Phenotype, serotype, and data-driven clustering reveal complementary dimensions of heterogeneity in ANCA-associated vasculitis: a multicentre Japanese cohort (J-CANVAS)

Anti-neutrophil cytoplasmic antibody (ANCA)-associated vasculitis (AAV) is a systemic autoimmune condition characterised by inflammation of small vessels, primarily driven by ANCAs targeting proteinase 3 (PR3) or myeloperoxidase (MPO). Historically, AAV subtypes, such as granulomatosis with polyangiitis (GPA) and microscopic polyangiitis (MPA), have been distinguished based on clinicopathological findings [1]. However, classification based on the ANCA serotype (PR3-AAV vs. MPO-AAV) has gained increasing prominence for several reasons. This approach is closely linked to underlying genetic variations, enables faster and more clinically relevant diagnoses, and provides independent prognostic insights into relapse risk, and shows strong concordance with the latest classification criteria [2‐5].

The optimal method for classifying AAV remains debated, largely because current systems, based solely on clinicopathological features or ANCA serotypes, may not fully capture patient heterogeneity [6, 7]. While clinicopathological grouping has been reported to be a strong prognostic indicator, significant overlap exists, as illustrated by cases such as MPO-ANCA-positive GPA. Additionally, patient outcomes vary widely within traditional categories; for example, overall survival and relapse-free survival in GPA differ according to ANCA status (PR3-positive, MPO-positive, or negative) [8]. This complexity suggests that relying solely on dichotomous classifications is inadequate and highlights the need for a more integrated framework to better reflect the full AAV spectrum.

Given the limitations of current AAV classification systems and the ongoing debate surrounding them, comparative evaluations of their prognostic value are warranted. Therefore, we aimed to assess the predictive value of traditional phenotype-based classification (GPA/MPA), ANCA serotype (PR3/MPO), and data-driven clustering approaches in a large Japanese AAV cohort. Specifically, we compared these classification methods in relation to key clinical outcomes, including all-cause mortality, renal outcomes, severe relapse, serious infection, and remission rates following induction therapy, using retrospective data analysis that incorporated survival modelling and cluster analysis.

This study addresses the ongoing debate regarding the optimal classification strategy for AAV, a challenge previously described as an “unmet need for a unifying view of the disease spectrum” [7]. In this extensive Japanese cohort, we discovered that the traditional classification based on phenotype (MPA vs. GPA) and ANCA serotype (MPO vs. PR3) each offer unique prognostic insights, indicating that they are complementary rather than competing. However, neither approach fully captured the heterogeneity among patients. In contrast, a data-driven clustering analysis identified four distinct subgroups, which demonstrated limited concordance with conventional classifications, underscoring the limitations of existing classification frameworks. These findings contribute to the ongoing debate on optimal AAV classification and emphasise the need for approaches that more accurately reflect the clinical diversity of patients.

While newer AAV classification criteria increasingly emphasise ANCA serotype [24, 25], our results highlight the enduring prognostic relevance of traditional phenotype classification. Notably, the phenotype distinction (MPA vs. GPA) more effectively stratified all-cause mortality risk in unadjusted analyses. This is particularly important because, while patients with MPA were on average older than those with GPA, the age difference between the MPO-ANCA and PR3-ANCA groups was even greater (Table 1). Despite this, phenotype classification still stratified mortality risk more effectively, suggesting it captures prognostic information not solely attributable to age. Serotype classification also demonstrated distinct associations with organ involvement patterns and relapse trends. Our finding that the clinicopathological classification serves as a powerful prognostic discriminator aligns with a previous cohort study [26]. However, our large cohort allows for a more nuanced interpretation, suggesting that the two systems are not competing but are, in fact, complementary. Regarding relapse risk, there was no clear superiority of either the phenotype or the serotype classification over the other; instead, the additional value of ANCA serotype lay in its ability to further stratify risk within a given clinicopathological phenotype. In contrast to the previous study that reported limited utility for such a combined approach [26], our findings suggest that this integrated perspective is essential for a comprehensive patient assessment.

