FAPM: functional annotation of proteins using multimodal models beyond structural modeling.

Area of Science:

• Biochemistry
• Bioinformatics
• Computational Biology

Background:

• Accurate protein property labeling (e.g., functional terms, catalytic activity) is difficult, particularly for proteins lacking homologs or with sparse data.
• Existing methods often rely on protein sequence features but neglect the semantic information within protein labels.

Purpose of the Study:

• To introduce a novel multimodal model for protein functional annotation.
• To improve the prediction of protein properties like Gene Ontology (GO) terms and catalytic activity.

Main Methods:

• Developed Functional Annotation of Proteins using Multimodal models (FAPM), a contrastive multimodal model.
• Integrated pretrained protein sequence and large language models to link natural language with protein sequence language.
• Enabled incorporation of additional text prompts, such as taxonomy information.

Main Results:

• FAPM demonstrates superior performance in understanding protein properties compared to sequence- or structure-only models.
• Achieved state-of-the-art results on public benchmarks and experimentally annotated phage proteins.
• Showcased enhanced predictive performance and explainability through the use of text prompts.

Conclusions:

• FAPM offers a powerful and flexible alternative for protein annotation, outperforming existing methods.
• The model's ability to leverage natural language enhances its accuracy and applicability, especially for under-annotated proteins.
• This multimodal approach represents a significant advancement in protein functional characterization.