BiMba: using Vision Mamba to predict protein sites that bind other proteins

Area of Science:

• Structural Biology
• Computational Biology
• Deep Learning

Background:

• Identifying protein binding sites is crucial for understanding protein interactions and designing drugs.
• Current computational methods struggle to integrate surface properties and residue information effectively.
• Recent advances in state-space models and vision-based deep learning offer new possibilities for modeling protein surfaces.

Purpose of the Study:

• To introduce BiMba (protein Binding site prediction using Vision Mamba), a deep learning framework for predicting protein binding sites.
• To leverage the Vision Mamba architecture for efficient modeling of long-range spatial dependencies on protein surfaces.
• To improve the accuracy and interpretability of protein binding site prediction.

Main Methods:

• BiMba represents 3D protein surfaces as 2D geometric or physicochemical grids.
• It integrates geometric and physicochemical information with residue-level descriptors for a unified representation.
• The framework utilizes the Vision Mamba architecture for efficient processing of surface data.

Main Results:

• BiMba demonstrates competitive performance on benchmark datasets, outperforming existing state-of-the-art methods.
• The model integrates spatial topology with biochemical context for accurate binding site prediction.
• Interpretability analyses reveal feature relevance and biologically meaningful residue clusters.

Conclusions:

• State-space models, like Vision Mamba, are effective for molecular surface learning.
• BiMba offers an efficient, interpretable, and scalable approach to protein binding site prediction.
• This work advances the application of deep learning in structural bioinformatics.