Pathogenic variation in insulin resistance genes is common in polycystic ovary syndrome (PCOS): a strategy for causal gene discovery using whole-exome sequencing (WES) in complex traits

Abstract

Polycystic ovary syndrome (PCOS) is a complex, multi-system, heritable endocrinopathy that is a common cause of anovulatory infertility in reproductive-aged women. While insulin resistance (IR) is not a diagnostic feature, it is widespread in women with PCOS, and often more severe than in women of similar age and BMI. Conversely, women with rare Mendelian disorders of IR also present with features of PCOS. We hypothesize that PCOS is driven by underlying IR, which can be evaluated through a genetic approach. We curated and stratified 310 genes related to three mechanisms of IR using molecular and clinical criteria. We evaluated protein-altering genetic variation in 102 insulin signaling genes, 29 obesity genes, and 22 dyslipidemia genes from whole-exome sequencing data from 675 PCOS patients. 40 insulin signaling genes, 12 obesity genes, and 10 dyslipidemia genes were significantly enriched for protein-altering variation in PCOS cases compared to healthy population controls. Variants in these 62 significantly enriched genes affected 51% of PCOS cases in our study cohort. The 15 highest ranked genes were selected for follow-up: LMNA, LEPR, KCNJ11, BSCL2, ACACA, NTRK2, GCK, ABCC8, SLC2A2, POMC, MC4R, TBC1D4, INSR, NR0B2, and GCKR. 50% of variants identified in these 15 genes were pathogenic, 35% were likely pathogenic, and only 15% were variants of uncertain significance. These findings support IR as a central pathway in PCOS. Furthermore, this study demonstrates that a candidate pathway approach with sufficient pre-processing can successfully identify functionally relevant variants and genes underlying complex traits.

Introduction

Polycystic ovary syndrome (PCOS) (MIM #184700) is a common, complex, multi-system endocrinopathy and the leading cause of anovulatory infertility in reproductive-aged women (1–3). The defining diagnostic criteria are hyperandrogenemia, irregular menses, and polycystic ovarian morphology. In addition to these cardinal features, PCOS patients exhibit metabolic impairments such as high rates of obesity, dyslipidemia, insulin resistance (IR), and type 2 diabetes (T2D) (4). Furthermore, IR in women with PCOS is more severe than in age- and BMI-matched women without PCOS (4). PCOS has a strong genetic component as established by twin studies and pedigree analyses (5, 6). Genome-wide association studies (GWAS) have revealed numerous candidate loci (7, 8); however, these only account for a small portion of the heritability. Thus, rare, high-effect variants may contribute to the missing heritability (9).

Genetic studies of complex traits are expanding their investigations beyond GWAS towards comprehensive sequencing technologies such as whole-exome sequencing (WES) to capture rare coding variants that may underlie disease etiology. Examples include Alzheimer’s disease (reviewed in (10)), chronic obstructive pulmonary disease (COPD) (reviewed in (11)), and endometriosis (reviewed in (12)). Notably, some complex diseases have related Mendelian, more severe forms. Highly penetrant, monogenic forms of a trait with well-characterized heritability can serve as informative models for the overarching complex trait. Such examples include monogenic familial hypercholesterolemia within the broader trait of high cholesterol levels (reviewed in (13)) and monogenic maturity-onset diabetes of the young (MODY) in relation to T2D (reviewed in (14)). Like these examples, PCOS may similarly encompass rare, monogenic forms of disease that contribute to its genetic and phenotypic heterogeneity. Genetic sequencing studies by our lab and others have identified several prominent genes with rare, high-effect variants accounting for a small portion of heritability (15–18). Importantly, we recently identified rare variants in LMNA significantly associated with PCOS cases in two independent cohorts (16). Rare pathogenic variants in LMNA are associated with a number of pathologies, but most relevant to PCOS is Dunnigan’s Familial Partial Lipodystrophy (FPLD2) (MIM #151660), a disorder of lipid storage and severe IR that also presents with hyperandrogenemia and irregular menses in female patients (19, 20). Therefore, we hypothesize that within the pool of patients presenting with the features of PCOS, there exists a portion of women who have underlying monogenic disorders, which could include other disorders of IR. Identifying these patients can aid in their targeted treatment and inform the overall genetic architecture of PCOS.

