TCNeKP: A Novel Deep Learning Architecture for Enzyme Catalytic Activity Prediction

Accurate prediction of enzyme kinetic parameters (<i>K</i>_cat and <i>K</i>_m) is crucial for enzyme rational design and engineering research. Based on a heterogeneous data set encompassing 17,893 <i>K</i>_cat and 24,585 <i>K</i>_m records across 8911 enzyme sequences from 7 EC classes and 5023 substrates, we introduce novel TCNeKP models for predicting <i>K</i>_cat and <i>K</i>_m values. Herein, enzymes' sequences were autoembedded and processed by a temporal convolutional network (TCN) module to extract the key features of catalytic and binding residues frequently located far apart in the primary sequences; substrates were encoded by a pretrained SMILES-Transformer language model; and catalytic conditions (pH and temperature) were encoded via radial basis function (RBF). The fused features were then fed into a fully connected network for single-task prediction of <i>K</i>_cat and <i>K</i>_m. Results demonstrate that TCNeKP-<i>K</i>_cat and TCNeKP-<i>K</i>_m models achieve robust performance across wild-type and mutant enzymes from 7 EC classes, outperforming state-of-the-art MPEK, UniKP, and DLKcat models (Table S3). Leveraging a cross-task dynamic parameter-sharing module with attention mechanism, we further developed a multitask TCNeKP model that achieves the highest <i>R</i>² values among the benchmark models for both <i>K</i>_cat (0.677) and <i>K</i>_m (0.657) prediction. These findings indicate that collaborative learning between <i>K</i>_cat and <i>K</i>_m prediction tasks enhances feature extraction for enzyme-substrate binding and catalysis, thereby significantly enhancing the predictive performance.