Abstract 3660: Machine learning-based combination prediction for Wee1 inhibitor

Abstract Motivation: Wee1 is the gatekeeper gene that allows proper DNA damage repair (DDR) during the G2/M cell cycle transition. Inhibiting the Wee1 gene will abrogate the DDR mechanism and lead cancer cells with high replication stress to premature catastrophic mitosis and eventually apoptosis. Several Wee1 inhibitors (Wee1i) are under active development in both preclinical and clinical studies. They have shown efficacy in multiple cancer types as monotherapy or as a combination with chemotherapies or targeted therapies. These studies suggest that a broader combination potential for Wee1i should be explored. Machine learning-based methods are promising and efficient approaches to model drug combination in a broader combination space. Here we have integrated two large combination studies and applied a tensor reconstruction based polynomial regression (comboLTR) method to predict the combination effects for Wee1i in expanded combinations. This work supports a machine learning approach in generating novel combination hypotheses in an expanded space from existing experimental data. Method: Two large published screening studies containing the Wee1i were normalized and integrated. This resulted in a combined screening study of 131 drugs and 53 cell line models. ComboLTR was applied to train drug combination response prediction models using collected known drug combination responses, drug molecular fingerprints, and model genomic features. The model performance was evaluated under different prediction scenarios, including the prediction of missing entries in a dose-response matrix and the prediction of entire dose-response matrices in new cell lines or with unknown drugs. The hyper-parameters were tuned separately for different prediction scenarios to reach the best performance. Three different synergy scores (HSA, Bliss, Loewe) were used to summarize the synergy effects of the predicted drug combinations. The average of the synergy scores was used to evaluate the combination partners for Wee1i. Results: We integrated drug combination response data from two publications and trained LTR models based on the data. In 5-fold cross-validations for the different prediction scenarios, new entry, new matrix, new drug combo, the model achieved a Pearson correlation of above 0.8 between predicted and measured response data. The final model was trained to predict responses of drug combinations between Wee1i and 130 other drugs in 53 cell lines. Based on these predictions, we calculated synergy scores for each drug combination and cell line triplet and ranked the synergy effect within each cancer type. Several compounds reported to be synergistic with Wee1i in selected cancer types were reproduced in the prediction, including dasatinib in ovarian cancer, SN38 in breast cancer, and CHK1i in ovarian cancer. The model also predicted these compounds as having synergistic effects with Wee1i in potential new cancer types which may be of future interest. Citation Format: Tianduanyi Wang, Juho Rousu, Lin Tang. Machine learning-based combination prediction for Wee1 inhibitor [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 3660.