GeoBind: segmentation of nucleic acid binding interface on protein surface with geometric deep learning

Area of Science:

• Computational biology
• Structural bioinformatics
• Machine learning in biochemistry

Background:

• Understanding protein-nucleic acid interactions is crucial for elucidating biological regulatory functions.
• Existing methods for identifying protein binding sites rely on handcrafted features, limiting their predictive power.
• There is a need for advanced computational approaches to accurately predict these interaction sites.

Purpose of the Study:

• To introduce GeoBind, a novel geometric deep learning method for predicting nucleic acid binding sites on protein surfaces.
• To evaluate GeoBind's performance against state-of-the-art predictors using benchmark datasets.
• To demonstrate the method's adaptability to other ligand binding site prediction tasks.

Main Methods:

• GeoBind utilizes geometric deep learning, processing protein surfaces as point clouds.
• It learns high-level representations by aggregating neighbor information within local reference frames.
• The method operates in a segmentation manner for precise site identification.

Main Results:

• GeoBind significantly outperforms existing state-of-the-art predictors on benchmark datasets.
• Case studies highlight GeoBind's effectiveness in analyzing complex protein structures, including those with multimer formation.
• The method achieved competitive performance when extended to predict five other types of ligand binding sites.

Conclusions:

• GeoBind offers a superior and more expressive approach for predicting nucleic acid binding sites compared to traditional methods.
• Its geometric deep learning framework provides powerful insights into molecular surface interactions.
• GeoBind demonstrates broad applicability and potential for advancing ligand binding site prediction across various biological contexts.