Authors’ Reply

We thank Matas et al. for their interest in our algorithm for reclassification of acute kidney transplant rejection.¹ We agree that data-driven approaches may add value to the iteratively developed Banff classification.² Moreover, our approach not only provides an easily accessible system for reclassification, it also overcomes the limitations of intermediate and mixed phenotypes, and adds information on disease severity, which is potentially relevant for clinical decision making.

In contrast to what Matas et al. conclude, we do not think that our algorithm’s value is diminished by the similarities between our clusters¹ and Banff categories.² On the contrary, our clusters and visual presentation¹ mathematically support the validity of the Banff classification, because both the Banff system and our approach distinguish rejection subtypes primarily on the basis of tubulointerstitial versus microcirculation inflammation, and donor-specific HLA antibodies (HLA-DSA). Furthermore, the similarities between our clusters and the Banff phenotypes enable clinical interpretation of our clusters (clusters 1 and 4 representing no rejection, cluster 3 representing T cell–mediated rejection, cluster 5 representing antibody-mediated rejection). The main, and clinically interesting, differences between our clusters and Banff are (1) the elimination of the disputed borderline category; (2) the addition of “mixed rejection” cluster 6; and (3) the addition of the novel cluster 2, representing the intriguing phenotype of HLA-DSA–negative microcirculation inflammation.³

How our clusters relate to the unvalidated clusters of the Deterioration of Kidney Allograft Function (DeKAF) study⁴ is more difficult to assess. In contrast to our k-mean clusters, specifically intended for reclassification of acute rejection on the basis of acute lesion scores,¹ DeKAF hierarchic clusters were essentially developed for differentiation of chronic injuries, mostly on the basis of features of chronicity.⁴ Moreover, we neither have access to the exact DeKAF algorithm nor data on the TATR (t-IFTA) lesion, which seems necessary for the DeKAF clustering.

Therefore, we concur that our algorithm, as any other, is not purely data driven but is dependent on the following human choices: (1) the clinical use, i.e., reclassification of acute rejection¹ (versus chronic injury⁴); (2) the method, i.e., k-means clustering, which allows reclassifying new biopsy specimens without the need for retraining¹ (versus hierarchic clustering,⁴ which is often unstable and leads to larger numbers of smaller clusters that are difficult to reproduce); (3) the included features, i.e., acute lesions and HLA-DSA¹ (versus selected, mainly chronic,⁴ lesions but not HLA-DSA⁴); and (4) the biopsy specimens included in training, i.e., early and late protocol and indication biopsy specimens¹ (versus later indication biopsy specimens⁴).

Additionally, and importantly, we chose semisupervised clustering. In our first unsupervised approach, only two main parameters contributed: (1) presence of HLA-DSA, and (2) the extent of tubulointerstitial inflammation (Supplemental Figure 1 and Supplemental Table 1).¹ In unsupervised clustering, the importance of microcirculation inflammation thus seemed underestimated, as compared with extensive literature. By weakly informing the clustering on the relation between individual lesions and graft outcome, which was our final semisupervised approach, we further improved cluster stability and predictive performance and clinical interpretability. We derived the number of clusters from the proportion of ambiguous clustering; we also did not interfere with the features or their weights to compute the clusters.¹

Obviously, our semisupervised approach led to some circular argument in the association between the clusters and outcome in our training cohort (Leuven); it was therefore essential to demonstrate the validity of our locked clustering algorithm, the visual presentation, and the added value beyond Banff, on totally independent cases (Paris and Lyon, in total 3835 biopsy specimens from 1989 patients; see detailed Supplemental Material).¹ Additional validation on non-European cohorts, using our online tool (https://rejectclass.pythonanywhere.com), would be very valuable to further strengthen confidence in the utility of our algorithm for reclassification of kidney transplant rejection.

Disclosures

M. Naesens reports serving as an advisor for the European Medicines Agency and on the editorial boards of several journals. O. Thaunat reports receiving research funding from bioMérieux, Bristol Myers Squibb, Immucor, and Novartis; receiving honoraria from Biotest and Novartis; serving as a scientific advisor for, or member of, the European Society for Organ Transplantation; and having consultancy agreements with Novartis. The remaining author has nothing to disclose.

Funding

None.