The pharmaceutical industry has a failure problem it can't seem to shake. Overall success rates from nonclinical development to approval hover between 7% and 12%, depending on which therapeutic area you're talking about. And the price tag for that brutal attrition? Try $244.9 million for a single Phase III oncology trial, according to 2024 federal data on cancer studies.
It's the kind of number that keeps CFOs awake at night.
A team of three founders believes they've spotted an opening in that expensive wreckage. Fresh from Y Combinator's Summer 2026 cohort, Atlas Discovery is making a bold claim: their AI can predict which patients will actually respond to drugs—before those patients ever enroll in a trial. The San Francisco startup, led by Sanjukta Bhattacharya, Shaamil Karim, and Christian Gensbigler, says their models hit 70-90% accuracy in separating responders from non-responders across multiple diseases. They measure this using AUROC, a standard statistical metric where anything above 0.7 starts to look interesting and 0.9 borders on remarkable.
If true—and that's a significant if—it's not just an incremental advance. It's the kind of breakthrough that could reshape how drugs get developed, tested, and brought to market. The industry's historical success rate from Phase I through final approval is roughly estimated between 7% and 12%, using data spanning 2011-2020. Even modest improvement in those odds would translate into billions saved and, perhaps more importantly, treatments reaching patients years faster.
But the space between a promising algorithm and actually changing how trials work? That's where things get complicated.
When Biology Meets the Balance Sheet
The fundamental problem isn't exactly news to anyone who's worked in drug development. A 2021 analysis of programs from 2011 to 2020 found that the ugliest failure rates happen at the Phase II to Phase III transition—precisely the point where costs explode. Even as the industry celebrated 79 novel active substances launched globally in 2025, per IQVIA's tracking, that underlying failure rate hasn't budged much.
Atlas Discovery frames the core issue with refreshing bluntness on their website: "No animal or cell in a dish can predict how a human will respond." Preclinical models, for all their utility, simply can't capture the messy biological complexity of actual patient populations. Which means Phase III trials—those massive, multi-year, continent-spanning studies—often become the first real test of whether a drug actually works in humans.
By then, pharma companies have already burned through hundreds of millions.
The financial pressure shows up in industry metrics, though not always where you'd expect. Deloitte's 2026 analysis of pharmaceutical innovation returns found that the internal rate of return for the top 20 biopharma companies crept up from 5.9% in 2024 to 7.0% in 2025. That bump was driven largely by obesity drug blockbusters—the semaglutides and tirzepatides that became household names. Strip those out, and the picture looks less rosy. R&D pipelines remain stubbornly inefficient.
The Foundation Model Bet
What Atlas Discovery is building centers on foundation models—those large neural networks trained on diverse datasets that can then be fine-tuned for specific tasks. It's the same underlying architecture that powers ChatGPT, but aimed at a very different problem: predicting patient biology rather than generating text.
In a June 2026 case study posted on their blog, the company detailed work on a 358-patient cohort treated with ustekinumab, an immunosuppressant used for conditions like psoriasis and Crohn's disease. Only 56 patients—about 15.6%—were actual responders. Their model achieved AUROC scores between 0.7 and 0.9, using what they describe as "interpretable gene attention" mechanisms. Translation: the system can highlight which biological features it thinks matter most for the prediction, rather than operating as a complete black box.
The company posted a bioRxiv preprint on June 15, 2026, and presented findings at ICLR, CSHL, and ICML 2026 workshops. The technical approach appears to pull from both structured preclinical datasets and actual clinical trial results—essentially learning patterns from past drug-patient interactions that might forecast future ones.
This isn't happening in isolation. A May 2026 paper in Scientific Reports showed that foundation model-derived features improved drug response prediction for inflammatory bowel disease using relatively small clinical datasets. A separate May 2026 review in Briefings in Bioinformatics surveyed deep learning methods for cancer drug response prediction, noting both genuine promise and persistent validation gaps that haven't been fully addressed.
The research momentum is real. So is the skepticism around whether models trained on historical data can actually generalize to new molecules and new populations. That question—generalization—is where AI in drug development tends to live or die.
A Market Already Crowded With Believers
Atlas Discovery is hardly alone in seeing this opportunity. The competitive landscape has gotten thick with well-funded players, each attacking patient response prediction from different technical angles.
