Decoding the hallmarks of allograft dysfunction with a comprehensive pan-organ transcriptomic atlas

Organ transplantation improves survival and quality of life for end-stage organ failure, yet long-term graft survival remains limited by allograft dysfunction, driven by ischemia–reperfusion injury (manifesting as delayed graft function, DGF), adaptive immune activation leading to rejection, and maladaptive repair culminating in fibrosis. Although organ-specific omics studies—especially in kidney transplantation—have identified molecular hallmarks and candidate biomarkers, their clinical adoption has been limited by lack of cross-organ consensus, technological heterogeneity (microarray, RNA-seq, NanoString), and reliance on observer-dependent histopathology. Prior attempts, such as the Banff Human Organ Transplant (BHOT) gene array, provided curated markers but lacked a quantitative framework for comparing molecular associations across organs and datasets. This study posits a ‘pan-organ allograft dysfunction’ concept, hypothesizing that key pathophysiological genomic signatures are organ-agnostic. The authors curated a large, multi-organ transcriptomic atlas (PROMAD) spanning DGF, acute rejection, fibrosis, and tolerance, and developed a transfer learning framework to identify conserved signatures and build organ-agnostic predictive models, ultimately aiming to enhance biomarker development and clinical monitoring across transplant types.

The paper synthesizes prior evidence that organ-specific omics signatures often fail to generalize across allografts, with discordant performance when applied to different organs. It highlights technological variability (microarray, RNA-seq, NanoString) complicating cross-study comparison and references the BHOT panel as a curated, consensus gene set reflecting allograft injury. It also cites clinical and molecular studies on rejection (kidney, heart, lung, liver) and fibrosis, underscoring persistent inflammation as a driver of dysfunction. Limitations of current diagnostics (histology, clinical parameters like creatinine/eGFR) and mixed external validation performance are noted, motivating comprehensive, quantitative integration across datasets and organs to resolve shared mechanisms and improve biomarker transferability.

Data curation: A systematic search of GEO and ArrayExpress (terms: kidney/liver/lung/heart transplant, allograft) identified 13,419 entries; after filtering for human studies, transcriptome-wide assays, adequate controls (stable grafts), sufficient features (≥8,000), sample size, and metadata availability, 150 transcriptomic datasets (microarray, bulk RNA-seq, single-cell RNA-seq) totaling 12,970 samples were included in the PROMAD atlas. Processed datasets (n = 168) are accessible via the PROMAD portal.

Preprocessing and differential expression: For microarrays, intensity data were log-transformed and quantile-normalized; differential expression (DE) used limma with eBayes and BH FDR 5%. For RNA-seq, unnormalized counts were filtered, TMM-normalized, and analyzed via limma-voom with BH FDR 5%. For scRNA-seq, counts were log-normalized; cells were annotated via Seurat/Azimuth, with T cell subtyping by PCA/Louvain and marker-based labeling. Single-cell batches were integrated using reciprocal PCA anchors and projected with t-SNE.

Cross-dataset integration of DE: Moderated test statistics per gene were transformed to z-scores and combined across datasets using Stouffer’s method (directPA). Expected overlaps of DE genes across organs were computed from marginal probabilities with chi-squared testing for enrichment beyond chance. Pathway enrichment used Wilcoxon rank-sum tests on combined P values (KEGG/Reactome), and GSEA for directional inference (clusterProfiler). Cell-type origins were inferred using Cepo statistics to compute enrichment of meta-analysis genes across annotated cell types, comparing acute rejection vs stable conditions by Wilcoxon tests.

Transferable Omics Prediction (TOP) framework: A reference-free, transfer learning approach (Bioconductor TOP) was developed to build cross-cohort, cross-platform models. Workflow: (1) Select top features (default top 50 DE genes) common across training datasets; (2) Construct log-ratio features (pairwise gene expression log differences) within each dataset to self-normalize and improve transferability; (3) Compute per-dataset log fold changes for each ratio relative to binary outcomes; (4) Combine across datasets using inverse-sample-size-weighted means and variances, with a fudge factor (90th percentile variance) to stabilize test statistics; (5) Derive transformed feature weights and apply organ/sample weights to balance organ representation (inverse of organ dataset counts); (6) Fit a weighted lasso regression on concatenated ratio matrices to yield sparse, interpretable models. Rationale: batch and organ are confounded; reference-free, ratio-based features avoid standard batch correction and enable cross-platform application.

