Biohub: The Future of Biology is Open-Source with Mark Zuckerberg, Priscilla Chan, and Alex Rives
In a Nutshell
Biohub's core mission is to accelerate biology by building open-source AI models and data-generation tools that empower the entire scientific community, rather than pursuing disease-specific cures directly. The strategy centers on creating hierarchical world models—from proteins to cells to systems—by integrating frontier AI with novel experimental methods that produce the large-scale data these models require. Key outcomes include releasing general protein models like ESM Fold that enable computational design of therapeutics and shifting from discovery-based to engineering-based biology through open, decentralized collaboration.
These notes were generated by AI and may contain inaccuracies.
The goal is to provide tools to the entire scientific community. The focus is on understanding how biology works at the individual level, including understanding a person's genetics and their risks for different illnesses. The objective is to treat individuals as individuals, understand mechanisms, and intervene effectively. The approach emphasizes open-source projects to get tools into more scientists' hands faster and accelerate progress across the entire scientific field rather than attempting to cure diseases directly.
The theory is not that Biohub will cure diseases. Instead, the goal is to accelerate the pace of progress for the whole scientific field. The team folded over 1.1 billion proteins and predicted their structures. They did not design a model specifically for antibodies or for binding one particular target. They designed a model that could understand proteins. If a protein can be designed to actually change physiology, then it becomes possible to cure someone.
Biohub in its current form is viewed as a good fit for what the organization brings to the table. The work started 10 years ago with the goal of building an organization that could cure, prevent, and manage all disease by the end of the century. Early meetings with Nobel Prize-winning scientists involved skepticism about this ambition. The response clarified that the organization would not be the one curing diseases. The goal was always to build tools that could accelerate the whole scientific field so the field collectively could cure all diseases. The original timeline of curing all disease by the end of the century is now considered too conservative.
Scientists identified core problems: working in silos, information not being shared after publication and instead being locked up for long periods, and lack of tooling. One example given was a great tool built by a postdoc in a lab that lives on their computer and disappears when they graduate. It was difficult to build shared tools and shared knowledge bases to move science faster.
The original Biohub model focused on long-term tool development by bringing together engineers and scientists across multiple universities. This approach worked. Over time, the science piece proved effective, leading to increased investment. Biohub is now the primary and main effort. The original San Francisco Biohub has expanded to New York and Chicago.
The unifying theme is the virtual biology initiative. This involves taking unique data sets that can be generated to model effectively, starting with the smallest pieces of proteins and eventually extending to cells and whole biological systems. The work combines frontier AI with frontier biology. Unlike language models where substantial data exists on the internet, biology lacks the necessary data. Existing academic data sets span decades, but much of the desired data does not exist yet.
New types of data require new approaches. Imaging work enables visualization of things people have not been able to see before. Cellular engineering work enables recording of processes occurring inside the body. The Chicago Biohub focuses on building devices to measure inflammation in ways that have not been possible. These efforts create new types of data sets that enable new types of models.
The primary need in the scientific field is tool development that empowers scientists to work faster. The long-term focus on tool development is the intended contribution. The first request for application was around single cell sequencing, focusing on RNA transcribed in individual cells. This was still early in understanding how different cells express their DNA.
Initial funding supported methods development so others could share methodology. This led to funding the Human Cell Atlas, now one of the largest databases of single cell transcriptomes. As annotation became difficult, Cell by Gene was built as a simple annotation tool for scientists. A community formed around Cell by Gene and began contributing additional data independently.
Cell by Gene is now a corpus of knowledge that many transcriptomic-based models are based on and is used regularly by the scientific community. Critics described the work as stamp collecting or gathering bits of data without extracting scientific knowledge and wisdom. The emergence of large language models provided a way to make sense of large amounts of data.
The vision is to understand how biology works and move from discovery-based science to engineering-based science. This would enable systematic understanding of how living beings and living cells work and why things go wrong. The moment large language models became prominent was recognized as a pivotal point where something significant could happen.
Alex Rives had been building models at EvolutionaryScale after starting at Meta FAIR. The pitch from Mark Zuckerberg and Priscilla Chan emphasized integration of frontier AI and frontier biology. This represents a new era of science made possible by artificial intelligence, operating at the scale of information theory where systems can predict the next token and learn world models from data, including biology.
