O-014 Clustering identifies distinct subtypes of PCOS – Towards a Rationale Approach to PCOS classification

Abstract

Abstract Study question Does unsupervised clustering identify biologically distinct subtypes in a cohort of women with polycystic ovary syndrome (PCOS) diagnosed by Rotterdam criteria? Summary answer This study demonstrates that unsupervised hierarchical clustering of eight pre-defined quantitative reproductive and metabolic traits identifies biologically distinct subtypes in women with PCOS. What is known already PCOS is a common, heterogeneous, endocrine disorder in women of reproductive-age. PCOS diagnosed by NIH or non-NIH Rotterdam criteria or by self-report is generally genetically similar. Using Hierarchical Clustering (HC), we have previously identified discrete, stable PCOS subtypes, which we designated reproductive (higher LH, FSH, SHBG) and metabolic (higher BMI, insulin, glucose), in a United States cohort of ∼900 PCOS women diagnosed by NIH criteria (Dapas et al. PLoS Med, 2020). The cases that did not cluster were designated “background subtype”. The subtypes appeared to capture biologically meaningful differences because they were associated with distinct and novel genome-wide significant loci. Study design, size, duration In the current study, we applied HC to the same traits (BMI, LH, FSH, DHEAS, SHBG, testosterone, fasting insulin and fasting glucose). We then assessed whether additional traits differed between the subtypes thus identified: anti-Müllerian hormone (AMH), total follicle count, modified Ferriman-Gallwey score, estrogen, TSH, DHEA, cortisol, androstenedione, prolactin, LDL, HDL, triglycerides and cholesterol. Participants/materials, setting, methods Women of European ancestry, aged 13-45 years, n = 2502, with PCOS according to the Rotterdam criteria were included; n = 1067 also fulfilled NIH criteria. All quantitative traits were log-transformed to approximate a normal distribution. Z-scores were used to compare the differences between the three clusters using ANCOVA, corrected for age. Pair-wise comparison of the different clusters was performed using Fisher’s least significance difference method and adjusted for multiple testing. Main results and the role of chance We replicated discrete subtypes in this large cohort of women with PCOS defined by the Rotterdam criteria. There were 1026 cases in the metabolic subtype, 450 cases in the reproductive subtype and 1026 in the background subtype. Cases in the reproductive subtype had significantly (all P < 0.001) higher serum AMH levels, follicle counts and HDL levels compared to the metabolic and background subtypes. These findings suggest that the reproductive subtype captures affected women with the alterations in folliculogenesis characteristic of PCOS, without using PCOM to define this subtype. In contrast, the cases in the metabolic subtype had significantly (all P < 0.001) higher triglyceride and LDL levels compared to the other subtypes providing further evidence that this subtype identifies cases with cardiometabolic risk. Androstenedione and TSH levels were significantly increased in the metabolic subtype compared to the background subtype (P < 0.001) and to both subtypes (P = 0.004), respectively. Cortisol and prolactin levels did not differ among the three subtypes. All results did not differ when the analysis was limited to NIH PCOS cases. Overall, our findings suggest that these PCOS subtypes have different etiologies and clinical outcomes. Subtyping may enable precision medicine approaches to the management of what is currently classified as PCOS. Limitations, reasons for caution A limitation of the study is that we have not replicated these findings in an independent cohort. We have not used an orthogonal method, such as genome-wide association, to confirm that the subtypes capture biologically distinct groups. Wider implications of the findings Taken together with our previous studies suggesting that the genetic architecture of these subtypes differs, the current study implies that PCOS consists of several etiologically distinct disorders. Our findings provide an example of the power of modern disease classification based on objective biologic differences rather than expert opinion. Trial registration number Not applicable