With the advent of sequencing technologies including WES, generation of sequence data has become increasingly accessible and rapid; however, the ability to interpret findings from this growing amount of genetic data remains a challenge, especially outside the context of Mendelian disease (21). This contributes to the accumulation of variants of uncertain significance (VUS) which have limited diagnostic or clinical utility (22). Because of this phenomenon, our efforts to apply WES to identify rare, monogenic contributors to PCOS requires substantial pre-processing of genes to ensure that the variants we identify are interpretable and relevant to the trait.

Given the phenotypic overlap between IR and PCOS, investigating the genetic architecture of IR may reveal novel genes relevant to PCOS etiology. IR is a complex condition attributable to multiple biological systems. Molecular mechanisms resulting in IR include direct impairment of the insulin signaling pathway, indirect effect through either impaired energy homeostasis resulting in obesity, or impaired adipose storage resulting in dyslipidemia (23, 24). Numerous genetic causes of severe IR have been identified which aid in these categorizations (24). While several studies have examined common GWAS variants associated with these pathologies in the context of PCOS (25–29), few have investigated IR genes for coding variants using WES. The most comprehensive analysis to date of IR genes in the context of PCOS was completed in a small cohort pre-selected for IR (15). The present study is the first to do so in a comprehensive, unbiased manner in a cohort not pre-selected for any metabolic features.

To test the hypothesis that genetic forms of IR contribute to the etiology of PCOS, we first generated a comprehensive list of 310 genes related to three classes of genetic IR: insulin signaling, obesity, and dyslipidemia. We then systematically prioritized the likely impact of these genes on IR etiology using molecular, clinical, and statistical data. Genes with sufficient a priori evidence were assayed for protein-altering variants in a cohort of 675 women with PCOS (Figure 1). Finally, we selected 15 PCOS candidate genes with the strongest evidence for a role in PCOS pathogenesis for genetic analyses at the variant level.

Figure 1. Summary of gene selection, prioritization, and association with PCOS.

Percentages represent positive genes relative to total unique genes included in each gene set, except for the PCOS cases affected by a variant, which represents the number of PCOS cases affected by a variant relative to total PCOS cases (n=675).

Subjects and Methods

Subjects

Our study includes subjects from two PCOS cohort studies. The first cohort was recruited for phenotypic and genetic studies of PCOS and was approved by the IRB of Massachusetts General Hospital (MGH) and University of Utah. All subjects were self-reported White of European ancestry, ages 18–45 years old and in good general health. Their inclusion and exclusion criteria have been described (30, 31). PCOS cases satisfied NIH criteria, requiring ≤9 menses per year and biochemical or clinical evidence of hyperandrogenism (30, 31). Normoandrogenic, regularly cycling controls were recruited and phenotyped alongside PCOS cases. 376 PCOS cases and all 235 control subjects originate from this study. The remaining 299 PCOS cases in our study are participants of the Pregnancy in Polycystic Ovary Syndrome II (PPCOSII) Trial recruited in Hershey, PA (32, 33). PCOS cases in PPCOSII fulfill a modified Rotterdam criteria (34), requiring ovulatory dysfunction and either hyperandrogenism or polycystic ovarian morphology (32, 33). All PCOS subjects from PPCOSII are of self-reported European ancestry. All participants provided informed consent for genetic studies.

Hormone and Biochemical Assays

Hormones, lipids, glucose, and insulin were measured from blood samples as previously described (30, 31, 33). Comparisons of variables were corrected for study cohort as a proxy for assay methodology.

Statistical Analysis

Statistical tests were computed in GraphPad Prism (version 10.0.3 for macOS, GraphPad Software, Boston, MA www.graphpad.com) and R Studio (version 2022.12.0.353). To compare cases and controls, each variable was modeled in a linear model assessing the contributions of age, BMI, case/control status, and study site. Homogeneity of residuals in the linear model were evaluated using the olsrr package in R. Variables with non-normal residuals were transformed to achieve normality.

To assess enrichment of variants in PCOS cases, PCOS case frequency was compared to gnomAD v2.1.1 non-Finish European population controls. Odds ratios (OR) and Fisher’s exact tests were computed as a gene-based burden. The 95% CI of the OR was computed by the Woolf logit interval. A 95% CI excluding 1 was considered significant. Fisher’s exact test was considered nominally significant if p<0.05, considered study-wide (308 genes) significant if p<1.6×10⁻⁴, and considered significant at the exome-wide (20,000 genes) level if p<2.5×10⁻⁶.

