MZSGO: multimodal zero-shot protein function annotation via evolutionary signals and textual semantics.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Current deep learning models for protein function prediction are limited by using narrow data modalities.
• Existing methods often overlook the semantic richness in protein domain definitions, hindering generalization to new functional labels.
• Protein function prediction requires integrating diverse data types for improved accuracy.

Purpose of the Study:

• To introduce MZSGO, a multimodal zero-shot framework for protein function prediction.
• To overcome limitations of current methods by fusing sequence-based and text-based protein data.
• To enable robust prediction of unseen protein functions and novel Gene Ontology (GO) terms.

Main Methods:

• Developed MZSGO, a multimodal zero-shot framework integrating protein language models and large language models (LLMs).
• Utilized an adaptive gated fusion mechanism to align sequence and text modalities.
• Leveraged evolutionary signals and semantic features from LLMs for enhanced protein representations.

Main Results:

• MZSGO demonstrates strong performance on supervised benchmarks.
• The framework shows a significant advantage over existing methods in zero-shot prediction tasks.
• MZSGO excels at identifying previously unseen, long-tail, and novel Gene Ontology (GO) terms.

Conclusions:

• MZSGO effectively bridges the semantic gap in protein function prediction by unifying representations and annotations.
• The multimodal approach enhances the ability to predict novel and long-tail functional labels.
• This framework offers a more robust and generalizable solution for protein function prediction.