MultiPredGO: Deep Multi-Modal Protein Function Prediction by Amalgamating Protein Structure, Sequence, and...

Area of Science:

• * Computational Biology
• * Bioinformatics
• * Molecular Biology

Background:

• * Proteins are crucial for cellular biochemical activities and biological regulation.
• * Accurate protein function prediction is essential for understanding biological systems.
• * Existing methods often rely on single data types, limiting prediction accuracy.

Purpose of the Study:

• * To introduce MultiPredGO, a novel multi-modal approach for protein function prediction.
• * To leverage protein sequence, secondary structure, and interaction data for improved predictions.
• * To develop and validate deep learning models tailored for different data modalities.

Main Methods:

• * Feature extraction using various convolutional neural network (CNN) architectures for sequence and structure data.
• * Development of distinct deep learning models for protein sequence and 3D structure data.
• * Integration of protein interaction information to enhance prediction efficiency.
• * Application of a neuro-symbolic hierarchical classification model mirroring Gene Ontology (GO) structure for dependent function prediction.

Main Results:

• * MultiPredGO demonstrated superior performance compared to uni-modal and existing multi-modal methods (INGA, DeepGO).
• * Achieved significant improvements in accuracy, F-measure, precision, and recall.
• * Specifically, MultiPredGO showed an average of 13.05% and 30.87% improvement over DeepGO for cellular component and molecular functions, respectively.

Conclusions:

• * The proposed MultiPredGO approach effectively integrates diverse protein data modalities for enhanced function prediction.
• * Multi-modal deep learning, combined with hierarchical classification, offers a powerful strategy for complex biological predictions.
• * MultiPredGO represents a significant advancement in predicting protein functions, particularly for cellular components and molecular functions.