KNOWLEDGE DISTILLATION-DRIVEN DEEP LEARNING FOR MOLECULAR BIOMARKER IDENTIFICATION AND DISEASE CLASSIFICATION

Abstract

The rapid advancement of artificial intelligence (AI) has transformed molecular diagnostics by enabling accurate and automated identification of disease-associated biomarkers. However, state-of-the-art deep learning models often require substantial computational resources, limiting their deployment in real-world clinical and laboratory settings. This study proposes a knowledge distillation-driven deep learning framework for efficient molecular biomarker identification and disease classification. A high-capacity teacher network is employed to transfer discriminative molecular feature representations to a lightweight MobileNetV2-based student model through a multi-level distillation strategy. The proposed framework integrates soft-label supervision and feature-based knowledge transfer to preserve classification accuracy while significantly reducing model complexity and inference time. Performance was evaluated using publicly available molecular biomarker datasets with standard classification metrics, including accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC). Experimental results demonstrate that the distilled student model achieves competitive predictive performance while requiring substantially fewer computational resources than conventional deep neural networks. The proposed approach offers an efficient and scalable solution for molecular diagnostics, precision medicine, and resource-constrained clinical environments. The findings highlight the potential of lightweight knowledge-distilled architectures for reliable biomarker-based disease prediction and support their integration into next-generation intelligent healthcare systems.

This study presents a novel lightweight convolutional neural network (CNN) framework for automated diabetic retinopathy (DR) classification, optimized for real-time deployment on portable devices. While deep learning has significantly advanced DR detection, most state-of-the-art models remain computationally intensive and unsuitable for widespread clinical use in resource-limited settings. Our approach introduces a lesion-aware knowledge distillation strategy that enhances the student model’s ability to recognize critical DR biomarkers by aligning attention maps from a high-capacity teacher model. In addition, a temperature-scaled divergence mechanism improves the transfer of fine-grained class relationships, addressing the limitations of traditional knowledge distillation methods. The proposed model achieves 93.7% accuracy and a 0.912 F1-score, closely matching the teacher model’s performance while using 88% fewer parameters and delivering a 5.8× speed-up in inference. Importantly, it demonstrates substantial gains in early-stage DR detection and lesion-specific sensitivity, enabling timely intervention and supporting human-AI collaboration in clinical settings. These results establish a practical path toward scalable, accurate, and efficient DR screening.