Development and Validation of HistoMX, a Gene-Expression Platform for Molecular Interpretation of Kidney Allograft Biopsies

Overview

This retrospective multicentre observational study developed and validated HistoMX, an automated gene-expression-based platform for multidimensional molecular interpretation of kidney allograft biopsies. The study included 2,410 archived post-transplant kidney allograft biopsy samples from adult kidney transplant recipients. Formalin-fixed paraffin-embedded biopsy tissue was profiled using the Banff Human Organ Transplant panel on the NanoString nCounter platform. HistoMX integrates locked preprocessing, quality control, single-sample normalization, diagnostic classification, Banff lesion-level modelling, molecular lesion-severity scores, composite injury indices, immune-cell estimation, pathway-enrichment analysis, and automated clinician-facing reporting. The platform was evaluated against reference histology based on the Banff classification. HistoMX was designed to complement, not replace, conventional histopathology by providing standardized molecular evidence across diagnostic, lesion-level, continuous injury, and biological dimensions of kidney allograft injury.

Kidney allograft biopsy interpretation remains central to post-transplant care, but conventional histology may not fully capture the biological continuum of alloimmune and non-alloimmune injury. Targeted transcriptomic profiling of formalin-fixed paraffin-embedded biopsy tissue using the Banff Human Organ Transplant panel provides a practical molecular layer aligned with routine pathology workflows. HistoMX was developed as an automated molecular interpretation platform intended to transform B-HOT nCounter expression data into standardized diagnostic probabilities, molecular lesion scores, composite injury indices, immune-cell estimates, pathway-enrichment outputs, and a clinician-facing report. HistoMX was developed and validated using 2,410 post-transplant kidney allograft biopsies from adult recipients across eleven transplant centers in Europe and North America. The population comprised a development cohort (including 7 centers: PITOR \[Saint-Louis, Necker\], Montpellier, Lille, Bordeaux, Toulouse, Barcelona, and Cedars-Sinai Medical Center) and three independent external validation cohorts (Arkana Laboratories, USA, Massachusetts General Hospital, USA, and University of Alberta, Canada). The development cohort included 693 adult kidney transplant recipients contributing to 764 biopsies collected between 2004 and 2021 across eight centres in Europe and North America. This cohort was split, stratified by main histological diagnosis, into a derivation set of 571 biopsies and an internal validation set of 193 biopsies. The derivation set was used for feature selection, model development, hyperparameter tuning, and estimation of preprocessing and normalization parameters. The internal validation set was kept independent from all model-development steps. External validation was performed in three independent cohorts: Arkana Laboratories, USA, with 1,309 biopsies; Massachusetts General Hospital, USA, with 241 biopsies; and the University of Alberta / Alberta University Hospital, Canada, with 96 biopsies. The same eligibility and assay-quality criteria were applied to the external validation cohorts. Repeated biopsies from the same recipient were retained where applicable because the analytical unit was the biopsy, consistent with the diagnostic objective of the study. Eligible samples were post-transplant kidney allograft biopsies from adult recipients submitted for transcriptomic profiling. Exclusion criteria were recipient age younger than 18 years at transplantation, multi-organ transplantation, pre-transplant biopsy, missing Banff reference pathology diagnosis, or failure of prespecified nCounter quality-control criteria. Clinical, histological, immunological, and virological data were extracted from local databases and transferred in anonymized or de-identified form. Biopsies were evaluated according to the Banff classification. For model development and validation, a single reference allograft diagnosis was assigned to each biopsy using the Banff Automation System adapted to the Banff 2022 kidney allograft classification, based on available Banff lesion scores, C4d results, donor-specific antibody status, and virological information. RNA was extracted from formalin-fixed paraffin-embedded biopsy tissue and profiled using the B-HOT panel on the NanoString nCounter platform. The panel includes 770 genes, comprising 758 transplant-relevant endogenous genes and 12 housekeeping genes. After gene-level filtering, 702 endogenous genes and six housekeeping genes were retained for downstream analysis. Validation samples were processed using locked derivation-cohort preprocessing and single-sample normalization parameters. This allowed each biopsy to be analysed independently and limited information leakage between derivation and validation datasets. Three molecular model domains were evaluated: binary classifiers for diagnostic endpoints and binary Banff lesion presence; ordinal classifiers for Banff lesion severity and simplified histological indices; and regression models for weighted continuous histological indices. Diagnostic endpoints included antibody-mediated rejection, T cell-mediated rejection, any rejection, BK virus nephropathy, acute tubular injury without rejection, and no active inflammatory complication. Banff lesion endpoints included glomerulitis, peritubular capillaritis, interstitial inflammation, tubulitis, intimal arteritis, transplant glomerulopathy, interstitial fibrosis, tubular atrophy, vascular fibrous intimal thickening, microvascular inflammation, total inflammation, and interstitial fibrosis/tubular atrophy. Feature selection was performed in the derivation cohort using Boruta. Multiple machine-learning algorithms were trained for each endpoint and combined into equally weighted ensembles. Model performance was evaluated in the internal validation set, the pooled external validation cohort, and cohort-specific external validation analyses. This study did not test a single confirmatory two-group hypothesis with a prespecified effect size. The sample size rationale was based on diagnostic prediction-model development and validation. All eligible B-HOT-profiled biopsies meeting quality-control and reference-diagnosis requirements were included to maximize precision, diagnostic representation, and generalisability. For prevalent diagnostic endpoints, the derivation dataset supported model estimation and cross-validation across major diagnostic categories. For rarer endpoints, including BK virus nephropathy, acute tubular injury without rejection, transplant glomerulopathy, and intimal arteritis, model training and interpretation explicitly accounted for lower event counts through adapted cross-validation and class-imbalance approaches. External validation used all available independent validation cohorts to assess performance, calibration, and transportability across settings. The immunophenotyping substudy included 32 paired biopsies and was used as exploratory supportive biological validation of transcriptomic cell-deconvolution estimates.