Arc Institute

In this featured lecture recorded at the Arc Symposium (February 4, 2025), Evo 2 co-author Patrick Hsu reveals how Arc researchers and collaborators combined machine learning with molecular biology to create a multimodel, multiscale biological foundation model, to decode the language of life. Watch here: https://lnkd.in/ewFtyC4r

New preprint: The first whole-body map of both DNA methylation and 3D genome organization at single-cell resolution across 16 human tissues, with in-depth analyses of cell type diversity of these epigenome modalities at unprecedented resolutions. The work was led by Arc Science Fellow Jingtian Zhou, along with colleagues from Salk Institute for Biological Studies and is now on Bioxriv: https://lnkd.in/ehzfTzKz The simultaneous profiling of these two modalities allowed the researchers to identify 35 major cell types and 206 cell subtypes. While DNA methylation and 3D genome organization are two epigenetic mechanisms known to relate to gene expression and cell fate, the team found extensive discrepancies can exist between them. This potentially reflects different cell types may rely on different epigenetic mechanisms to maintain their identity and function. 

Our work using AI agents to design genetic perturbation experiments will be presented at ICLR 2025! 🇸🇬 We explored the use of 𝗟𝗟𝗠𝘀 𝗳𝗼𝗿 𝗮𝘂𝘁𝗼𝗻𝗼𝗺𝗼𝘂𝘀 𝗱𝗲𝘀𝗶𝗴𝗻 𝗼𝗳 𝗯𝗶𝗼𝗹𝗼𝗴𝗶𝗰𝗮𝗹 𝗲𝘅𝗽𝗲𝗿𝗶𝗺𝗲𝗻𝘁𝘀, achieving improved accuracy, interpretability and robustness over existing approaches. BioDiscoveryAgent designs experiments using only an LLM + tools. In each round, the agent constructs a prompt that combines task description + experimental results. The LLM response identifies genes to perturb in the next round and provides reasoning and long-term strategy 𝗥𝗲𝘀𝘂𝗹𝘁𝘀: BioDiscoveryAgent detects 21% more hits in 1-gene perturbation screens, 46% more non-essential hit genes. This is *without* being specifically trained for the task, unlike Bayesian optimization baselines. It also shows 2x improvement over a random baseline in 2-gene perturbations. The agent uses both prior knowledge + experimental results for decision-making. Early rounds rely more on prior knowledge while later rounds focus on experimental results. Agent with access to both consistently outperforms. 𝗦𝗶𝗴𝗻𝗶𝗳𝗶𝗰𝗮𝗻𝗰𝗲: With the growing interest in active experimental design for biological discovery, our work highlights how LLMs could become core components of these systems in the future. Stay tuned for our upcoming work on how strategic tool design and selection can further optimize performance for large LLMs! Collaboration with Andrew Lee, Qian Huang, Jian Vora, Alex Marson, Percy Liang, Jure Leskovec Paper: https://t.co/rl08ZIZHnI Code: https://t.co/C3ixUxChDL

Innovation Investigator Awards: Applications Due March 15! Faculty at Stanford University, University of California, Berkeley, and University of California, San Francisco: We're seeking visionary researchers in immunology, neurobiology, and machine learning to join our community of scientific innovators. Awardees maintain their labs and employment at their university while benefitting from access to Arc resources and facilities and collaboration with Arc scientists. As an Innovation Investigator, you'll receive: 💸 $1 million in unrestricted funding over 5 years 🔓 Freedom to pursue your most ambitious research questions 🤝 Collaboration with technology centers and labs at Arc Applications due March 15. Join a network where curiosity-driven research thrives. Apply now: https://lnkd.in/gjUcRG75 #ScienceFunding #Research #Innovation #Immunology #Neurobiology #MachineLearning

Same spot, two years apart. 🧬 Exciting to see how far we've come!