Exome Sequencing

Genomic DNA extracted from peripheral blood was sequenced by WES at CIDR as described previously (17). Library preparation and enrichment were prepared using a low-input library prep protocol developed at CIDR. DNA was processed and amplified using Roche’s Kapa Hyper prep kit and Kapa HiFi enzyme kit. Sequencing was performed on the Illumina NovaSeq 6000 platform using 100 base pair paired end runs using Illumina NovaSeq 6000 S4 Reagent Kit and NovaSeq Xp 4-Lane kit workflow.

Alignment to reference genome build 38 and variant calling were completed using the CIDRSeqSuite pipeline (17). Quality control was performed at CIDR using Illumina InfiniumQCArray-24v1–0 array to confirm sex and ancestry, and to identify DNA sample contamination, unexpected duplicates, or relatedness. Annotation of called SNVs and indels was executed using SnpEff and SnpSift (35, 36). The Genome Aggregation Database (gnomAD) v3.1.2 (37) MAF was ascertained. Annotated variants were filtered based on Phred quality score ≥ 20, read depth ≥ 25,000, mapping quality ≥ 58.75 and ≤ 61.25, variant quality score log odds ≥ 7.81, exonic- region, variant type (predicted missense, nonsense, frameshift, and splice site variants), and Rare Exome Variant Ensemble Learner (REVEL) (38) score ≥ 0.6.

Gene Selection

Four sets of genes were identified for prioritization in this study. First, “100 random genes” were identified using a random number generation to generate 100 numbers between 1–22394, corresponding to line numbers in a spreadsheet of all RefSeq Select genes (genome.ucsc.edu). The remaining gene sets refer to the three molecular mechanisms of IR investigated in this study: insulin signaling, obesity, and dyslipidemia. All three gene sets were identified by database or literature search and through known GWAS associations. The GWAS Catalog (ebi.ac.uk/gwas) was queried on 8/12/2024 to identify GWAS loci. Loci were required to have one “qualifying” GWAS association that fulfilled any of the following strength-based criteria: Z score increase or decrease by ≥ 3, ≥ 10% change in measurement (i.e. insulin or triglycerides), or OR > 1.20. Finally, identified loci were evaluated for eQTL or disease association in the Online Mendelian Inheritance in Man (39) (OMIM) (omim.org) database ensuring a strong relationship between the GWAS annotated gene and a phenotype. GWAS-associated genes needed to fulfil both a strength-based criterion (Z ≥ ±3, ≥ 10% change, or OR > 1.20) and have an OMIM trait association to be evaluated further. To identify the gene set known as “insulin signaling,” the Kyoto Encyclopedia of Genes and Genomes (KEGG) (genome.jp/kegg) was queried for two KEGG pathways on 4/11/2024: first “insulin signaling pathway” and then “insulin resistance.” Genes in the KEGG IR gene set were only added if they were unique and did not duplicate genes from the KEGG insulin signaling gene set. Qualifying GWAS loci associated with three traits were added to the insulin signaling gene set: T2D, IR, and insulin level. The “obesity” and “dyslipidemia” gene sets were identified through literature search in July 2024. Qualifying GWAS loci associated with BMI and obesity were included in the obesity gene set, and GWAS loci for triglyceride levels were included in the dyslipidemia gene set.

Gene Prioritization

Genes were evaluated for their weighted candidate gene confidence level using a 4-point system (Figure 2). One point was awarded for a high degree of evolutionary constraint on missense variation (37), defined as Z > 3 in the Genome Aggregation Database (40) (gnomAD) (gnomad.broadinstitute.org) v4 on 8/9/2024. One point was awarded for any OMIM gene-phenotype relationship queried on 8/9/2024. An additional point was possible if the gene-phenotype relationship in OMIM was an IR-related phenotype, or had features of metabolic impairment, glucose dysregulation, hyperinsulinemia, hyperphagia, obesity, dyslipidemia, menstrual dysfunction, or other related phenotypes. Finally, one point was awarded for tissue-specificity of mRNA expression reported in the Human Protein Atlas (41) (HPA) (proteinatlas.org) queried on 8/12/2024. If mRNA was enriched in reproductive tissues, brain, adipose, skeletal muscle, liver, or pancreas, this point was awarded.

Figure 2.

Candidate genes weighted confidence ranking strategy.

The 4-point system was used to create 5 categories of genes with varying degrees of confidence: 0 points as “none,” 1 point as “weak,” 2 points as “moderate,” 3 points as “strong,” and 4 points as “robust.” Genes with 2–4 points were selected for further investigation.

