A model trained on the history of oncology programs predicted which trials companies would launch next more accurately than language AI systems

A model trained on the history of oncology programs predicted which trials companies would launch next more accurately than language AI systems

On August 4, a preprint described offline learning, in which a model learns from decisions that have already been made. The authors tested whether a model could use the information available at each point in the past to predict the set of trials that an oncology company would launch over the next six months.

Clinical development begins long before the first patient enters a trial. A team must choose the disease, the phase, the comparator treatment, and the study design. Each decision shapes the years that follow, including participant recruitment, trial conduct, and the wait for results.

The authors note that models for clinical trials usually evaluate studies that have already been designed or help recruit patients for them. In the preprint, they proposed training a model to address an earlier decision:

“We study an earlier strategic question: given the state of a program and its external environment at time t, which trials should be launched next?”

The authors compiled 31,7 thousand public records for training, including trial registries, regulatory reviews, sponsor reports, drug utilization data, and epidemiological data. These records yielded 881 six month episodes across 45 programs. For each starting date, the model proposes the next set of studies. The researchers then compare its proposal with the portfolio that the company actually launched during the following six months. The score measures how closely the prediction matches that historical choice.

During fine tuning, the authors assigned more weight to episodes in which a trial later supported FDA approval or the program achieved higher annual revenue. They reduced the weight of episodes involving failed trials. The model therefore learned more often from trajectories associated with these historical outcomes.

Across 24 historical timepoints corresponding to decisions made after August 2025, this version achieved 46,2% agreement on the indication, meaning the disease for which the trial was launched, compared with 25,0% for the best tool equipped language agent. Under a stricter comparison that considered the indication, phase, and strategy together, the scores were 14,2% and 2,1%. Every system had access only to information that was available on the relevant historical date.

The test measures how well the model can reproduce strategies that companies chose in the past. The choice of the next trial determines which biomedical hypotheses proceed to testing in humans.