A Preregistered Multi-Cohort Evaluation of the FATHOM AI System for Molecular Testing Prioritization to Support Clinical Trial Enrollment

Overview

Many clinical trials evaluating cancer treatments require patients to undergo testing for specific molecular markers as part of eligibility screening, typically using immunohistochemistry or sequencing. Because relatively few patients may carry a required marker, trial investigators often test large numbers of patients to identify the few who may ultimately qualify for enrollment. Pathology laboratories routinely produce hematoxylin-and-eosin (H\&E) slides during cancer diagnosis. Pathology foundation models-large neural networks pretrained on millions of histology images-have shown promise in predicting molecular characteristics from these slides. Researchers can use these models to build classifiers that predict specific molecular markers and prioritize patients for confirmatory testing. This study evaluates FATHOM (Facilitating Accrual through Tumor Histology and Omics Matching), an autonomous research system powered by large multimodal models. Its agents read registered clinical trial records, identify molecular markers used as enrollment criteria, build prediction models using pathology foundation models, select the individual models or model combinations that best meet prespecified criteria, set their decision thresholds, and determine whether to deploy them. Together, a prediction model, its decision threshold, and the decision to deploy it constitute an AI prediction policy. Before FATHOM runs, the investigators preregister the clinical trial records that its agents may read, the cutoff date that defines which trial information they may use, the rules governing the agents, and the analysis plan. The system timestamps and locks each policy immediately after an agent produces it. The investigators then apply the policies to archived patient slides and compare their predictions with existing molecular marker results. The primary outcome is the proportion of prespecified evaluation scenarios in which an agent-generated policy, compared with universal molecular testing, either enriches the population selected for confirmatory testing with marker-positive patients or safely spares patients from confirmatory testing while meeting prespecified performance criteria. This study analyzes existing pathology images and clinical trial records only. It does not enroll or contact patients, influence patient care, or affect participation in any clinical trial.

WHAT IS REGISTERED. This study evaluates screening policies generated by autonomous research agents using archived pathology slides and existing molecular marker profiles. The study does not prospectively enroll or contact patients. The investigators register the clinical trial corpus that the agents may read, the trial-record cutoff date, the information that the agents may access, the rules governing their work, and the methods used to score their policies. THE AGENTS. Each agent may access the registered clinical trial corpus, published literature, out-of-fold performance estimates for its own classifiers, and any development data identified in its manifest. The agent receives no results from any sealed evaluation cohort. For visual recognition, each agent uses pathology AI models as feature extractors and trains classifiers on the extracted features. The agent determines which molecular markers to model, which model to use, how to set each operating threshold, and whether to deploy the resulting policy. The agent records each decision and its rationale. THE MANIFEST. Each agent run produces a timestamped manifest listing every policy generated and each policy's final deployment decision. On the study start date, the investigators designate the autonomous-agent approach and its comparators for the primary evaluation. TRIAL DEMAND CUTOFF. Trials first posted before January 1, 2026, define retrospective trial demand. Trials first posted on or after January 1, 2026, are used for the temporal generalization evaluation. SEALED ANALYSIS RULE. Each policy result corresponds to an evaluation scenario, defined as one cohort paired with one molecular marker. For each scenario, the study logs and publishes the date on which investigators first compare any model output with the ground-truth marker result. COHORTS. The study uses archived institutional and consortium cohorts containing routine diagnostic H\&E slides linked to molecular profiles. No clinician uses model output from this study to make patient-care decisions.