AMY-SA-XGBoost: A Machine Learning Model to Predict Amylase Specific Activity

Abstract Amylase is an important enzyme preparation, and its enzyme activity directly determines its catalytic efficiency in starch hydrolysis. However, current experimental methods for determining amylase specific activity are time-consuming and costly, and experimentally measured specific activity data are sparse and noisy, with limited experimental data currently available. We developed AMY-SA-XGBoost, a machine learning-based approach for predicting amylase specific activity. We systematically constructed a high-quality dataset comprising 178 samples suitable for studying amylase catalytic efficiency by collecting and screening amylase activity data from enzyme databases and literature. The XGBoost-based model exhibited strong predictive performance for amylase activity, with an R2 of 0.9378 and an RMSE of 0.4329. The AMY-SA-XGBoost can capture specific activity changes in mutated enzymes and identify amino acid residues with substantial impact on specific activity. In addition, we screened 14 amylase mutants with the highest predicted increases in specific activity and experimentally validated each mutant. Experimental results confirm that AMY-SA-XGBoost accurately predicts the specific activity of amylase mutants based on their amino acid sequences, successfully identifying variants with higher activity. Among the tested mutants, S359N demonstrated the most significant improvement, achieving a specific activity of 9573 ± 21.30 U/mg. This work provides both theoretical foundation and practical insights for enhancing the catalytic efficiency of amylase in pharmaceutical and food industry applications.