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

Organ transplantation is a life-saving therapy for end-stage organ failure, but long-term graft survival is threatened by allograft dysfunction, encompassing delayed graft function (DGF), acute rejection, and fibrosis. While organ-specific studies have identified molecular hallmarks, a unifying understanding across different organs is lacking. This study aims to address this gap by creating a comprehensive pan-organ transcriptomic atlas to identify conserved genomic signatures associated with allograft dysfunction and develop a robust prediction framework. The researchers hypothesized that despite organ-specific differences, common pathophysiological genomic signatures exist across different transplanted organs, a concept termed 'pan-organ allograft dysfunction'. This hypothesis was tested by curating publicly available transcriptomic datasets from heart, lung, liver, and kidney transplants, representing DGF, acute rejection, fibrosis, and tolerance. A novel transfer learning framework, named Transferable Omics Prediction (TOP), was developed to leverage information across organs and improve predictive accuracy. The resulting comprehensive dataset, named Pan-organ Resource for Molecular Allograft Dysfunction (PROMAD), was made publicly available to facilitate further research and biomarker development.

Previous research using 'omics' technologies has focused primarily on individual organs, leading to knowledge gaps and inconsistencies in understanding the molecular mechanisms of allograft dysfunction. Studies have shown that markers predictive of dysfunction in one organ often fail to predict dysfunction in others. Although technological advances in multi-omics approaches offer global assessments of tissue microenvironments, limited consensus among organs has hindered the translation of these findings into routine clinical practice. The Banff Human Organ Transplant (BHOT) gene array represents an attempt to address this limitation, but lacks a quantitative capacity for comparing molecular associations across transplant datasets. This study directly addresses this limitation by creating a large-scale dataset that allows for the comparison of gene expression changes across different organs.

The researchers curated publicly available transcriptomic datasets from four major solid organ transplant types (kidney, heart, liver, lung), including microarray, bulk tissue RNA sequencing (RNA-seq), and single-cell RNA-seq data. The resulting PROMAD atlas contained 150 datasets and over 12,000 samples. Data preprocessing involved log transformations, quantile normalization for microarrays, TMM normalization for RNA-seq, and Seurat for single-cell RNA-seq data. Differential gene expression analysis was performed using limma-voom and the Benjamini-Hochberg procedure to control false discovery rate. The directPA package was used to combine test statistics across datasets to identify consistently differentially expressed genes associated with each pathology (DGF, acute rejection, fibrosis, tolerance). The TOP framework, a novel transfer learning approach, was developed to construct robust and organ-agnostic predictive models. This framework leverages a key feature engineering step that uses log-ratio matrices of the most differentially expressed genes and a weighted mean of the log fold changes, accounting for dataset sizes and organ representation. The TOP model was evaluated using leave-one-dataset-out cross-validation (LOOCV), with the area under the receiver operating characteristic curve (AUC) used as a performance metric. The findings were validated using a prospective, single-center kidney and kidney-pancreas transplant cohort (AUSCAD) with clinical, histopathological, and transcriptomic data.

The study identified a set of 158 genes consistently associated with acute rejection across all four organ types, significantly exceeding chance expectation. These genes included chemokines (CXCL9, CXCL10, CXCL11), granzymes (GZMA, GZMB), and cell surface receptors (CD2, CD8A, CD53). Single-cell RNA-seq analysis revealed myeloid cell subsets as the primary cellular source of this pan-organ rejection signal. Analysis of whole blood samples identified 77 genes associated with acute rejection across organs, predominantly involved in inflammation and immune regulation. The TOP framework demonstrated superior classification performance for pan-organ models compared to organ-specific models in predicting acute rejection, achieving mean AUCs of 0.63, 0.74, and 0.71 for heart, kidney, and liver datasets, respectively, compared to 0.55, 0.70, and 0.55 for organ-specific models. Validation in the AUSCAD cohort showed the pan-organ model outperformed both clinical parameters and kidney-specific models in predicting rejection (AUC = 0.81 vs. 0.58 and 0.70, respectively). The study also identified 57 genes associated with allograft fibrosis across organs, with enriched inflammatory pathways. Analysis of a data-driven gene set (500 genes) for global allograft dysfunction outperformed the BHOT panel in predicting DGF in the AUSCAD cohort (AUC = 0.89 vs. 0.79). Finally, the study identified gene signatures associated with allograft tolerance and demonstrated that pre-implantation biopsies could predict DGF, particularly highlighting the beneficial impact of normothermic machine perfusion (NMP) in reducing DGF risk.

This study provides compelling evidence for the existence of shared molecular signatures of allograft dysfunction across different transplanted organs. The development and validation of the pan-organ TOP model demonstrate the power of a transfer learning approach to enhance the accuracy and generalizability of predictive models for allograft pathologies. The identification of myeloid cells as a key player in acute rejection suggests potential therapeutic targets for modulating alloimmunity. The findings related to allograft fibrosis highlight the role of persistent immune activation in chronic transplant dysfunction, opening new avenues for research into targeted immunosuppression strategies. The predictive power of pre-implantation biopsies for DGF, further refined by the inclusion of NMP data, underscores the potential for improving graft outcomes through early assessment and intervention. The establishment of PROMAD as a publicly available resource will greatly facilitate future research in this area.

This study successfully established a comprehensive pan-organ transcriptomic atlas (PROMAD) of allograft dysfunction, identifying conserved molecular signatures across different organs and developing a superior transfer learning prediction model. The findings have significant implications for improving diagnosis, risk stratification, and treatment of allograft dysfunction. Future research should focus on validating these findings in larger, more diverse cohorts, exploring the identified gene signatures as potential biomarkers, and investigating the identified therapeutic targets to further improve the long-term success of organ transplantation.

The study's reliance on publicly available data introduces limitations related to data heterogeneity, varying quality, and potential biases in sample annotations and clinical information. While the LOOCV strategy and AUSCAD validation provide strong support for the findings, further prospective validation across various transplant centers and organ types is needed. The lack of detailed phenotypic data (immunosuppression regimens, donor histories) restricts a more comprehensive analysis of these variables. Despite efforts to standardize data processing, the absence of cross-dataset normalization could potentially influence the findings. The interpretation of rejection could benefit from more detailed phenotyping, which is an ongoing challenge in the field.