Machine learning models for predicting enzymatic hydrolysis yields of lignocellulosic biomass after various pretreatments

This study presents a comparative computational framework of six machine learning architectures to predict enzymatic bioconversion efficiency of lignocellulosic biomass, utilizing multidimensional inputs spanning compositional signatures and pretreatment operational parameters. Support vector regression (SVR) demonstrated superior performance among all models, with R² values exceeding 0.90 for test sets of target variables. After optimization, SVR finally achieved coefficients of determination (R²) of 0.95 and 0.99 for glucose and xylose yields, respectively. Solution pH was identified as the dominant factor influencing chemical composition, structural characteristics, and solid yield during pretreatment. Through SHapley Additive exPlanations (SHAP) and gradient-based importance quantification, first-principal interpretations of feature-response relationships were established, revealing nonlinear interdependencies between pretreatment-induced structural modifications and subsequent enzymatic accessibility. A software tool was engineered to precisely predict glucose and xylose yields following different pretreatments. This study offers novel insights into critical determinants and their synergistic relationships affecting the pretreatment process and enzymatic hydrolysis yields for lignocellulosic biomass.