SynVerse: a modular framework for building and evaluating deep learning-based drug synergy prediction models.

Area of Science:

• Computational biology
• Pharmacology
• Artificial intelligence in medicine

Background:

• Synergistic drug combinations are crucial for cancer therapy, but experimental screening is costly.
• Deep learning (DL) offers a computational approach to predict drug synergy, but faces challenges like data leakage and limited validation.
• Existing DL models often lack rigorous evaluation of their generalizability and the factors influencing their performance.

Purpose of the Study:

• To develop and evaluate SynVerse, a comprehensive framework for assessing DL model generalizability in predicting drug pair synergy.
• To conduct systematic ablation studies to understand the contribution of different features and model components.
• To identify limitations of current DL approaches and guide future development for reliable computational drug synergy prediction.

Main Methods:

• Developed SynVerse, a framework with four data-splitting strategies for evaluating DL model generalizability.
• Implemented three ablation studies: module-based, feature shuffling, and a novel network-based approach.
• Evaluated sixteen DL models using eight drug/cell line features, five preprocessing techniques, and two encoders.

Main Results:

• No evaluated DL model surpassed a baseline using one-hot encoding.
• Biologically relevant drug/cell line features and drug-drug interactions did not significantly improve predictive performance.
• All models demonstrated poor generalization capabilities when applied to unseen drugs and cell lines.

Conclusions:

• Current DL models for predicting drug synergy lack robustness and generalizability.
• SynVerse highlights critical shortcomings in computational drug synergy prediction methods.
• Substantial advancements are required for these predictors to be reliably integrated into experimental and clinical cancer treatment strategies.