Explainable machine learning and multimodal single cell analysis link NR4A2 to CD8+ T-cell terminal exhaustion and a CD8-adjusted prognostic signature in hepatocellular carcinoma.

BACKGROUND: Terminal exhaustion of CD8+ tumour-infiltrating lymphocytes constrains the benefit of immune-checkpoint blockade in hepatocellular carcinoma (HCC). The orphan nuclear receptor NR4A2 (Nurr1) belongs to a transcription-factor family implicated in murine T-cell exhaustion, yet its behaviour in the human HCC microenvironment has not been characterised across data modalities. We asked whether NR4A2 is specifically coupled to the terminal-exhaustion programme of human HCC CD8+ T cells, and whether that coupling is reflected in chromatin, tissue topography and bulk-tumour outcome data. METHODS: We reanalysed publicly available human HCC datasets spanning four modalities: single-cell RNA sequencing (scRNA-seq; GSE149614, GSE98638, GSE151530, GSE125449), single-cell ATAC sequencing (scATAC-seq; GSE227265), imaging-based spatial transcriptomics (10× Xenium human liver-cancer section) and bulk RNA sequencing with overall-survival annotation (TCGA-LIHC). NR4A2 was profiled across cell compartments and, at higher resolution, across CD8+ T-cell states defined from curated exhaustion, progenitor, effector and naïve gene modules. We tested NR4A2-exhaustion coupling at single-cell and pseudobulk level, examined NR4A2 motif accessibility, traced NR4A2 along a diffusion-pseudotime exhaustion axis, evaluated spatial co-localisation of exhausted CD8+ T cells with tumour-associated macrophages (TAMs), benchmarked machine-learning classifiers of the exhausted state, and compared the human NR4A2 co-expression direction with the published NR4A1/2-knockout phenotype. We additionally trained interpretable gradient-boosted-tree models with SHAP attribution to characterise the exhausted and NR4A2-high states, scored a NR4A2 co-expression regulon, related the programme to immunotherapy-relevant signatures, and modelled overall survival using a CD8-infiltration-adjusted exhaustion score. RESULTS: NR4A2 was broadly expressed across HCC compartments and was highest in myeloid cells rather than being T-cell restricted; at whole-compartment and bulk-tumour resolution it was not a clean exhaustion marker. At the resolution of CD8+ T-cell states, however, NR4A2 was sharply enriched in terminally exhausted cells (mean 3.62 versus 1.05; Mann-Whitney p ≈ 3.4 × 10-83) and correlated with the exhaustion module (single-cell Spearman ρ = 0.45) but not with the effector module (ρ = 0.06), arguing that NR4A2 is not a generic activation gene. NR4A2 expression increased monotonically along a diffusion-pseudotime axis toward terminal exhaustion (ρ = 0.37). In Xenium tissue, exhausted CD8+ T cells were modestly but significantly enriched in the neighbourhood of TAMs relative to other CD8+ T cells (permutation p = 0.005). The human NR4A2 co-expression direction was 83% concordant with the published NR4A1/2-knockout differential-expression direction. In bulk TCGA-LIHC, NR4A2 was lower in tumour than adjacent liver and only weakly exhaustion-correlated, consistent with hepatocyte-dominated bulk signal and reinforcing the need for compartment-resolved analysis. Interpretable machine learning attributed NR4A2-high status to an immediate-early/AP-1 activation programme, a NR4A2 co-expression regulon was selectively active in exhausted cells, and the programme tracked the T-cell-inflamed phenotype (ρ = 0.87); although raw bulk NR4A2 was non-prognostic, a CD8-infiltration-adjusted exhaustion score independently predicted shorter overall survival (hazard ratio 1.42 per standard deviation, p < 0.001). CONCLUSIONS: A NR4A2 co-expression regulon was most active in exhausted cells; explainable modelling linked NR4A2-high status to an immediate-early/AP-1 activation programme; the programme marked immune-hot tumours; and although bulk NR4A2 was non-prognostic, a CD8-infiltration-adjusted exhaustion score was independently associated with worse overall survival (HR 1.42 per SD, p < 0.001). Across four independent data modalities, NR4A2 behaves as a resolution-dependent marker of the terminal-exhaustion programme in HCC CD8+ T cells that is spatially associated with the myeloid niche. The convergence supports NR4A2 as a candidate component of an HCC exhaustion signature and a hypothesis-generating target for functional study, while underscoring that its association is masked in bulk and whole-compartment data.