Somewhere inside Big Pharma's spreadsheets, there's a number that haunts every clinical program: $800,000. That's roughly what each day of drug development costs, according to an August 2024 white paper from the Tufts Center for the Study of Drug Development. The clock starts in Phase 1. It keeps ticking through Phase 2 and Phase 3. More often than anyone in the industry likes to admit, it ticks all the way to failure.
The success rates tell a sobering story. Data tracking clinical programs from 2011 through 2020—compiled by BIO, the industry's lobbying arm—show overall odds from Phase 1 to approval hovering in the single digits across many therapeutic areas. The math works out brutally: companies spend years and hundreds of millions testing compounds on patients who, by sheer statistics, won't benefit.
But something shifted in the period around 2025 and 2026. Call it a credibility threshold. A confluence of regulatory approvals, peer-reviewed breakthroughs, and actual pharma-scale deployments suggests that AI's ability to predict human drug response has moved from speculative promise to operational reality. It's infrastructure now, not hype.
Fewer Patients, Faster Answers
Here's the problem clinical trials can't escape: you need control groups. But recruiting them burns time and money, while exposing real people to placebos. Unlearn, a San Francisco outfit, figured out a workaround they call "digital twins"—machine learning models that forecast what would have happened to participants in a control arm, drawing on medical histories and baseline measurements.
It sounds like science fiction. Regulators, however, are buying in. The European Medicines Agency issued a formal Qualification Opinion for Unlearn's PROCOVA methodology back in September 2022. The FDA released draft guidance in January 2025 specifically addressing AI models supporting regulatory decision-making for drugs and biologics. That's not tentative interest—that's a green light.
The payoff, according to Unlearn's own claims: control-arm reductions hovering around 33 percent, with enrollment timelines shrinking by more than four months. In February 2026, the company announced it would deploy digital twins in the PIONEER-ALS Phase 1/2 study, a collaboration with VectorY. Unlearn had raised a $50 million Series C in February 2024, and has stacked partnerships across neurodegeneration—deals with remynd and Trace Neuroscience through 2025.
An academic paper published in May 2026 laid out the deployment mechanics: instead of recruiting 100 patients for a control arm, you recruit 70 and generate AI-predicted counterfactuals for the remainder. If the model holds up, you get the statistical power at lower cost and faster speed. The EMA qualification and FDA guidance shifted this from theory to practice.
Reading the Responders
While digital twins optimize trial logistics, another class of AI aims deeper—predicting which patients will respond before the trial even starts.
Early July 2026 brought a noteworthy data drop in Nature Medicine: results from a model called COMPASS, which predicted immunotherapy outcomes across 1,133 patients and seven cancer types using pretreatment RNA sequencing. In a held-out Phase 2 urothelial carcinoma cohort, COMPASS beat both PD-L1 immunohistochemistry and tumor mutational burden, delivering a hazard ratio of 4.7 with statistical significance.
That paper landed amid a wave of multi-omic response predictors. A Nature Cancer study from September 2024 demonstrated deep-learning frameworks imputing transcriptomics from histopathology to forecast drug response. A June 2024 Nature Communications paper explored single-cell resistance predictions. An October 2025 article in npj Genomic Medicine tackled polygenic risk score transfer learning for pharmacogenomics. The momentum is unmistakable.
Atlas Discovery, which emerged from Y Combinator's Summer 2026 batch, positions itself squarely in this foundation-model lineage. A June 2026 blog post from the company described a case study on the UNIFI trial—ustekinumab in ulcerative colitis—claiming it predicted response from baseline biopsies with an area under the curve of 0.76, using pretrained embeddings and a supervised learning head.
The pitch: train foundation models on millions of preclinical and clinical samples to amortize biological understanding, then fine-tune trial-specific models to separate responders from non-responders. Done right, you shrink sample sizes. Atlas lists Y Combinator, Pear, and Glasswing among its backers, though specific funding amounts hadn't been disclosed publicly as of mid-2026.
Toxicity Before Trials
Predicting toxicity before human trials opens yet another front.
Emulate's human Liver-Chip—peer-reviewed in Communications Medicine in late 2022, with follow-ups through 2024—achieved 87 percent sensitivity and 100 percent specificity for drug-induced liver injury across a blinded set of 27 drugs. The company's economic model pegged potential productivity gains at roughly $3 billion annually if broadly adopted in small-molecule pipelines. On September 24, 2024, the FDA's ISTAND pilot program accepted the first organ-on-chip DILI (drug-induced liver injury) letter of intent, carving out a formal pathway for microphysiological systems to inform regulatory decisions.
Quris-AI pairs miniature multi-organ systems with AI to predict both safety and efficacy—what the company calls "clinical-trial-on-a-chip." On January 7, 2025, Merck KGaA (the Darmstadt, Germany operation, not the New Jersey one) announced it had integrated Quris' Bio-AI platform for preclinical small-molecule safety evaluation. Quris raised approximately $37 million in cumulative seed financing as of early 2023, per earlier disclosures. The aim: reduce reliance on animal models that routinely fail to predict human outcomes.
Broader institutional support is building. The IQ MPS Affiliate within the IQ Consortium is advancing pharma-aligned standards. NIST and Advanced Biology published fit-for-purpose and standardization recommendations through 2024 and 2025. The FDA's NAMs (New Alternative Methods) Roadmap, updated through 2026, outlines qualification steps for microphysiological systems and other human-relevant models—backstopped by the FDA Modernization Act 2.0, signed in December 2022, which authorizes non-animal alternatives in nonclinical testing.
