Molecular profiling and machine learning risk stratification of graft rejection in kidney transplantation

Overview & Synopsis

Long-term survival of kidney allografts is mainly limited by the occurrence of rejection. To improve the diagnosis of rejection and the characterization of the renal lesions, the Banff classification has become the gold standard in renal transplantation. Based on histological evaluation of renal biopsies, combined with clinical and immunological parameters, the classification now includes molecular biomarkers. Our aim is to improve the molecular characterization of rejection and histological lesions by studying the whole transcriptome in order to enhance the performance of molecular classifiers generated or other transcriptomic methods. Total RNA from 770 biopsies, derived from 540 patients recruited from the KTD-Innov (NCT03582436) and EU-TRAIN (NCT03652402) clinical trials, was extracted and bulk-sequenced by Illumina double-strand sequencing. Genes associated with antibody-mediated rejection (AMR), T-cell mediated rejection (TCMR) and all histological lesions in the Banff classification were determined by differential analysis, along with their associated signaling pathways. Variable selection was performed on the molecular signatures to train 4 machine learning models (Naïve Bayesian Classifier, Linear Support Vector Machine, Extreme Gradient Boosting, K-Nearest Neighbors). An archetypal analysis was then performed on the predictions of the best histological classifiers to obtain unsupervised sample profiles. Finally, a function inference analysis of associated long non-coding RNAs predictive of AMR and TCMR was performed after inference of regulatory networks. Differential analysis identified 6,141 and 8,478 transcripts associated with AMR and TCMR, respectively, of which 603 (9.8%) and 1186 (14%) are described for the first time and absent from the known B-HOT and microarray gene panels. Pathway analysis showed that the B-HOT panel was associated with the main immunological processes involved in AMR and TCMR, whereas the microarray specifically integrates metabolic functions and cell cycle progression processes. Variable selection reduced the number of genes in the range 52 - 867 predictors. NGS-based classifiers demonstrated robust performance (PRAUC of 0.708 to 0.980) for the prediction of Banff lesions, AMR and TCMR. Archetype analysis revealed eight distinct phenotypes, each characterized by distinct clinical, immunological and histological features. The 8 NGS-based archetypes showed distinct allograft survival profiles with additional graft loss rates between archetypes, ranging from 90% to 56% 7 years after evaluation (p