SpatConv Enables the Accurate Prediction of Protein Binding Sites by a Pretrained Protein Language Model and an...

Area of Science:

• Biochemistry
• Structural Biology
• Computational Biology

Background:

• Protein interactions are vital for biological processes, and identifying protein binding sites is key to understanding diseases like cancer.
• Current prediction models struggle with information loss and limited learning, failing to capture complex residue binding patterns.

Purpose of the Study:

• To introduce SpatConv, a general network for predicting binding residues of proteins, peptides, and metal ions.
• To overcome limitations of traditional models by integrating sequence and structure features effectively.

Main Methods:

• SpatConv utilizes sequence features from large protein language models and structure features from a local coordinate framework.
• A novel graph-free bio-spatial convolution characterizes residue spatial environments to learn binding patterns.

Main Results:

• SpatConv significantly outperforms state-of-the-art predictors, revealing insights into binding sites and physicochemical properties.
• The method shows robust performance on both predicted and experimental protein structures.
• Applied to SARS-CoV-2 spike protein, SpatConv identified antibody binding sites, supporting drug development.

Conclusions:

• SpatConv offers a powerful and reliable approach for predicting protein binding sites across various molecule types.
• The tool enhances understanding of disease mechanisms and facilitates new drug discovery efforts.
• An accessible online server is available for SpatConv.