The molecular mechanisms operating in human organ transplant rejection are best inferred from the mRNAs expressed in biopsies because the corresponding proteins often have low expression and short half-lives, while small non-coding RNAs lack specificity. Associations should be characterized in a population that rigorously identifies T cell-mediated (TCMR) and antibody-mediated rejection (ABMR). This is best achieved in kidney transplant biopsies, but the results are generalizable to heart, lung, or liver transplants. Associations can be universal (all rejection), TCMR-selective, or ABMR-selective, with universal being strongest and ABMR-selective weakest. Top universal transcripts are IFNG-inducible (eg, CXCL11 IDO1, WARS) or shared by effector T cells (ETCs) and NK cells (eg, KLRD1, CCL4). TCMR-selective transcripts are expressed in activated ETCs (eg, CTLA4, IFNG), activated (eg, ADAMDEC1), or IFNG-induced macrophages (eg, ANKRD22). ABMR-selective transcripts are expressed in NK cells (eg, FGFBP2, GNLY) and endothelial cells (eg, ROBO4, DARC). Transcript associations are highly reproducible between biopsy sets when the same rejection definitions, case mix, algorithm, and technology are applied, but exact ranks will vary. Previously published rejection-associated transcripts resemble universal and TCMR-selective transcripts due to incomplete representation of ABMR. Rejection-associated transcripts are never completely rejection-specific because they are shared with the stereotyped response-to-injury and innate immunity.

We previously reported a system for assessing rejection in kidney transplant biopsies using microarray‐based gene expression data, the Molecular Microscope® Diagnostic System (MMDx). The present study was designed to optimize the accuracy and stability of MMDx diagnoses by replacing single machine learning classifiers with ensembles of diverse classifier methods. We also examined the use of automated report sign‐outs and the agreement between multiple human interpreters of the molecular results. Ensembles generated diagnoses that were both more accurate than the best individual classifiers, and nearly as stable as the best, consistent with expectations from the machine learning literature. Human experts had ≈93% agreement (balanced accuracy) signing out the reports, and random forest‐based automated sign‐outs showed similar levels of agreement with the human experts (92% and 94% for predicting the expert MMDx sign‐outs for T cell–mediated (TCMR) and antibody‐mediated rejection (ABMR), respectively). In most cases disagreements, whether between experts or between experts and automated sign‐outs, were in biopsies near diagnostic thresholds. Considerable disagreement with histology persisted. The balanced accuracies of MMDx sign‐outs for histology diagnoses of TCMR and ABMR were 73% and 78%, respectively. Disagreement with histology is largely due to the known noise in histology assessments (ClinicalTrials.gov NCT01299168).

Discrepancy analysis comparing two diagnostic platforms offers potential insights into both without assuming either is always correct. Having optimized the Molecular Microscope Diagnostic System (MMDx) in renal transplant biopsies, we studied discrepancies within MMDx (reports and sign‐out comments) and between MMDx and histology. Interpathologist discrepancies have been documented previously and were not assessed. Discrepancy cases were classified as “clear” (eg, antibody‐mediated rejection [ABMR] vs T cell–mediated rejection [TCMR]), “boundary” (eg, ABMR vs possible ABMR), or “mixed” (eg, Mixed vs ABMR). MMDx report scores showed 99% correlations; sign‐out interpretations showed 7% variation between observers, all located around boundaries. Histology disagreed with MMDx in 37% of biopsies, including 315 clear discrepancies, all with implications for therapy. Discrepancies were distributed widely in all histology diagnoses but increased in some scenarios; for example, histology TCMR contained 14% MMDx ABMR and 20% MMDx no rejection. MMDx usually gave unambiguous diagnoses in cases with ambiguous histology, for example, borderline and transplant glomerulopathy. Histology lesions or features associated with more frequent discrepancies (eg, tubulitis, arteritis, and polyomavirus nephropathy) were not associated with increased MMDx uncertainty, indicating that MMDx can clarify biopsies with histologic ambiguity. The patterns of histology‐MMDx discrepancies highlight specific histology diagnoses in which MMDx assessment should be considered for guiding therapy.

BackgroundThe relationship between the donor-derived cell-free DNA fraction (dd-cfDNA[%]) in plasma in kidney transplant recipients at time of indication biopsy and gene expression in the biopsied allograft has not been defined.MethodsIn the prospective, multicenter Trifecta study, we collected tissue from 300 biopsies from 289 kidney transplant recipients to compare genome-wide gene expression in biopsies with dd-cfDNA(%) in corresponding plasma samples drawn just before biopsy. Rejection was assessed with the microarray-based Molecular Microscope Diagnostic System using automatically assigned rejection archetypes and molecular report sign-outs, and histology assessments that followed Banff guidelines.ResultsThe median time of biopsy post-transplantation was 455 days (5 days to 32 years), with a case mix similar to that of previous studies: 180 (60%) no rejection, 89 (30%) antibody-mediated rejection (ABMR), and 31 (10%) T cell–mediated rejection (TCMR) and mixed. In genome-wide mRNA measurements, all 20 top probe sets correlating with dd-cfDNA(%) were previously annotated for association with ABMR and all types of rejection, either natural killer (NK) cell–expressed (e.g., GNLY, CCL4, TRDC, and S1PR5) or IFN-γ–inducible (e.g., PLA1A, IDO1, CXCL11, and WARS). Among gene set and classifier scores, dd-cfDNA(%) correlated very strongly with ABMR and all types of rejection, reasonably strongly with active TCMR, and weakly with inactive TCMR, kidney injury, and atrophy fibrosis. Active ABMR, mixed, and active TCMR had the highest dd-cfDNA(%), whereas dd-cfDNA(%) was lower in late-stage ABMR and less-active TCMR. By multivariate random forests and logistic regression, molecular rejection variables predicted dd-cfDNA(%) better than histologic variables.ConclusionsThe dd-cfDNA(%) at time of indication biopsy strongly correlates with active molecular rejection and has the potential to reduce unnecessary biopsies.Clinical Trial registration number:NCT04239703