Predicting Enzyme Turnover Numbers and Enabling Rational Enzyme Evolution

Enzyme turnover number ( $k_{\text{cat}}$ ) is a central kinetic parameter for biocatalysis, but experimental determination is low-throughput and existing computational methods inadequately model enzyme-reaction interplay. Here, we introduced MCKcat, a deep learning framework that integrates multi-scale convolutional feature extraction and cross-attention to enable deep, reciprocal fusion of enzyme sequence and reaction fingerprint representations for accurate $k_{\text{cat}}$ prediction. We constructed MCKcat-DB, a large-scale dataset comprising 33 396 enzyme-reaction-based $k_{\text{cat}}$ data points, covering both wild-type enzymes and a large number of mutants. MCKcat demonstrated competitive performance across diverse prediction scenarios on two benchmarks. A two-step strategy was developed to engineer Bacillus aryabhattai laccase with synergistically improved thermostability and catalytic activity. Rational design generated 217 candidate thermostable mutants, followed by MCKcat-based screening that identified 20 hits. Validation showed 15 mutants (75% positive rate) exhibited simultaneous enhancements in both thermostability and $k_{\text{cat}}$ . A user-friendly web server was provided to facilitate broad adoption. MCKcat establishes a robust, generalizable strategy for data-driven $k_{\text{cat}}$ prediction and artificial intelligence-assisted enzyme engineering.