Comparative Assessment of Protein Large Language Models for Enzyme Commission Number Prediction.

Area of Science:

• Bioinformatics
• Computational Biology
• Enzyme Engineering

Background:

• Protein large language models (LLMs) extract enzyme sequence representations for function prediction (EC numbers).
• A comprehensive comparison of LLMs for enzyme function prediction is lacking.
• LLMs and sequence alignment methods (e.g., BLASTp) have not been extensively compared as individual predictors.

Purpose of the Study:

• To assess the performance of ESM2, ESM1b, and ProtBERT language models in predicting enzyme commission (EC) numbers.
• To compare LLMs against BLASTp and one-hot encoding-based models.
• To evaluate the complementary potential of LLMs and sequence alignment methods.

Main Methods:

• Evaluated ESM2, ESM1b, and ProtBERT language models for EC number prediction.
• Compared LLM performance against BLASTp and models using one-hot amino acid sequence encodings.
• Assessed deep learning models combined with LLMs against traditional methods.

Main Results:

• LLMs combined with fully connected neural networks outperformed one-hot encoding models.
• BLASTp showed marginally better overall performance, but LLMs and BLASTp results are complementary.
• ESM2 demonstrated the best performance among tested LLMs, excelling in difficult annotations and for enzymes without homologs.

Conclusions:

• LLMs require further improvement to supersede BLASTp in routine enzyme annotation.
• LLMs provide valuable predictions for challenging enzyme annotations, particularly with <25% sequence identity.
• Combining BLASTp and LLM models enhances overall prediction effectiveness.