A Novel Approach for Discovering New Biomarkers in Autoimmune Diseases Combining Graph Neural Networks with Literature Mining
JOURNAL OF UNIVERSITY OF BABYLON for Pure and Applied Sciences · 30.09.2026 · 34 (3) · 816-831
DOI: 10.29196/jubpas.v34i3.6915 ↗Kurz gefasst
Fachartikel in der Zeitschrift JOURNAL OF UNIVERSITY OF BABYLON for Pure and Applied Sciences, online erschienen am 30.09.2026. Verfasst von Hasanain Flayyih Hasan und 2 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Background
Emerging autoimmune disorders, including post-viral autoimmune sequelae and immune-checkpoint-inhibitor-induced toxicities, present a major clinical challenge because their underlying molecular biomarkers are largely uncharted. Conventional discovery pipelines built on genome-wide association studies and prospective clinical cohorts evolve too slowly to keep pace with the rapid emergence of new immunological phenotypes.
Objective
This study aims to design and evaluate a two-stage Graph Neural Network (GNN) framework that mines large-scale biomedical literature to predict novel biomarker–phenotype associations and to provide mechanistic explanations for each prediction. Materials and Methods: A heterogeneous Biomedical Knowledge Graph (BKG) was constructed from approximately 35 million MEDLINE records using a transformer-based Named Entity Recognition (NER) and Relation Extraction (RE) pipeline. In Phase 1, a Graph Attention Network (GAT) is used to make the initial link prediction. In Phase 2, we use a Heterogeneous Graph Transformer (HGT) re-ranker with a GNNExplainer module to generate a composite confidence score, incorporating structural, semantic and literature-frequency signals. The model was tested using a purely temporal split (trains the relation from before 2024 and tests it from 2024 to 2026).
Results
The Area Under the Curve (AUC) and Precision at 10 (P@10) in Phase 1 are 0.932 and 0.88 respectively, and in Phase 2 are 0.951 and 0.91 respectively, beating the baselines of Adamic-Adar, TransE, GraphSAGE, and Graph Convolutional Network (GCN) statistically. Also, framework generated clinically interpretable mechanistic sub-graphs for top ranked candidates.
Conclusion
The two-stage GNN pipeline is a scalable, interpretable, and accurate strategy to discover in-silico biomarkers in emerging autoimmune diseases, and can significantly speed up the shift from computational hypotheses to experimental testing.
Zitieren
APA
Hasan, H. F., Hussyn, S., & Idan, H. M. (2026). A Novel Approach for Discovering New Biomarkers in Autoimmune Diseases Combining Graph Neural Networks with Literature Mining. JOURNAL OF UNIVERSITY OF BABYLON for Pure and Applied Sciences, 34(3), 816-831. https://doi.org/10.29196/jubpas.v34i3.6915
Vancouver
Hasan HF, Hussyn S, Idan HM. A Novel Approach for Discovering New Biomarkers in Autoimmune Diseases Combining Graph Neural Networks with Literature Mining. JOURNAL OF UNIVERSITY OF BABYLON for Pure and Applied Sciences. 2026;34(3):816-831. doi:10.29196/jubpas.v34i3.6915
BibTeX
@article{Hasan2026,
author = {Hasan, Hasanain Flayyih and Hussyn, Saja and Idan, Hasan Mohammed},
title = {A Novel Approach for Discovering New Biomarkers in Autoimmune Diseases Combining Graph Neural Networks with Literature Mining},
journal = {JOURNAL OF UNIVERSITY OF BABYLON for Pure and Applied Sciences},
year = {2026},
volume = {34},
number = {3},
pages = {816--831},
doi = {10.29196/jubpas.v34i3.6915}
}
EUVIMED