COMBINING SHAP AND GENETIC ALGORITHM OPTIMIZATION TO ENHANCE RANDOM FOREST EXPLAINABILITY IN PREDICTING β-GALACTOSIDASE IMMOBILIZATION

Enzyme immobilization can be supported by Machine Learning (ML) models, which often function as black boxes. This study employs SHAP and Genetic Algorithms to interpret and optimize β-galactosidase immobilization using a trained Random Forest model. The dataset included input variables related to the enzyme, immobilization conditions, and buffer properties. At the same time, the outputs were the optimum temperature, optimum pH, and the number of cycles required to maintain 60% activity. SHAP analysis revealed that some features had a greater influence on predictions, aligning with experimental findings. Optimization yielded output values of 65.4°C for the optimum temperature, 8.3 for the optimum pH, and 15 cycles of stability. The method also provided the ideal immobilization conditions to attend to these metrics. Future works will focus on experimental validation of the model’s predictive capability.