Fig 1.
Overview of the study workflow.
The study workflow consists of three main stages: (1) Unsupervised construction of co-expression networks (GCN) for liver, heart and kidney transplant cohorts; (2) Comparative network analysis to identify conserved functional modules across the three organs; (3) Development and validation of a pan-organ gene signature for allograft rejection using machine learning approaches.
Table 1.
Dataset characteristics and sample distribution across transplant organs and rejection phenotypes used in this analysis. TCMR: T-cell mediated Rejection; ABMR: Antibody-mediated rejection; pABMR: Possible ABMR; pTCMR: Possible TCMR.
Fig 2.
Cross-organ network and module conservation.
(A) Venn diagram showing the overlap of network genes between liver, kidney, and heart. (B) Module similarity heatmaps showing pairwise overlaps between modules: Heart vs. Kidney, Kidney vs. Liver, and Liver vs. Heart, quantified using the Jaccard index. (C) Module similarity network, where nodes represent modules (colored by tissue type) and edges are weighted by their Jaccard indices, illustrating cross-organ module conservation.
Fig 3.
Three-way conserved immune pathway subgroups.
(A) A bar graph showing the number of genes in each of the six three-way conserved subgroups (C1-C6), categorized by their functional enrichment. (B) A bar graph depicting the fold enrichment for each subgroup, highlighting the statistical significance of their conservation across liver, kidney and heart.
Fig 4.
Gene expression heatmaps of cell cycle-related genes across organs.
Heatmaps showing the expression levels of the 24-cell cycle-related genes (BC2) in heart, kidney, and liver biopsy samples.
Table 2.
Predictive performance of the six three-way conserved immune subgroups. The table reports the best classification performance achieved by each conserved immune signature (C1-C6) across three machine learning models: Random Forest (RF), Lasso Regression (Lasso), Support Vector Machine (SVM). Reported metrics include Accuracy (ACC) and Area Under the Curve (AUC), with corresponding standard deviations.
Table 3.
Benchmarking conserved gene signatures against the RATs standard. The predictive performance of different subsets of the conserved gene signature was compared with the established Rejection Associated Transcripts (RATs) panel. Accuracy (ACC) and Area Under the Curve (AUC) are reported for Random Forest models evaluated independently on the liver, heart, and kidney transplant dataset. The 95% confidence intervals (CIs) were calculated as: estimate ± 1.96 × standard error.
Fig 5.
Identification and validation of a minimal universal rejection signature.
(A) Consensus feature importance scores for the top universal biomarkers. (B) The panel showing that each gene is significantly upregulated in rejecting samples across liver, kidney, and heart, along with their overlap with the established RATs panel (shown as *). (C-E) Boxplots of module eigengene activity for the 20-gene signature in liver (C), kidney (D), and heart (E) samples, illustrating clear separation between pathological and normal groups.
Fig 6.
Stratification of lung transplant biopsies using the 20-gene conserved immune signature.
(A-B) Heatmap of z-score-normalized expression of the 20 consensus-ranked immune genes across lung transplant biopsies grouped as (A) Normal vs Pathological, (B) Normal vs pathological subgroups. (C-D) Eigengene expression (first principal component of the 20-gene set) shown for (C) Normal vs Pathological and (D) Normal vs pathological subgroups (T cell injury, inflammatory injury, rejection).