Alex Rives

don't miss out this unique opportunity to join forces on benchmarking time series foundation models on different downstream validation tasks: classification, prediction, anomaly detection. A community effort led by Thomas Moreau from the #MIND team (Centre Inria de Saclay, Inria & CEA Joliot). Still a pending question: Do these model perform well for online real-time forecasting such as in BCI applications, like our play field in BrainSync. Alexandre Gramfort Pedro L. C. Rodrigues Théo Gnassounou Mathieu Dagréou Lucas Struber Jean Faber Joseph Paillard Merlin Dumeur Mathurin Massias

26

Boltzdesign1 is out! Invert Boltz/AlphaFold to design binders against proteins, small molecules, RNA, DNA, and metals! Congratulations to Yehlin Cho, Martin Pacesa, Zhidian Zhang, Bruno E. Correia, and Sergey Ovchinnikov for the release “BoltzDesign1: Inverting All-Atom Structure Prediction Model for Generalized Biomolecular Binder Design” We've seen great sucess from methods like ColabDesign and its descendants, hopefully Boltzdesign1 will do same in designing enzymes, and better protein binders against diverse targets, including non-proteins. Excited to see what the community builds on top. Paper: https://lnkd.in/gn8gucaX Try it out on Tamarind Bio: https://lnkd.in/g9yc83nU Learn more in our longer write-up: https://lnkd.in/gfYwbR7S Please note that the authors warn that they've not validated any of the binder designs experimentally, use with caution. Abstract Deep learning in structure prediction has revolutionized protein research, enabling large-scale screening, novel hypothesis generation, and accelerated experimental design across biological domains. Recent advances, including RoseTTAFold-AA and AlphaFold3, have extended structure prediction models to work with small molecules, nucleic acids, ions, and covalent modifications. We present BoltzDesign1, which inverts the Boltz-1 model, an open source reproduction of AlphaFold3, to enable the design of protein binders for diverse molecular targets without requiring model finetuning. By utilizing only the Pairformer and Confidence modules, our method significantly reduces computational costs while achieving outstanding in silico success rates and diversity in binder generation. Optimizing directly on the distogram allows us to shape the probability distribution of atomic distances, rather than adjusting a single structure, steering the design toward sequences that yield robust structures with well-defined energy minima. By leveraging a fully atomic model trained on a wide variety of macromolecules, we can generate diverse heterocomplexes with flexible ligand conformations—a capability not currently matched by existing methods. This approach enables the design of novel protein interactions with potential applications in biosensors, enzyme engineering, therapeutic development, and biotechnological innovations.

431

15 Comments

Interactive feedback changes how code LLMs perform and rank, a new study finds. Led by senior author Valerie Chen, a Carnegie Mellon PhD student and recent visiting researcher at CDS, the project also included Courant PhD student Jane Pan, CDS PhD student Jacob Pfau, and CDS Assistant Professor He He. The team simulated developer interactions to replace static coding benchmarks like HumanEval and MBPP. When models received human-style feedback—code edits, single-sentence prompts, or paragraph-length instructions—they sometimes improved by up to 9.3 percentage points in accuracy and moved as many as four spots in ranking. “Testing on new problems the models had never seen made feedback even more powerful,” said Chen. Read more: https://lnkd.in/eAPTP5NR

5

We report new work from our group in collaboration with Google's genomics team. A challenge with genetic discovery for heart failure is that its classification within large scale biobanks is extremely vague. This has led to the surprising situation where, in 2026, there are only 2 validated GWAS loci for heart failure with preserved ejection fraction, a sub-group that accounts for HALF of the total heart failure patients, heart failure being the single most common admitting diagnosis in patients over the age of 65 (more than 30 million people globally). We need to do better. One approach is to use machine learning to develop a precision phenotype (for this you need rich health care system data from many patients, not biobank data). Then, you can take that probabilistic framework and apply it to biobanks which, while less enriched for patients, provide the power of large scale and deep genomics and proteomics. This was the idea that Jack W O'Sullivan MD, PhD and a large collaborating team sought to pursue in this work. The findings went beyond all our expectations. First, the team increased the number of loci confidently associated with HFpEF almost 50 fold! We report 99 loci, up from 2. Second, by integrating multi-omics in a causal inference framework, we prioritize gene-protein pathways as candidates for therapeutic development while de-prioritizing non-causal targets, reclassifying them as biomarkers. How could we tell we were on track? We noted first that three clinical trials for one of our deprioritized targets (MPO) were negative (genomics really can save money in drug discovery). We also validated one of our prioritized targets in a novel clinical study. In conclusion, we hope this work accelerates our understanding of heart failure. We believe this approach to precision phenotyping can extend the reach of our global biobanks in many directions. Finally, we believe incorporating causal inference in a multiomic framework can directly inform investment decisions in drug discovery pipelines. Congratulations to the entire team and special thanks to our Google collaborators who are the co-leaders of this work. Taedong (Ted) Yun Cory McLean Andrew Carroll Paper: https://lnkd.in/gamstJQM

