GenBio AI

At the ADIA Lab Symposium in Abu Dhabi, our Co-Founder and CTO, Le Song, delivered a keynote on AIDO Cell, discussing how a multiscale simulator can track continuous biological state across an entire experimental workflow. Thank you to the ADIA Lab team for hosting a fantastic event. Learn more and join the AIDO Cell waitlist: https://lnkd.in/eG3CGKw7

Could a virtual cell eventually replace clinical trials? AlAin News put that question to CTO Le Song and he believes the proper framing is reducing the number of clinical trials. AIDO Cell is there to cut the number of experiments researchers need to run and to make sure the ones they do run are the right ones. A team can narrow a thousand possibilities in simulation, then take only the most promising few into the lab. He was also direct about validation. Checking a model against data that already exists shows it agrees with biology we have already observed. It does not yet show the model can stand in for an experiment no one has run, and the field has no standard for that today. For now we compare AIDO Cell's predictions with published research, and new predictions are being tested in the lab. Thank you to Mohamed Maher and Al Ain News for the conversation. The full interview, in Arabic: https://lnkd.in/eTYg587f

The ADIA Lab Symposium opens in Abu Dhabi on Monday, October 26. Le Song, our Co-Founder and CTO, will be delivering a keynote on day two. The Symposium’s theme is Design: Engineering Systems with Reliable Intelligence, which is a fair description of the problem Le and our team have spent the last two years on. His talk is AIDO Cell: A General-Purpose Simulator for Cell Biology, built on what we launched in August. He will be clear about what version 1.0 of the simulator is for, and will close on what would have to be true to treat a virtual cell as one composable unit inside something larger. Abu Dhabi is home to one of GenBio AI's three global offices, alongside Paris and our headquarters in Palo Alto.

When a drug fails in clinical trials, the reason is usually one of two things. It didn't work well enough, or it caused unexpected side effects. In a new Forbes piece by Ray Fernandez on how biotech and pharma are adopting AI, Le Song, our Co-Founder and CTO, traces both back to one place. "Both problems largely come from the same gap: no existing tool can predict how a molecule will behave inside the full complexity of a living cell," he said. AIDO Cell is our answer to that gap. It is the first world model of the human cell, a stateful, multi-scale simulation of how a cell responds to a disease signal or a therapeutic intervention. Researchers can apply a candidate molecule in simulation and read how the cell's state changes, which gives them an earlier look at whether it does what it should and what else it does. The article follows AI from discovery through clinical trials, manufacturing and distribution, with founders from across that chain. https://bit.ly/3VjUoGn

Le Song, our Co-Founder and CTO, is on the opening panel at Molecules, Models & Agents in San Francisco on Monday, October 5. The session, From Data to Discovery: AI Across Digital Health and Digital Biology, moderated by Vega Shah of NVIDIA NVentures. Le is joined by Michael Aratow, Co-Founder and Chief Medical Officer of Ellipsis Health and Deniz Kavi, Co-Founder and CEO of Tamarind Bio. Each of them has to answer the same question at a different scale: what counts as evidence that a prediction is right?

Co-Founders Eric Xing and Le Song have a Perspective out today in Cell Press, on what a virtual cell has to be before it can stand in for an experiment. They propose defining it as a world model: a system that holds a persistent cell state, treats a genetic, chemical or environmental intervention as an action on that state, and decodes the result into expression, structure, interaction networks and morphology that all describe the same cell at the same moment. Because the state carries forward between interventions, chaining several together generates coherent, multi-modal trajectories of cell states, rather than isolated predictions. The paper also puts a request to the field. Current perturbation leaderboards score one narrow slice of cellular state, so a high score says little about whether a simulation holds together across modalities or over a long sequence of interventions. Eric and Le call for a shared “virtual cell arena": standardized multi-modal tasks, long-horizon measurements, and goal-oriented experimental data where perturbations are designed to reach a stated phenotype. The Perspective is open access, read it here: https://lnkd.in/eE3Pv8Nd Want early access to the platform built on these ideas? Join the waitlist for AIDO Cell: https://lnkd.in/eG3CGKw7

The PAMI Mark Everingham Prize went to the FGVC Challenge Series at ECCV last week, and one of the recipients is our own Elijah Cole. FGVC is more than 15 years of benchmarks and competitions built by computer scientists and domain experts working side by side. The prize exists to recognize that kind of service to a research community. Elijah leads multimodal modeling of cells and tissue at GenBio AI. Congratulations to him and the team that shares this award.

