<![CDATA[Drug-induced autoimmunity (DIA) presents a complex obstacle in pharmacological safety due to its rare occurrence and unpredictable manifestation, often compounded by class imbalance in clinical datasets. This study investigates the influence of three loss functions, Binary Cross-Entropy (BCE), Focal Loss, and Dice Loss, on the performance of deep learning architectures comprising Multi-Layer Perceptron (MLP), Convolutional Neural Network (CNN), and 2-Layer Neural Network (SimpleNN). Models were trained using numerical molecular descriptors from the publicly available DIA dataset. The architectures were chosen based on their complementary properties: MLP is suitable for high-dimensional tabular descriptor data, CNN was examined to explore whether 1D convolutions can capture localized feature interactions among correlated descriptors, and 2-Layer Neural Network served as a lightweight baseline for comparison. A stratified 5-fold cross-validation strategy was employed to ensure statistical robustness. The results demonstrate that the MLP model, optimized with Focal Loss, consistently delivered the highest classification performance, achieving average scores of 94% accuracy, 93% precision, 95% recall, 94% F1-score, and an AUC of 0.97. In contrast, CNN and SimpleNN architectures yielded less favorable outcomes under the same loss configurations. These findings highlight the importance of aligning loss function choice with model complexity in the context of imbalanced biomedical data. The insights from this work contribute to the development of more reliable computational frameworks for early-phase immunogenicity screening and support the advancement of precision pharmacovigilance strategies.]]>
Copyrights © 2025