Variant Filtering and Interpretation

Variants mapping to genes with at least moderate (2–4 points) evidence for candidate gene strength were filtered from the WES dataset containing protein-coding variants (missense, nonsense, frameshift). Variants with REVEL (38) score ≤ 0.6 and variants affecting control subjects only were eliminated from the dataset. In-frame insertions and deletions were assessed by the Sorting Intolerant From Tolerant (SIFT) coding indel classification algorithm (42), and variants considered “neutral” were discarded.

Variants selected for further analysis were classified using variant interpretation guidelines by the American College of Medical Genetics and Genomics (ACMG) and the Association for Molecular Pathology (AMP) (43). All missense variants in the dataset have REVEL ≥ 0.6 and therefore all have computational support for deleterious effect (ACMG/AMP-PP3). Nonsense and frameshift variants were predicted to be null (ACMG/AMP-PVS1). Genes with high constraint on missense scores (Z > 3.0) were expected to be genes with low rates of benign missense variants (ACMG/AMP-PP2). Statistical overrepresentation in PCOS cases compared to controls (ACMG/AMP-PS4) was assessed by OR using gnomAD v2.1.1 NFE population controls. An OR with a 95% confidence interval (CI) excluding 1 was considered significant. Functional data was assessed to determine if variants map to functional domains of the protein (ACMG/AMP-PM1) as defined in UniProt (uniprot.org) or had previous reports of deleterious effect in functional studies (ACMG/AMP-PS3). Previous reports of pathogenicity (ACMG/AMP-PS1) derived from literature search for case reports and/or genetic studies of disease cohorts were considered if related to the biology of IR.

Selection of Highlighted Genes

Fifteen genes were selected for detailed variant analysis. To select the genes to highlight, genes were ranked on three scales: weight of evidence for candidacy, degree of association with PCOS, and number of PCOS cases affected by variants in the gene. A combined rank score from these three rankings was computed. The top 15 genes in the combined rank score were further investigated for their relationship to PCOS.

Results

Study participant characteristics

The PCOS cases in our study cohort had significantly higher testosterone levels (p<2×10⁻¹⁶) and BMI (p<2×10⁻¹⁶) than controls and did not differ significantly in age (Table 1). Metabolically, PCOS cases had lower SHBG (p=4.27×10⁻¹¹) and HDL cholesterol (p=0.00754), and higher LDL cholesterol (p=3.97×10⁻⁴) and triglyceride levels (p=6.04×10⁻⁶) when compared to controls (Table 1). PCOS cases had higher waist-to-hip ratios (WHR) than controls (p<2×10⁻¹⁶). Fasting glucose levels were lower in PCOS cases than controls (p=5.04×10⁻¹⁰), which could indicate hyperinsulinemia preceding IR. PCOS cases had higher fasting insulin levels than controls, although the difference is not significant (Table 1). Overall, the phenotypic features of the PCOS cohort do not indicate an overtly insulin resistant subtype of PCOS; rather, they reflect a heterogenous, generic PCOS cohort.

Table 1.

Clinical and biochemical features of PCOS cases and controls in this study.

	PCOS Cases (n=675)		Controls (n=235)
	n	Median (IQR)	n	Median (IQR)	p value
Age (years)	675	28 (25–32)	235	26 (22–32)	n.s.
BMI (kg/m2)	674	31.3 (24.8–39.0)	234	23.1 (21.1–25.0)	<2×10−16
Testosterone (ng/dl)	582	52 (37–72)	224	32 (22–41)	<2×10−16
DHEAS (ng/ml)	284	1950 (1313–2608)	229	1930 (1375–2635)	n.s.
SHBG (nM)	579	31.0 (20.5–46.0)	228	61.0 (44.3–83.8)	4.27×110−11
HOMA-IR	593	2.02 (0.94–4.25)	216	1.05 (0.71–1.44)	n.s.
WHR	347	0.88 (0.83–0.93)	270	0.82 (0.77–0.85)	<2×10−16
Fasting Glucose (mg/dl)	612	87 (81–94)	217	94 (87–101)	1.91×10−7
Fasting Insulin (uIU/ml)	609	9.3 (4.5–19.2)	225	4.4 (3.1–6.3)	n.s.
HbA1c (%)	306	5.3 (5.1–5.5)	218	5.2 (5.0–5.4)	n.s.
Total cholesterol (mg/dl)	635	179 (155–203)	225	169 (150–192)	n.s.
HDL cholesterol (mg/dl)	634	44 (35–56)	225	61 (52–70)	0.00754
LDL cholesterol (mg/dl)	633	113 (92–135)	225	95 (80–111)	3.97×10−4
Triglycerides (mg/dl)	634	96 (66–143)	225	59 (49–76)	6.04×10−6

Variables evaluated in a linear model assessing the contributions of age, BMI, case/control group, and assay or study site where applicable. P value represents the contribution of case/control group to the variation in the data.

