Annotating protein functions via fusing multiple biological modalities.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Protein function annotation is crucial for understanding disease mechanisms and identifying therapeutic targets.
• Current deep learning methods struggle to integrate diverse biological data, leading to suboptimal protein representations and inaccurate function prediction.
• Sparse label representations in existing models hinder convergence to optimal solutions.

Purpose of the Study:

• To develop a robust framework for fusing heterogeneous biological modalities for accurate protein function annotation.
• To address the limitations of existing deep learning methods in multimodal data integration and sparse label representation.
• To enhance the quality of protein representations for improved biological insights.

Main Methods:

• Proposed MIF2GO (Multimodal Information Fusion to infer Gene Ontology terms), a sequential, multi-step approach to fuse up to six biological modalities.
• Implemented a novel deep learning architecture designed for effective integration of diverse biological data types.
• Validated the method on seven benchmark datasets across different species.

Main Results:

• MIF2GO significantly outperformed state-of-the-art methods in protein function annotation.
• The method demonstrated high robustness and generalizability across various species.
• Achieved powerful protein representations through effective fusion of multimodal biological data.

Conclusions:

• MIF2GO provides a scalable and robust framework for integrating multimodal biological data for protein function annotation.
• The findings facilitate advancements in precision medicine and the discovery of novel therapeutic strategies.
• Offers valuable biological insights into the associations between different biological modalities and protein functions.