James Zou

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

🚨 Milestone for AI research: DeepSeek-R1 featured on the cover of Nature On September 17, DeepSeek’s work on reinforcement learning for reasoning in LLMs became the first large language model ever to pass independent peer review and be published in Nature. Why this matters: 🔹 Scientific breakthrough – DeepSeek-R1 shows that reinforcement learning without human-annotated reasoning traces can trigger the emergence of advanced reasoning behaviors like reflection, verification, and strategy adaptation. 🔹 Performance – DeepSeek-R1-Zero surpassed human competitors in the AIME 2024 math Olympiad benchmark, and excelled in coding and STEM tasks. 🔹 Open science – Alongside the paper, the team released smaller distilled models, enabling the community to study and apply these reasoning methods. 🔹 Transparency & rigor – The Nature publication includes a 64-page peer review file and a comprehensive safety report. DeepSeek directly addressed concerns about data contamination and “distillation,” detailing decontamination steps and clarifying that reasoning abilities were learned independently, not borrowed from other models. 🔹 Broader impact – With >10 million downloads on Hugging Face, DeepSeek-R1 has become the most widely adopted open-source reasoning model. Its combination of open release + peer review sets a new bar for transparency and reproducibility in AI research. As Nature noted in its editorial: in an industry often dominated by unverified claims, DeepSeek has taken a remarkable step toward transparency and credibility. 👉 Full paper (open access): https://lnkd.in/eVSH6pT5 👉 Peer review file: https://lnkd.in/eTV2nhPF

21

Yubei Chen, fresh from a CDS postdoc under CDS founding director Yann LeCun and Meta FAIR, brought insights from his computational‑neuroscience and self‑supervised learning research at UC Berkeley into founding Aizip. Chen drew on techniques originally used for visualizing word embeddings and concept neurons to shape Aizip’s focus on tiny, interpretable models designed for edge‑device deployment. Aizip’s AI nanofactory and Gizmo model line (300M–2B parameters) enable on‑device tasks like face recognition, ECG/EEG analysis, keyword spotting, and chatbots—all without the cloud, reducing power and latency. Chen noted that most AI innovation was pushing model scale, but real‑world use cases often demand opposite constraints: “high efficiency, low power consumption, and minimal latency.” By optimizing algorithms for resource‑constrained hardware, Aizip is enabling scientific rigor to meet deployment needs. Read the full interview: https://lnkd.in/eBrtb9cS #EdgeAI #TinyML #OnDeviceAI #AIInterpretability

65

The study introduces GPN-Star, a novel genomic language model (gLM) designed to learn functional constraints in DNA sequences across evolutionary timescales by explicitly incorporating phylogenetic information and whole-genome alignments into its architecture. Traditional gLMs adapted from natural language processing often demand enormous model size and computation yet underperform compared to classical evolutionary models, especially in complex genomes such as humans. GPN-Star overcomes these limitations by leveraging species relationships through phylogeny-aware mechanisms and training on whole-genome alignments spanning vertebrate, mammalian, and primate divergences. When benchmarked on a wide range of variant effect prediction tasks, GPN-Star achieves state-of-the-art performance in both coding and non-coding regions, outperforming previous models in prioritizing pathogenic variants, enriching complex trait heritability, and increasing power in rare variant association tests. The framework also generalizes to other organisms — including mouse, chicken, fruit fly, nematode, and Arabidopsis — demonstrating robustness and broad applicability. Overall, GPN-Star presents a scalable, flexible tool for genome interpretation that efficiently integrates evolutionary signals and modern deep learning to improve the functional annotation of genetic variation. Read: https://lnkd.in/g2Driw6c

1

🔬 New Open-Access Article in Computational Biomedicine ❗ Drug‑target affinity prediction based on multi-source information and graph convolutional network Lei X, Tang X, Zhang Y. Computational Biomedicine 2026;1:202520 🔗 Read here: https://lnkd.in/geup-vpd 💡 This study introduces a novel deep learning framework for drug–target affinity (DTA) prediction, integrating graph convolutional networks with multi-source molecular and protein features. Key innovations include: ✨ Multi-source feature integration – combining molecular graphs, protein structures, and biological descriptors for richer representation. ✨ Graph neural network-based modeling – using GCN and graph isomorphism networks to capture local and global interactions. ✨ Enhanced predictive performance – significant improvements on benchmark datasets (Davis & KIBA), demonstrating superior accuracy and reliability. Corresponding author: Prof. Xiujuan Lei, Shaanxi Normal University 💡 If your research focuses on AI-driven drug discovery, computational pharmacology, or graph-based bioinformatics, we invite you to check out this work, share it with your network, and cite it if relevant! 📩 CBM also welcomes submissions of high-quality studies in these areas – join us in advancing computational biomedicine! #CBM #DrugTargetAffinity #GraphNeuralNetworks #ComputationalDrugDiscovery #AIinBiomedicine #Bioinformatics

