In-silico performance, validation, and modeling of the Nanostring Banff Human Organ transplant gene panel using archival data from human kidney transplants

Abstract

Background RNA gene expression of renal transplantation biopsies is commonly used to identify the immunological patterns of graft rejection. Mostly done with microarrays, seminal findings defined the patterns of gene sets associated with rejection and non-rejection kidney allograft diagnoses. To make gene expression more accessible, the Molecular Diagnostics Working Group of the Banff Foundation for Allograft Pathology and NanoString Technologies partnered to create the Banff Human Organ Transplant Panel (BHOT), a gene panel set of 770 genes as a surrogate for microarrays (~ 50,000 genes). The advantage of this platform is that gene expressions are quantifiable on formalin fixed and paraffin embedded archival tissue samples, making gene expression analyses more accessible. The purpose of this report is to test in silico the utility of the BHOT panel as a surrogate for microarrays on archival microarray data and test the performance of the modelled BHOT data. Methods BHOT genes as a subset of genes from downloaded archival public microarray data on human renal allograft gene expression were analyzed and modelled by a variety of statistical methods. Results Three methods of parsing genes verify that the BHOT panel readily identifies renal rejection and non-rejection diagnoses using in silico statistical analyses of seminal archival databases. Multiple modelling algorithms show a highly variable pattern of misclassifications per sample, either between differently constructed principal components or between modelling algorithms. The misclassifications are related to the gene expression heterogeneity within a given diagnosis because clustering the data into 9 groups modelled with fewer misclassifications. Conclusion This report supports using the Banff Human Organ Transplant Panel for gene expression of human renal allografts as a surrogate for microarrays on archival tissue. The data modelled satisfactorily with aggregate diagnoses although with limited per sample accuracy and, thereby, reflects and confirms the modelling complexity and the challenges of modelling gene expression as previously reported.

A Decentralized Kidney Transplant Biopsy Classifier for Transplant Rejection Developed Using Genes of the Banff-Human Organ Transplant Panel

“…In our study, the B-HOT panel appeared to be a high-quality and reliable gene-set for the classification of kidney transplant biopsies, supporting the findings of other research groups investigating the discriminatory value of this knowledge-based panel for kidney transplant diagnostics ( 15 ). An alternative model built with feature selection performed over the whole gene-set failed to outperform the B-HOT panel-based model.…”

Thirty years of the International Banff Classification for Allograft Pathology: the past, present, and future of kidney transplant diagnostics

“…31,[78][79][80]86 Beyond the advantage of being performed on the same sample used for light microscopy, FFPE-based technologies offer the opportunity for large retrospective and longitudinal analyses of archived samples in the setting of decentralized multicenter studies, that is, allow for retrospective randomization with survival end points available. 87 The nCounter system is approved for clinical diagnostics and paired with user-friendly analytical software, thus representing a simple, relatively fast (24-hour turnaround time), automated platform that is well poised to be integrated into the routine diagnostic workflows in existing pathology laboratories while making results reproducible and comparable between laboratories.…”

The MMDx®diagnostic system: A critical re‐appraisal of its knowledge gaps and a call for rigorous validation studies

“…The oncology literature has shown that microarray based assays can result in misclassification rates of 31–49% 13 . Bioinformatics analyses of public transcriptomics data derived from the human allograft kidney generates sample misclassification rates of 27.9–46.9% 14 . Notably, these estimates are in the same range as reported discrepancies between MMDX and histology.…”

“…By comparison, oncology studies that that take into account the interplay of all the steps in the diagnostic pipeline (vertical red arrow) report interinstitutional sample misclassification rates in the 31–49% range 13 . For renal allograft transcriptomics data available in public databases overall classification error rates vary between 27.9 to 46.9% 14 . .…”