A systematic evaluation of deep learning methods for the prediction of drug synergy in cancer

• DOI: 10.1371/journal.pcbi.1010200
• Publication Date: 2023-03-23T18:26:28Z
• Authors: Delora Baptista; Pedro G. Ferreira; Miguel Rocha
• Keywords: Matthews correlation coefficient; Feature (linguistics)

One of the main obstacles to successful treatment cancer is phenomenon drug resistance. A common strategy overcome resistance use combination therapies. However, space possibilities huge and efficient search strategies are required. Machine Learning (ML) can be a useful tool for discovery novel, clinically relevant anti-cancer combinations. In particular, deep learning (DL) has become popular choice modeling effects. Here, we set out examine impact different methodological choices on performance multimodal DL-based synergy prediction methods, including input data types, preprocessing steps model architectures. Focusing NCI ALMANAC dataset, found that feature selection based prior biological knowledge positive impact—limiting gene expression or response-specific genes improved performance. Drug features appeared more predictive response, with 41% increase in coefficient determination (R 2 ) 26% Spearman correlation relative baseline used only cell line identifiers. Molecular fingerprint-based representations performed slightly better than learned representations—ECFP4 fingerprints increased R by 5.3% 2.8% w.r.t best representations. general, fully connected feature-encoding subnetworks outperformed other DL ML methods 35% 14% (Spearman). Additionally, an ensemble combining top models about 6.5% 4% Using state-of-the-art interpretability method, showed learn associate response biologically meaningful way. The explored this study will help improve development computational rational design effective combinations therapy.