Abbreviations: IQR, inter-quartile range; n.s., not significant; BMI, body-mass index; DHEAS. dehydroepiandrosterone sulfate; SHBG, sex hormone binding globulin; HOMA-IR, homeostatic model assessment of IR; WHR, waist-to-hip ratio; HbA1c, glycated hemoglobin A1c, HDL, high-density lipoprotein; LDL, low-density lipoprotein.

Gene Selection & Prioritization

Gene selection in the three IR gene sets yielded a total of 310 unique genes to be considered (Fig 1, Table S1). 235 genes comprised the insulin signaling gene set, 49 genes comprised the obesity gene set, and 26 genes comprised the dyslipidemia gene set. Two genes had duplicate rationale for inclusion within their gene set: TBC1D4 was included in the insulin signaling gene set both by KEGG IR and T2D GWAS evidence, and LEPR was included in the obesity gene set both by monogenic obesity and obesity GWAS evidence (Table S1).

All 310 IR genes as well as the 100 randomly selected gene controls were assessed for their weighted strength as PCOS candidate genes (Table S2–S3). Just over half of the 100 randomly selected genes had no evidence for candidate weight, and only 16 of these genes received scores of moderate or above (Fig 3). In contrast, all three IR gene sets had higher proportions of genes receiving higher weighted confidence scores than the randomly selected genes (Fig 3). Sixteen of the randomly selected genes, 102 insulin signaling genes, 29 obesity genes, and 22 dyslipidemia genes received scores of moderate, strong, or robust. 153 IR genes across the three gene sets were therefore carried forward in the study (Fig 1, Table S2).

Figure 3. Distributions of weighted confidence rankings for each gene set.

The number of genes with each level of weighted confidence was plotted to show the relative distribution of candidate strength. (None=0 points, Weak=1, Moderate=2, Strong=3, Robust=4)

Prioritized Genes are Enriched for Variation in PCOS Cases

Nonsense, frameshift, and predicted damaging in-frame indels (SIFT indel “damaging”) and missense variants (REVEL ≥ 0.6) were identified in the 153 tested genes (Table S4). 59 genes did not contain any variants meeting these cutoffs. We identified 356 variants across 94 genes affecting 515 PCOS cases in our study (Fig 1, Table S4). We collapsed variants into association tests by gene to test for enrichment of variation in PCOS cases compared to population controls. 40 insulin signaling genes, 12 obesity genes, and 10 dyslipidemia genes were significantly enriched for variation in PCOS cases (Fig 1, 4, Table 2, Table S5). Additionally, all three pathways are significantly enriched for variation in PCOS cases when variants were pooled by pathway, while the random genes were not (Fig 1, 4, Table 2). By Fisher’s exact test, 39 insulin signaling, 10 obesity, and 8 dyslipidemia genes were nominally significant (p<0.05) (Table 2). Four insulin signaling genes (KCNJ11, GCK, ACACA, ABCC8), one obesity gene (ITPR3), and one dyslipidemia gene (BSCL2) were significant at the study-wide level (p<1.6×10⁻⁴), and 4 insulin signaling genes (CPT1A, PAX4, TH, CD36), one obesity gene (LEPR), and one dyslipidemia gene (LMNA) were further significant at the exome-wide level (p<2.5×10⁻⁶) (Table 2, Fig 4). Variants in genes significantly associated with PCOS affected a total of 341 PCOS cases across all three IR gene sets, amounting to 51% of the PCOS cohort (Fig 1).

Figure 4. Enrichment of variation in PCOS cases compared to population controls.

OR of PCOS cases relative to gnomAD non-Finnish European population controls and corresponding 95% CIs were plotted on a logarithmic axis. Genes are sorted by the lowest 95% CI limit and OR dots were sized according to the number of PCOS cases with a variant in the gene for gene-based ORs (A-C). (A) Insulin signaling, (B) Obesity, and (C) Dyslipidemia gene sets. Asterisks indicate genes significant at the exome-wide level (p<2.5×10⁻⁶) by Fisher’s exact test. (D) ORs computed as a pooled burden for each pathway.

