EZpred: improving deep learning-based enzyme function prediction using unlabeled sequence homologs

Abstract

Features extracted from sequence homologs significantly enhance the accuracy of deep learning-based protein structure prediction. Indeed, models such as AlphaFold, which extracts features from sequence homologs, 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 protein function prediction. Although a small number of models also incorporate function labels from sequence homologs, they cannot utilize features extracted from sequence homologs that lack function labels. To address this gap, we propose EZpred, which is the first deep learning model to use unlabeled sequence homologs for protein function prediction. Starting with the target sequence and homologs identified by MMseqs2, EZpred extracts sequence features using the ESMC protein language model. These features are then fed into a deep learning model to predict the Enzyme Commission (EC) numbers of the target protein. For 753 enzymes, the F1-score of EZpred EC number prediction is 4% higher than a similar model that does not use sequence homologs and at least 10% higher that 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. Significance StatementMultiple sequence alignment (MSA) of homologous sequences is the most important source of input feature for deep learning-based protein structure prediction. Yet, this kind of feature is rarely used for protein function prediction. We propose the first deep learning model that significantly improves protein function prediction using features from sequence homologs without functional labels. This shows the utility of an important set of features that are often overlooked by previous studies.