AI Model Predicts Clinical Trial Success From Registration-Time Features
High attrition remains one of the most expensive failures in drug development, and this preprint tackles the problem at the point where many programs first become vulnerable: trial registration. inClinico Agent is built to estimate whether a study will reach its primary endpoint before outcome data exist, using only information available at registration. The design choice matters because endpoint prediction is often distorted by leakage from post hoc data, protocol amendments, or features that only become obvious after a trial has already started to unfold.
The system separates trials by primary endpoint type, building one model for efficacy and another for safety or toxicity. That split is clinically sensible. Efficacy trials and safety-driven studies fail for different reasons, and a shared predictor can blur the biological and operational signals that distinguish a promising development program from one drifting toward attrition. Each trial receives a continuous Predictive Trial Index, or PTI, scored from 0 to 1 as a probability-like estimate of success.
The most consequential piece is not the machine learning architecture alone, but the way the models were constrained. Training relied on registration-time variables such as target and mechanism-of-action precedent, drug modality and novelty, trial design, operational characteristics, and frozen literature-derived target-choice embeddings. Those inputs are the kinds of details trial teams actually know when a protocol is finalized, which makes the approach more realistic than systems that implicitly borrow from future information. In a field where retrospective prediction can look impressive while failing prospectively, the avoidance of lookahead bias is the central technical safeguard.
Prospective Performance on Later-Starting Trials
On a test set of trials initiated in 2021 or later, the efficacy model achieved an AUROC of 0.828 with a 95% confidence interval of 0.812 to 0.845 across 2,347 trials. The safety and toxicity model reached an AUROC of 0.809, with a 95% confidence interval of 0.787 to 0.831 across 1,485 trials. Those are strong discrimination metrics for a domain as noisy as drug development, where heterogeneous sponsors, endpoints, indications, and trial designs can make prediction unusually difficult.
The separation between efficacy and safety prediction is especially meaningful because the reasons a study fails are often different from the reasons a drug fails in the clinic. A therapy can look mechanistically compelling and still falter because dose selection is wrong, the operational design is weak, or the target lacks enough human precedent to support the hypothesis. Conversely, a study may be biologically promising but collapse under toxicity, narrow tolerability, or safety monitoring constraints that were underestimated during development planning.
Clinical trial attrition is not only a discovery problem; it is a capital allocation problem.
In practical terms, a model that can rank programs by projected success probability before enrollment begins could influence portfolio triage, protocol refinement, and monitoring intensity. A low PTI would not prove a trial will fail, but it may flag combinations of target novelty, modality risk, and design complexity that deserve extra scrutiny before scarce resources are committed. A high PTI, by contrast, could support prioritization of programs with stronger precedential support and more robust trial architecture.
Why Registration-Time Features May Be More Useful Than Late-Stage Signals
The model’s feature set reflects a straightforward clinical logic: some trials are difficult from the beginning. Target precedent, mechanism-of-action history, and modality are proxies for how much human evidence already exists around the biological hypothesis. Design and operational characteristics capture whether the protocol is likely to generate interpretable results, while literature-derived embeddings attempt to encode the accumulated context around target choice in a way that can be used computationally.
This is where predictive biology and operational science meet. A first-in-class mechanism may offer scientific upside, but it also carries a higher burden of uncertainty, especially when the target has little clinical precedent. A more established mechanism may travel with lower uncertainty, though not necessarily with better efficacy. The model appears to treat those tradeoffs as structured inputs rather than noise, which is exactly what a registration-time prediction engine should do.
The authors also embed the models inside inClinico Agent, described as an agentic AI system combining frontier large language models with deterministic scoring engines, retrieval tools, and specialized clinical skills. That hybrid design is important because pure language models are not enough for regulated clinical forecasting. Deterministic components can stabilize output, retrieval can anchor the system to source material, and specialized skills can separate mathematical scoring from open-ended text generation.
ClinBench Adds a Leakage-Free Prospective Test
The paper goes beyond static model evaluation by introducing ClinBench, a prospective benchmark of 3,311 ongoing trials scored by inClinico Agent and five frontier LLMs. The purpose is not merely to compare models, but to create a leakage-free framework for watching predictive accuracy as trials mature and eventually complete. Prospective benchmarking is one of the few credible ways to determine whether a system that looks good on historical data can survive contact with live development pipelines.
The benchmark also changed during revision, with ClinBench expanding from 2,040 trials to 3,311 trials. That larger benchmark matters because it widens the prospective sample space and should reduce the chance that performance estimates depend on a narrow subset of trial types. A system tested only on old or overly curated cohorts may look better than it would in daily use, especially when confronted with modern trial designs, newer modalities, or changing sponsor behavior.
For drug developers, the immediate question is not whether an AI score is elegant. It is whether the score can help identify which programs deserve more caution, more resources, or a second look at the protocol before the first participant is enrolled. A model trained only on registration-time information cannot resolve all uncertainty, but it can formalize risk in a way that is auditable and repeatable.
Limits of a Preprint and the Remaining Validation Gap
This remains a preprint and has not undergone peer review, so the findings should be treated as preliminary. The work is also developed entirely within a single commercial AI organization, and all authors are affiliated with Insilico Medicine. That structure does not invalidate the results, but it does place added weight on independent replication, external benchmarking, and transparent comparison against nonproprietary baselines.
The database used to train the models is not publicly available, which limits outside inspection of feature construction, curation decisions, and class balance across the full training set. The prospective ClinBench predictions are available in supplementary material, but the core training corpus remains closed. Without broader external validation, the field still does not know how the model will perform across therapeutic areas, sponsor types, and regions that may differ from the development set.
The central unanswered issue is whether prospective discrimination will hold as more trials complete and as the benchmark continues to expand. If it does, inClinico Agent would represent a useful decision-support layer for development teams seeking to reduce wasted trial starts and better allocate scarce capital. If it does not, the result would still be informative: it would show how far registration-time data can go before the hidden complexity of clinical development overwhelms a predictive score.