Graph Neural Network-Driven Discovery Of Exhaustion-Associated Gene Signatures In Single-Cell Rna Sequencing Data From Cancer Immunotherapy

T-cell exhaustion represents a major obstacle to successful cancer immunotherapy, yet identifying robust, patient-agnostic gene signatures from heterogeneous single-cell RNA sequencing (scRNA-seq) data remains challenging due to inter-patient variability, sparse expression patterns, and the need for interpretability. We developed ExhaustionGNN, an end-to-end graph neural network (GNN) pipeline that integrates patient-aligned gene co-expression networks with multi-modal feature aggregation for exhaustion signature discovery. ExhaustionGNN achieved superior gene-level AUC (۰.۸۹ ۰.۰۴) and patient-level discrimination (AUC = ۰.۸۲) compared to baselines. It recovered ۱۸/۲۰ known exhaustion markers in the top ۵۰ ranks and identified ۱۲ novel candidates (CXCR۶, NR۴A۱, ITGAE) validated in independent cohorts. Network analysis revealed three exhaustion modules: inhibitory receptors, transcription factors, and effector molecules. GNNExplainer identified PDCD۱-LAG۳-TIGIT as the core predictive triplet (edge importance ۰.۸۵). Heatmaps demonstrated consistent elevation of top-ranked genes in non-responders. ExhaustionGNN provides a scalable, interpretable framework for scRNA-seq signature discovery, outperforming traditional methods by ۱۳–۲۲% in recovery of known markers and revealing biologically coherent modules missed by differential expression alone. The pipeline is directly applicable to personalized immunotherapy design and multi-omics integration.

Keywords:

single-cell RNA-seq , T-cell exhaustion , graph neural networks , graph attention network , biomarker discovery , immunotherapy , network medicine