Table 2.

Gene-based associations for all genes assayed in this study.

Gene Set	Gene	OR (95% CI)	Fisher’s Exact p value	Nominally Significanta	Study-Wide Significantb	Exome-Wide Significantc
Insulin Signaling	CPT1A	250.9 (26.1–2413.4)	6.57×10−6	*	*	*
	KCNJ11	125.6 (21.0–752.4)	1.62×10−5	*	*
	GCK	404.1 (19.4–8421.5)	1.50×10−4	*	*
	SLC2A2	166.9 (19.4–1841.3)	4.18×10−4	*
	SREBF1	89.4 (12.6–635.4)	7.24×10−4	*
	PAX4	25.7 (12.2–54.2)	5.87×10−10	*	*	*
	TH	32.2 (12.4–91.8)	2.03×10−6	*	*	*
	MAPK10	251.4 (10.2–6174.8)	1.18×10−2	*
	WDR72	250.9 (10.2–6162.0)	1.18×10−2	*
	TBC1D4	249.8 (10.2–6133.5)	1.19×10−2	*
	IKBKB	249.2 (10.1–6118.7)	1.19×10−2	*
	TSC1	28.0 (9.0–86.8)	3.15×10−5	*
	PHKA1	36.2 (9.6–135.2)	2.29×10−4	*
	KRAS	41.7 (7.6–227.7)	2.05×10−3	*
	CD36	10.7 (6.5–17.8)	4.30×10−12	*	*	*
	ACACA	15.6 (6.4–40.3)	3.79×10−5	*	*
	G6PC	19.4 (5.5–68.1)	8.16×10−4	*
	MAP2K2	84.3 (5.3–1348.8)	2.33×10−2	*
	HK1	83.8 (5.2–1340.7)	2.35×10−2	*
	KCNU1	31.3 (5.5–153.3)	4.19×10−3	*
	ANKH	83.5 (5.2–1335.0)	2.36×10−2	*
	PRKCE	82.4 (5.2–1318.6)	2.38×10−2	*
	PRKCQ	23.9 (5.0–115.4)	4.75×10−3	*
	SGCG	10.5 (3.6–26.7)	1.88×10−4	*
	GYS1	10.2 (3.6–28.8)	9.36×10−4	*
	PCK1	9.9 (3.5–27.8)	1.04×10−3	*
	ABCC8	5.7 (3.1–10.5)	8.48×10−6	*	*
	MTOR	12.9 (2.9–57.3)	1.32×10−2	*
	ADCY5	28.0 (2.9–269.0)	4.63×10−2	*
	DEPDC5	27.9 (2.9–267.9)	4.65×10−2	*
	EPHB2	7.1 (2.9–17.8)	9.77×10−4	*
	CACNA1D	6.4 (3.0–14.1)	5.37×10−4	*
	NEUROG3	21.0 (2.3–187.8)	5.76×10−2
	PRKAG3	3.6 (2.0–6.3)	1.37×10−4	*
	PYGM	2.0 (1.4–2.8)	1.84×10−4	*
	SPTB	2.7 (1.4–5.0)	5.94×10−3	*
	HNF4A	4.3 (1.4–13.8)	3.60×10−2	*
	BPTF	2.1 (1.1–3.9)	3.24×10−2	*
	WFS1	1.4 (1.0–1.9)	3.83×10−2	*
	ANXA5	4.0 (1.0–16.5)	9.54×10−2
	PKLR	1.6 (1.0–2.7)	9.69×10−2
	PHKB	1.2 (1.0–1.4)	0.153
	RASGRP1	6.9 (0.9–53.3)	0.144
