Abstract 3659: Biology-driven, machine learning-based development of a biomarker to predict response to WEE1 inhibitor Debio 0123

Abstract Introduction: The WEE1 kinase is a central regulator of G2/M and S phase cell cycle checkpoints which is potently activated in response to DNA damage. To date, a number of biomarker candidates have been proposed for WEE1 monotherapy sensitivity, including TP53 alterations, however, there is currently limited clinical evidence that biomarker-positive patient populations are significantly enriched in response to WEE1 inhibitors. Debio 0123 is a WEE1 inhibitor that has previously shown significant preclinical efficacy as a single agent and is currently being investigated in phase 1 clinical studies either as a monotherapy or in combination with different therapies. Herein we present a first-generation digital biomarker utilizing the ResponderID™ framework and the GenialisTM Supermodel to configure a model capable of predicting responders from non-responders to the Debio 0123 WEE1 inhibitor. Methods: The initial model is based on an extra-trees regression algorithm spanning five biological pathways related to DNA damage response as features. It was primarily trained on public and proprietary datasets to build a foundation DDR biology model, prior to further training on Debio 0123 patient-derived organoid RNA-seq datasets. The model was then tested on Debio 0123 datasets to measure model performance utilizing leave-one-out cross-validation. Results: The first-generation DDR biomarker showed potential to predict Debio 0123 IC50 values for viability (prediction accuracy of 0.7). Further refinement of optimal activity thresholds, tuning of features by tissue transferability, and establishment of a classifier-based model yielded an improved prediction accuracy (0.8 vs. null model accuracy of 0.5). Furthermore, this optimized model accurately predicted independent PDX model responses. Specifically, the model accurately predicted a non-responder (tumor growth inhibition <5%) in vivo and strikingly, an in vivo responder that was considered a non-responder by ex vivo organoid IC50. Conclusion: This biology-driven, machine learning-based classifier predicts response to Debio 0123 in patient derived organoid and xenograft models at an accuracy significantly greater than that of a dummy model and serves as a basis for further optimization and retrospective analyses of Debio 0123 clinical studies. Citation Format: Jeannette Fuchs, Luke Piggott, Kristian Urh, Matjaž Žganec, Eva Lavrenčič, Mark Uhlik, Carolina Haefliger. Biology-driven, machine learning-based development of a biomarker to predict response to WEE1 inhibitor Debio 0123 [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 3659.