"The goal of the virtual cell, at least the way we see it, is to reduce the number of experiments you’re doing." — GenBio AI Co-Founder and Chief Scientific Officer Ziv Bar-Joseph By capturing biology as a holistic system rather than isolated tasks, AIDO Cell allows researchers to run multi-step, sequential experiments in silico before ever stepping into the lab. Read the full piece by Bree Foster, PhD in Drug Discovery News to explore how world models are redefining cellular simulation. https://lnkd.in/ghdgn4Ea

Just one week ago we released a preview of AIDO Cell, a world model of the human cell. Here's some of the coverage on what we built and how it's been received. STAT covered the launch, including this framing from our Co-founder David Baker on the shift from modeling a single protein to modeling how an entire cell's machinery responds to a drug or a gene knockout: "With AlphaFold, you're predicting the structure of one house or learning about one person in that city, whereas in AIDO Cell, you're sort of predicting how that whole city works." https://lnkd.in/eWDgcz59 "Can AI put the cell back together?" That's the question Longevity.Technology asked in their piece on what makes AIDO Cell different. As our Co-founder, Chief Scientist and President Eric Xing puts it, “What world model architecture enables, and what makes our system genuinely new, is that it’s both multi-scale and stateful. The simulation remembers what you’ve done to the cell. You can run entire sequences of perturbations, where each one builds on the last, and see the evolving, cumulative effect in the cell, just as you would in a real (wet lab) workflow.” https://lnkd.in/eSUBQt8g "A meaningful foundation to build on, even at this early stage," is how David Baker described AIDO Cell's potential impact on drug discovery, toxicology and disease research, in FirstWord Pharma's coverage of the launch. https://lnkd.in/e47fd_aX AIDO Cell also appeared in TLDR, BigDATAwire, SynBioBeta, The Innermost Loop, and BAIO among other newsletters across the AI and biotech communities. To the researchers and journalists who dug into the work this week, we thank you and we’re excited to share more as it develops. Join the waitlist for early access: https://lnkd.in/eG3CGKw7

GenBio was founded with the audacious goal of simulating all of biology. The natural starting point for biological simulation is the cell: the smallest unit of life. Today we're sharing a preview of our first version of that simulator, AIDO Cell. It is based on a world model architecture, which enables some fundamentally different capabilities than we're used to: 1. Sequential Experimentation: Each perturbation becomes the starting point for the next 2. Coherent Multimodal Readouts: Many measurement types are decoded from one shared state 3. Multi-scale Simulation: Molecular interventions propagate through protein complexes, signaling networks, transcriptomics, protein abundance, and cell morphology, all in one go. Some other highlights from the technical report: 1. We benchmark on 5 task families covering 31 tasks across molecular perturbations, gene knockouts, structure prediction, molecular complexes, and functional genomics. AIDO Cell is the only method to cover all 5 families, and it matches or exceeds SOTA on 24/31 total tasks. 2. We perform case studies replicating known biology in prototype virtual cells for K-562 and Hep-G2, and show how AIDO Cell can be used to design drugs to achieve a specific cell state, and perform genetic rescues in-silico. 3. We explore cross-context transfer, where AIDO Cell could be scaled and used to simulate primary cells and disease cells that are not amenable to massive-scale data generation. This is a major milestone for GenBio and for the field. There are many problems to solve now that we know this is possible, and many opportunities that general-purpose biology simulators enable. This is still a prototype, and we can't finish this journey alone. We're actively looking for partners to work with on using this prototype version, and building out a full virtual cell bank for simulating all cell types. - Early access waitlist: https://lnkd.in/gP_TGMwK - Technical Blog: https://lnkd.in/gYjA8AV6 - Full Technical Report: https://lnkd.in/gffedNGa - STAT News Coverage: https://lnkd.in/gT6zG8mW - Nature Medicine perspective: https://lnkd.in/gcu5M_UU Eric Xing Le Song Emma Lundberg Ziv Bar-Joseph Fred Hu David Baker Eran Segal Debora Marks Hongyi Wang Elijah Cole Euxhen Hasanaj Christoph Feinauer Xingyi Cheng Hugo Ly Aaron Lifland, PhD Oz Huner Liz A. Bilige S. Janice Jia Zuo