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

Protein sequences encode rich structural and functional information that governs how organisms respond to genetic variation, environmental challenge, and disease. However, existing computational methods typically rely on a single information source, whether sequence, structure, or functional annotation, and their predictive power is substantially reduced for low-homology proteins or orphan proteins. Here we present ProSSF (Protein Sequence-Structure-Function), a unified multimodal pretraining framework that performs masked pretraining on large-scale protein sequences, encodes three-dimensional structural information via a Geometric Vector Perceptron Graph Neural Network (GVP-GNN), integrates Gene Ontology (GO) semantics through a dual-path hierarchical encoder, and aligns all three modalities into a shared representation space via cross-modal attention. Evaluated across three downstream tasks, ProSSF achieves a Spearman correlation of 0.74 ± 0.009 on the TAPE protein stability benchmark, a mean Micro-F1 of 84.60% ± 0.9% on the SHS148K protein-protein interaction dataset under the stringent DFS partition, and comparable or superior Fmax and AUPR relative to state-of-the-art baselines across all three GO sub-ontologies. Ablation analyses demonstrate that structural geometry and GO functional semantics contribute complementary and task-dependent information, with the largest performance gains observed under low-homology conditions. Attention-based interpretability analyses further reveal that the model preferentially attends to biologically meaningful regions, such as kinase catalytic domains, without explicit supervision. This study provides a unified multimodal pretraining framework and demonstrates that jointly encoding sequence, structure, and functional semantics substantially improves the generalizability of protein property prediction. Future studies should validate this framework on larger, taxonomically diverse protein datasets and explore its potential applications in the functional annotation of disease-associated proteins and the identification of novel drug targets.