Autoregressive enzyme function prediction with multi-scale multi-modality fusion.

Area of Science:

• Bioinformatics
• Computational Biology
• Enzymology

Background:

• Accurate enzyme function prediction is vital for biological research and industrial applications.
• Current deep learning models often use only sequence or structural data and fail to leverage the hierarchical nature of Enzyme Commission (EC) numbers.

Purpose of the Study:

• To introduce a novel multi-modal, multi-scale, and autoregressive deep learning model for predicting enzyme function.
• To address the limitations of existing methods by integrating both protein sequence and 3D structural information.
• To leverage the hierarchical structure of EC numbers for more precise function prediction.

Main Methods:

• Developed the Multi-scale multi-modality Autoregressive Predictor (MAPred) model.
• MAPred integrates primary amino acid sequences and 3D structural tokens using a dual-pathway approach.
• Employs an autoregressive network to sequentially predict EC number digits, respecting their hierarchical classification.

Main Results:

• MAPred demonstrated superior performance compared to existing models on benchmark datasets (New-392, Price, New-815).
• The model effectively captures comprehensive protein characteristics and identifies critical local functional sites.
• Achieved significant advancements in the reliability and granularity of protein function prediction.

Conclusions:

• MAPred represents a significant advancement in predicting enzyme function by integrating multi-modal data and respecting EC number hierarchy.
• The model's ability to leverage both sequence and structural information enhances the accuracy and detail of enzyme function prediction.
• This approach offers a more robust tool for bioinformatics research and enzyme-related applications.