Tahoe Therapeutics
Tahoe Therapeutics, formerly Vevo Therapeutics, generates perturbation data on single cells and trains machine learning models on it for drug discovery. The South San Francisco company produced Tahoe-100M, single cell measurements from cancer models exposed to drugs, released openly in February 2025 as the first contribution to Arc Institute's Virtual Cell Atlas, with its methods and observations published in Cell in 2026, and it now generates data across primary cells, organoids, iPSC derived cells and in vivo models. Tahoe-x1 is an openly released family of models for representing cell states, and Rhaister, introduced in June 2026 with the Emerald Bay data release, predicts how cells respond to perturbations they have not seen, evaluated against laboratory measurements. Tara is a research agent that queries Tahoe's datasets from a scientist's own AI assistant, generating hypotheses and checking them against the data.
In January 2026, Tahoe and Alloy Therapeutics formed a joint venture to develop antibody drug conjugates for hard to treat cancers, with Tahoe contributing tumor targets and Alloy the drug development capabilities, and Tahoe joined Arc Institute and Biohub to generate a larger perturbation dataset for public release. The company raised $30 million in August 2025 after a $12 million seed in 2022, has collaborated with NVIDIA on biological foundation models, and added drug discovery leaders Wayne Spevak and Ben Powell in November 2025.
Capability Axes
An AI Health Index grade measures what a buyer can verify from public sources on the date shown. It is not a rating of how good the product is. A vendor can build an excellent system and grade low on an axis because it publishes nothing an outsider can check. How grades read
The discovery platform runs models trained on experimental cell data to study drug response in oncology: Tahoe-x1 represents cell states for downstream research, and Rhaister predicts responses in biological settings outside its input experiments, trained in seconds and run in milliseconds. The Alloy venture uses that discovery work to pick tumor targets for an oncology program, and Tara puts an agent over the datasets to generate and test hypotheses. The atlas is also an output other researchers use on its own, but within the discovery business the models are how Tahoe turns measurements into predictions and research leads.
Tara is sold as an autonomous research agent that generates hypotheses, validates them against the data and surfaces findings, working through a scientist's own AI assistant, and the models return research predictions that scientists compare with existing observations before taking them into experiments. Examples show agents returning analyses a scientist can inspect and challenge.
No approval gate, confidence threshold or escalation step applies to outputs before they inform further work, and the use stays in research on drug response and target selection, with later development decisions made by research teams.
Tahoe-x1 is a named model family with weights released on Hugging Face and GitHub, a description of the cell datasets behind it, separate checkpoints and benchmark comparisons for cancer research tasks. Rhaister is a separate family of machine learning models with its own research account and comparison against laboratory measurements, so an outside researcher can see which release and model type produced a result. Version and update documentation covers those releases, not the third party assistants that query or summarize Tahoe data through Tara.
Tahoe-x1 and Rhaister are Tahoe's own research releases that a researcher can run locally, so those workflows send no input to a Tahoe inference service. Tara plugs into assistant environments the user chooses, including Codex and Claude Code, and the demos show particular tool combinations without listing every place a research query or its output may be processed. The Arc Institute and Biohub collaboration generates experimental inputs, and no party list covers prompts or partner research data across the discovery offering.
The Tahoe-100M dataset has a peer reviewed paper in Cell documenting the experimental measurements and analyses across cancer cell lines. The Tahoe-x1 work reports model comparisons on cancer research benchmarks, including how well the models represent cell states, and Rhaister was evaluated against laboratory observations in the Emerald Bay release, with a research preprint explaining the comparison. The Cell paper validates the data and experimental contribution rather than a model, and none of the work establishes independently evaluated treatment outcomes or clinical response prediction.
The main public workflow uses experimental measurements from cancer cell models, which researchers can run through released models in their own environments or query through Tara. Tahoe explains how it curates and publishes the atlas, but not how it would handle identifiable clinical information in a joint discovery project or through a connected assistant.
No retention schedule for partner inputs, policy on training with identifiable information or procedure for reporting an unsafe AI output is given, and the homepage carries no privacy or security statement, gaps that would matter once a partner links experimental data to patient records.
The published oncology atlas holds measurements from cell models and is offered as a research resource, and research collaborations run separately with pharmaceutical and technology partners. Nothing says whether Tahoe acts as a business associate when a partner supplies identifiable health data or whether any product handles such data under an agreement, and no service privacy document governs that workflow. A public atlas built on laboratory data settles nothing contractual for a private program that brings in patient records.
