Evaluating large language models for annotating proteins.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• UniProtKB contains over 251 million proteins, but only 0.25% are annotated with Pfam family domains.
• Current Pfam annotation methods, while effective, struggle to keep pace with the rate of protein discovery.
• Deep learning models for Pfam annotation require substantial training data, posing challenges for underrepresented protein families.

Purpose of the Study:

• To develop and evaluate a novel transfer learning protocol for enhancing protein domain annotation.
• To leverage protein large language models (LLMs) and their sequence embeddings for improved Pfam classification.
• To address the data scarcity issue in annotating poorly populated protein families.

Main Methods:

• Utilized protein large language models (LLMs) trained with self-supervision on large unannotated datasets to generate sequence embeddings.
• Applied supervised learning on small, annotated datasets using these embeddings for specialized protein domain prediction tasks.
• Evaluated multiple state-of-the-art protein LLMs and machine learning architectures within the proposed protocol.

Main Results:

• The novel protocol achieved significantly better results than state-of-the-art methods for protein families classification.
• Demonstrated a 60% reduction in prediction error compared to standard protein annotation techniques.
• Successfully showed the practical application of LLM embeddings for protein annotation.

Conclusions:

• Transfer learning with protein LLMs offers a powerful and efficient approach to protein domain annotation.
• The proposed method substantially enhances the accuracy and efficiency of Pfam classification, particularly for challenging families.
• The readily available pipeline and code facilitate the adoption of this advanced annotation strategy.