AI clinical trials drug development has moved well past pilot projects into everyday use for patient selection, adaptive design, and trial monitoring. Sponsors now apply machine learning across enrollment, biomarker analysis, and interim decision-making, but whether AI is improving trial success rates, rather than merely trial efficiency, remains a harder question worth answering honestly.
Key takeaways
- Roughly 90 percent of drug candidates that enter clinical trials still fail, and AI has not yet demonstrably closed that gap.
- Machine learning models can match patients to trials with accuracy approaching expert-level performance while cutting screening time substantially.
- Biomarker-driven stratification remains one of the most direct levers for improving trial enrollment quality, not just trial speed.
- Adaptive and Bayesian trial designs, now the subject of the FDA's formal guidance, let sponsors adjust trials using accumulating data.
- Real-world data from electronic health records (EHRs) and registries increasingly informs trial design and post-market safety monitoring.
Why trials fail: The translational gap
Most clinical trial failures trace back to a translational gap between how a candidate performs in preclinical models and how it performs in human patients. A 2025 analysis of more than 20,000 clinical development programs found that clinical trial success rates declined for roughly 2 decades before recently plateauing and ticking upward again, yet even that recent uptick leaves overall success in the single digits to low double digits, nowhere close to the gains AI has delivered in discovery-stage speed.
That persistence matters because AI clinical trials drug development has so far concentrated its gains on efficiency: faster patient identification, faster site selection, faster data cleaning, rather than on the underlying biological uncertainty that drives Phase 2 and Phase 3 attrition. Understanding why the translational gap in AI drug discovery persists, and specifically why preclinical predictions fail in humans even when a model performs well on training data, is a prerequisite for evaluating what AI can and cannot do once a candidate reaches the clinic.
Most attrition happens in the transition from Phase 1, which primarily tests safety, into Phase 2, which tests whether a mechanism that looked sound preclinically actually produces a clinical benefit. AI models trained on preclinical or early biomarker data inherit the same species differences, data gaps, and correlation-versus-causation limitations that have long constrained traditional drug development, which is why faster preclinical screening has not, by itself, translated into a materially higher approval rate.
This distinction between operational speed and biological success runs through every section below. The sections that follow examine where AI is genuinely changing clinical development, starting with patient selection, and where the translational gap still limits how much that progress can move the needle on overall trial success, all within the broader arc of AI-enabled drug development, spanning target identification through clinical translation.
AI in patient stratification
AI-based patient stratification improves trial enrollment quality by matching candidates to eligibility criteria and likely biological response with a precision manual chart review cannot match at scale. A large language model system evaluated for zero-shot patient-to-trial matching achieved a patient-matching accuracy of 87.3 percent against expert-level judgments while cutting screening time by 42.6 percent, across more than 1,000 patient-criterion pairs.
That kind of gain is meaningful because manual eligibility screening has long been one of the slowest, most labor-intensive steps in trial startup, particularly for complex oncology and rare disease protocols with dozens of inclusion and exclusion criteria. Natural language processing models trained to parse unstructured clinical notes can surface eligible patients that keyword-based searches miss entirely, expanding the effective pool of candidates without loosening protocol criteria.
Faster, more accurate matching does not, on its own, guarantee a successful trial. It shortens enrollment and reduces screen-fail rates, but the deeper question of whether the enrolled population will actually respond to the intervention depends on the biological hypothesis behind patient selection, which is where biomarker-driven patient stratification becomes the more consequential lever.
Site-level adoption of these tools also varies considerably. Academic medical centers with structured research databases and dedicated clinical trial informatics teams tend to see the largest gains from AI-based matching, while community sites with less digitized records often still depend on manual chart review, a gap that shapes how evenly these efficiency improvements reach a trial's enrolling sites.
Biomarker-driven trial enrichment
Biomarker-driven enrichment improves trial success rates by ensuring enrolled patients are biologically positioned to respond to the mechanism being tested, rather than simply meeting broad diagnostic criteria. An analysis of more than 400,000 trial records found that programs using a biomarker-based patient-selection strategy succeeded roughly twice as often as those without one, and nearly seven times as often in oncology specifically.
Machine learning extends that enrichment strategy by combining multiple biomarker signals (genomic, proteomic, and imaging-derived) into a single predictive model of likely response, rather than relying on a single companion diagnostic cutoff. That multi-modal approach is especially valuable in heterogeneous diseases where no single biomarker reliably predicts treatment benefit across the full patient population.
The tradeoff is that enrichment strategies narrow the eligible population, which can slow enrollment even as it improves the odds of a positive trial outcome. Sponsors increasingly manage that tension through adaptive designs that revisit enrollment criteria as biomarker data accumulate during the study itself.
Companion diagnostics remain the most regulated form of biomarker-driven enrichment, since a diagnostic used to select patients for a specific therapy typically requires its own authorization from the FDA alongside the drug it supports. Multi-biomarker AI models complicate that regulatory picture because they often combine several data types into a single composite score rather than a single binary diagnostic result, which is prompting regulators to reconsider how a companion diagnostic pathway should apply to model-based enrichment strategies. Biomarker-driven patient stratification is where this tension is most visible in current trial practice.
Where AI-enabled enrichment differs from conventional patient selection:
- Conventional selection relies on a small number of pre-specified diagnostic criteria fixed before the trial begins.
- AI-enabled enrichment can combine dozens of biomarker and clinical variables into a single, continuously updated response-probability score.
- Conventional selection treats all eligible patients as clinically interchangeable within a diagnostic category.
- AI-enabled enrichment explicitly models heterogeneity within a diagnostic category, prioritizing patients most likely to benefit from the specific mechanism under study.
