GOProFormer: A Multi-Modal Transformer Method for Gene Ontology Protein Function Prediction.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning in Biology

Background:

• Protein Language Models (PLMs) excel at learning sequence representations for various biological predictions.
• However, their utility for accurate protein function prediction, specifically Gene Ontology (GO) annotation, remains an open challenge.
• Existing methods for GO prediction often suffer from issues in dataset construction, leading to unreliable performance evaluations.

Purpose of the Study:

• To introduce a novel method for protein function prediction using PLMs and transformer architectures.
• To address limitations in current training and testing dataset methodologies for evaluating protein function prediction models.
• To establish a new benchmark dataset for rigorous evaluation and advancement of GO annotation methods.

Main Methods:

• Developed a dual-transformer approach: a sequence transformer for protein encoding and a graph transformer for GO term hierarchy representation.
• Integrated sequence and GO term representations for multi-label classification of protein functions.
• Investigated and proposed a novel strategy for constructing training and testing datasets to accurately assess model generalization.

Main Results:

• The proposed dual-transformer method demonstrated superior performance compared to existing representative GO prediction techniques.
• The study revealed that current dataset construction approaches can significantly misestimate model generalization capabilities.
• A new benchmark dataset was created, offering a more robust platform for evaluating and comparing protein function prediction models.

Conclusions:

• The novel transformer-based method effectively enhances protein function prediction and GO annotation accuracy.
• Rigorous dataset construction is critical for reliable evaluation of protein function prediction models.
• The developed benchmark dataset will facilitate future research and accelerate progress in the field of automated protein annotation.