Explainable Artificial Intelligence and Gene-Network Analysis Prioritise a TREM2/LGALS3/GPNMB Scar-Associated Macrophage Programme in Human Liver Cirrhosis

Macrophages organise the fibrotic niche of liver cirrhosis. The scar-associated macrophage state is well described, but the genes that define it have rarely been ranked in an unbiased, model-based way, and the statistical limits of such re-analyses are seldom reported. Here we set out to re-derive the scar-associated macrophage programme without supervision, to rank its constituent genes by convergent computational evidence, and to state explicitly what a cohort of this size can and cannot support. We re-analysed 58,358 single-cell transcriptomes from five uninjured and five cirrhotic human livers (GEO GSE136103; Edinburgh DataShare DS_10283_3433). After quality control, Harmony batch integration and Leiden clustering, the mononuclear-phagocyte compartment (9869 cells) was re-clustered. We applied donor-level compositional testing stratified within the CD45+/CD45− sorted fractions, donor-level pseudobulk differential expression, three classifiers with SHapley Additive exPlanations (SHAP) under donor-grouped cross-validation, transcription-factor differential expression, diffusion pseudotime, ligand–receptor modelling and CellOracle in silico perturbation. The donor, never the cell, was the unit of analysis for every condition comparison. A discrete macrophage cluster (390 cells; 60.5% cirrhotic; contributed by all 10 donors) carried a coherent lipid/scar programme comprising LGALS3, GPNMB, TREM2, CD9, FABP5, APOE, APOC1 and CTSD, and was recovered in all 22 clustering configurations tested. The integrated prioritisation was robust: seven genes remained in the top 20 in 100% of 1000 Dirichlet re-weightings. MAFB, MITF and TFEC were significantly enriched transcription factors, and pseudotime placed the state downstream of a monocyte root. Three findings temper these results. No lineage showed a significant compositional change after multiple-testing correction, and every test was underpowered for its observed effect. No gene reached significance in donor-level pseudobulk differential expression. Classifier performance of 0.99 reflects the circularity of predicting a cluster defined from the same matrix; the non-circular contrast of cirrhotic versus uninjured cells gave 0.61–0.65. Overall, a documented, fully reproducible framework re-derives and ranks a TREM2/LGALS3/GPNMB scar-associated macrophage programme and nominates galectin-3, GPNMB and the SPP1 axis as prioritised, clinically unproven candidates. The prioritisation is a hypothesis requiring independent, spatial, protein-level and functional validation.