BioHub

Biohub is a 501(c)(3) biomedical research organization combining AI research and experimental biology to understand how cells operate and interact as systems. Its work includes virtual-cell models, biological imaging, inflammation instrumentation and immune-cell engineering. On November 6, 2025, it announced the unification of its scientific teams under the Biohub name and the addition of EvolutionaryScale’s researchers.[10][4]

The organization develops scientific tools and foundational models intended to help other researchers cure and prevent disease. In a November 2025 interview, Mark Zuckerberg described its role as “helping scientists do it”; Priscilla Chan emphasized that clinical translation requires partners. This AI-and-biology program grew from an earlier cross-university research institute established in 2016.[1][5]

In September 2016, Priscilla Chan and Mark Zuckerberg announced a $600 million commitment to create the Chan Zuckerberg Biohub. The institute brought together scientists and engineers from Stanford, UC San Francisco and UC Berkeley, alongside the Chan Zuckerberg Initiative’s engineering team. Its initial leaders were Stanford’s Stephen Quake and UCSF’s Joe DeRisi. The announced headquarters were beside UCSF’s Mission Bay campus, with a satellite site at Stanford; its open-access model would let researchers at the partner universities and elsewhere use its technology and collaborate with its scientists.[5][6]

The first major research projects were announced as five-year efforts: participation in the Human Cell Atlas, an international project to map and characterize healthy human cells, and an Infectious Disease Initiative exploring tools for drugs, diagnostic tests and vaccines. The Biohub commitment was distinct from Chan and Zuckerberg’s broader pledge of $3 billion over ten years to their science initiative.[5][6]

On December 7, 2021, CZI announced a Biohub Network intended to bring scientific and technology institutions together around grand challenges on 15-year horizons. It also extended the San Francisco institute’s operations through 2031. Quake would leave the institute’s co-presidency to lead the network, while DeRisi would assume sole leadership in San Francisco. Alongside this expansion, CZI announced an advanced biomedical imaging institute to develop new technologies through long-term, high-risk research.[7]

In April 2025, CZI released TranscriptFormer, a transformer-based generative model pretrained on single-cell transcriptomics from 12 species spanning 1.5 billion years of evolution. Its training sources included CZ CELLxGENE, Tabula Sapiens, ZebraHub and other public datasets. The researchers reported applications in annotating cell types across species, identifying disease states and prompting the model to explore gene interactions, including cell-type identification in species absent from pretraining, such as rhesus macaque and marmoset.[8]

Computational biologist Sara Simmonds and AI researcher James Pearce were two of TranscriptFormer’s builders. In their April 2025 interview, they described jointly selecting meaningful training data, with computational biologists handling collection, processing and task evaluation, AI researchers building and training models, and engineers supporting infrastructure. Pearce explained the model’s biological tokens: “In the case of TranscriptFormer, an individual gene from an individual cell is a token.”[9]

Pearce envisioned a virtual cell as interacting models in a shared representation space: TranscriptFormer could represent single-cell molecular data, while other models could represent microscopy. This was a proposed architecture for combining different views of biology. In November 2025, Chan described the experimental feedback it would require: test model predictions in the wet lab, determine whether the predicted outcome occurred, and feed the results back into the model.[9][1]

In the November 13, 2025 interview, Chan described a corpus of 125 million cellular RNA transcriptomes and estimated that their own efforts had contributed about 25% of the data, with the wider ecosystem supplying 75%. The corpus illustrated the importance of community data sharing; it was not a count of cells generated by Biohub alone.[1]

The November 6, 2025 announcement united the scientific teams as Biohub and said EvolutionaryScale’s team would join. EvolutionaryScale co-founder and chief scientist Alex Rives would become Head of Science, leading an integrated program across experimental biology, data and AI. Biohub also announced a compute expansion target of 10,000 GPUs by 2028.[4]

Biohub made three additional models freely available on its virtual cells platform: VariantFormer, translating genetic variation into tissue-specific gene activity patterns; CryoLens, analyzing structural similarity in cryo-electron tomography data; and scLDM, generating synthetic single-cell data. They complemented GREmLN, a graph-aware model of gene regulatory networks, and rBio, a conversational language model designed for biological reasoning.[4]

The same announcement launched the Virtual Immune System project to model immune-system complexity, with aims including simulating immune therapies and reprogramming dysfunctional cells. Chan offered a concrete cellular-engineering example from New York Biohub: research toward cells that could detect problematic plaques in the heart, record a signal in DNA, then self-lyse so the signal could be read from cell-free DNA. She presented this as research under way, rather than an available clinical diagnostic.[4][1]

On April 29, 2026, Biohub announced the five-year Virtual Biology Initiative, anchored by a $500 million commitment. It allocated $100 million to external research and coordinated data generation, and $400 million to internal measurement, imaging, engineering, infrastructure and data generation. Biohub committed to making the data it generates open and freely available. The commitment anchors a larger global effort requiring additional funding.[10]

Announced coordinating collaborators included the Allen, Arc, Broad and Wellcome Sanger institutes, alongside the Human Cell Atlas and Human Protein Atlas. NVIDIA joined as a technology partner supporting computing infrastructure, software and technical expertise; Renaissance Philanthropy joined to catalyze additional funding. The announcement described the Billion Cells Project, launched in 2025, as coordinating 17 projects with industry collaborators 10x Genomics and Ultima Genomics. That project count describes the effort’s organization, rather than completion of a billion-cell dataset.[10]

On May 27, 2026, Biohub released ESMC, ESMFold2 and ESM Atlas, making all three freely available to the scientific community. ESMC learns protein representations; ESMFold2 transforms those representations into atomic-resolution predictions of proteins and biomolecular complexes; ESM Atlas makes learned relationships among proteins searchable. Biohub reported approximately 2.8 billion sequences in ESMC’s training data, distinct from the atlas’s coverage of 6.8 billion sequences and 1.1 billion predicted structures.[11]

Biohub reported using ESMFold2 to design binders against five cancer and immunology targets: EGFR, PDGFRβ, PD-L1, CTLA-4 and CD45. Its May 2026 release reported laboratory-confirmed binding, with hit rates of 36–88% for compact minibinders and 15–29% for antibody-derived formats. It also reported that designed PD-L1 binders restored T-cell signaling in laboratory tests. These were laboratory findings, not clinical success rates or evidence of approved treatments.[11]