532

17 Comments

A team of researchers led by You Wu and Lei Xie of Northeastern University, with co-author Dean Phil Bourne, has proposed a bold new model for drug discovery. Their study, “AI-powered Programmable Virtual Humans Toward Human Physiologically-Based Drug Discovery,” published in Drug Discovery Today, introduces a new concept: programmable virtual humans, detailed computational models that allow scientists to simulate how new drug compounds behave inside the body before human trials begin. Read more about this innovative approach to reimagining the future of drug discovery here: https://bit.ly/4oSbJ2q

72

2 Comments

🎉 Preprint alert! Our stellar Ph.D. students Chujun He and Jiaqi Zhang, under the supervision of Eric and Wendy Schmidt Center Director Caroline Uhler and collaborator Munther Dahleh, have developed MORPH, a MOdular framework for predicting Responses to Perturbational cHanges. MORPH takes #SingleCell data from control cells, along with prior knowledge of gene perturbations to predict the distribution of cellular responses to unseen #perturbations in #PerturbSeq or #OpticalPooledScreen. We demonstrated that MORPH outperforms the state-of-the-art models in the following tasks: 1) Predicting transcriptional responses to genetic perturbations that were not present during training at the single-cell level. 2) Transferring the effects of perturbations to untested cell lines. 3) Predicting the outcomes of unseen combinatorial gene perturbations and identifying the types of interactions occurring between the perturbed genes. 4) Guiding the optimal design of the most informative perturbations. MORPH is interpretable – beyond predicting the perturbation effects, MORPH can be used to obtain valuable insights into the underlying gene regulatory programs. To the best of our knowledge, MORPH is the only method that has been demonstrated on genetic perturbation data measured in various modalities, including not only transcriptomics but also imaging. This project is in line with our ongoing effort to establish the state-of-the-art in perturbation prediction, the Cell Perturbation Prediction Challenge (CPPC). 🔗 Find the preprint: https://lnkd.in/eqAJytAB 💻 View the Code: https://lnkd.in/eVeNVNCW 📖 Read more about the Cell Perturbation Prediction Challenge (CPPC): https://lnkd.in/e2aRWi9B and our first challenge: https://lnkd.in/evbCP4fv Broad Institute of MIT and Harvard, MIT Institute for Data, Systems, and Society (IDSS) #SchmidtCenter #MIT #LIDS #IDSS #MORPH #Genomics #MachineLearning #AIinBiology #ComputationalBiology #ScienceNews

29

🧬 Last Week in AI × Life Science Our newsletter dove into three papers showing how AI is transforming biology: University of California, Berkeley and Caltech's NucleusDiff: Reduces atomic collisions in drug design by 100% by incorporating van der Waals constraints, physics-informed AI generating better molecules. CeMM's CellWhisperer: Ask your single-cell data questions in plain English. Trained on 1M+ transcriptomes, it lets you query datasets with natural language and get biologically meaningful answers. Genentech's Nona: One unified framework for genomic prediction, interpretation, AND design through clever masking. Also revealed an unexpected privacy risk: ATAC-seq data can identify individuals with perfect accuracy. 💡The trend? Multimodal approaches that integrate different data types are unlocking new capabilities across computational biology. 🔗Full breakdown in our weekly newsletter: https://lnkd.in/ej5FHwqN #AIinBiology #MachineLearning #ComputationalBiology

12

Join us this Friday in-person in San Francisco (Jackson Square) for "Agents in Academia". We have two talks from PhD researchers at Berkeley (Sky Lab) and Stanford on agent memory / continual learning and local inference. 1. Learning to Use External Memory for Long-Context LLMs (Dacheng Li, UC Berkeley Sky Lab) Exploring external-memory–augmented LLMs trained with reinforcement learning (RL) - how can we post-train models to enable learned use of the external memory systems? 2. Intelligence per Watt: Measuring Intelligence Efficiency of Local AI (Jon Saad-Falcon, Stanford NLP / Hazy Research) Now that small LMs are capable of achieving comparable performance to closed large LMs... can local inference viably redistribute demand from centralized infrastructure? RSVP on Luma, space is limited: https://lnkd.in/gF2rymBe

37

1 Comment