Self-supervised domain adaptation of protein language model based solely on positive enzyme-reaction pairs

There is growing interest in developing predictive models of enzyme catalytic properties that leverage activity data spanning diverse enzyme families. A fundamental challenge lies in the inherent biases of public biochemical databases. These databases predominantly catalog valid enzyme activities, rarely include negative instances, and report quantitative catalytic parameters for only a relatively small subset of enzymes. Such limitations pose a major obstacle to supervised learning of enzyme catalytic properties. One existing approach for model training involves generating synthetic negative enzyme-activity pairs by recombining existing enzymes and their activity information, particularly substrates or chemical reactions, that were not originally associated within datasets. However, it remains unclear whether the generated negative examples are truly inactive or merely unobserved active instances. To build a model that captures functional properties across diverse enzyme families while avoiding reliance on negative examples, this paper introduces a self-supervised domain adaptation methodology for pre-trained protein language models, solely based on positive enzyme-reaction pairs. The enzyme representations obtained from the adapted protein language model achieved superior or at least competitive performance compared to those from an existing method that relies on synthetic negatives, in both the turnover number prediction task for natural reactions of wild-type enzymes and the activity prediction task for family-wide enzyme-substrate specificity screening datasets. Overall, our approach represents a methodological advancement that eliminates the need for synthetic negatives and provides a scalable framework for leveraging the growing enzyme activity data in biochemical databases.