Phenotypes and model evaluation: Rejection was defined as a composite of TCMR, ABMR, and mixed phenotypes. Models were trained and evaluated with leave-one-dataset-out cross-validation (LOOCV). Organ-specific models (trained only on datasets from the target organ) were compared to pan-organ models (trained on all organs), with AUC as performance metric. Weighting schemes were tested to balance organ contributions and avoid dominance by kidney datasets. Cross-platform robustness (microarray vs RNA-seq) was compared against naive normalization and ComBat.

Validation cohorts: AUSCAD prospective cohort (Westmead Hospital, Australia): kidney and kidney-pancreas recipients with paired 3-month protocol biopsies and blood RNA-seq (n=70 with blood; additional biopsy cohorts at implantation and 3 months; total samples across analyses detailed in Methods). Histology followed Banff 2019 criteria. External validation also included pre-implantation biopsies (n=279 across 7 datasets) for DGF prediction and paired samples before/after normothermic machine perfusion (NMP) to assess predicted risk changes.

Gene panels: Performance of the curated BHOT panel (770 genes) was compared with a data-derived 500-gene panel identified from PROMAD as overexpressed across DGF, rejection, and fibrosis. Panels were used as predefined feature sets within TOP models for AUSCAD validation of DGF, rejection, and fibrosis classification.

• Resource: PROMAD aggregates 150 datasets and 12,970 samples across kidney (n=8,853), heart (n=1,160), lung (n=1,241), and liver (n=1,216) using microarray, bulk RNA-seq, and single-cell RNA-seq; data and portal publicly available.
• Acute rejection (biopsies): Across 54 datasets (40 kidney, 5 lung, 5 liver, 4 heart), 158 genes were consistently differentially expressed across all four organs—~20-fold more than the 8 expected by chance (P = 5.44 × 10^−11). Key markers included chemokines (CXCL9, CXCL10, CXCL11), granzymes (GZMA, GZMB), and receptors (CD2, CD8A, CD53). Single-cell analyses localized the pan-organ rejection signature predominantly to myeloid subsets (monocytes/macrophages) with enrichment in rejecting grafts.
• Acute rejection (liquid biopsy): In 23 blood datasets (18 kidney, 3 heart, 2 liver), 77 genes were consistently associated with rejection across organs at combined P < 1×10^−5. Genes implicated inflammation (CASP1, CASP4, IRF4) and immune regulation (CD28, CD36, FCER1G). Single-cell blood data showed overexpression mainly in CD14+ monocytes.
• Transfer learning (TOP) performance for blood-based rejection: Pan-organ models outperformed organ-specific models. Mean AUCs: organ-specific heart 0.55, kidney 0.70, liver 0.55 vs pan-organ heart 0.63, kidney 0.74, liver 0.71. Effective models required as few as 50 gene ratios. Equal organ weighting improved performance and mitigated kidney-dominance. TOP’s ratio-based normalization enhanced cross-platform transferability compared with naive normalization and ComBat.
• AUSCAD validation (blood, 3 months): Pan-organ liquid biopsy model predicted rejection with AUC = 0.81, outperforming clinical parameters (creatinine, eGFR, albumin; AUC = 0.58) and kidney-specific models (AUC = 0.70).
• Fibrosis: From 14 allograft datasets (kidney, liver, lung), 57 genes were differentially expressed across organs at combined P < 1×10^−7, featuring inflammation and immune recognition (CASP1, TLR7, TNFAIP8, CD27, CD52, CD74) and HLA. Prospective datasets showed concordance between genes predictive of future fibrosis and those expressed in established fibrosis (R^2 = 0.21, p < 0.0001), indicating conserved processes. Comparison with native organ fibrosis revealed distinct immune-related pathway activation in allografts (e.g., interferon signaling, TCR activation). Single-cell analyses in kidney fibrosis showed enrichment in T cells and macrophages.
• Global dysfunction gene set vs BHOT: BHOT separated rejection and fibrosis but had limited DGF differentiation. A data-derived 500-gene panel (400 not in BHOT) captured DGF, rejection, and fibrosis across organs. In AUSCAD biopsy validation: DGF AUC 0.89 (data-derived) vs 0.79 (BHOT); rejection AUC 0.93 vs 0.90; fibrosis AUC 0.81 vs 0.83.
• Tolerance: Across eight datasets (5 whole blood, 3 PBMC), 38 genes (blood) and 45 genes (PBMC) characterized tolerance, implicating suppression of immune responses and regulation of T cell proliferation. Pan-organ tolerance models outperformed organ-specific models.
• Pre-implantation prediction of DGF: TOP trained on seven pre-implantation biopsy datasets (n=279) achieved AUC = 0.89 for predicting DGF/primary non-function (kidney and liver). Top predictors included immune surface markers (CD3D, CD48, CD52, CD72). Predicted DGF risk decreased significantly after brief (<2 h) NMP compared with longer (>6 h) NMP across paired datasets, indicating potential therapeutic modulation captured by the model.

