Robust Prediction of Protein-Ligand Binding Potency with Multi-modal Customized Gate Control

The main protease (Mpro) is a critical target in the design of antiviral drugs against coronaviruses, while accurately predicting the binding affinity between small molecules and this target remains a key challenge. In the recent Polaris challenge of blind drug-potency prediction targeting SARS-CoV-2 and MERS-CoV Mpro, we developed a multimodal multitask graph attention network based on the customized gate control framework (abbreviated as MultiMolCGC). Our team achieved top performance among all participating teams in the blind prediction challenge. In this paper, we detail the model development and further explorations in terms of pretraining, adjusting the model architecture, and many others. Our model consistently outperforms traditional machine learning baselines, demonstrating the effectiveness of end-to-end deep learning in capturing complex molecular interactions. Integrating multimodal representations proved essential, and the multitask specialized gating architecture outperformed both single-task and nonspecialized multitask variants, highlighting the value of tailored knowledge sharing. While auxiliary loss weighting and hyperparameter tuning offered modest improvements, incorporating predicted structural data unexpectedly reduced performance, likely due to structural uncertainty. Notably, pretraining on large-scale synthetic docking data sets significantly enhanced performance in low-data scenarios, reducing dependence on experimental pIC₅₀ data. The numerical results highlight the potential of MultiMolCGC as a robust and accurate deep-learning framework for protein-ligand binding in future studies.