Jean-Philippe Vert

Excited to share that the Broad Institute of MIT and Harvard Machine Learning in Drug Discovery Symposium is coming up. Join for a gathering of leaders pushing the boundaries at the intersection of AI in drug discovery. The full agenda is now live: https://lnkd.in/eKZmPwnZ I am happy to be presenting work from Psivant Therapeutics that leverages AI/ML, physics, and wet lab experiments to drug challenging I&I targets. While the AI/ML field continues to gain momentum, we are beginning to learn more about the true value and limitations of AI/ML. In short, AI/ML approaches are great for interpolation but quite poor for extrapolation. When you have the appropriate training data, use AI/ML and you will generally get good predictions, fast. However, much of drug discovery involves finding novel drugs for undrugged targets (i.e. not great data). For those scenarios, proceed with caution when using AI/ML methods. It is easy to fool yourself into thinking your method can extrapolate when in reality you have data in your training set that looks a lot like your test set (which does not look like the novel drugs you want to design). For example, see recent publications on protein co-folding from Eva Nittinger et al. at AZ https://lnkd.in/eWmw8-w3, Pat Walters https://lnkd.in/ekEjhhnA and the SIB Swiss Institute of Bioinformatics https://lnkd.in/eG5puMUd At the same time, we are also familiar with the limitations of physics-based approaches (complex setup; computationally intensive; sampling local/global minima; force fields; tautomers; stereochemistry; waters; expert knowledge; and more). Yet physics-based methods can truly extrapolate and in fact do not need any training data when embarking on a new target or class of ligands. Our philosophy: * Build a computational ecosystem that fully leverages state-of-the-art methods in both AI/ML and physics. * Run at scale. * Automate as appropriate. * Keep humans in the loop. * Design an intuitive data-centric user interface for experienced drug hunters. * Hire expert developers who understand the strengths and limits along with a firm grasp of drug discovery such that the focus is on the highest value problems within the project. This might sound obvious, but for all the people at the interface of in silico and wet lab approaches, you know how hard it can be. The magic happens when you combine the speed of AI/ML, the extrapolative power of physics, the right experiments, and the judgment of experienced drug hunters. Registration is open, but this one tends to sell out, so sign up now: https://lnkd.in/epuJfyNT Looking forward to joining great speakers, such as Thomas Miller, Marinka Zitnik, and more.

303

6 commentaires

Massive foundational datasets – often built without specific hypotheses in mind, only to provide fodder for further research – form the backbone of today's AI revolution in biomedicine. Case in point: When Google DeepMind revealed their AlphaGenome model, they cited two NIH-funded, Broad-led resources, the ENCODE and GTEx catalogs, as two of their main sources of training data for AlphaGenome. In a Broad Q&A, we asked Kristin Ardlie and Bradley Bernstein, who led these resources' development, about how datasets like these fit in the AI age, AI's role in revealing how the genome works, and how to make AI as useful as possible for genomic discovery. 🔗: https://lnkd.in/eEa6Kkns #BroadInstitute #Science #AI #ArtificialIntelligence

192

🚀 Introducing Boltz-ABFE – accelerating accurate predictions without crystal structures in AI drug discovery. 🔹 A new paper from the research team at Recursion and Valence Labs demonstrates how we’ve combined the open-source model we developed with Massachusetts Institute of Technology, Boltz-2, with our own absolute binding free energy protocol to create Boltz-ABFE – a pipeline that makes gold-standard protein-ligand binding affinity predictions possible without the need for experimental crystal structures. 🔹 Key takeaways: ▪️ No Crystal Structure Needed: Traditionally, the most accurate physics-based methods for predicting drug binding affinity (FEP) required a high-quality 3D crystal structure of the protein-drug complex, which is often unavailable in the early stages of discovery. ▪️ Accurate AI-Predicted Structures: We demonstrate that AI models like Boltz-2 can generate protein-ligand structures that are accurate enough to be used as a starting point for these high-precision simulations. ▪️ A Robust New Pipeline: We developed Boltz-ABFE, a method that combines AI structure prediction with an automated protocol to correct common errors and prepare the structures for absolute binding free energy (ABFE) calculations. ▪️ Accelerating Discovery: By enabling accurate, structure-based affinity estimation much earlier in the process, this work paves the way for faster and more efficient identification and optimization of new drug candidates. 💥 This paper represents another important step for early-stage drug discovery. Congrats to the team! Stephan Thaler, William (Zhiyi) Wu, William G., Richard T. Bradshaw, Gail Bartlett, Prudencio Tossou, and Geoffrey Wood. *Link to the paper in comments! #DrugDiscovery #AI #FEP #TechBio 