	FBP2	1.1 (0.8–1.6)	0.458
	INSR	1.3 (0.8–2.2)	0.299
	SACS	1.2 (0.8–1.8)	0.425
	PYGL	1.0 (0.8–1.3)	0.941
	CREB3L3	5.6 (0.7–42.2)	0.173
	MASP1	1.8 (0.7–4.7)	0.290
	TG	0.9 (0.7–1.3)	0.804
	FBP1	4.9 (0.7–37.1)	0.192
	ACACB	0.8 (0.6–1.0)	3.90×10−2	*
	PRKAR1B	4.3 (0.6–32.5)	0.214
	KCNQ3	1.3 (0.6–2.9)	0.481
	ACHE	1.6 (0.5–5.1)	0.438
	MTNR1B	1.4 (0.5–3.7)	0.542
	AGT	1.1 (0.4–2.5)	0.816
	SGCD	2.2 (0.3–15.9)	0.374
	PPP1R3A	0.6 (0.3–1.3)	0.284
	LPIN2	0.6 (0.2–1.9)	0.643
	GYS2	0.4 (0.1–1.7)	0.345
	Insulin Signaling	2.3 (2.0–2.7)	<1.0×10–10−15	*	*	*
Obesity	LEPR	585.1 (30.2–11332.3)	1.66×10−6	*	*	*
	NTRK2	167.6 (15.2–1849.6)	4.15×10−4	*
	MC4R	47.7 (7.5–304.7)	2.47×10−3	*
	NR0B2	24.0 (5.0–115.6)	4.73×10−3	*
	BBS2	12.9 (4.5–37.1)	4.12×10−4	*
	PCSK1	42.0 (3.8–463.1)	3.49×10−2	*
	BBS4	21.0 (2.3–187.9)	5.75×10−2
	ITPR3	3.5 (2.2–5.7)	9.61×10−6	*	*
	NPC1	3.2 (1.3–7.7)	2.43×10−2	*
	POMC	2.2 (1.3–3.8)	1.14×10−2	*
	FBN2	1.6 (1.0–3.8)	4.80×10−2	*
	TFAP2B	6.0 (0.8–46.0)	0.163
	MC3R	5.2 (0.7–39.5)	0.183
	TUB	0.9 (0.7–1.3)	0.809
	BBS1	1.2 (0.6–2.6)	0.521
	MKKS	1.0 (0.6–1.7)	>0.99
	SLC30A8	4.2 (0.6–31.3)	0.220
	KCNH5	1.8 (0.6–5.5)	0.249
	SLC7A14	1.1 (0.5–2.2)	0.712
	Obesity	1.8 (1.5–2.1)	2.9×10−8	*	*	*
Dyslipidemia	LMNA	63.3 (19.3–207.5)	1.63×10−7	*	*	*
	NEUROD2	245.0 (10.0–6015.8)	1.21×10−2	*
	BSCL2	13.1 (5.7–31.6)	1.71×10−5	*	*
	DNAH17	31.7 (5.6–155.2)	4.09×10−3	*
	LIPE	83.9 (5.2–1342.7)	2.34×10−2	*
	ZPR1	41.9 (3.8–462.9)	3.49×10−2	*
	APOA5	11.9 (2.7–52.6)	1.51×10−2	*
	PCYT1A	10.5 (2.4–45.7)	1.87×10−2	*
	AGPAT2	16.3 (1.9–140.0)	7.04×10−2
	CAV1	10.4 (1.3–83.5)	0.102
	MFN2	3.4 (1.1–11.0)	6.12×10−2
	GCKR	1.8 (0.8–4.0)	0.165
	CIDEC	4.4 (0.6–32.9)	0.212
	LPL	2.2 (0.3–16.1)	0.370
	Dyslipidemia	4.9 (3.4–7.1)	1.3×10-12	*	*	*

