An Adaptive Evolutionary Algorithm Based on Multi‐Granularity Search and Application in Feature Selection Problem of Metabolic Stability of Liver Microsomes

ABSTRACT In drug discovery and development, machine learning facilitates the screening of drug candidates by quickly identifying features related to drug metabolic stability through classification and prediction. Aiming at the premature convergence and post‐stagnation problems of the pelican optimization algorithm (POA), this paper proposes an improved pelican optimization algorithm (CADEPOA) based on a multi‐granularity dynamic search strategy. CADEPOA is further discretized and combined with a KNN classifier to construct a wrapper feature selection model for classifying liver microsome metabolic stability datasets. Specifically, the multi‐granularity model fuses coarse‐grained breadth exploration and fine‐grained depth mining to search different levels of the decision space, and the dynamic switching mechanism balances exploration and exploitation. In addition, a Sine chaos mapping factor improves the initial population quality, and a differential evolution mechanism with a dynamic scaling factor enhances population diversity through mutation, crossover, and selection. On the CEC2022 test suite, CADEPOA obtains the best average fitness on 11 of 12 benchmark functions and the lowest standard deviation on 10 of 12 functions in the ablation comparison. On ten UCI datasets, BCADEPOA achieves the best or tied‐best accuracy in all cases. For the human liver microsomes and liver microsomes datasets, BCADEPOA achieves accuracies of 76.40% and 85.57%, F‐scores of 66.20% and 85.21%, and G‐mean values of 72.76% and 85.53%, while retaining about 51.13% and 51.32% of the original features, respectively. These results demonstrate that CADEPOA improves optimization stability and that BCADEPOA is effective for wrapper‐based feature selection in metabolic stability prediction.