EZpred: Improving Deep Learning-Based Enzyme Function Prediction Using Unlabeled Sequence Homologs

Features extracted from sequence homologs significantly enhance the accuracy of deep learning-based protein structure prediction. Indeed, homolog-aware models such as AlphaFold generally produce more accurate protein structures compared to single sequence-based methods like ESMfold. In contrast, features from sequence homologs are seldom employed for deep learning-based prediction of the biological functions of a protein, such as Gene Ontology terms or Enzyme Commission (EC) numbers. Although a small number of models also incorporate function labels from sequence homologs, they cannot utilize features extracted from homologs that lack function labels. To address this gap, we developed EZpred, which is the first deep learning model to use sequence homologs by extracting their sequence features while ignoring any function labels. Starting with the target sequence and homologs identified by MMseqs2, EZpred extracts sequence features using the ESM-C protein language model. These features are then fed into a deep learning model to predict the EC numbers of the target protein. For test enzymes with < 30% sequence identity to any training proteins, the F1-score of EZpred EC number prediction is at least 34% higher than state-of-the-art EC number prediction models. These results demonstrate the strong positive impact of sequence homologs in deep learning-based enzyme function prediction.