A Large Molecular Model (LMM)-Based Predictor of Clinical Response to the WEE1 Inhibitor Debio 0123 + Carboplatin Therapy

E-Poster Presented at ESMO AI

Jeannette Fuchs¹, Luke Piggott¹, Christophe Mas¹, Esteban Rodrigo Imedio¹, Kristian Urh³, Marcel Levstek³, Matjaž Žganec², Eva Lavrencic Pavlic³, Mark Uhlik², Anna Pokorska-Bocci¹

1 Debiopharm International S.A., Lausanne, Switzerland
2 Genialis Inc., Boston MA, United States
3 Genialis d.o.o., Ljubljana, Slovenia

Summary

Debio 0123 is an investigational, orally available, highly selective, and brain-penetrant adenosine triphosphate (ATP)-competitive inhibitor of the WEE1 tyrosine kinase, currently in phase I/II clinical trials either as a monotherapy or in combination with various therapies. Inhibition of WEE1 presents an opportunity as a therapeutic target in cancer therapy, either in cells relying on cell cycle checkpoints regulated by WEE1 or to potentiate DNA damaging agents.

We previously described a first-generation digital biomarker that accurately predicted response to Debio 0123 in both patient-derived organoid and in vivo xenograft models. This biology-driven, machine learning–based classifier outperformed the baseline model, underscoring its potential for clinical application2. Building on that foundation, we now present a second-generation, clinically relevant, biology-informed machine learning predictor of response to Debio 0123 and carboplatin (CB) combination therapy. Developed using the Genialis ResponderID™ and Supermodel platforms, this model was built upon clinical data from patients enrolled in the Debio 0123-101 clinical trial (NCT03968653). Using a logistic regression model with ElasticNet regularization that was trained on diverse and biologically relevant features (biomodules) and their interactions, our predictor has shown excellent performance on the 24-week patient response in the Debio 0123-101 cohort, with robust AUROC (area under the receiver operating characteristic curve) (0.95), accuracy (0.80), and effective separation of patients with and without treatment benefit. This second generation model re-captures the same biological pathways as previously identified in preclinical models and uncovered additional biologies predictive for patient response.

These findings highlight the potential of a machine learning-driven approach to refine patient selection for WEE1 inhibitor therapies, providing a strong foundation for further clinical validation of Debio 0123.