AI and drug development · Clinical response prediction
Atlas Discovery
A three person company in San Francisco building foundation models that try to say, before a patient is dosed, whether a drug will work. Two months out of stealth, with preprints and a preclinical partner, and no regulatory or financial record.
The trial that could have been a third the size
UNIFI was the phase 3 program that established ustekinumab in moderate to severe ulcerative colitis. Its induction phase randomized 961 patients. A substudy took baseline colon biopsies from 550 of them, and that data was later released publicly as GSE206285.
Years later a three person startup pulled the data down and asked a different question of it: before anyone was dosed, could you have told who would respond?
Atlas Discovery embedded the biopsies with a foundation model of patient biology, kept the 358 ustekinumab-treated patients that passed its filters, 56 of whom met the response endpoint at week 8, and trained a small supervised head on the embeddings. Five-fold cross validation returned an AUROC of 0.760.
Then it did the arithmetic a drug developer would care about. Holding the power UNIFI actually achieved fixed, a biomarker that accurate would have allowed roughly 91 patients per arm instead of 320, a saving of more than 450 people.
The company states the limits itself. The predictor was trained and cross validated on the same cohort it is scored against, and enrichment cannot shrink the part of a trial that exists to characterize safety.
The problem it is aimed at
The gap is clinical translation: the expensive distance between a promising result in cells or animals and evidence that a drug works in people.
Its statistical form is narrow. The information needed sits in molecular assays on biopsies, but one trial of a few hundred patients is too small to learn a rule across roughly 20,000 genes. The field answered with low dimensional companion diagnostics, several of which stayed in clinical use at AUROCs of only 0.55 to 0.65 because nothing richer could be fitted at trial scale.
The bet is that pretraining moves the learning off the trial. A model absorbs the structure of biology once, from millions of samples, and each new trial learns only which directions in that latent space separate responders from non responders. Atlas then wants to use those predictions to repurpose approved drugs into diseases nobody has funded.
Two Dartmouth engineers and an abandoned PhD
Shaamil Karim studied computer science at Dartmouth and worked on applied AI at Etched and RunPod. Christian Gensbigler studied mathematics and biology at Dartmouth and researched theoretical biology at Johns Hopkins. Those biographies come from YC general partner Ankit Gupta, reported by RuntimeWire; Dartmouth addresses appear in the footnotes of the preprints.
Sanjukta Bhattacharya spent months at Los Alamos National Laboratory on sample efficient learning, took a master’s by research at Edinburgh, and started a PhD in AI at the University of Bristol before leaving it. Jon Lees of Bristol is the fourth author on both preprints. Gupta, the YC partner attached to the company, co-founded Reverie Labs, which Ginkgo Bioworks acquired in 2024.
Seven months of public record
February 19, 2026: Discrete Cell Models goes up on bioRxiv, a diffusion framework that models single cell gene expression in discrete rather than continuous space. June 15: a second preprint introduces ExpressionVAE, and the clinical response case study appears on the company blog.
Late July: Atlas leaves stealth through a post by Gupta on X, picked up by RuntimeWire. August: three posts on repurposing agents, trial prediction and ClinicBench, an evaluation drawn from more than 500,000 real patient records. On August 2 Alliance to Cure Cavernous Malformation says it is running preclinical tests on a candidate the agents found. September: Demo Day.
What is proven, and what is still claimed
| Evidence | What the record shows | Source type |
|---|---|---|
| Preprint, Feb 2026 | Discrete Diffusion for Single-Cell Gene Expression Modeling, bioRxiv, Feb 19, 2026, by Bhattacharya, Gensbigler, Karim and Lees. Reports a 5-fold gain in MMD2RBF over scLDM on the Dentate Gyrus benchmark. Not peer reviewed. | Public record |
| Preprint, Jun 2026 | Elucidating the Design Space of Generative Models for Single-Cell Perturbation Prediction, bioRxiv, Jun 15, 2026, same four authors. The public code repository titles the work ICML GenBio 2026. Not peer reviewed. | Public record |
| Peer reviewed papers | None found in PubMed for the company as of Sep 23, 2026. | Not found |
| Response prediction | AUROC 0.760 on 358 patients from the public GSE206285 set, five-fold cross validation on that same cohort. Retrospective. No prospective use in a live trial reported. | Company-stated |
| Trial success prediction | Trained on a 100,000 trial corpus, evaluated on a held out 10,000, AUROC 0.858 for phase 3 to approval. The company notes its baselines were scored on different test sets. | Company-stated |
| RepurposingBench | The posted results score frontier models only: 35% to 41% weighted recall on 143 rediscovery pairs, 5.6 to 8.4 of 100 on 10 prospective pairs. The claim of nearly 5x improvement by Atlas’s own agents sits in the launch post, not in the benchmark write up. | Claim not in the posted results |
| Repurposing backtest | On a temporally restricted heart failure test the harness ranked dapagliflozin 2nd, against 4,388 of about 8,000 for the knowledge graph model TxGNN. Company run and company scored. | Company-stated |
| Preclinical partner | Alliance to Cure Cavernous Malformation is beginning preclinical tests of an approved drug the agents surfaced for CCM. The drug is not named. No results yet. | Partner-stated |
| Commercial traction | Six figures in contracted revenue, an eight figure letter of intent with an AI lab, three rare disease foundations testing surfaced drugs. No counterparty named. | Company-stated |
| Founding year | YC gives 2026 and a team of 3; a vendor directory lists 2024 and 2 founders. | Directories disagree |
| Financing | An SEC EDGAR company search on Sep 23, 2026 returned no matching companies, so no Form D. No priced round reported. | Public record |
Read plainly: everything Atlas has shown is retrospective, computational and scored by Atlas. The work is unusually well documented for a company this young, with limitations sections that argue against its own headline numbers. But no patient has been enrolled on one of its predictions, no drug it named has produced a public animal result, and neither preprint has been through peer review. The first external test is the one running now in a patient foundation’s lab.
