Normalization of nCounter Gene Expression Data Alters Molecular Diagnostics in Kidney Transplantation

Normalization methods affected the performance of gene-based diagnostic classifiers for kidney transplant rejection across a large multicenter cohort. Complex normalization reduced classifier performance by overcorrecting signal, while simpler methods preserved rejection signature and key robustness. By identifying optimal normalization methods, this work advances a standardized preprocessing framework for molecular diagnostics in transplantation.

The Banff 2022 classification endorses intragraft gene expression profiling using the Banff Human Organ Transplant (B-HOT) consensus gene panel for rejection diagnosis. However, lack of standardized analytical pipelines, including data normalization, limits clinical implementation, with its effect on diagnostic performance yet to be determined.

We evaluated ten normalization methods in 868 kidney allograft biopsies from nine European and North American centers, all Banff-graded and B-HOT-profiled on nCounter, comprising derivation ( n =441), internal ( n =186), and external ( n =241) validation cohorts. Each method was assessed through its downstream impact on ( 1 ) gene count stability, ( 2 ) differential expression and cross-platform concordance with RNA sequencing (RNA-seq) data, and ( 3 ) discrimination and calibration of predictive models for antibody-mediated rejection (AMR) and T -cell-mediated rejection (TCMR).

Most methods improved count stability and showed high concordance with RNA-seq for overall gene expression. They also produced robust differential expression signatures consistent with those detected by RNA-seq, except for RUVSeq and RCRNorm , which identified fewer differentially expressed genes and showed lower concordance. In the overall validation cohort ( n =427), diagnostic performance was consistently high across nSolver -based approaches, nanostringr , NanoStringDiff , MetaNorm , and RCRNormFast (AMR, area under the ROC curve [AUROC], 0.88-0.91; area under precision-recall curves [AUPRC], 0.86-0.89; TCMR AUROC, 0.90-0.92; AUPRC, 0.78-0.83). Performance declined with RCRNorm (AMR AUROC/AUPRC, 0.55/0.41; TCMR, 0.53/0.18) and, for TCMR, with RUVSeq (AUROC, 0.84-0.85; AUPRC, 0.64-0.65). Calibration was satisfactory for most methods, except for RCRNorm and for TCMR models after RUVSeq .

Normalization choice significantly impacted gene expression profiles and diagnostic classifier performance. Most methods, including nSolver-based pipelines, achieved robust discrimination for both AMR and TCMR. Complex methods, including RCRNorm, and RUVSeq for TCMR, reduced performance, with simpler approaches consistently outperforming them for B-HOT-based molecular diagnostics.