In this study, unsupervised clustering identified four distinct clinical subgroups (Fig. 4), which showed limited concordance with traditional phenotype or serotype classifications (Online Resource 1, Supplementary Fig. 8). The emergence of data-driven subgroups distinct from established classifications aligns with findings from other clustering studies in AAV, which have reported variable cluster numbers, ranging from four to seven, depending on the cohort and analytical methods used [11, 27‐28]. Collectively, these findings suggest that AAV heterogeneity extends beyond simple binary classification. The limited concordance in our cohort suggests that clinical profile–based clustering reveals heterogeneous dimensions not captured by conventional classification systems. This highlights the limitations of established criteria for research stratification. These robust, clinically-derived endotypes provide multifaceted clinical utility. They define distinct clinical archetypes that may aid diagnostic recognition, and these patterns in turn suggest different underlying biological endotypes that can generate hypotheses for future research. Finally, these different endotypes were associated with distinct clinical outcomes, with statistically significant differences observed in all-cause mortality and composite renal outcomes (Supplementary Table 4). The pattern of this prognostic association may differ from that of other studies, possibly because our analysis included a broad spectrum of both GPA and MPA, in contrast to studies that have found prognostic stratification within a more focused MPA-only cohort [29]. The identification of these robust clinical patterns within a predominantly MPO-ANCA positive cohort is a key finding in itself, demonstrating that the heterogeneity within MPO-AAV is structured and cannot be captured by serotype alone. The clinical relevance of the four identified clusters requires further validation.

Our findings regarding classification comparisons and disease heterogeneity have broader implications for AAV conceptualisation and management. While binary classifications remain practical, they may be inadequate for capturing the full spectrum of patient diversity, particularly among heterogeneous MPO-ANCA-positive patients and those with discordant phenotype-serotype profiles. This highlights the need for more refined stratification approaches and supports the ongoing shift toward precision medicine in AAV [30]. Enhanced stratification may require the integration of multiple data dimensions, including phenotype, serotype, data-driven clusters, and biomarkers, potentially augmented by multi-omics approaches to uncover deeper biological heterogeneity linked to clinical outcomes [31]. Although the direct clinical implications of this descriptive study are preliminary, identifying these distinct patterns lays the groundwork for future research aimed at disentangling AAV heterogeneity to optimise patient care.

This study has a number of limitations. First, its retrospective design and the use of a single-country cohort may limit the generalisability of the findings. Given the known differences in AAV presentation, such as the predominance of MPO-AAV in Japan compared to the PR3-AAV predominance in many European and North American cohorts [32], our findings, including the cluster profiles, may not be directly applicable to other ethnic groups and require validation in diverse populations. However, the large, multicenter nature of our cohort supports the high generalizability of our findings within the Japanese AAV population. Second, the identified clusters require external validation, and the small sizes of certain subgroups, particularly the PR3-MPA group (n = 16), have reduced statistical power for these comparisons. Therefore, the results involving these subgroups should be interpreted with extreme caution, and no definitive conclusions should be drawn from them. Third, the pronounced age differences between groups represent a potential source of confounding in our unadjusted survival analyses. Conversely, our comparative effectiveness analyses for treatment response were designed to mitigate such confounding, using an IPW model that successfully balanced for age (Supplementary Table 3). Fourth, given the retrospective, multicentre design, there is potential for information bias, as the quality and completeness of data abstracted from clinical records may have varied between participating institutions. Despite these limitations, the findings emphasise the importance of future research aimed at validating the clinical relevance of these clusters, investigating the biological underpinnings of MPO-AAV heterogeneity (e.g., through multi-omics approaches), and conducting comparative studies across more diverse populations. Such efforts are essential for integrating multi-dimensional data to improve stratification and enable personalised management for this complex and heterogeneous disease.

In conclusion, this large Japanese cohort demonstrated that traditional phenotype and ANCA serotype classifications provide distinct prognostic insights. However, data-driven clustering revealed additional clinical heterogeneity, particularly within the MPO-ANCA-positive group, that was not captured by conventional classification systems. These findings highlight the need for a more integrative approach to patient stratification in AAV. Incorporating phenotypic, serotypic, and data-driven cluster information is essential to deepen our understanding of disease complexity and advance personalised management for this heterogeneous condition.