Presentation Information
[1ENZ-13]Machine learning–guided dual optimization of the substrate specificity and thermostability of yeast alcohol dehydrogenase
○Yuki Ogawa1, Tomokazu Shirai1, Akihiko Kondo1,2 (1. RIKEN (Japan), 2. Kobe Univ. (Japan))
Keywords:
machine learning,substrate specificity,thermostability,yeast alcohol dehydrogenase,enzyme engineering
[Purpose]
Engineering enzymes that achieve both high catalytic activity and robustness remains challenging. In particular, improving substrate specificity and thermostability often involves trade-offs, as stability-enhancing mutations frequently reduce catalytic activity. Yeast alcohol dehydrogenase (YADH) is a useful biocatalyst, but its limited substrate scope and moderate thermal stability restrict broader applications. This study aimed to establish a machine learning–guided strategy to improve the conversion of the aging-related odor compound trans-2-nonenal to the less odorous trans-2-nonenol while simultaneously optimizing substrate specificity and thermostability of YADH.
[Method]
Docking simulations between YADH and trans-2-nonenal identified five residues around the substrate-binding pocket for mutagenesis targets. Approximately 200 variants, including single and multiple mutants, were constructed. Enzyme activities were evaluated by absorbance-based assays, and thermostabilities were assessed using a thermal shift assay. These data were used to train machine learning models: a support vector classifier (SVC) for activity classification, a support vector machine (SVM) regression model for activity prediction, and an SVM-based penalty model for thermostability reduction. Model outputs were integrated multiplicatively to prioritize variants.
[Results]
The machine learning–guided approach identified variants with improved properties. Four variants showed increased specific activity toward trans-2-nonenal compared to the wild type, and one variant also exhibited improved thermostability. The best-performing variant exhibited more than a twofold increase in specific activity and increased thermal resistance. Variants were also obtained that lost activity toward the native substrate acetaldehyde while gaining higher activity toward trans-2-nonenal, indicating altered substrate specificity. These results demonstrate efficient identification of beneficial mutations from a limited dataset.
[Consideration]
The results highlight the effectiveness of combining structure-based design and machine learning for multi-objective enzyme engineering. The thermostability penalty model, designed to penalize variants with decreased thermoratbilities, effectively constrained the predictions, with most selected variants showing less than 10 °C reduction in thermostability. However, some trade-offs remained, reflecting a complex fitness landscape. Model performance depends on data quality, and iterative refinement may further improve accuracy
[Conclusion]
This study presents a machine learning–guided framework integrating docking and multi-model evaluation for simultaneous optimization of substrate specificity and thermostability in YADH. The approach enables efficient identification of functional variants and supports data-driven enzyme design.
Engineering enzymes that achieve both high catalytic activity and robustness remains challenging. In particular, improving substrate specificity and thermostability often involves trade-offs, as stability-enhancing mutations frequently reduce catalytic activity. Yeast alcohol dehydrogenase (YADH) is a useful biocatalyst, but its limited substrate scope and moderate thermal stability restrict broader applications. This study aimed to establish a machine learning–guided strategy to improve the conversion of the aging-related odor compound trans-2-nonenal to the less odorous trans-2-nonenol while simultaneously optimizing substrate specificity and thermostability of YADH.
[Method]
Docking simulations between YADH and trans-2-nonenal identified five residues around the substrate-binding pocket for mutagenesis targets. Approximately 200 variants, including single and multiple mutants, were constructed. Enzyme activities were evaluated by absorbance-based assays, and thermostabilities were assessed using a thermal shift assay. These data were used to train machine learning models: a support vector classifier (SVC) for activity classification, a support vector machine (SVM) regression model for activity prediction, and an SVM-based penalty model for thermostability reduction. Model outputs were integrated multiplicatively to prioritize variants.
[Results]
The machine learning–guided approach identified variants with improved properties. Four variants showed increased specific activity toward trans-2-nonenal compared to the wild type, and one variant also exhibited improved thermostability. The best-performing variant exhibited more than a twofold increase in specific activity and increased thermal resistance. Variants were also obtained that lost activity toward the native substrate acetaldehyde while gaining higher activity toward trans-2-nonenal, indicating altered substrate specificity. These results demonstrate efficient identification of beneficial mutations from a limited dataset.
[Consideration]
The results highlight the effectiveness of combining structure-based design and machine learning for multi-objective enzyme engineering. The thermostability penalty model, designed to penalize variants with decreased thermoratbilities, effectively constrained the predictions, with most selected variants showing less than 10 °C reduction in thermostability. However, some trade-offs remained, reflecting a complex fitness landscape. Model performance depends on data quality, and iterative refinement may further improve accuracy
[Conclusion]
This study presents a machine learning–guided framework integrating docking and multi-model evaluation for simultaneous optimization of substrate specificity and thermostability in YADH. The approach enables efficient identification of functional variants and supports data-driven enzyme design.
Comment
To browse or post comments, you must log in.Log in