AI Learns to Plan Trials from Pharma’s History, Outperforming General-Purpose Agents in Predicting Oncology Clinical Development

A new study attempts to answer one of the most expensive and difficult-to-standardize questions in drug development: Given efficacy signals, the competitive landscape, and limited resources, which clinical trials should be added to the portfolio next? Rather than asking general-purpose AI to reason on the spot, the researchers trained models to learn development strategies directly from the historical decisions and subsequent outcomes of dozens of oncology drug programs.

William Bolton and Philip Torr divided each drug program into six-month decision windows and asked the models to predict which trials the sponsor would initiate next based on the information publicly available at the time. The dataset covers 45 oncology drug programs and 881 decision segments, compiled from 31,700 public records, including trial registrations, regulatory reviews, company filings, drug-utilization data, and cancer epidemiology data. Each prediction also had to complete 14 fields, including indication, phase, trial strategy, comparator arm, region, and endpoint.

The best-performing model used “reward-weighted behavior cloning.” It did more than imitate decisions previously made by pharmaceutical companies: during training, it assigned greater weight to trials that later obtained FDA approval and generated revenue, while reducing the influence of failed or terminated trials. In a test set covering the period after August 2025, which in theory was less likely to be contaminated by existing training data, the model achieved an indication F1 score of 46.2% and a strict F1 score of 14.2%. The strongest general-purpose tool agent on each metric achieved only 25.0% and 2.1%, respectively. Strict F1 requires the indication, trial phase, and strategy to match simultaneously, making it a closer approximation of a complete development decision.

This work had previously been selected for a spotlight presentation and a poster at two ICML 2026 workshops, respectively. The version released in May used a dataset of the same size; at that time, the offline-trained model achieved an indication F1 score of 39.9% in the time-split test, compared with 11.2% for the tool agent. The August preprint incorporated more training methods and evaluations and reported the updated result of 46.2%. The study also found that providing the specialized model with time-bounded data retrieved by the tool agent increased indication F1 and strict F1 further to 59.0% and 18.6%, respectively, although this hybrid test was likewise based on a very small sample.

A critical boundary remains behind these numbers: The model learned “what sponsors historically went on to do,” not that any particular decision was necessarily better in causal terms. The quality of the drug itself, pricing, company resources, and undisclosed information can all affect approval and revenue. Moreover, the latest test set used to isolate the model from existing memorized information contained only 24 decision windows. Although 46.2% exceeded the comparison groups, it also means that more than half of the indication configurations did not match the historical answers, while the absolute strict F1 score was even lower.

In addition, public development records are skewed toward large pharmaceutical companies, common cancer types, and assets that successfully reached the market. This may lead the model to reproduce existing industry priorities while overlooking rare cancers, pediatric research, or research directions with weaker commercial appeal. The preprint is therefore better suited as a proof of concept for AI-assisted scenario analysis than as a tool for automatically determining trial, investment, or regulatory strategies. Before it can truly enter pharmaceutical decision-making workflows, it will still require prospective validation, bias audits, and joint oversight by clinical, statistical, and regulatory teams.