The released Tahoe models can run inside a researcher's own computing environment, and Tara reaches the datasets through an assistant the user picks, paths that put security in the hands of whoever runs them. Tahoe documents the scientific provenance of its published data, but no security examination covers the hosted research tools or a partner workspace, and no SOC 2 report, ISO 27001 certificate or independent test is claimed for the company. Local deployment gives researchers control of their infrastructure while leaving Tahoe's hosted services unexamined.
Tahoe-x1 and Rhaister are offered for research on cell states and drug response, and the Alloy venture develops oncology drug candidates from targets found in that work, so the outputs inform early choices about what to study experimentally. No clinical indication or patient facing diagnostic use is claimed for the models, and no FDA clearance is cited for a research tool. Clinical development and regulatory work for a future therapy would be a separate undertaking from releasing a model or a dataset used to choose what to study.
The published model work compares performance across experimental conditions and against other methods, and the open releases let researchers reproduce the stated benchmarks or examine a result. The cell atlas describes its cancer models, which shows what experimental evidence is represented.
No formal governance program covers private projects, and no performance comparison by patient demographic group is offered; results across cancer cell lines say nothing about whether a drug response prediction holds as well for people with different clinical backgrounds.
Researchers can check a prediction or an assistant answer against the public atlas and the benchmark material, since the dataset has a published methods paper and each model release a research account, a way to challenge a result before taking it into a laboratory or development program. No output warranty, guaranteed response, contractual correction obligation or compensation for a misleading research result is offered. Tara has a contact address, with no defined route for a partner or downstream user to obtain a remedy for an AI output.
No clinical record integration is offered for Tahoe's products, which start from experimental measurements in a discovery setting. The research models handle single cell experiment data, Tara reaches the datasets through research assistants, and the Alloy partnership carries discovery outputs into oncology development.
No named connection to Epic, Oracle Health or another record system, and no HL7 or FHIR exchange for bringing longitudinal records in or writing results back, is described, and no laboratory information system or research data platform connector is named either.
A research group can download the open Tahoe-x1 weights and run the model in its own environment, which keeps the computation and any local research inputs where the group chooses. Tara works differently, connecting a scientist's AI assistant to Tahoe's datasets through a plugin.
Where the hosted service runs, where partner data or logs rest and how private research projects are kept apart are not stated, so the local model release is the one deployment path with a location the customer controls.
Research assets are released openly, including the public cell atlas and model weights on GitHub and Hugging Face, and the drug discovery work runs through separately negotiated collaborations: the Alloy agreement seeded a joint company for oncology programs, and the Arc and Biohub work is a funded research collaboration. Tara installs as a plugin to a researcher's assistant, with no price stated. No price is published for a discovery project, dedicated research access, program milestones or a venture like Alloy's, and the free research releases sit apart from any commercial access.
The published evidence centers on preclinical cancer research: Tahoe-100M spans many cancer cell lines, the models study drug response and other cell behavior, and the Alloy venture targets hard to treat cancers. The Arc and Biohub effort and Tahoe's newer data across primary cells, organoids, iPSC derived cells and in vivo models widen the experimental settings represented. Validated behavior covers those laboratory settings, with nothing for a diagnostic specialty, a hospital workflow or treatment in people.
Alternatives to Tahoe Therapeutics
Other products that do the same kind of work as Tahoe Therapeutics, and what each one does.
South San Francisco company founded in 2023 and launched publicly in April 2024 with more than $1 billion in committed funding, co led by ARCH Venture Partners and Foresite Labs.
Noetik builds AI models of how tumors behave and licenses them to drug companies for cancer research. The models help researchers find targets, run virtual experiments and predict which patients will respond to a therapy.
Immunai occupies a different position from most of this category. It is not developing its own drugs.
Deep Genomics is a Toronto AI drug discovery company launched in 2015 by Brendan Frey out of his University of Toronto work on the genetic determinants of disease.
Nabla Bio designs antibodies on a computer for pharmaceutical partners, then makes and tests them in its own lab.
Cellarity is a Flagship Pioneering company in Somerville, Massachusetts, founded in 2019, that designs drugs against the behavior of a whole cell rather than against a molecular target.
Pricing
Vendor-published figures are labeled as such. Figures labeled “Estimated” are derived from third-party sources and have not been confirmed by the vendor.
| Entry Price | Pricing Basis | BAA Tier | Implementation | Source |
|---|---|---|---|---|
|
No public commercial rates
|
Research releases are open; private discovery work is partnership based | No business associate agreement is offered. | Partner terms are negotiated | Vendor Published |
Research models and data are released openly. Financial terms for joint discovery programs, partner access and Tara are not published.