Deep learning framework for the diagnosis of systemic sclerosis using serum-derived microbial signatures

Abstract

Systemic sclerosis (SSc) is a rare and severe autoimmune disease of unknown etiology, currently lacking effective treatment. Therefore, early diagnosis and timely intervention are critical for improving clinical outcomes. Accumulating evidence implicates the human microbiome as a key contributor to SSc pathogenesis, with distinct microbial alterations identified in patients. Given that circulating anti-microbial immunoglobulin M(IgM) antibodies reflect recent and systemic microbial perturbations across multiple organ systems, comprehensive profiling of serum IgM antibodies against diverse microbial species may offer a promising strategy for predicting early disease development through the identification of SSc-associated microbial signatures. However, predictive diagnostic models incorporating microbiome derived serological data in SSc have not yet been developed. In this study, we introduce a novel and, interpretable deep learning framework based on Graph Attention Networks (GAT) and Convolutional Neural Networks (CNN) to enable accurate diagnosis of SSc using serum anti-microbial IgM antibody data and uncover valid inter-bacterial relationships. We analyzed a dataset comprising 126 individuals (76 SSc and 50 healthy controls). To capture complex inter-species correlations often overlooked by conventional tabular data, we first converted the microbial profiles into a graph structure. The constructed ensemble model, combining GAT for relational feature extraction and CNN for hierarchical pattern recognition, achieved a high diagnostic accuracy of 0.9051, representing an improvement of up to 10.96 percentage points over recent comparable studies. To interpret the model's diagnostic rationale, we applied Association Rule Mining (ARM) to the latent representations from the proposed framework's intermediate layer. This analysis revealed significant clustering patterns, distinguishing groups with specific microbial combinations common in SSc patients, and these patterns that were rarely observed in healthy controls. Our work presents a novel methodology for interpreting deep learning predictions in healthcare and provides initial evidence for the diagnostic potential of this approach in a rare disease setting, using only a small set of core microbial combinations. These findings suggest the potential to enhance screening efficiency for SSc and provide hypothesis-generating insights into biological mechanisms underlying the disease, warranting further validation in larger, independent cohorts.