Fine-tuning protein embeddings for functional similarity evaluation.

Area of Science:

• Bioinformatics
• Computational Biology
• Protein Science

Background:

• Proteins with unknown functions are often compared to known proteins using sequence similarity or learned embedding spaces.
• Protein sequence embeddings aid in annotation, clustering, and discovering protein families.
• The potential for deliberately designing embeddings to enhance downstream tasks remains underexplored.

Purpose of the Study:

• To investigate whether fine-tuning protein language models can improve the quality and utility of protein embeddings for downstream tasks.
• To assess the impact of fine-tuning on functional annotation and protein family discovery.

Main Methods:

• Fine-tuning pre-trained protein language models using a classification loss on Gene Ontology (GO) terms.
• Evaluating the performance of fine-tuned embeddings in K-nearest neighbor classifiers for GO annotation.
• Assessing the quality of fine-tuned embeddings for protein family discovery using clustering.

Main Results:

• Direct fine-tuning of language models significantly enhances protein embedding quality for functional annotation.
• Fine-tuned embeddings achieve superior performance in GO annotation compared to standard methods, even outperforming directly fine-tuned classifiers.
• The improved embeddings maintain interpretability through protein similarity comparisons and perform well in clustering tasks for protein family rediscovery.

Conclusions:

• Fine-tuning protein language models is an effective strategy to improve protein embedding quality.
• Enhanced embeddings facilitate more accurate functional annotation and protein family discovery.
• This approach offers a powerful tool for bioinformatics research, combining interpretability with improved performance.