Enhancing Enzyme Commission Number Prediction With Contrastive Learning and Agent Attention

The accurate prediction of enzyme function is crucial for elucidating disease mechanisms and identifying drug targets. Nevertheless, existing enzyme commission (EC) number prediction methods are limited by database coverage and the depth of sequence information mining, hindering the efficiency and precision of enzyme function annotation. Therefore, this study introduces ProteEC-CLA (Protein EC number prediction model with Contrastive Learning and Agent Attention). ProteEC-CLA utilizes contrastive learning to construct positive and negative sample pairs, which not only enhances sequence feature extraction but also improves the utilization of unlabeled data. This process helps the model learn the differences in sequence features, thereby enhancing its ability to predict enzyme function. Integrating the pre-trained protein language model ESM2, the model generates informative sequence embeddings for deep functional correlation analysis, significantly enhancing prediction accuracy. With the incorporation of the Agent Attention mechanism, ProteEC-CLA's ability to comprehensively capture local details and global features is enhanced, ensuring high-accuracy predictions on complex sequences. The results demonstrate that ProteEC-CLA performs exceptionally well on two independent and representative datasets. In the standard dataset, it achieves 98.92% accuracy at the EC4 level. In the more challenging clustered split dataset, ProteEC-CLA achieves 93.34% accuracy and an F1-score of 94.72%. With only enzyme sequences as input, ProteEC-CLA can accurately predict EC numbers up to the fourth level, significantly enhancing annotation efficiency and accuracy, which makes it a highly efficient and precise functional annotation tool for enzymology research and applications.