Systemic Lupus Erythematosus prediction using Epistatic-Quantile Fusion Transformer network with integrated multi-omics and clinical data

Abstract

Systemic Lupus Erythematosus (SLE) is a complex autoimmune disorder with heterogeneous symptoms and overlapping clinical presentations, making early prediction extremely difficult. Traditional models often fail to integrate high-dimensional multi-omics data and EHR records effectively, primarily due to their inability to handle biological variability, data imbalance, and complex feature dependencies. To address these gaps, the study proposes Epistatic-Quantile Fusion Transformer (EQF-T), a unified framework that introduces multiple novel components. Initially, for pre-processing, the Beta-Variational Rank-ordered Quantile Autoencoder (Beta-VARQA) is used, which combines Beta-divergence, Rank-ordered Quantile Filtering, and Variational Autoencoding to denoise and normalize heterogeneous inputs, retaining biologically significant patterns. For feature extraction, the framework incorporates Epistatic Attention fused Multi-Omics Laplacian Transformer (EA-MLT), which captures intricate dependencies and Epistatic Synergistic effects, essential for understanding the dynamic progression of SLE. This EA-MLT employs Epistatic Attention to capture higher-order gene-gene interactions and integrates the Multi-Omics Laplacian Transformer (MOLT), which uses a Laplacian Attention Mechanism to model structural dependencies across omics layers. The final classification is performed by SLE-Net (SLE Prediction Network), an end-to-end deep learning model designed to analyze fused data and provide interpretable outputs. Together, these components enable EQF-T to effectively learn from complex, high-dimensional biological and clinical data. Further, the proposed model achieves superior performance with 99.82 % accuracy, 99.78 % precision, 99.76 % recall, 99.77 % F1-score, and 99.8 % ROC-AUC, demonstrating its reliability and potential for precise SLE prediction.