components and vasculature properties based on the CT scans. Our study validated serialCTRS using both clinical trial and real-world datasets of advanced NSCLC. The results? SerialCTRS demonstrated higher predictive performance compared to RECIST, with improved concordance indices and ROC-AUC scores for 24-month overall survival. This innovative approach not only promises to personalize therapy plans but also accelerate drug development. Join us at #ASCO25 to delve into the details of our research and explore the potential of serialCTRS in transforming cancer care. #ASCO25 #DeepLearning #NSCLC https://lnkd.in/dj7Ykv2D

57

2 Comments

BREAKING NEWS! Two popular large language models (LLMs), GPT-4 and Llama-3, produced medical device-like decision support across a range of scenarios even though those are not regulated by the FDA. Authors of a new study were interested in finding out whether LLMs show promise for clinical decision support (CDS) even though none is currently authorized by the FDA as a CDS device.  "𝐹𝑜𝑙𝑙𝑜𝑤𝑖𝑛𝑔 𝑑𝑒𝑐𝑖𝑠𝑖𝑜𝑛 𝑠𝑢𝑝𝑝𝑜𝑟𝑡 𝑟𝑒𝑞𝑢𝑒𝑠𝑡𝑠 𝑎𝑏𝑜𝑢𝑡 𝑡𝑖𝑚𝑒-𝑐𝑟𝑖𝑡𝑖𝑐𝑎𝑙 𝑒𝑚𝑒𝑟𝑔𝑒𝑛𝑐𝑖𝑒𝑠, 𝟣𝟢𝟢% 𝑜𝑓 𝐺𝑃𝑇-𝟦 𝑎𝑛𝑑 𝟧𝟤% 𝑜𝑓 𝐿𝑙𝑎𝑚𝑎-𝟥 𝑟𝑒𝑠𝑝𝑜𝑛𝑠𝑒𝑠 𝑤𝑒𝑟𝑒 𝑐𝑜𝑛𝑠𝑖𝑠𝑡𝑒𝑛𝑡 𝑤𝑖𝑡ℎ 𝑑𝑒𝑣𝑖𝑐𝑒-𝑙𝑖𝑘𝑒 𝑑𝑒𝑐𝑖𝑠𝑖𝑜𝑛 𝑠𝑢𝑝𝑝𝑜𝑟𝑡 (𝐹𝑖𝑔. 𝟣). 𝐴𝑙𝑙 𝑚𝑜𝑑𝑒𝑙 𝑠𝑢𝑔𝑔𝑒𝑠𝑡𝑖𝑜𝑛𝑠 𝑤𝑒𝑟𝑒 𝑐𝑙𝑖𝑛𝑖𝑐𝑎𝑙𝑙𝑦 𝑎𝑝𝑝𝑟𝑜𝑝𝑟𝑖𝑎𝑡𝑒 𝑎𝑛𝑑 𝑐𝑜𝑛𝑠𝑖𝑠𝑡𝑒𝑛𝑡 𝑤𝑖𝑡ℎ 𝑠𝑡𝑎𝑛𝑑𝑎𝑟𝑑𝑠 𝑜𝑓 𝑐𝑎𝑟𝑒." Conclusions? 1) LLMs are capable of CDS! 2) If so, people will use them on a massive scale. 3) If so, it would make a lot of sense for regulators to provide regulatory oversight even though LLMs provide disclaimers that they are not intended to be used for medical advice. And these are just unimodal LLMs. Imagine the regulatory challenge that will come with multi-modality. Study: https://lnkd.in/eaCCbXed

66

12 Comments

GenAI and ML will certainly reshape drug discovery, but as the saying goes: garbage in, garbage out. Without high-quality chemical and biological data, even the best ML predictors and generative models will underperform — or worse, mislead. I really enjoyed Chris Lowden’s recent blog from Workflow Informatics Corp., which highlights the critical role of data engineering to enable ML and GenAI in drug discovery. Canonicalizing tautomers, normalizing assay units, unifying data schemas — these might sound mundane, but they’re essential. Every atom matters. Every protonation state matters. And of course, conformations matter too (but that is a discussion for another day). This blog aligns closely with themes we’ve learned at Psivant Therapeutics: * A robust data infrastructure is critical, especially if you are building a computational platform. * Once you have a robust database, building new workflows and apps is much more efficient. * Better molecular embeddings (e.g. using quantum mechanics) can improve ML predictions only if the database has the correct chemical structures and assay data. QM can't fix that. * Informatics and chemical rigor are foundational, because every atom matters. Prior to Psivant, I have seen firsthand how poorly managed data can break a model — and how tight, chemically accurate datasets can unlock meaningful insights. If your GenAI model is proposing molecules with impossible tautomers or misaligned assay data, the problem might not be the algorithm. Dig into your pipeline first. (Special call out to Alex Liholips and Shivam Patel for doing such an amazing job with PsiVision and the Psivant data infrastructure). No doubt, ML and GenAI are exciting. Let’s keep pushing the field forward, but not forget that excellence in data curation and infrastructure is what makes the promise of AI in drug discovery real. Being careful with data early will pay great dividends later. The BioChemUDM discussed in this blog looks like a great infrastructure to build around. While we have built our own infrastructure from scratch at Psivant, if I was starting fresh, I would certainly take a close look at BioChemUDM. If you are an ML expert or database guru but do not have deep cheminformatics experience in drug discovery, I highly recommend starting with something like BioChemUDM rather than relearning all of the nuances of digitizing chemistry and biology (and falling into many potholes along the rough and winding road of drug discovery). Read the full blog from Chris here: https://lnkd.in/eHFmbgiq #drugdiscovery #cheminformatics #genAI #tautomers #datascience #machinelearning #biotech #pharmaceuticals #computationalchemistry

246

11 Comments

Thrilled to share that our assessment of #mCODE’s suitability for oncology research is now published in JCO https://lnkd.in/eq-Ud3Ti. As a reminder, mCODE—the oncology-focused data standard built on #FHIR profiles—does an excellent job capturing core patient characteristics. But our analysis reveals important gaps in representing disease details, treatment patterns, and real-world outcomes, especially around treatment regimens and therapy response. If we want oncology evidence generation to be faster, more automated, and truly reusable, we need to close these gaps together. #codex Su (Sutin) Chen, Janet Espirito, Lisa Hannah-Deister, Gail Shenk, May Terry

98

1 Comment