The study addresses the central hypothesis that key molecular drivers of allograft dysfunction are conserved across organs. By quantitatively integrating 150 datasets spanning four solid organs and multiple technologies, the authors identified shared transcriptomic signatures for acute rejection, fibrosis, DGF, and tolerance. The findings that monocyte/macrophage lineages are primary sources of the pan-organ rejection signal align with mechanistic data from preclinical models and targeted human studies, consolidating the role of myeloid activation and APC–T cell interactions in rejection. Blood-based signatures mirrored tissue findings, enabling minimally invasive monitoring.

The TOP transfer learning framework demonstrated that cross-organ feature learning yields more accurate and generalizable diagnostics than organ-specific models, with robust performance across platforms using ratio-based, self-normalizing features. This pan-organ strategy outperformed clinical metrics and existing curated panels (e.g., BHOT) in several applications, including rejection detection and DGF prediction from pre-implantation biopsies, and captured changes associated with NMP, highlighting translational potential for clinical decision support, donor organ assessment, and therapeutic optimization.

Importantly, the study disentangles transplant-associated fibrosis from native organ fibrosis, emphasizing persistent immune activation as a driver of chronic allograft injury and suggesting immunomodulatory interventions. The PROMAD atlas and associated web resource support standardized evaluation of biomarkers and facilitate discovery and validation efforts, potentially accelerating the development of pan-organ diagnostics and risk stratification tools.

This work delivers a comprehensive pan-organ transcriptomic atlas (PROMAD) and a transfer learning framework (TOP) that together reveal shared molecular hallmarks of allograft dysfunction across heart, lung, liver, and kidney transplants. The authors identify conserved rejection and fibrosis signatures with a myeloid-cell origin, develop organ-agnostic blood-based predictors that surpass organ-specific models and standard clinical measures, and validate performance in a prospective cohort. A data-driven global dysfunction gene set complements and, in some tasks, outperforms the BHOT panel. The ability to predict outcomes from pre-implantation biopsies and to detect changes with NMP underscores clinical utility. Future research should prospectively validate pan-organ models across diverse centers and organs, perform experimental mechanistic studies to refine cell-type-specific targets (particularly within myeloid and T cell compartments), and integrate advanced normalization/embedding methods to uncover additional transcriptomic associations. PROMAD provides a foundational resource for developing pan-organ biomarkers, diagnostics, and therapeutics.

• Reliance on publicly available datasets limits phenotypic granularity (e.g., immunosuppression regimens, donor histories) and sample annotations, particularly for detailed rejection subtypes, potentially introducing heterogeneity.
• Absence of cross-dataset normalization due to confounding of batch and organ; although TOP mitigates this via ratio-based features, alternative normalization/embedding approaches might reveal additional associations.
• Findings are primarily computational; limited experimental validation and prospective multi-organ clinical validation beyond AUSCAD constrain generalizability.
• Variations in pathology classification across studies can affect comparative biomarker performance; while the authors used composite definitions and LOOCV, standardized prospective datasets are needed.
• Kidney datasets dominate the atlas; although organ weighting was applied, residual imbalances may influence some signals.