Virtual Cohorts, Real Decisions
QuantHealth and Nova In Silico—operating the jinkō platform—simulate entire virtual patient populations to optimize trial design before the first real patient enrolls.
QuantHealth claims it has run more than 600 simulations with accuracy reaching 90 percent across roughly 30 diseases, according to press statements from mid-2026. In October 2025, Sanofi Ventures made a strategic investment in the company. Nova In Silico announced an AI-powered simulation collaboration with Fujitsu in June 2025 and has pointed to retrospective validations of trials like FLAURA2 in materials from 2024 and 2025.
These are vendor-reported figures. Methodologies and independent validation vary. But pharma is voting with contracts.
Aitia, which combines causal inference with multi-omic patient data in what it calls "Gemini Digital Twins," announced oncology collaborations with Servier in October 2024 and Gustave Roussy in June 2025. The positioning isn't about replacing real trials—it's about pre-screening designs to avoid expensive mid-study pivots.
PrecisionLife and Ovation announced a partnership in February 2026 to identify genetic biomarkers predicting GLP-1 response. Scipher Medicine's PrismRA test, which predicts inadequate TNF inhibitor response in rheumatoid arthritis, received Medicare LCD coverage in September 2023—proof that payers will reimburse validated predictive diagnostics. The EXCYTE-2 trial (NCT06648512), last updated in August 2025, is using an AI-based precision medicine platform to link ex vivo drug response with AML outcomes—response prediction moving into prospective trials rather than retrospective analysis.
The Big Money Moving In
Owkin, a New York–based AI biotech, announced a multi-year collaboration with Sanofi on June 5, 2026, to co-develop "AI agents" for biopharma and license Owkin's K Pro platform. This builds on Sanofi's $180 million equity investment in Owkin from November 2021, plus continued expansions through immunology and oncology programs in 2024 and 2025.
VeriSIM Life formalized a research collaboration with the FDA's National Center for Toxicological Research on June 29, 2026, to advance its mechanistic-AI BIOiSIM platform and a "Translational Index" for human outcomes prediction. When a startup gets the FDA to co-develop methodology, that's not window dressing.
McKinsey weighed in on January 9, 2025, pegging generative AI's value across discovery-to-clinical operations in the tens of billions annually and calling AI in clinical development a board-level priority. IQVIA's Global R&D Trends 2026, published in April, flagged what it termed a "credible signal" on AI-enabled programs—and noted that macro policy headwinds like the Inflation Reduction Act's price negotiations and pharmacy benefit manager reforms are raising the premium on early derisking.
An ASCO abstract presented in June 2026 cataloged 117 AI-enabled therapeutic assets entering human trials by July 1, 2025, establishing a baseline cohort for outcomes tracking. ACRO's March 2026 State of the Industry survey showed broadening AI use across contract research organizations' activity in 2025. The tools are moving from pilot programs to production pipelines.
Regulators Set the Rules
Early 2026 brought something potentially more significant than any single technology: the EMA and FDA jointly announced common principles for AI in medicine development, signaling transatlantic alignment on transparency, data quality, bias mitigation, and validation.
The EMA had finalized its AI reflection paper on September 30, 2024, laying out expectations across the product lifecycle. The FDA's January 6, 2025, draft guidance on AI supporting regulatory decisions noted that more than 500 submissions with AI components had been filed since 2016 and provided a framework for credibility assessment.
These aren't blanket approvals, to be clear. Regulators are qualifying methods case by case—Unlearn's PROCOVA for prognostic covariate adjustment, Emulate's Liver-Chip DILI methodology via ISTAND. But the existence of formal qualification pathways means companies can invest in AI tools with reasonable confidence regulators won't reject them on principle. That's no small shift.
Stanford's AI Index 2026, released in April, reported a rapid rise in medical digital twin research and RCT-backed AI deployments. Perhaps more telling: ethics content in medical AI publications more than doubled in 2025. The technology is maturing past the "can we?" phase into "how should we?"
What Comes Next
The next 18 months will likely see wider adoption of digital twins and external comparators in early-phase and rare-disease trials, where control-arm recruitment is especially painful. Foundation-model approaches trained on multi-omics plus clinical context should expand beyond immunotherapy—where published evidence is strongest—into metabolic and neurological indications.
Microphysiological systems submissions via ISTAND and similar qualification tracks will grow as the FDA's NAMs roadmap milestones roll forward through 2026. More payer-recognized predictive diagnostics, building on precedents like Scipher's PrismRA Medicare coverage, could emerge as companion diagnostics for enrichment strategies, tightening the link between AI-predicted responders and trial eligibility.
Some consolidation among vendors seems inevitable. Pharma will standardize on a smaller set of validated platforms, and EMA/FDA guidance will likely evolve from high-level principles into method-specific requirements.
The 2011–2020 clinical success data that motivated much of this work is aging. A fresh multi-year analysis—especially one parsing AI-enabled versus traditional programs—would clarify whether these tools are genuinely shifting the odds or just optimizing around the margins. The ASCO cohort of 117 AI-designed assets entering trials by mid-2025 offers a natural dataset for longitudinal tracking.
In the meantime, the industry is moving. Every trial accelerated by a digital twin, every patient spared a non-responder arm by an omics predictor, every toxic compound caught by an organ chip before Phase 1. These aren't moonshots anymore. They're line items in 2026 R&D budgets, backed by regulatory frameworks and pharma partnerships.
The race isn't to prove AI can predict drug response anymore. It's to scale the systems that already do—and to figure out, perhaps more urgently than the industry expected, what happens when the math on that $800,000 daily burn rate finally starts to change.