12

2 Comments

𝗘𝘃𝗼𝗹𝘂𝘁𝗶𝗼𝗻 𝗶𝘀 𝗱𝗲𝗳𝗶𝗻𝗶𝘁𝗲𝗹𝘆 𝗯𝗮𝗰𝗸, 𝗣𝗮𝗿𝘁 𝗜𝗜 At a recent "Workshop on AI in the Cloud" at UC Berkeley organized by Industry-Academia Partnership (IAP, see https://lnkd.in/gJAqvdyA), Ion Stoica (UC Berkeley Professor and a co-founder of Databricks, Anyscale, LMArena) gave a fantastic, and fantastically clear, overview of the *few* things he has done. Some examples include "Ray, a distributed framework for scaling AI workloads, vLLM and SGLang, two high-throughput inference engines for LLMs, and LMArena, a platform for accurate LLM benchmarking." But what got me is his answer to someone asking for what to focus on in research. His answer: reinforcement learning and 𝗲𝘃𝗼𝗹𝘂𝘁𝗶𝗼𝗻𝗮𝗿𝘆 𝗮𝗹𝗴𝗼𝗿𝗶𝘁𝗵𝗺𝘀! When I asked him about evolution after his talk, he said: (1) "we love them, they work really well" and (2) his team is a big user of OpenEvolve, an open source implementation of Google DeepMind's AlphaEvolve. From Asankhaya Sharma's blog on Hugging Face (https://lnkd.in/gN_aR-eF): "AlphaEvolve represents a significant advancement in this field by:  1. Using LLMs to generate sophisticated code modifications  2. Evaluating these modifications with automated metrics  3. Using an evolutionary framework to improve promising solutions  4. Evolving entire codebases, not just individual functions OpenEvolve implements these principles in a flexible, configurable open-source package" Evolutionary algorithms are having a rebirth and can be combined with LLMs and other forms of 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝘃𝗲 models to produce powerful novel concepts, hypotheses or algorithms. Until now, they have not really been deployed at the kind of scale that LLMs and other neural networks have been. I can't wait to see what happens when we get to that scale. Google Deepmind and Sakana AI have been working on exciting applications, a breath a fresh air in a rather stuffy technical monoculture. If Ion Stoica is enthusiastic about them, it is a sign! Joshua Knowles, Danny Hillis, Llion Jones, Thomas Wolf

11

1 Comment

🚀 Large language models (LLMs) are rapidly reshaping the landscape of clinical and translational science. The latest Translational Bytes post by James Lu for #CTSjournal explores how these tools can be leveraged for research, communication, and discovery—while also raising important questions about responsible use, transparency, and equity. 📖 Read here:  https://bit.ly/3IawfvA Whether you're curious about integrating AI into your workflows or considering the broader implications for the field, this piece offers a thoughtful primer on both the opportunities and challenges ahead. What role do you see LLMs playing in advancing translational science? #AI #LLM #TranslationalScience #ClinicalResearch #Innovation #ASCPTjournals

13

Excited to share our study in Neuron led by Travis Goode, a K99 postdoctoral fellow (on job mkt), other members of Sahay lab & Collaborators @ UW Seattle /Hopkins/UCSF, defining a neural circuit that links experience with feeding. Open Access: Download from https://lnkd.in/dh5wQZv #eatingdisorders #obesity Significance: Remember that great experience you had at that restaurant and how you want to go back. Food consumption or motivation to eat is dependent on internal states (ghrelin, incretins) and external cues or “contexts”. Our environment or context exerts a powerful influence on where we eat and don’t choose to eat. The brain circuits and cell-types underlying how our prior experiences reinforce or calibrate eating at specific locations are poorly understood.  Defining such mechanisms may shed light on therapeutics to treat disordered eating in humans such as binge eating that may arise from loss of contextual control or calibration of eating. The hippocampus, a seahorse shaped memory center in the brain, is crucial for forming memories of our experiences. Growing recognition from human imaging studies that hippocampal activity and hippocampal connectivity with feeding centers of the brain like the hypothalamus are altered in individuals who are binge eaters and obese. We show that the hippocampus recruits Prodynorphin producing neurons in the dorsolateral septum (DLS Pdyn) to instruct an evolutionarily conserved feeding circuit module in the lateral hypothalamus to confer experience-dependent calibration of feeding. What we found: Using snRNAseq (29K cells) and in situs we mapped neuropeptidergic cells in the dorsolateral septum and identified a sparse population of DLS Pdyn neurons that is conserved from mice to humans. Using activity-dependent imaging we found that DLS Pdyn neurons exhibit context-dependent activity, with larger changes in activity happening when mice were in a food-reinforced context. Using optogenetic methods to silence DLS Pdyn neurons or their inputs from the hippocampus or prodynorphinergic inputs to the lateral hypothalamus, we found that context-conditioned feeding was impaired. i.e mice ate comparable amounts of food in reinforced context and a novel context. We observed similar outcomes when Pdyn was genetically deleted from these neurons, suggesting that kappa opioid receptors may have a role in contextual regulation of feeding. Implications for PharmacoTx for weight loss and eating disorders: Dysfunction of PDYN production or impairments in physiology or circuitry of these cells may contribute to disordered eating. Our instantiation of a circuit for contextual control of eating refines our circuit-based understanding of weight loss- and eating-targeted pharmacotherapies. Since GLP1R is expressed in DLS Pdyn neurons and stimulation of these cells suppresses feeding and triggers aversion behavior, it maybe that GLP1R agonists exert their effects on food intake, in part, through dysphoria.

194

3 Comments