Unlearn, which has raised significant venture funding, focuses on disease-specific digital twins that forecast placebo trajectories—allowing trials to shrink their control arms. In 2026 alone, the company announced collaborations with VectorY (for an ALS trial), Acumen (Alzheimer's), and SOLA Biosciences. Their Parkinson's Disease Digital Twin v1.2, trained on more than 2,100 participants, forecasts longitudinal placebo outcomes using Bayesian methods. Different technical bet, different commercial angle.
Owkin, coming off multi-year deals with Sanofi in 2026 and retrospective subgroup work with Servier (announced December 2025), is building what it calls "agentic AI" for biomarker discovery. Tempus expanded its strategic collaboration with Bristol Myers Squibb specifically to "enhance the probability of success across clinical development programs in oncology and neuroscience." Recursion, in its 2025 shareholder letter, signaled a strategic pivot toward "ClinTech"—AI-enabled trial design and patient stratification—as its pipeline matures and it needs to actually prove these drugs work.
Each company is making different trade-offs. Quris-AI uses "patient-on-a-chip" technology plus AI for preclinical safety prediction; Merck KGaA adopted it in January 2025 after a two-year validation showing it could flag drug-induced liver injury. VeriSIM Life is collaborating with FDA's National Center for Toxicological Research on mechanistic-AI models. Lantern Pharma's RADR platform proposes mutation-specific enrichment strategies—like targeting EGFR L858R mutations in its LP-300 HARMONIC Phase 2 trial.
What sets Atlas Discovery apart, at least in how they position themselves, is the foundation model architecture trained explicitly to predict clinical response rather than optimize trial logistics or generate synthetic controls. Whether that distinction matters commercially depends entirely on validation at scale—and that's still ahead of them.
Regulators Aren't Moving Fast, But They're Moving
The regulatory door is opening. Slowly.
On January 14, 2026, the FDA and EMA jointly released "10 Guiding Principles for Good AI Practice in Drug Development." The principles emphasize transparency, validation, and ongoing monitoring—a clear signal that AI tools won't get a pass just because they're novel or exciting.
The first concrete proof point came in December 2025, when the FDA qualified PathAI's AIM-NASH tool to assist liver biopsy scoring in MASH/NASH trials. The EMA followed with a qualification opinion in March 2025. It's a narrow use case—the tool assists pathologists in scoring biopsies, it doesn't replace clinical endpoints—but it established a pathway that didn't exist before.
The EMA's September 2024 Reflection Paper on AI in the medicinal product lifecycle laid out expectations for model development, validation, and post-market surveillance. The agency is now developing a separate reflection paper on external controls for regulatory decision-making, expected sometime in 2025-2026. The FDA has published guidance on decentralized trials, real-world evidence for device decisions, and maintains an active AI for Drug Development hub tracking submissions.
ICH E6(R3), the updated Good Clinical Practice guideline adopted in January 2025, supports risk-based quality and modern data methods—which could facilitate AI-enabled trial monitoring and adaptive designs. But regulatory acceptance of AI-derived predictions for actual trial enrichment or endpoint definition? That remains case-by-case, molecule-by-molecule.
Atlas Discovery has not provided any documented information about regulatory interactions as of mid-2026. At three people and fresh out of YC, they're likely focused on building validation datasets and early pharma pilot projects rather than navigating qualification pathways. But the regulatory landscape will ultimately shape how quickly—and in what form—pharma actually adopts their technology.
Market Sizing as Creative Fiction
Market research firms love projecting hockey-stick growth for AI in clinical trials. They also can't seem to agree on the numbers.
Fortune Business Insights valued the 2025 market at $3.8 billion and projects $77.3 billion by 2034—a 39.1% compound annual growth rate that would make this one of the fastest-growing sectors in healthcare. Research & Markets sees $7.6 billion in 2026 growing to $46.9 billion by 2031. MarketGlass, more conservatively, forecasts $1.6 billion in 2025 reaching $4.1 billion by 2032 at 14% CAGR.
The spread reflects definitional chaos. What actually counts as "AI in clinical trials"? Patient recruitment algorithms? Digital twins? Natural language processing for protocol design? Data monitoring platforms? The answer depends on who's selling the report.
Adjacent markets offer some context, though. Companion diagnostics—tests that identify patients likely to respond to specific therapies—are projected to grow from roughly $8.7-10.3 billion in 2026 to $20-22.5 billion by 2033-2035. Predictive biomarkers, used broadly in drug development, represent a $32 billion market in 2026 with projections near $63.5 billion by 2030.