Models are being developed at different scales: incremental molecular modeling and protein folding, cell-based work, and larger-scale biological systems. Modeling cellular behavior differs from protein folding in terms of data and modeling approaches. Each layer is qualitatively different. Understanding protein interactions is necessary to understand how cells work. Understanding systems like the immune system or interactions between cells requires first understanding cells.
High-level abstraction might simulate a system, but real understanding requires building simulations hierarchically at each level. The approach starts with building blocks and proteins. Different types of data will be collected for each layer, and modeling techniques will continue advancing.
A unique aspect is that AI efforts and wet lab efforts are treated as a single effort. Work has been done to bring them together. Data is gathered strategically to connect across the hierarchy. Examples include looking at transcriptomics with spatial information within a cell, examining translucent zebrafish to observe development across different cells including when the brain develops, and using sensors to examine cell-cell communication and different molecules.
Experiments and data collection are designed to bridge across layers and provide connective tissue that supports modeling. The developmental biologists and molecular biologists traditionally worked separately with a reductionist view versus a systems view. The effort focuses on bridging these perspectives through an integrative approach across different layers of biology.
Mechanistic interpretability, traditionally applied to large language models to understand representation space and computation, offers a rich toolkit. For biology, protein language models are trained on protein sequences. Anything they learn about biology emerges from the token prediction training task. These models have demonstrated they can learn biological structure and biological function.
Models are trained on billions of protein sequences covering both known and unknown biology. They develop representations that capture aspects corresponding to the reductive picture of biology built over centuries. This connects proteins where little is known with proteins where more is known through underlying structure grammar in the model's representation space.
At the extreme, this could enable understanding of body systems not previously understood or the mechanism of action for new treatments by interrogating model representations. The goal is to learn the underlying basis for predictions, opening the black box to understand the biology the model represents.
The decision to operate as open source stems from wanting to give tools to the whole scientific community and achieve bigger impact by getting tools into more scientists' hands faster. It is not clear that the work could not be run as a business, but open source was chosen for greater reach.
Data is a greater constraint than compute. Unlike language models, biology data cannot be produced by paying a factory. New novel scientific approaches must be invented, such as the cellular engineering in New York or devices in Chicago. Frontier biology requires real science to advance biological methods to observe what creates the data for models.
The scale of ambition, the 10-to-15-year time horizon required for building these complicated tools, and the capital required make the nonprofit structure appropriate. Not having to think strategically about monetization simplifies the work. The goal is to release models as open source and get them into people's hands to accelerate the whole scientific field.
The neutral nonprofit nature helps harness more people to enter the effort. Achieving the mission of understanding the totality of human biology and curing, preventing, and managing all disease requires the entire academic and biotech industry to work together in a unified way. Talent should not be excluded.
There is a super long tail of diseases. Common diseases like heart disease, cancer, and neurodegenerative diseases contain many subcategories. Rare diseases often get orphaned. Decentralizing the effort and putting tools in many people's hands allows individuals with deep interest in specific conditions, such as spinal muscular atrophy, to make progress. Understanding niche disease processes can unlock broader knowledge about how the human body works.
The goal is to understand how biology works at the individual level. This includes understanding a person's genetics, their risks for different illnesses, and the mechanistic connection between a gene variant, a protein, and a disease process. With that chain understood, a protein or drug can be designed bespoke to the individual for intervention.
Currently, when facing non-standard conditions, people search PubMed, review papers and supplements, examine methods, and assess whether they are represented in the study. This leads to guesses without mechanistic understanding, trying drugs that impact pathways thought to be implicated. The aim is to treat the individual as an individual, understand mechanisms, and intervene effectively. Different diseases are at different stages.
Different diseases exist at varying stages of understanding and intervention capability. For some conditions, the priority is determining which gene variants cause disease versus which do not, which can be highly empowering for patients. Beyond genetic understanding, some diseases allow researchers to grasp the full causal chain but lack the ability to intervene and modify specific protein function. Designing a protein to alter physiology represents a pathway to actual cures, holding equal excitement to foundational discoveries about disease origins.
The approach centers on developing generalizable tools that deliver highly personalized outcomes for individual patients. Large models built on broad data can then be applied across diverse contexts. The vision of preventing or curing all diseases within 100 years may accelerate given AI advances, though the timeline remains dynamic as fixing existing issues reveals new challenges requiring attention.
