Comparative assessment of protein large language models for enzyme commission number prediction

Abstract

Background: Protein large language models (LLM) have been used to extract representations of enzyme sequences to predict their function, which is encoded by enzyme commission (EC) numbers. However, a comprehensive comparison of different LLMs for this task is still lacking, leaving questions about their relative performance. Moreover, protein sequence alignments (e.g. BLASTp or DIAMOND) are often combined with machine learning models to assign EC numbers from homologous enzymes, thus compensating for the shortcomings of these models predictions. In this context, LLMs and sequence alignment methods have not been extensively compared as individual predictors, raising unaddressed questions about LLMs performance and limitations relative to the alignment methods. In this study, we set out to assess the performance of ESM2, ESM1b, and ProtBERT language models in their ability to predict EC numbers, comparing them with BLASTp, against each other and against models that rely on one-hot encodings of amino acid sequences. Results: Our findings reveal that combining these LLMs with fully connected neural networks surpasses the performance of deep learning models that rely on one-hot encodings. Moreover, although BLASTp provided marginally better results overall, DL models provide results that complement BLASTps, revealing that LLMs better predict certain EC numbers while BLASTp excels in predicting others. The ESM2 stood out as the best model among the LLMs tested, providing more accurate predictions on difficult annotation tasks and for enzymes without homologs. Conclusions: Crucially, this study demonstrates that LLMs still have to be improved to become the gold standard tool over BLASTp in mainstream enzyme annotation routines. On the other hand, LLMs can provide good predictions for more difficult-to-annotate enzymes, particularly when the identity between the query sequence and the reference database falls below 25%. Our results reinforce the claim that BLASTp and LLM models complement each other and can be more effective when used together. © The Author(s) 2025.