ProSSF: integrating sequence, structure, and gene ontology for prediction of protein stability, interaction, and...

Area of Science:

• Computational Biology
• Structural Biology
• Bioinformatics

Background:

• Protein sequences contain vital information for organismal responses to genetic variation, environmental factors, and disease.
• Current computational methods often rely on single data types (sequence, structure, or function), limiting predictive power for low-homology or orphan proteins.

Purpose of the Study:

• To introduce ProSSF (Protein Sequence-Structure-Function), a unified multimodal pretraining framework.
• To integrate sequence, 3D structural, and Gene Ontology (GO) functional information for enhanced protein property prediction.

Main Methods:

• Utilized masked pretraining on large-scale protein sequences.
• Employed a Geometric Vector Perceptron Graph Neural Network (GVP-GNN) for 3D structural encoding.
• Integrated GO semantics via a dual-path hierarchical encoder and aligned modalities using cross-modal attention.

Main Results:

• Achieved a Spearman correlation of 0.74 ± 0.009 on protein stability prediction (TAPE benchmark).
• Reached a mean Micro-F1 of 84.60% ± 0.9% on protein-protein interaction prediction (SHS148K dataset).
• Demonstrated complementary contributions of structural geometry and GO semantics, particularly for low-homology proteins.

Conclusions:

• Jointly encoding sequence, structure, and functional semantics significantly improves protein property prediction generalizability.
• ProSSF preferentially attends to biologically relevant regions, indicating effective learning without explicit supervision.
• The framework holds potential for functional annotation of disease proteins and drug target identification.