Enzyme function prediction learns from unlabeled homologs: EZpred lifts four-digit EC F1 to 0.786 on 720 newly annotated enzymes
Synopsis
EZpred carries the homolog-based features long used in protein structure prediction into enzyme function prediction, taking only sequence features from homologs and ignoring their function labels. Most deep learning models for this task read features from the target sequence alone, and the few that use homologs only aggregate labeled ones. On 720 enzymes with new EC numbers in UniProt 2024_02, EZpred reaches an F1 of 0.786 for full four-digit EC numbers, 27% and 31% above CLEAN and GraphEC.
Interpretation
Enzyme function prediction now draws on homologs that carry no function labels, not only on the target sequence or on labeled homolog templates. Earlier deep learning enzyme function models mostly extracted features from the target sequence alone, and the few that used homologs aggregated only labeled ones, leaving unlabeled homologs unused. On 720 enzymes with new EC numbers in UniProt 2024_02 and no such annotation in 2023_02, EZpred reaches an F1 of 0.786 for four-digit EC numbers, 27% and 31% above CLEAN and GraphEC, with Wilcoxon p-values of 3.51E-20 and 2.89E-31.
The pipeline finds homologs of the target enzyme with MMseqs2, then encodes the target and its top four homologs with the ESM C protein language model. Each sequence takes the last three ESM C layers, is mean-pooled along the sequence, and the target and homolog vectors are concatenated and passed to a multilayer perceptron that outputs EC number probabilities. Removing the homolog search drops the same architecture from F1 0.782 to 0.756; using only the last ESM C layer gives 0.763 against 0.782 for the last three.
The final prediction combines template retrieval, over both sequence and structure, with the deep learning output under a weight. The template part searches sequence templates with MMseqs2 and structure templates with Foldseek against the AlphaFold structure database, then merges its score with the deep learning score. The template and deep learning parts each reach F1 0.782 alone, and combining them adds only 0.5%; on very hard targets with TM-score below 0.2 the deep learning part outperforms the template part.
Perspective
The demonstrated setting is 720 enzymes with new EC numbers in UniProt 2024_02 and none in 2023_02, plus 625 non-enzymes with new experimental molecular function annotations; the authors present the model as a guide for follow-up experiments and state it should not be used to predict mutational effects.
The authors report that combining the template and deep learning parts adds only 0.5% F1 and attribute this to both parts already sitting near a ceiling; on very hard targets the deep learning part beats the template part below TM-score 0.2, while the reverse holds between 0.2 and 0.6. Mean pooling leaves the model insensitive to active-site mutations, and weighting active-site residues is named as future work.