Novel Phenotypes for Acute Kidney Transplant Rejection Using Semi-supervised Clustering.

Abstract

We note that Vaulet et al. confirmed the Deterioration of Kidney Allograft Function (DeKAF) studies that (1) identified Banff histologic clusters with differing postbiopsy graft survival; and (2) showed that markers of antibody-mediated rejection affect graft survival, even after adjusting for inflammation (e.g., Banff inflammation, tubulitis, glomerulitis [g], peritubular capillaritis scores). In spite of methodologic similarities, Vaulet et al. neglected to cite DeKAF publications that provide insights into cluster phenotypes of late allograft pathologies. The clustering studies differed as follows: (1) DeKAF considered all histologic lesions (acute and chronic) for clustering, whereas Vaulet et al. only considered those associated with acute inflammation plus C4d, donor-specific antibodies, and thrombotic microangiopathy; (2) DeKAF included indication biopsy specimens, whereas Vaulet et al. included both indication and surveillance biopsy specimens; (3) for DeKAF, the median time from transplant to biopsy was 5.7 years, whereas 83.3% of indication biopsies were in the first year in the Vaulet et al. study (median, 22 days post-transplant); and (4) DeKAF’s unsupervised clustering identified six primary clusters, whereas Vaulet et al. identified four. However, because they found the “histologic and clinical relevance” of their unsupervised clusters to be unclear, the authors then used a semisupervised clustering approach— weighing the histologic features with survival information— and identified six clusters. There are striking similarities between clusters in the studies. Both studies identified a cluster withmild fibrosis but little inflammation, both identified a cluster consistent with acute T cell–mediated rejection, and both identified a cluster with “g” as the predominant histologic score. For Vaulet et al., these three clusters were similar to the remaining three, except for the absence of donor-specific antibodies; for DeKAF, differences in both acute and chronic lesions differentiated clusters. There are aspects of Vaulet et al.’s findings that limit their clinical relevance. First, the identified clusters were not purely data driven, but were determined on the basis of phenotypes they expected to exist. They chose to disregard results with three or fewer clusters because they were “not helpful to describe different phenotypes.” Indeed, the authors rejected their unsupervised clusters because they did not “reflect the clinical reality andprevious knowledgeon the relevance of these lesions [g, peritubular capillaritis, and C4d] and [antibodymediated rejection].” Hence, the resulting clusters can only be interpretedwithin the rigid framework imposed by the authors. Secondly, we question the validity of their conclusion that they “showed statistically improved prediction of graft failure with the clustering approach compared with using the Banff categories.” The cluster features were weighted on the basis of graft survival, and their prognostic significance canonly be evaluated in external data. Although we agree with the authors that phenotypes beyond the Banff categories may exist, the clusters identified by Vaulet et al. only confirm the importance of lesions already known to be associated with graft failure.