McKinsey estimated in January 2024 that generative AI could unlock $60-110 billion in annual value across pharmaceutical R&D, manufacturing, and medical functions. A January 2026 BCG analysis emphasized that "AI agents" are transforming healthcare operations including trial monitoring and recruitment—but warned against scattershot pilots in favor of focused, scaled deployments.
The numbers are directional at best, possibly aspirational. What matters more is whether pharma R&D budgets are actually shifting toward AI-enabled approaches. And the evidence, so far, suggests they are.
The Validation Problem Nobody Talks About Enough
Here's where skepticism is warranted. Atlas Discovery's claimed AUROC of 0.7-0.9 is promising—if it generalizes beyond the datasets it was trained on. AUROC measures how well a model distinguishes between classes (responders versus non-responders), with 0.5 being random chance and 1.0 being perfect separation. Scores in the 0.7-0.9 range suggest real predictive signal.
But AUROC on a retrospective dataset is one thing. Prospective validation in actual trials—where you make predictions before you know the outcomes—is something else entirely.
A June 2026 arXiv landscape analysis of AI trials in medicine found that multimodal AI accounts for roughly 34% of AI trials, with clinical text and NLP trials up sevenfold since 2018. The field is shifting from retrospective studies toward more prospective evaluations. But many still lack long-term outcome data, and even fewer have been tested across multiple sites or diverse populations.
Foundation models in genomics and biology face well-documented challenges: distribution shift when applied to new populations, batch effects across datasets, and the persistent risk of learning spurious correlations rather than actual causal biology. A June 2026 arXiv paper on human genetic evidence and drug approval odds found that targets backed by genetic validation had roughly 3.25 times higher approval odds—suggesting that biology-driven selection beats purely data-driven pattern matching, at least for now.
Atlas Discovery's founders argue their models provide "interpretable gene attention," meaning the system highlights which genes or pathways drive predictions. If that holds up, it's valuable—both for generating biological hypotheses and for engaging skeptical pharma scientists who've seen too many black-box AI promises fail. But interpretability doesn't guarantee the predictions are correct. It just means you can see why the model thinks what it thinks.
The company has raised funding from Y Combinator, Pear VC, and Glasswing Ventures. The amount hasn't been disclosed publicly as of mid-2026. For a three-person team, the immediate challenge isn't market size or competitive positioning. It's building credibility through pilot studies with pharma partners willing to test these predictions in real trial designs, with real money at stake.
What Happens Next
The pharmaceutical industry is consolidating around fewer, larger AI partnerships rather than running dozens of disconnected pilots. Recursion's deals with Roche/Genentech, Owkin's agreements with Sanofi, Tempus's expansion with Bristol Myers Squibb—these are multi-year collaborations embedding AI into core R&D workflows. Pharma companies have largely concluded they can't build this capability in-house fast enough, or cheaply enough, to compete.
For Atlas Discovery, the path forward likely involves securing one or two lighthouse pharma partnerships—preferably in therapeutic areas where response heterogeneity is well-documented and the cost of failure is painfully high. Oncology is the obvious target. So are autoimmune diseases and neurology. The June 2026 blog post on ustekinumab response suggests they're already working in the autoimmune space, which makes sense. Crohn's disease and ulcerative colitis patients show wildly variable responses to biologics, making them ideal test cases.
Digital twins, external control arms, and AI-derived biomarker signatures are expected to proliferate in Phase II trials for enrichment and power calculations, according to regulatory and academic analyses published throughout 2026. The path to Phase III acceptance and full regulatory approval remains case-by-case, requiring transparent validation and continuous monitoring aligned with FDA-EMA principles.
Foundation models trained across modalities—genomics, pathology, electronic health record data, imaging—are increasingly being used to propose responder biomarker signatures, stabilize endpoints through AI-assisted scoring, and simulate trial arms. Recent literature emphasizes cross-site harmonization of model embeddings and domain-shift mitigation as key technical challenges that haven't been fully solved.
The question is whether Atlas Discovery's specific bet—foundation models trained explicitly for clinical response prediction—gives them a durable advantage, or whether larger, better-capitalized platforms with broader data moats will simply absorb the approach and dominate. At three people with undisclosed seed funding, they're early. Very early.
The industry's $244 million Phase III problem isn't going away. But solving it will take more than impressive AUROC scores on a blog post. It will require prospective validation, regulatory navigation, and the willingness of risk-averse pharma executives to bet real trial budgets on predictions from a machine.
That's a harder sell than the technology itself.
