Simulated Annealing Identifies Five Shared Drivers of T Cell Exhaustion Across Four Human Cancers

T cell exhaustion is a major barrier to effective cancer immunotherapy. While individual exhaustion markers such as PDCD1 and TOX have been extensively studied, the shared drivers across different cancer types remain poorly defined. Identifying conserved exhaustion drivers could provide pan-cancer therapeutic targets. We analyzed single-cell RNA-seq data from four human cancers: hepatocellular carcinoma (HCC), colorectal cancer (CRC), melanoma, and non-small cell lung cancer (NSCLC). We applied three complementary computational approaches weighted gene co-expression network analysis (WGCNA), XGBoost-based feature importance, and simulated annealing (SA) for optimal gene subset selection. Pathway enrichment analysis was performed using KEGG, Reactome, and Gene Ontology (GO) databases. We developed a simulated annealing framework that outperformed WGCNA and XGBoost in identifying conserved exhaustion drivers. SA identified five shared drivers (TOX, PDDC1, HAVCR2, TIGIT, CXCL13) across all four cancers, and twenty novel candidate genes including ITM2A, TNFRSF1B, COTL1, SLA, and PTPN22 that have not been previously linked to exhaustion. Notably, NR4A1 a widely reported exhaustion driver was not selected in any cancer type, challenging its role as a shared driver. HCC showed a distinct exhaustion signature compared to other cancers. Our study provides a new computational framework (SA) for identifying exhaustion drivers and reveals twenty novel candidate genes. The five shared drivers represent pan-cancer targets, while the novel genes open new avenues for research. The unexpected exclusion of NR4A1 suggests tissue-specific rather than shared roles in exhaustion.