AI-based predictive biomarker discovery via contrastive learning retrospectively improves clinical trial outcome

Modern clinical trials can capture tens of thousands of clinicogenomic measurements per individual. Employing manual approaches to discover predictive biomarkers, as differentiated from prognostic markers, is a challenging task. To address this challenge, we present an automated neural network framework based on contrastive learning, which we have named the predictive biomarker modeling framework (PBMF). This general-purpose framework explores potential predictive biomarkers in a systematic and unbiased manner, as demonstrated in simulated “ground truth” synthetic scenarios resembling clinical trials. Applied retrospectively to real clinicogenomic data sets, particularly in the complex field of immunooncology (IO) predictive biomarker discovery, our algorithm successfully found biomarkers that identify IO-treated individuals who survive longer than those treated with chemotherapy. In a retrospective analysis, we demonstrated how our framework could have contributed to a phase 3 clinical trial ([NCT02008227][1]) by uncovering a predictive biomarker based solely on early study data. Patients identified with this predictive biomarker had a 15% improvement in survival risk, as compared to those of the original trial. This improvement was achieved with a simple, interpretable decision tree generated via PBMF knowledge distillation. Our framework offers a rapid and robust approach to inform biomarker strategy, providing actionable outcomes for clinical decision-making. ### Competing Interest Statement G.A.-A., D.B., G.J.S., E.K., K.M.S., and E.J. are employees of AstraZeneca with stock ownership, interests, and/or options in the company. ### Funding Statement This study was funded by AstraZeneca ### Author Declarations I confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained. Yes The details of the IRB/oversight body that provided approval or exemption for the research described are given below: All datasets with human data are publicly available at the following URLs: (1) https://www.uniklinik-freiburg.de/imbi/stud-le/multivariable-model-building.html (2) https://www.nature.com/articles/s41591-018-0134-3 I confirm that all necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived, and that any patient/participant/sample identifiers included were not known to anyone (e.g., hospital staff, patients or participants themselves) outside the research group so cannot be used to identify individuals. Yes I understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance). Yes I have followed all appropriate research reporting guidelines, such as any relevant EQUATOR Network research reporting checklist(s) and other pertinent material, if applicable. Yes All human data used are available online at the following URLs: (1) https://www.uniklinik-freiburg.de/imbi/stud-le/multivariable-model-building.html (2) https://www.nature.com/articles/s41591-018-0134-3. Simulation data are available upon reasonable request to authors. [1]: /lookup/external-ref?link_type=CLINTRIALGOV&access_num=NCT02008227&atom=%2Fmedrxiv%2Fearly%2F2024%2F02%2F03%2F2024.01.31.24302104.atom