Halmos Labs, a Y Combinator startup that builds AI agents for drug discovery, says its software has moved a cancer vaccine and an antibiotic into physical testing, with laboratory results still pending. The materials reviewed did not disclose the terms of any funding round.
That qualifier matters. No independent wet-lab data, trial registration or regulatory authorization for either candidate turned up in the research, and the claim lands in a sector that is still waiting for clinical proof that agent-driven research tools can deliver.
What Halmos says it built
Lund University described Halmos on Oct. 8 as "autonomous R&D as a service," a company that initially designs experiments for biotech customers. The university reported work on antimicrobial peptides and cancer vaccines, along with contract research and manufacturing organizations, which it did not name.
Y Combinator lists the company in Lund, Sweden. Dealroom, the startup data platform, gives San Francisco as its headquarters while noting its Lund origins. The research could not establish why the two differ.
According to its homepage, Halmos's tools work with Claude, ChatGPT, Codex and Cursor, and the company offers an MCP interface, a standard that lets AI assistants connect to outside software. A person sits at the approval stage. The pitch, in other words, is human-authorized laboratory coordination rather than unsupervised experimentation.
A cancer-vaccine demo that stays on the computer
The homepage describes an in-silico mRNA design exercise, meaning one run entirely in software, that includes retrospective ranking against known immune-response labels. A separate example on the site walks through a tumor-versus-normal comparison, expression assessment, peptide ranking, reverse translation, mRNA assembly, folding and hazard screening. The output is called a "manufacturable payload."
The phrase is a computational description. It does not establish that anything has been synthesized, delivered or expressed in a body, and it says nothing about immunogenicity, safety or antitumor efficacy.
Halmos's research page also lists "Branchpoint: Structured agentic multiverse analysis," by Mattias Akke, Carl Bringle and Linus von Ekensteen Löfgren, with an accepted poster at a NeurIPS 2026 workshop. A workshop acceptance, useful as it is for credibility among machine-learning researchers, is not drug validation.
Rivals are further along
Insilico Medicine announced Sept. 10 that it had dosed the first patient in Genesis-IPF-3. The research cites that announcement for the company's Phase 3 status, a clinical stage Halmos has not approached.
Isomorphic Labs announced a funding round on May 12 with Thrive, Alphabet, GV, MGX, Temasek, CapitalG and the UK Sovereign AI Fund. The company said the proceeds would scale its IsoDDE drug-design engine, move therapeutic programs toward the clinic and pay for hiring.
An Oct. 9 review in npj Drug Discovery by Krish Ramadurai and Abhirup Banerjee found that no AI-designed small molecule had received FDA marketing approval as of mid-2026. The authors describe an emerging architecture that combines multimodal biological representations, generative chemistry and laboratory feedback, and they set it apart from earlier systems that did not continuously fold experimental results back into the design process.
Skeptics, and a high bar for cancer vaccines
Andreas Bender, Jack W. Scannell, David A. Shaywitz and co-authors wrote in Nature Reviews Drug Discovery on Aug. 7 that clinically relevant impact remains limited. They pointed to conditional biological data, weak clinical translation, underspecified use cases, and benchmarks that measure model performance rather than better research decisions. They also cautioned that companies often disclose too little about the human work of lead optimization to credit final candidates solely to AI.
The bar for cancer vaccines is steep. Merck said June 1 that five-year Phase 2b follow-up of its mRNA vaccine intismeran autogene, given with Keytruda to patients with high-risk melanoma, showed a 49% reduction in recurrence or death and a 59% reduction in distant metastasis or death. Merck and Moderna are due to present Phase 3 InterPath-001 results at ESMO 2026.
Regulators are still finding their footing. The FDA's page on AI in drug development says its 2025 guidance remains in draft form. It cites more than 500 submissions containing AI components from 2016 to 2023, a figure that does not translate into a count of AI-discovered drugs or approvals.
What comes next
The test now is the pending lab data on the vaccine and the antibiotic. Halmos has not published results, and nothing in the research places it ahead of any clinical-stage competitor.