Adaptive trial design and Bayesian methods
Adaptive trial design allows sponsors to modify a study, including dose selection, enrollment criteria, or sample size, based on accumulating data rather than fixing every parameter before the trial starts. The FDA has formalized this shift with draft guidance on Bayesian methodology for clinical trials of drug and biological products, which lays out expectations for pre-specifying success criteria, evaluating operating characteristics through simulation, and justifying the prior distributions used in an analysis.
Bayesian methods are a natural complement to adaptive designs because they update the probability of a trial outcome as new data arrive, rather than waiting for a single fixed endpoint analysis. Machine learning contributes to this process by generating the predictive models that inform interim decisions, such as which dose arm to carry forward or which biomarker-defined subgroup shows the strongest signal.
The following framework outlines how sponsors typically evaluate whether an adaptive, AI-informed design is appropriate for a given program:
- Confirm that the trial has a clear pre-specified decision rule for each planned adaptation before data collection begins.
- Simulate the design's operating characteristics, including type I error and power, under a range of plausible scenarios.
- Document the source and justification of any prior distribution or external data used to inform the analysis.
- Establish a data monitoring process capable of executing interim analyses without unblinding bias.
- Prepare documentation sufficient for regulatory reproducibility of the Bayesian or adaptive analysis.
Adaptive platform trials, which test multiple candidate therapies against a shared control arm and can add or drop treatment arms over time, illustrate how far this approach has moved from a purely statistical exercise into an operational one. Running a platform trial well requires infrastructure for continuous data cleaning, real-time model updating, and cross-functional coordination between statisticians, clinical operations, and regulatory affairs, capabilities that increasingly depend on the same machine learning pipelines used for patient stratification and biomarker analysis elsewhere in the trial.
Real-world data in AI trial design
Real-world data, including EHRs, claims data, and disease registries, increasingly informs trial design decisions that once relied solely on prospective data collection. The FDA has emphasized in its guidance on real-world evidence that the strength of any real-world evidence submission depends heavily on whether the underlying data are fit for the specific regulatory question at hand.
AI methods make that real-world data usable at scale by extracting structured variables from unstructured clinical notes, standardizing coding across institutions, and identifying eligible patient populations for external control arms or post-market safety monitoring. That capability is central to how real-world evidence and AI are reshaping decisions that once depended entirely on prospective, randomized data collection.
Real-world data still carries limitations that AI cannot fully resolve on its own: it is often incomplete, inconsistently coded across health systems, and biased toward populations with better healthcare access. Sponsors that treat real-world evidence as a complement to, rather than a replacement for, randomized data are best positioned to use it credibly in regulatory submissions.
External control arms built from real-world data are one of the more consequential applications of this shift, particularly for rare diseases and oncology indications where randomizing patients to a placebo or standard-of-care arm raises ethical or feasibility concerns. Constructing a credible external control arm requires matching real-world patients to trial-eligible patients on the same inclusion and exclusion criteria, adjusting for confounding differences in disease severity or prior treatment, and documenting that process transparently enough for a regulator to assess whether the comparison is scientifically sound.
How the FDA and the EMA regulate AI-informed trials
Regulators are actively building frameworks to evaluate AI-informed trial designs and AI-generated evidence rather than treating them as a special exception to existing rules. The FDA has issued draft guidance on AI model credibility, a risk-based framework used to support regulatory decisions, tying model credibility to a defined context of use rather than a generic validation checklist.
| Regulatory tool | What it covers | Why it matters for AI-informed trials |
|---|---|---|
| Bayesian methodology guidance | Pre-specification, prior justification, and operating characteristics for Bayesian analyses | Formalizes how AI-informed adaptive decisions can support primary inference |
| AI credibility framework | Risk-based evaluation of AI model credibility for regulatory decision-making | Ties AI model trust to a specific context of use rather than a blanket standard |
| Real-world evidence guidance | Study design expectations for non-interventional studies | Governs how AI-extracted real-world data can support effectiveness or safety claims |
These frameworks matter because sponsors increasingly need to justify not just a trial's statistical design but the AI methods feeding into that design, from the patient-matching algorithm to the biomarker model to the real-world data pipeline. Programs that document each of these components with the same rigor as a traditional statistical analysis plan are better positioned for regulatory review, whether the question at hand involves adaptive dosing or AI-powered drug repurposing built on an existing safety record.
International alignment on these questions is still emerging. The International Council for Harmonisation (ICH) has begun discussing how existing multiregional trial guidance intersects with AI-informed design elements, though region-specific frameworks, such as those developing separately at the FDA and the European Medicines Agency (EMA), remain the primary reference points sponsors use today. That divergence means a global development program often needs to satisfy the more conservative of the applicable regional standards for any AI-informed design element.
Where AI clinical trials drug development goes from here
AI clinical trials drug development has genuinely improved how sponsors find patients, design adaptive studies, and use real-world data, but it has not yet demonstrably moved the overall clinical success rate. That distinction between operational efficiency and biological success is the honest throughline connecting patient stratification, adaptive design, and the translational gap.
The programs most likely to benefit are those that pair AI-driven efficiency gains with rigorous biomarker hypotheses, transparent adaptive design documentation, and real-world data used as a complement to, not a substitute for, randomized evidence. As regulatory frameworks for Bayesian methods, AI credibility, and real-world evidence continue to mature, AI clinical trials drug development is likely to become a documented, auditable part of standard trial design rather than a novel add-on.
For translational scientists and R&D directors evaluating where to invest, the practical priority is documentation discipline: capturing how sponsors validated a patient-matching model, how they derived a biomarker score, and how they pre-specified an adaptive decision rule, in a form that can withstand regulatory scrutiny. That discipline, more than any single algorithmic advance, is what will determine whether AI's efficiency gains in clinical trials eventually translate into a measurable improvement in overall success rates.
This article was produced under Drug Discovery News' AI Editorial Guidelines.
