CatPred: A comprehensive framework for deep learning in vitro enzyme kinetic parameters kcat, Km and Ki

Abstract

Quantification of enzymatic activities still heavily relies on experimental assays, which can be expensive and time-consuming. Therefore, methods that enable accurate predictions of enzyme activity can serve as effective digital twins. A few recent studies have shown the possibility of training machine learning (ML) models for predicting the enzyme turnover numbers (kcat) and Michaelis constants (Km) using only features derived from enzyme sequences and substrate chemical topologies by training on in vitro measurements. However, several challenges remain such as lack of standardized training datasets, evaluation of predictive performance on out-of-distribution examples, and model uncertainty quantification. Here, we introduce CatPred, a comprehensive framework for ML prediction of in vitro enzyme kinetics. We explored different learning architectures and feature representations for enzymes including those utilizing pretrained protein language model features and pretrained three-dimensional structural features. We systematically evaluate the performance of trained models for predicting kcat, Km, and inhibition constants (Ki) of enzymatic reactions on held-out test sets with a special emphasis on out-of-distribution test samples (corresponding to enzyme sequences dissimilar from those encountered during training). CatPred assumes a probabilistic regression approach offering query-specific standard deviation and mean value predictions. Results on unseen data confirm that accuracy in enzyme parameter predictions made by CatPred positively correlate with lower predicted variances. Incorporating pre-trained language model features is found to be enabling for achieving robust performance on out-of-distribution samples. Test evaluations on both held-out and out-of-distribution test datasets confirm that CatPred performs at least competitively with existing methods while simultaneously offering robust uncertainty quantification. CatPred offers wider scope and larger data coverage ([~]23k, 41k, 12k data-points respectively for kcat, Km and Ki). A web-resource to use the trained models is made available at: https://tiny.cc/catpred