Research prioritizes understanding broader biological systems rather than isolated diseases. Inflammation represents a key area of emphasis at the Chicago Biohub, with extensive data linking it to multiple conditions. Studying inflammation broadly enables other organizations and companies to develop targeted therapies using shared tools. The immune system serves as another valuable model for cellular engineering work, spanning proteins to cells to dynamic whole-body systems, given its ability for cells to travel throughout the body and its role in addressing various diseases.
The strategy focuses on building toolsets that accelerate work by biotech companies and academic researchers studying specific conditions. External signs of progress that doctors and patients should monitor include how quickly science advances following major model releases.
Translating research discoveries to the clinic requires changes in how clinical research is conducted. Shortening the distance between laboratory findings and patient outcomes remains a priority, though multiple steps demand creative thinking from clinicians about safe deployment methods. Partnerships such as the Jennifer Doudna Cures program at UCSF are exploring how research deployment must evolve given accelerating discovery rates.
The recent ESM Fold release provides an open system for scientific discovery in protein biology. This language model, trained on billions of protein sequences, learns emergent representations of protein biology and enables atomic-resolution structure predictions at high speed. The model balances speed and accuracy on the Pareto frontier for structure prediction tasks. It has characterized over 1.1 billion proteins by predicting their structures and identifying mechanistic connections across the protein universe through interpretability methods.
The model functions as a general world model of protein biology, allowing users to search design space for new proteins. It achieves state-of-the-art performance across structure prediction benchmarks, particularly on protein-protein interactions and protein-antibody interactions critical for therapeutic design. Users can design proteins and single-chain antibodies digitally, then select from hundreds of thousands of trajectories for synthesis and testing in small experimental cycles such as 96-well plates. Nanomolar binders have been identified through this process.
The model was not specifically trained for antibodies or particular targets; protein design emerged as a property of general protein understanding. The open-source release enables anyone to build upon the system, shifting from traditional laboratory screening of hundreds of thousands or millions of antibodies to computational generation followed by minimal experimental validation.
Experimental characterization remains essential. A structural biology center equipped with powerful cryo-EM microscopes enables biophysical and functional assessment of designed proteins. Proteins designed for therapeutically relevant targets have had their function confirmed, with atomic-resolution structures revealing binding interfaces.
Conventional drug development timelines average 15 years at approximately $1.5 billion total cost, with roughly $50 million allocated to molecule discovery and pre-clinical work. The remaining $1.45 billion and decade-plus timeline cover clinical development, influenced by regulatory requirements, recruitment challenges, and trial failures related to absorption, toxicity, and similar factors. The primary focus remains on basic biology and initial molecule creation, though comprehensive cellular models may eventually predict off-target effects.
Single-cell atlases covering all cell types can identify receptors expressed in tissues beyond the intended target, enabling prediction of downstream effects such as renal toxicity before human trials. Transcriptomic models help understand how different cells respond to interventions. Delivery mechanisms and patient care considerations influence which diseases are selected first based on delivery feasibility and risk-reward profiles.
The case of baby KJ at CHOP demonstrated successful CRISPR therapeutic delivery to edit a mutation that would have caused significant neurotoxicity. The disease was selected carefully because liver cells could be targeted effectively, illustrating the importance of choosing initial applications that match delivery capabilities.
Lowering barriers to molecule design and drug development enables programmable biology approaches where medicines can be created for individual patients. This shift carries major implications for drug development processes and the future of medicine. Virtual clinical trials based on personalized models represent one potential regulatory evolution. Stronger predictive capabilities regarding organ-specific impacts and toxicity may encourage more programs to advance.
The Rare One program addresses the long tail of rare diseases where conventional economics discourage company investment. Patient groups can self-organize, pool resources, and express willingness to participate in experimental treatments, improving the economics of developing therapies for smaller populations. Patient registries, natural history studies, biobanks, and patient-organized clinical trials have accelerated gene therapy development for certain rare diseases to 3-5 years rather than decades.
Allowing patients to opt into frontier trials provides a mechanism to test more candidates while maintaining historical vetting standards for the general population. Open ecosystems apply across biology and AI development. The philosophy emphasizes building tools that empower individuals rather than concentrating advancement within few centralized institutions. No central superintelligence is expected to solve all scientific problems; instead, progress occurs through empowering individuals to pursue ideas outside mainstream thinking.
