Going beyond labels: Semi-supervised learning for enzyme kinetic parameter prediction
Enzyme kinetic parameters, such as the turnover number (kcat), underpin enzyme engineering and metabolic modeling. Machine learning models can predict these parameters from sequence and substrate representations, but their accuracy is limited by the scarcity and narrow coverage of experimentally labeled data, particularly for out-of-distribution enzymes. Existing strategies to exploit unlabeled data rely on homologous sequences, pseudo-labels or synthetic perturbations of labeled samples, limiting their potential to fully leverage unlabeled data for model improvement. Here, we show that regularization-based semi-supervised learning improves model performance by leveraging biologically meaningful unlabeled samples alongside labeled data during task-specific training, without assigning pseudo-labels to the unlabeled samples. Specifically, enzyme-substrate pairs lacking kcat, but characterized for other kinetic parameters, can serve as unlabeled data that improve kcat prediction. Using the clustered CataPro dataset at 40% sequence identity as a stringent benchmark, neural consistency regularization and graph-Laplacian regularization each performed better when unlabeled records were added, beyond the gains from regularization on labeled data alone. The benefit required sufficient labeled data, saturated at a few thousand unlabeled records for the neural model, and depended on the density of the unlabeled data in the feature space rather than on their proximity to labeled data. Gains extended to wild-type and mutant enzymes and to a local low-data adenylate kinase landscape. These findings show that incomplete kinetic records specifically, and unlabeled data points generally, are a valuable resource for enzyme kinetics prediction.