M3Site: multiclass multimodal learning for protein active site identification and classification.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Drug Discovery

Background:

• Accurate protein active site identification is vital for understanding protein function, enabling drug design, and advancing synthetic biology.
• Existing methods often use binary classification and single data types, limiting their predictive power and scope.
• There is a need for advanced computational tools that can leverage diverse data modalities for more comprehensive active site analysis.

Purpose of the Study:

• To develop M3Site, a novel multimodal framework for residue-level, multiclass protein active site prediction.
• To integrate protein sequence embeddings, structural graph representations, and functional text annotations for enhanced prediction accuracy.
• To provide an accessible and practical tool for researchers in structural biology and drug discovery.

Main Methods:

• Developed M3Site, a multimodal framework integrating sequence, structure, and functional annotations.
• Utilized pretrained protein language models, equivariant graph neural networks, and biomedical language models for feature extraction.
• Implemented a function-informed cross-attention module for cross-modal fusion and an adaptive weighted fusion mechanism.

Main Results:

• M3Site demonstrated significantly superior performance compared to existing state-of-the-art methods in active site prediction.
• The framework successfully integrates diverse data modalities, improving the accuracy and granularity of predictions.
• An interactive application was developed, facilitating practical use for predictions and visualizations.

Conclusions:

• M3Site represents a significant advancement in protein active site prediction by effectively leveraging multimodal data.
• The framework's ability to perform multiclass, residue-level predictions offers enhanced utility for biological and medical research.
• The public availability of the dataset, code, and application promotes further research and development in the field.