Open source represents one instantiation of technology empowerment, placing capabilities in everyone's hands. In science, this approach aligns with deep institutional commitments, though biosafety considerations require careful balancing. The overarching vision treats technology as a tool placed in individual hands to drive societal progress.
AI researchers at Biohub could work on language models at major commercial labs, but the combination of frontier biology with frontier AI creates unique opportunities unavailable elsewhere. The mission to prevent and cure disease within a 100-year timeline attracts scientists motivated by ambitious, achievable goals. The organization provides resources and positioning to pursue this mission seriously.
Progress in AI does not require hundreds or thousands of researchers; strong groups of approximately a dozen to two dozen people can drive meaningful advances. Finding researchers aligned with the mission is feasible because the goal resonates broadly. Different individuals pursue different missions, and enabling exploration across varied interests accelerates overall progress.
Moving from structure prediction models to reliable zero-shot clinical applications requires solving challenges historically requiring entire careers, including drug optimization, pre-clinical development, and early safety assessment. New scientific paradigms can simplify previously intractable problems through emergent model capabilities. Digital simulation of drug distribution and binding across the human body offers one pathway to addressing toxicity prediction.
Following the ESM Fold release, users have integrated the model with agentic systems for automated protein design, combining frontier AI reasoning capabilities with biological world models. The next major research challenge centers on developing a virtual cell model. Input from the broader ecosystem regarding model usage informs development priorities alongside internal views on structural coverage and capability expansion.
The discussion addresses the virtual cell model and what inputs and outputs to expect. The goal is a system that models each level of biological complexity, including the proteomic layer, genetic layer, and transcriptomic layer, and connects these to phenotype. The model needs enough generality to answer questions about new interventions in contexts it has not been trained on. The key gap to close is making predictions that generalize, which will require an enormous effort to generate data.
Research decisions involve normal constraint management. Every lab feels compute constrained. Ongoing debates center on whether to double down on advancing the protein piece or do more cellular work. Model training decisions also depend on how far along the scaling curve teams want to operate, with model size dependent on the scale of available data. The process is one of constraint optimization: making enough progress to do world-class work at one thing at a time while planting seeds that can blossom over the next couple of years.
This period represents the most dynamic time in technology seen over the course of a career, with rapid weekly advances in AI. Participants describe feeling both invigorated and exhausted, characterizing the moment as a manic phase. Things are very unpredictable, making it difficult to know what is coming next. Early signs of exponentiation appear on the model side with agentic flows, and models are beginning to help more with building other models, though this remains in very early days.
Looking back five years from now, the primary goal is building a hierarchical set of world models around biology while doing the highest quality work in the world. The organization is set up with a world-class AI research team alongside the collection of biohubs, which function as world-class life sciences research organizations. This combination represents a setup no other organization in the world has. Success means producing work that is meaningfully better than what other labs or efforts can approximate, delivering a unique intellectual contribution. Strong faith exists that idea generation from people using the models will materialize, so the focus remains on ensuring world-class work, after which the rest will largely take care of itself.
The biggest update from the last year is the formalization of Biohub as the main focus of philanthropy. Alex Rives and the team joining has brought a world-class group that knows each other and works well together, providing stability that carries underestimated compounding benefits. Prior to Alex leading the effort, previous Biohub leaders were primarily biologists interested in technology. The current structure flips this: while Alex has a biology background, he is primarily an AI researcher with a background in AI applied to biology. This reflects the expectation that this approach will drive more value going forward. The industry overall remains on the exponential curve, with growth continuing to accelerate rather than merely continuing, validating the decision to make a very big investment in these areas.
A critical aspect of the work is closing the loop with actual biology. Code and research operate as closed-loop systems that iterate very fast, whereas biology has historically been an open-loop system. Closing that loop is considered crucial to progress. With the new strategy and Alex at the helm, teams that previously moved generally in the same direction with understanding of potential collaborations are now moving together in a more directed, interlocked way. This represents a shift from having work at a level of maturation where interconnections make sense to actively playing off each other toward shared goals. The change has taken significant work and required teams to reach a level of maturity where interlocking is productive.
Keep No Priors: AI, Machine Learning, Tech, & Startups in your library
Save the videos and channels worth coming back to, and find them again in one place.
