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

Area of Science:

• Computational chemistry and drug discovery
• Artificial intelligence in molecular modeling
• Antiviral drug development

Background:

• The main protease (Mpro) is a crucial target for antiviral drug design against coronaviruses.
• Accurately predicting small molecule binding affinity to Mpro is a significant challenge.

Purpose of the Study:

• To develop and detail a novel deep learning model for blind drug-potency prediction targeting SARS-CoV-2 and MERS-CoV Mpro.
• To evaluate the model's performance against traditional baselines and explore various optimization strategies.

Main Methods:

• Development of a multimodal multitask graph attention network (MultiMolCGC) using a customized gate control framework.
• Integration of multimodal molecular representations and a specialized multitask gating architecture.
• Exploration of pretraining strategies, model architecture adjustments, and the impact of predicted structural data.

Main Results:

• The MultiMolCGC model achieved top performance in a blind drug-potency prediction challenge.
• The model demonstrated superior performance compared to traditional machine learning baselines.
• Pretraining on synthetic docking data significantly improved performance in low-data conditions.

Conclusions:

• The MultiMolCGC framework shows significant potential as a robust and accurate deep learning tool for protein-ligand binding affinity prediction.
• Tailored knowledge sharing through specialized multitask gating is valuable for improving prediction accuracy.
• Pretraining offers a viable strategy to enhance model performance, especially when experimental data is limited.