A unified multimodal model for generalizable zero-shot and supervised protein function prediction.

Area of Science:

• Bioinformatics
• Computational Biology
• Artificial Intelligence in Biology

Background:

• Protein function prediction is crucial but challenging, often limited by traditional methods using single or few data types.
• Existing approaches struggle with predicting novel function terms due to reliance on preselected Gene Ontology (GO) terms.
• Integrating diverse biological data modalities is key to capturing complex functional relationships.

Purpose of the Study:

• To develop a multimodal AI model, FunBind, for enhanced protein function prediction.
• To enable the prediction of previously unseen protein functions.
• To improve accuracy by jointly learning from diverse biological data.

Main Methods:

• FunBind utilizes five modalities: protein sequences, textual descriptions, domain annotations, structures, and GO terms.
• Employs self-supervised pretraining with contrastive learning to create a unified latent space for sequence and other modalities.
• Incorporates supervised fine-tuning for comprehensive and accurate function classification.

Main Results:

• FunBind demonstrates effective zero-shot generalization to novel function terms.
• The model achieves superior performance compared to single-modality models and current state-of-the-art methods.
• Joint multimodal fine-tuning significantly enhances prediction accuracy.

Conclusions:

• FunBind offers a powerful approach for accurate protein function prediction by integrating multimodal data.
• The model's zero-shot capabilities address the challenge of predicting novel protein functions.
• FunBind represents a significant advancement in computational biology for understanding protein roles.