inClinico Agent for clinical trial endpoint prediction and prospective evaluation

Chen, Duan, Gou and others proposed a machine-learning system for predicting whether Phase I to Phase IV clinical trials will hit their primary endpoints, modeling efficacy endpoints separately from safety/toxicity endpoints. The authors say the models were trained only on features available at registration, including target and mechanism-of-action precedents, drug modality and novelty, trial design and operational features, and frozen literature-derived target-selection embeddings, to avoid look-ahead bias and outcome leakage. Each prediction outputs a continuous Pre-Trial Index (PTI) between 0 and 1. On a test set of trials launched in 2021 and later, the efficacy model had an AUROC of 0.828 (95% CI 0.812-0.845; n=2,347), and the safety/toxicity model 0.809 (95% CI 0.787-0.831; n=1,485). The authors say the models have been embedded in the inClinico Agent, and they launched ClinBench, a prospective benchmark covering 3,311 ongoing trials.