Enhancing Protein Function Prediction Performance by Utilizing AlphaFold-Predicted Protein Structures.

Area of Science:

• Biochemistry
• Computational Biology
• Bioinformatics

Background:

• Protein structure is critical for function, but experimentally determined structures are limited.
• Data-driven models for protein function prediction require large datasets.
• AlphaFold2 provides a large dataset of predicted protein structures, offering a potential solution.

Purpose of the Study:

• To evaluate if AlphaFold-predicted structures improve structure-based protein function prediction models.
• To compare model performance using real structures versus predicted structures for training.

Main Methods:

• Constructed a benchmark dataset.
• Implemented a state-of-the-art structure-based approach.
• Trained models with real structures only, predicted structures only, and combined datasets.

Main Results:

• Model performance improved across all Gene Ontology (GO) term categories when predicted structures were added.
• A model trained solely on AlphaFold-predicted structures achieved performance comparable to a model trained on experimentally solved structures.

Conclusions:

• Structure-based protein function prediction models benefit from incorporating AlphaFold-predicted structures.
• Predicted protein structures are nearly as effective as experimental structures for function prediction training.