What to watch
- Preclinical results from Alliance to Cure Cavernous Malformation, and whether the drug is ever named.
- Any prospective use of the response predictor to enrich enrollment in a live trial.
- Peer review of the two bioRxiv preprints.
- Whether the eight figure letter of intent becomes a signed contract with a named counterparty.
- A first Form D, due within 15 days of the first sale in a Regulation D offering.
In their words
“Our thesis is that AI will drop the cost of drug development enough to push the viable threshold from ~$150M toward ~$15M, opening a ~$500B market incumbents have structurally ignored.”
Atlas Discovery launch post, Y Combinator, 2026 · Company
“The bottleneck is no longer supply of compounds; it’s the judgment needed to search that space and run the trials, and the removal of human coordination costs.”
Atlas Discovery launch post, Y Combinator, 2026 · Company
“In biology, a model can look strong on the data it was built around and still fail on a new patient, a new assay, or a harder question.”
Sanjukta Bhattacharya, co-founder, personal site, 2026 · Founder
“This is one indication, drug, and cohort, with a model pretrained on a fraction of existing data. What we have demonstrated is a lower bound on what is achievable.”
Atlas Discovery, clinical response prediction write up, 2026 · Company
“We’re excited about the potential of our partnership with Atlas Discovery to dramatically speed up the path to real treatments for our patients,”
Connie Lee, CEO, Alliance to Cure Cavernous Malformation, 2026 · Partner
“The result does not establish how the model would perform prospectively in a live trial, but it illustrates the economic and operational case Atlas is making to drug developers.”
Ryan Merket, RuntimeWire, July 2026 · Independent
“Biotech will eventually be judged by biology rather than benchmark scores. For a company this young, Atlas has already created several independent ways to find out whether the system works.”
RuntimeWire, Demo Day ranking, September 2026 · Independent
“We mean to help move medicine toward a world in which no biologically-tractable disease goes untreated merely for want of a market.”
Atlas Discovery, Pushing the Economic Frontier of Medicine, 2026 · Company
Related companies
Sources
- Public recordDiscrete Diffusion for Single-Cell Gene Expression Modeling
- Public recordElucidating the Design Space of Generative Models for Single-Cell Perturbation Prediction
- Public recordsc-evae code repository, listed as ICML GenBio 2026
- Public recordEDGAR company search, no matching companies
- Not foundFDA, ClinicalTrials.gov, NIH RePORTER, NSF and PubMed searches
- IndependentAtlas Discovery launches AI models to predict patient response in clinical trials
- IndependentYC S26’s top decile: the 24 companies RuntimeWire would call first
- IndependentWhat Y Combinator’s September 2026 Demo Day reveals
- IndependentYC S26 Companies, full startup list
- DirectoryAtlas Discovery company and product profiles
- CompanyAtlas Discovery: AI-native pharma company, launch post
- CompanyClinical Response Prediction with Foundational Models of Patient Biology
- CompanyClinical Trial Prediction with Agents
- CompanyAgents for Drug Repurposing
- CompanyRepurposingBench: A Benchmark for Drug Repurposing
- CompanyPushing the Economic Frontier of Medicine
- PartnerAlliance to Cure Cavernous Malformation partnership announcement
- FounderSanjukta Bhattacharya personal site
Profile researched and written by Healthcare Discovery. Last updated September 29, 2026.
