Arc Institute
| Arc Institute | |
|---|---|
| Formation | December 15, 2021 |
| Founders | Silvana Konermann, Patrick Hsu, Patrick Collison |
| Type | Independent nonprofit research institute |
| Headquarters | Palo Alto, California |
| Website | arcinstitute.org |
Arc Institute is an independent nonprofit research institute in Palo Alto, California. It gives investigators long, flexible grants and builds machine-learning models of biology: Evo and Evo 2 read DNA, and State predicts how a cell responds to a drug or a gene knockout. It opened on December 15, 2021, with Stanford University, the University of California, Berkeley, and the University of California, San Francisco as partners.[1]
Molecular Reality Corporation uses the term brain lab for companies that build large models of science. Arc, a nonprofit, is not one of the five companies on its chart.[2]
Founding and structure
Arc was founded by Silvana Konermann, then an assistant professor of biochemistry at Stanford, Patrick Hsu, then an assistant professor of bioengineering at Berkeley, and Patrick Collison, then chief executive of the payments company Stripe. The three had worked together on Fast Grants, the rapid-funding program for COVID-19 research started in 2020.[3][1] The Berkeley announcement said the founding donors had pledged "more than $650 million".[1]
Funding goes to investigators, who get years of support without a project plan attached. Core investigators run laboratories at the Palo Alto site on "renewable, eight-year terms". Innovation Investigators keep their university posts and receive five years of unrestricted funding, and Science Fellows move from doctoral training straight to leading a group.[1][3]
By October 2026 the institute listed 14 investigator laboratories, 5 technology centers and more than 400 staff, whom it calls "Arconauts". Two institute-wide initiatives sit above the laboratories: the Virtual Cell Initiative, and an Alzheimer's disease initiative that looks for the genes and pathways behind the disease.[3]
Genome language models
Evo, described in Science on November 15, 2024, has 7 billion parameters and reads up to 131 kilobases of DNA at single-nucleotide resolution. It was trained on 2.7 million prokaryotic and phage genomes. The authors used it to generate synthetic CRISPR-Cas complexes and entire transposable systems, and to produce coding-rich sequences up to 650 kilobases long.[4]
Evo 2 followed as a preprint on February 19, 2025. It comes in 7-billion and 40-billion-parameter sizes and was trained on 9.3 trillion base pairs from over 128,000 genomes, with a context window of 1 million nucleotides. Its architecture is a hybrid that Arc calls StripedHyena 2, and the team says it allowed 30 times more training data than Evo 1.[5][6] The preprint reports that the model learned, from sequence alone, to mark exon-intron boundaries, transcription factor binding sites and prophage regions, and that without task-specific training it ranks the effect of variants in the human gene BRCA1. Arc's announcement puts its accuracy at separating benign from potentially pathogenic variants at over 90 percent.[5][6]
The weights, the training and inference code and the OpenGenome2 dataset were released with the preprint.[5]
The models have been tested most concretely on a virus. In Science on August 6, 2026, an Arc team reported the first generative design of complete bacteriophage genomes. The template was ΦX174, a phage whose genome of 5,386 nucleotides became the first DNA genome ever sequenced, in 1977.[7] Evo 1 and Evo 2 wrote new genomes in its image, and laboratory testing yielded 16 viable phages. One of them builds its capsid with a DNA packaging protein from a distant relative. A cocktail of the generated phages overcame three strains of Escherichia coli that had resisted ΦX174.[8][9]
Bridge RNAs
Bridge RNAs, reported in Nature on June 26, 2024, are non-coding RNAs that bring a recombinase to a target site and a donor sequence at once, so that both can be chosen by changing the RNA.[10]
Virtual cell work
The Virtual Cell Initiative aims at models that predict what a cell does when it is perturbed by a drug, a gene edit or a signaling molecule. State, first posted as a preprint on June 27, 2025, learned from observational data on 167 million cells and perturbation data on more than 100 million cells across 70 human cell contexts.[11][12] The Virtual Cell Atlas gathers public single-cell data reprocessed by a framework called scBaseCount, and carries Tahoe-100M, a perturbation dataset of about 100 million cells contributed by Tahoe Therapeutics.[11][13]
In June 2025 Arc also announced the Virtual Cell Challenge, an annual contest to predict gene expression after perturbation, with a $100,000 grand prize.[a] A second round was announced in August 2026.[11]
See also
Notes
References
- ^ a b c d "UC Berkeley partners with new Arc Institute to tackle complex diseases". University of California, Berkeley. December 15, 2021.
- ^ Dev Okonkwo-Hale (August 21, 2026). "The Startup That Says Everyone Else Is Building the Brain". The Picoampere.
- ^ a b c "About Arc Institute". Arc Institute. Retrieved October 25, 2026.
- ^ Nguyen, Eric et al. (2024). "Sequence modeling and design from molecular to genome scale with Evo". Science 386 (6723). doi:10.1126/science.ado9336.
- ^ a b c Brixi, Garyk et al. (2025). "Genome modeling and design across all domains of life with Evo 2". bioRxiv. doi:10.1101/2025.02.18.638918.
- ^ a b "Evo 2". Arc Institute. February 19, 2025.
- ^ Sanger, F. et al. (1977). "Nucleotide sequence of bacteriophage φX174 DNA". Nature 265: 687-695. doi:10.1038/265687a0.
- ^ King, Samuel H. et al. (2026). "Generative design of bacteriophages with genome language models". Science. doi:10.1126/science.aec2657.
- ^ King, Samuel H. et al. (2025). "Generative design of novel bacteriophages with genome language models". bioRxiv. doi:10.1101/2025.09.12.675911.
- ^ Durrant, Matthew G. et al. (2024). "Bridge RNAs direct programmable recombination of target and donor DNA". Nature 630 (8018): 984-993. doi:10.1038/s41586-024-07552-4.
- ^ a b c "Arc's Virtual Cell Initiative". Arc Institute. Retrieved October 25, 2026.
- ^ Adduri, Abhinav K. et al. (2025). "Predicting cellular responses to perturbation across diverse contexts with State". bioRxiv. doi:10.1101/2025.06.26.661135.
- ^ Zhang, Jesse et al. (2025). "Tahoe-100M: A giga-scale single-cell perturbation atlas for context-dependent gene function and cellular modeling". bioRxiv. doi:10.1101/2025.02.20.639398.
- ^ Roohani, Yusuf H. et al. (2025). "Virtual Cell Challenge: Toward a Turing test for the virtual cell". Cell 188 (13): 3370-3374. doi:10.1016/j.cell.2025.06.008.