The paper introduces CataPro, a neural network-based framework for predicting enzyme kinetic parameters, such as kcat and Km. The framework uses a combination of language models and molecular fingerprints to encode enzyme and substrate information. The model is trained on an unbiased dataset of enzyme kinetic data and demonstrates improved performance compared to existing methods, such as DLKcat and UniKP. The paper also presents a strategy for predicting kcat/Km by incorporating a neural network-based correction term. The model is evaluated on a range of enzymes and substrates, and is shown to outperform existing methods in terms of metrics such as Pearson’s correlation coefficient and root-mean-squared error.
The paper then demonstrates the application of CataPro in enzyme discovery and engineering, using the decarboxylation of ferulic acid to produce 4-VG as an example. The model is used to select candidates for the oxidation of 4-VG, and five representative enzymes are identified. The paper also presents the use of CataPro in enzyme engineering, using the SsCSO enzyme to improve its activity and selectivity. The results demonstrate the effectiveness of CataPro in enzyme discovery and engineering, with the potential to accelerate the discovery of new enzymes and improve their activity and selectivity.