721

10 commentaires

Have you heard of openproblems.bio? 🧬 It's kind of like Kaggle for single-cell biology. And it's a brilliant model of how the ML and bioinformatics communities collaborate. Single-cell analysis sits at a fascinating intersection: massive, tabular datasets meet cutting-edge ML meets the messy reality of biology. But here's the challenge: ML researchers love Jupyter notebooks, while biology demands robust, reproducible pipelines for processing data. Different tools, different cultures. OpenProblems.bio bridges this gap brilliantly. It transforms core single-cell challenges into living, community-run benchmarks where anyone can: ✅ Submit methods in Python or R (no pipeline expertise needed) ✅ Compare results fairly with standardized datasets and metrics ✅ Access everything openly, including published results and methods The secret sauce? A powerful three-part ecosystem: * Viash converts scripts into reproducible pipeline components * Nextflow orchestrates complex workflows * Seqera provides the platform for elastic cloud execution on AWS What excites me most: this is **FAIR open science done right** Complete transparency, version control, quality checks built in, and a framework that welcomes contributions from both communities without forcing either to abandon their tools. Tasks range from dimensionality reduction to perturbation prediction, with new benchmarks continually added by the community. If you're working in computational biology or ML for life sciences, this is infrastructure worth knowing about. The future of biomedical AI needs exactly this kind of collaborative, rigorous foundation. Thanks to Robrecht Cannoodt and the openproblems.bio community for collaborating on the blog: https://lnkd.in/gpNJNd2r #Bioinformatics #MachineLearning #SingleCell #OpenScience #Benchmarking

173

Today, the Diagnostics Accelerator (DxA) announced two major milestones: the release of SpeechDx’s first tranche of data and a new partnership with Callyope, expanding the ecosystem advancing speech-based diagnostics. Callyope is the second partner to license the dataset. By leveraging #SpeechDx data to train and validate its AI speech-foundation model, Callyope can accelerate the development of tools for mental health and neurodegenerative diseases, including #Alzheimer’s and #Parkinson’s. Through harnessing the power of #AI, SpeechDx is creating reliable, affordable diagnostic tools, a foundational goal of DxA. This gold standard dataset will power a new generation of AI tools to forecast Alzheimer’s risk, track progression, and guide interventions before symptoms appear – paving the way for precision medicine. https://lnkd.in/ebRdX-xz

64

6 commentaires

BPGbio is building the most advanced biology-first AI platform in the industry by connecting tissue architecture with molecular readouts. Our NAi® Interrogative Biology® platform doesn’t just predict — it explains and de-risks drug development. This layered workflow shows how we integrate biology-first evidence into AI-driven insights — creating decision-grade data that informs drug distribution, target engagement, and biomarker discovery. Here’s what each layer means: 🟣 Histology – The starting point: visualizing tissue architecture to understand cellular context. 🟠 Spatial Metabolomics – Mapping how metabolites are distributed across tissue, revealing the biochemical state of disease and therapy response. 💛 Immunoproteomics – Pinpointing proteins and immune markers within their spatial environment to track therapeutic activity and immune response. 💙 Ion Segmentation – Breaking complex tissue signals into meaningful clusters, separating healthy, diseased, and drug-affected regions. ⚫ Multivariate Data Analysis – Integrating multiple data types to uncover causal patterns, not just correlations. 🌐 Immune Metabolic Integration – Linking immune activity with metabolic states to capture the full biological picture of how therapies act in the body. At BPGbio, we believe the future of medicine is driven by biology-first AI, powered by real patient data and multi-omics integration. #BiologyFirstAI #SpatialOmics #MALDI #timsTOF #DrugDiscovery #PrecisionMedicine #MitochondrialDisease #AIDrug

41

💊 AI is transforming #drugdiscovery — but where are the actual drugs? Startups, such as Google DeepMind, Insilico Medicine, and Recursion, are racing to bring the first AI-designed medicines to market. Some have already entered clinical trials with promising results—but the journey from molecule to medicine is still long and complex.

442

20 commentaires

Alert: Zifo Launches AI Antibody Engineering App! Zifo's new AI Antibody Engineering application is now available on the Snowflake Marketplace. This no-code GenAI platform empowers scientists to fine-tune protein models and handle the entire discovery pipeline, from sequence generation to 3D structure prediction. Read the Full story here: https://lnkd.in/gQrd9iSN The development of new antibody-based therapies is traditionally a complex, time-consuming, and expensive process. Zifo's Antibody Engineering app directly confronts this challenge by integrating state-of-the-art AI into a seamless, end-to-end workflow that promises to improve the hit rate for high-affinity therapeutic antibodies. "We are entering a new era of drug discovery, one where AI and data converge to solve biology's most complex puzzles," said Chris McClure, Head of Solution Services for Zifo in North America. "Zifo Antibody Engineering is more than just a tool; it's a strategic partner for scientists. By building it as a Snowflake Native App, we're giving researchers the power to innovate faster, reduce R&D costs, and ultimately bring life-saving therapies to patients sooner. We've effectively democratized access to complex AI, removing the need for deep machine learning expertise." #LifeSciences #Innovation #AIinHealthcare #AntibodyEngineering

182