Biohub

Biohub is a 501(c)(3) biomedical research organization supported by the Chan Zuckerberg Initiative (CZI). Its AI-powered biology program combines laboratory technologies, data generation and predictive models to study proteins, cells and immune systems. Its purpose is to equip scientists with tools to understand, cure or prevent disease.[4][5]

On November 6, 2025, founders Mark Zuckerberg and Priscilla Chan announced that Biohub would become their primary philanthropic effort, receiving “the vast majority of our resources.” Their founders’ update retained other philanthropic efforts and assigned CZI an infrastructure and support role. The decision established an allocation priority without specifying a monetary amount.[3]

The accompanying initiative announcement said EvolutionaryScale’s team would join Biohub and its co-founder and chief scientist, Alex Rives, would become head of science, integrating experimental biology, data and AI while retaining affiliations with the Broad Institute and MIT. By May 2026, Rives was identified as Biohub’s head of science. Biohub also announced a tenfold compute expansion to 10,000 GPUs by 2028—a future capacity target.[4][5]

The program set four research challenges: a unified AI model of the cell; improved biological imaging; instrumentation to monitor and modulate inflammation in real time; and immune-system reprogramming for disease detection, prevention and treatment. Its Virtual Immune System initiative aims to model interactions within the human immune system and support simulation of immune therapies.[4]

In the November 2025 founder interview, Chan described the intended feedback cycle: model predictions must be tested in wet-lab experiments, with observed results fed back into the model. Zuckerberg envisioned models generating hypotheses that scientists select for testing. Chan also distinguished Biohub’s fundamental research from clinical delivery: “We’re building the basics. We’re understanding, like, cells and molecules. Someone else will do it.” Partnerships would be needed to carry that work into medicine.[1]

The November 2025 initiative announcement made three models freely available: VariantFormer, which translates personal genetic variations into tissue-specific gene activity patterns; CryoLens, a pretrained cryo-electron tomography model for unsupervised structural similarity analysis; and scLDM, which generates synthetic single-cell data. They complemented GREmLN, a graph-aware model of gene regulation, and rBio, a conversational language model for biological reasoning.[4]

Chan reported in the November 13, 2025 interview that the cell-transcriptome corpus had reached 125 million cells. She estimated that their organization supplied approximately 25% of the data and the wider ecosystem supplied 75%, crediting CELLxGENE with making contribution easier. These RNA measurements provide one view of cellular activity; she described spatial imaging and dynamic imaging as additional dimensions needed to understand cells.[1]

On May 27, 2026, Biohub released a protein-biology modeling system comprising ESMC, a protein language model; ESMFold2, a structure-prediction and design model; and ESM Atlas, a navigable map of protein representations. Biohub reported approximately 2.8 billion training sequences for ESMC, versus atlas coverage of 6.8 billion protein sequences and 1.1 billion predicted structures. All three were made freely available; the released models carry an MIT license.[5][6]

The ESM research program predates the team’s integration into Biohub, tracing its first transformer protein language model to 2019. Early models learned by predicting randomly masked amino acids from the rest of a sequence. Biohub’s technical explanation describes ESMC’s learned representations as a foundation for predicting protein structure and function.[6][4]

ESMFold2 repeatedly passes representations through shared transformer parameters, refining a structural prediction with each loop. Training propagates gradients through that recurrent process; additional loops at inference time allocate more computation without retraining. The model uses ESMC representations directly and optionally accepts multiple-sequence alignments. Biohub credits NVIDIA’s TransformerEngine and cuEquivariance kernels in training, and collaboration with NVIDIA on context-parallel and accelerated-inference kernels.[6]

Biohub reported laboratory-tested binder designs against five cancer and immunology targets: EGFR, PDGFRβ, PD-L1, CTLA-4 and CD45. Reported hit rates were 36–88% for compact minibinders and 15–29% for antibody-derived formats. Its technical account describes generating candidates, ranking them with model confidence scores and screening the top 84 designs for each target, format and compute scale. The computational search took days, addressing the earliest stage of binder discovery.[5][6]

For PD-L1, Biohub reported that a designed single-chain antibody fragment bound with a measured affinity of 4.3 nM and relieved suppression of T-cell signaling in cell-based assays. It also used cryo-electron microscopy to validate the predicted location and orientation of a designed minibinder bound to EGFR. These were functional and structural laboratory results, rather than demonstrated treatments in patients.[6][5]

The expanded AI program builds on an earlier institutional model: bringing scientists and engineers together to develop biological tools and shared data. That model began with the original Chan Zuckerberg Biohub in 2016.[7][3]

Priscilla Chan’s September 23, 2016 announcement described a $600 million commitment to create Biohub, bringing together researchers from Stanford, UCSF and Berkeley with CZI’s engineering team. Stanford’s Stephen Quake and UCSF’s Joe DeRisi were its initial leaders.[7]

Its first major projects were the Human Cell Atlas and the Infectious Disease Initiative, announced for the following five years. The Human Cell Atlas was an international effort to map and characterize cells in the healthy human body. The infectious-disease program proposed tools and approaches for drugs, diagnostic tests and vaccines, including responses to HIV, Ebola and emerging threats such as Zika.[7]

By October 2023, the network included Biohubs in San Francisco and Chicago, the imaging institute in Redwood City, and the newly announced New York Biohub. The New York launch brought together Columbia University, The Rockefeller University and Yale University under announced leader Andrea Califano. Its research agenda centered on engineering immune cells to detect early disease, report molecular changes and eventually make targeted repairs.[8]

The October 18, 2023 announcement specified a planned $250 million investment from CZI and $10 million contributions each from the New York City Economic Development Corporation and Empire State Development. Those itemized commitments totaled $270 million.[8]