Today, we're launching the Arc Virtual Cell Atlas, a growing resource for computation-ready single-cell measurements. As the initial contributions, Vevo Therapeutics has open sourced Tahoe-100M, the world's largest single-cell dataset, mapping 60,000 drug-cell interactions, and we’re announcing scBaseCamp, the first RNA sequencing data repository curated using AI agents. Tahoe-100M includes 100 million cells and measures cellular response across 50 cancer cell lines to 1,200 drug perturbations. It was generated using Vevo’s Mosaic Technology, the first platform to make pan-cancer testing of drugs at single cell resolution scalable, and with support from Parse Biosciences's GigaLab leveraging its single-cell RNA sequencing capabilities. For scBaseCamp, Arc scientists trained AI agents to mine observational data from more than 200 million cells representing 21 different species sourced from public repositories, and processed them to a standardized form. This enables better integration between different types of single-cell data tailored for virtual cell model pretraining. Combined, the release includes data from over 300 million cells. Learn about the Vevo-Arc partnership: https://lnkd.in/gJS9QjYi  Read about Tahoe-100M: https://lnkd.in/gUVdCaQr Read about scBaseCamp: https://lnkd.in/g9_6jvxZ Read about the Arc Virtual Cell Atlas: https://lnkd.in/gp7-XX-Z Visit the Arc Virtual Cell Atlas: https://lnkd.in/g4mC2nBr

We're excited to announce Arc Institute's search for a Chief Scientific Officer to help advance our mission of accelerating scientific progress on complex human diseases. Arc combines cutting-edge experimental biology with computational methods—from our Evo 2 biological foundation model to unravelling cellular dynamics in Alzheimer's disease. We're looking for a visionary individual to support our experimental technical staff who will partner closely with our CTO and other scientific leaders at Arc to drive our flagship, institute-wide projects forward. Are you energized by ambitious, collaborative research that no single lab could tackle alone? We'd love to connect. Visit https://lnkd.in/eEFq2ZiZ to learn more.

Introducing Evo 2, the largest AI model for biology to date, enabling machines to read, write, and think in the language of life itself. Trained on 9.3 trillion nucleotides from over 128 thousand archaeal, prokaryotic, and eukaryotic genomes, Evo 2’s deep understanding of biological code means that it can identify patterns in gene sequences across disparate organisms that experimental researchers would need years to uncover. The model can accurately identify disease-causing mutations in human genes and is capable of designing new genomes that are as long as the genomes of simple bacteria. A collaboration between Arc Institute, NVIDIA Healthcare, Stanford University, University of California, Berkeley, and University of California, San Francisco, Evo 2 is fully open source - including training data, code, and model weights. - Read the preprint: https://lnkd.in/eu2bxrpU - See the announcement: https://lnkd.in/esEdyF5G - Try the user-friendly interface Evo Designer to generate DNA by sequence, species, and more: https://lnkd.in/eV6PiyQ2 - Access the code from Arc’s GitHub: https://lnkd.in/e6SErXQ5 - Explore via the NVIDIA BioNeMo framework: https://lnkd.in/efWKEKNN - Use the mechanistic interpretability visualizer to uncover the key biological features and patterns the model learns to recognize in genomic sequences, developed in partnership with Goodfire: https://lnkd.in/eqDZTggf

Congrats to Isha Jain, Arc Institute Core Investigator and Gladstone Institutes / University of California, San Francisco, Associate Professor and Investigator, for her lab's development of "oxygen in a pill" therapies, reported in Cell by Cell Press. The Jain Lab is using this not just for mitochondrial disorders but a surprising set of complex human diseases (stay tuned!). The Jain Lab is also hiring if you're excited about oxygen or redox cofactor (vitamin) metabolism: https://lnkd.in/exdqPftc

Congratulations to Maayan Levy, Arc's newest Innovation Investigator in Residence, for receiving The Vilcek Foundation 2025 Prize for Creative Promise in Biomedical Science. We're so happy you joined us! The Levy Lab seeks to understand, develop, and apply the concept of metabotherapy—the usage of metabolites as vehicles and targets to prevent and treat disease. https://lnkd.in/et-gY9Vk