Adaptive Ensemble LS-SVM with Stability Feature Selection for Drug-Induced Autoimmunity Prediction

Abstract

Drug-Induced Autoimmunity (DIA) is a serious side effect that can occur due to the use of certain drugs and has the potential to disrupt the body's immune system. Early identification of compounds that are at risk of causing DIA is crucial to improve drug safety and support the pharmaceutical development process. However, DIA prediction faces challenges such as the high dimensionality of molecular descriptors, limited sample size, and imbalanced class distribution. This study proposes a new approach in the form of an integration of Stability Feature Selection (SFS) and Adaptive Ensemble Least Squares Support Vector Machine (AE-LS-SVM) to improve the stability of feature selection and the generalization capability of the DIA prediction model. The dataset used is derived from the Drug-Induced Autoimmunity Prediction Dataset consisting of 477 training data sets and 120 independent test data sets with 195 RDKit-based molecular descriptors. Stability Feature Selection is built using a combination of ReliefF, Mutual Information, and Boruta to obtain consistently selected features, while Adaptive Ensemble LS-SVM utilizes an adaptive weighted voting mechanism to combine several LS-SVM classifiers. The results showed that the number of features was successfully reduced from 195 to 20 stable features. The proposed model achieved an Accuracy of 80.00%, Precision of 71.40%, Recall of 33.30%, F1-Score of 45.50%, ROC-AUC of 79.40%, and MCC of 0.390. These results demonstrate improved performance compared to a single SVM and produce a more stable model on high-dimensional data. However, the relatively low recall value indicates that sensitivity to DIA cases remains a challenge that needs to be improved in future research.