MPBind: a multitask protein binding site predictor using protein language models and equivariant GNNs.

Area of Science:

• Computational Biology
• Biochemistry
• Structural Biology

Background:

• Proteins are fundamental to biological functions, mediating interactions with diverse molecules like DNA, RNA, ligands, ions, and lipids.
• Understanding these protein interactions is crucial for cellular communication, metabolic and gene regulation, and maintaining structural integrity.
• Accurate prediction of protein interaction sites is essential for deciphering protein function and biological mechanisms.

Purpose of the Study:

• To introduce MPBind, a novel multitask method for predicting protein binding sites.
• To develop a versatile tool capable of predicting binding sites for various molecular interaction categories.
• To achieve state-of-the-art accuracy in protein binding site prediction across multiple interaction types.

Main Methods:

• Integration of protein language models (PLMs) to extract sequence-based structural and functional information.
• Utilization of equivariant graph neural networks (EGNNs) to capture geometric features of protein 3D structures.
• Application of multitask learning to predict binding sites for five categories of binding partners: proteins, DNA/RNA, ligands, lipids, and ions.

Main Results:

• MPBind demonstrates state-of-the-art accuracy across five molecular classes.
• Achieved AUROC scores of 0.83 for protein-protein binding site prediction and 0.81 for protein-DNA/RNA binding site prediction.
• Outperformed existing general and task-specific binding site prediction methods, highlighting its versatility and effectiveness.

Conclusions:

• MPBind is a highly accurate and versatile tool for predicting protein binding sites.
• The method's ability to generalize across diverse molecular interaction types makes it valuable for biological research.
• MPBind advances the field of protein interaction prediction, offering a robust solution for understanding protein function.