We annotated the top transcripts associated with kidney transplant rejection by p-value, either universal for all rejection or selective for T cell–mediated rejection (TCMR) or antibody-mediated rejection (ABMR; ClinicalTrials.gov NCT01299168). We used eight class-comparison algorithms to interrogate microarray results from 703 biopsies, 205 with rejection. The positive comparators were all rejection, TCMR, or ABMR; the negative comparators varied from normal biopsies to all nonrejecting biopsies, including other diseases. The universal algorithm, rejection versus all nonrejection, identified transcripts mainly inducible by interferon γ. Selectivity for ABMR or TCMR required the other rejection class as well as nonrejection biopsies in the comparator to avoid selecting universal transcripts. Direct comparison of ABMR versus TCMR yielded only transcripts related to TCMR, the stronger signal. Transcripts highly associated with rejection were never completely specific for rejection: Many were increased in biopsies without rejection, reflecting sharing between rejection and injury-induced innate immunity. Union of the top 200 transcripts from universal and selective algorithms yielded 454 transcripts that permitted unsupervised analysis of biopsies in principal component analysis: PC1 was rejection, and PC2 was separation of TCMR from ABMR. Appreciating rejection-associated molecular changes requires a diverse case mix, accurate histologic classification (including C4d-negative ABMR), and both selective and universal algorithms.
Transcripts associated with kidney allograft rejection, either universal for all rejection or selective for T cell–mediated or antibody-mediated rejection, are highly conserved when the algorithm, the case mix, the definitions of rejection, and the technology platform are conserved, but they are never completely specific for rejection.