MMISA-KM: a deep-learning method using multi-modal information and self-attention mechanisms for the prediction of Michaelis constants

The Michaelis constant (K<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">m</inf>) is a critical parameter in enzyme kinetics, reflecting the affinity between enzymes and substrates. Accurately predicting this value is crucial for the reconstruction of metabolic networks, which provides the foundation for designing and optimizing cellular factories. However, accurately predicting K<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">m</inf> remains challenging due to the complexity of enzyme-substrate interactions and the scarcity of experimental data. To address this, we have developed MMISA-KM, a novel deep learning model that integrates multi-modal information, including protein sequences, protein graphs, substrate SMILES strings, and molecular graphs, combined with a self-attention mechanism for feature fusion. The results demonstrate that MMISA-KM outperforms existing methods, showing superior accuracy and generalization ability in both wild-type and mutant enzymes. Ablation studies confirm the importance of its multi-modal architecture, with each component contributing to enhanced performance. The model not only predicts K<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">m</inf> with high precision but also offers insights into enzyme promiscuity. The results across diverse datasets establish it as a powerful tool for predicting enzymatic reaction kinetic parameters. The data and code presented in this study are openly available at https://github.com/kaiwang-group/MMISA-KM.