Implementation of Bat Algorithm-Support Vector Machine in Predicting Caco-2 Permeability of Drugs
Abstract
Caco-2 permeability is an important indicator for evaluating the oral absorption potential of drug compounds, especially since approximately \~{}90% of marketed drugs are administered through the oral route. However, experimental Caco-2 permeability measurement can be time-consuming and affected by laboratory variability, making in silico prediction useful as an early screening approach. This prediction task is also affected by high-dimensional molecular features and imbalanced permeability classes, which can influence classification performance. This study implements a Bat Algorithm with Support Vector Machine approach to predict multiclass Caco-2 permeability using Morgan fingerprint features from 5,246 molecules. Variance thresholding was first applied to remove low-variance features, followed by feature selection using BA to obtain a more compact feature subset for each SVM kernel. The selected features were then used for SVM classification, and Grid Search was applied for hyperparameter tuning. The results show that BA reduced the feature space across several kernels. After hyperparameter tuning, the RBF kernel obtained a cross-validation macro F1-score of 0.803. In the final model evaluation, the polynomial kernel achieved a Class 2 F1-score of 0.6129 and an accuracy of 0.6705. These findings present the performance characteristics of the BA-SVM approach in imbalanced molecular classification tasks across multiple SVM kernels.