^ap<0.05

^bp<1.6×10⁻⁴

^cp<2.5×10⁻⁶

Top 15 genes selected to highlight

The 15 highest ranked genes were selected for in depth analysis. In these 15 genes, 48 variants were identified affecting 78 PCOS cases (Table 4–5, S6). 24 of these variants were classified as pathogenic, 17 as likely pathogenic, and only 7 as VUS (Table 5). Variants in the highlighted genes have been associated with a number of related pathologies, including dominant traits such as MODY (Leu135Val, Cys418Arg, Ala1536Val in ABCC8; Ala259Thr in GCK), and monogenic obesity (Thr150Ile in MC4R; His143Gln, Arg236Gly in POMC) and recessive traits such as congenital hyperinsulinemia (Gly92Ser, Leu135Val, Asp310Asn, Cys418Arg, Leu511Pro, Ala847Thr, and Ala1536Val in ABCC8; Ala96Val and Lys222Gln in KCNJ11) (Table 5).

Table 4.

Variants identified in 15 highlighted PCOS genes.

Gene	Chr.	Nucleotide Change (g.)	Protein Change (p.)	Variant ID	NFE MAF	Count Cases	PCOS Frequencya	Odds Ratio (95% CI)
ABCC8	11	17474971G>T	Pro69Thr	11–17474971-G-T	0	1	0.0007	237.7 (9.7–5836.8)
	11	17474902C>T	Gly92Ser	rs780870376	8.79×10−6	1	0.0007	79.2 (5.0–1267.6)
	11	17470110G>C	Leu135Val	rs368450282	6.16×10−5	1	0.0007	11.3 (1.4–92.1)
	11	17460571C>T	Asp310Asn	rs769569410	0	1	0.0007	237.4 (9.7–5831.0)
	11	17448596A>G	Cys418Arg	rs67254669	1.09×10−3	1	0.0007	0.6 (0.1–4.6)
	11	17442818A>G	Leu511Pro	rs797045206	0	1	0.0007	237.7 (9.7–5838.9)
	11	17442813C>T	Ala513Thr	rs761748692	7.93×10−5	1	0.0007	8.8 (1.1–69.4)
	11	17413447G>T	Gln808Lys	rs202189540	6.15×10−5	1	0.0007	11.3 (1.4–92.1)
	11	17412683C>T	Ala847Thr	rs561593131	8.98×10−6	1	0.0007	77.6 (4.9–1241.8)
	11	17408439_17408440insA	Glu925fs	11–17408439-C-CA	0	1	0.0007	232.8 (9.5–5717.7)
	11	17393698G>A	Ala1536Val	rs745918247	0	1	0.0007	237.5 (9.7–5833.2)
ACACA	17	37263836C>T	Arg393His	rs187880919	1.67×10−4	2	0.0015	8.3 (1.9–35.8)
	17	37149964C>T	Arg1860Gln	rs1435154885	0	1	0.0007	237.3 (9.7–5827.9)
	17	37130146T>C	Thr1918Ala	17–37130146-T-C	0	1	0.0007	237.3 (9.7–5827.9)
	17	37113227C>A	Val2105Leu	rs1376108098	0	1	0.0007	237.3 (9.7–5827.9)
BSCL2	11	62707142G>T	Cys18Ter	rs761790583	1.58×10−5	1	0.0007	44.2 (2.8–706.9)
	11	62705346T>C	Tyr120Cys	rs370905417	1.68×10−4	1	0.0007	4.1 (0.6–31.0)
	11	62694572C>T	Arg209His	rs1388984096	0	1	0.0007	236.5 (9.6–5807.4)
	11	62692394G>A	Ala282Val	rs185341934	4.40×10−5	3	0.0022	47.6 (11.4–199.3)
GCK	7	44147738C>T	Ala259Thr	rs1375656631	0	1	0.0007	237.0 (9.7–5821.3)
	7	44145621G>C	Arg377Gly	7–44145621-G-C	0	1	0.0007	237.0 (9.7–5821.3)
GCKR	2	27498276G>A	Val103Met	rs146175795	1.76×10−5	1	0.0007	42.1 (3.8–465.1)
	2	27506512C>T	Arg301Ter	rs767008458	8.79×10−6	1	0.0007	84.3 (5.3–1348.6)
	2	27507240C>T	Arg358Ter	rs201754753	1.58×10−4	1	0.0007	4.7 (0.6–35.1)
	2	27507302C>CA	Thr379fs	rs573498430	3.10×10−5	3	0.0022	1.0 (0.3–3.3)
INSR	19	7117400T>A	Val1012Met	rs1799816	8.39×10−3	14	0.0104	1.2 (0.7–2.0)
	19	7125507C>T	Met1269Leu	rs375197837	1.42×10−4	1	0.0007	4.9 (0.7–37.1)
KCNJ11	11	17387805G>A	Ala96Val	11–17387805-G-A	0	1	0.0007	236.8 (9.6–5815.0)
	11	17387428T>G	Lys222Gln	rs747090537	1.77×10−5	1	0.0007	39.5 (3.6–435.6)
	11	17386948_17386949insC TTGGG	Pro380_Lys381dup	rs1440128889	0	1	0.0007	236.8 (9.6–5815.0)
LEPR	1	65601563G>A	Ser389Asn	rs780534740	0	1	0.0007	237.4 (9.7–5829.2)
	1	65610286T>C	Leu662Ser	rs1363533488	0	1	0.0007	237.4 (9.7–5829.2)
	1	65610293G>C	Trp664Cys	rs1239306320	0	1	0.0007	237.4 (9.7–5829.2)
LMNA	1	156114992G>T	Arg25Leu	rs61578124	2.05×10−5	1	0.0007	33.9 (3.1–374.4)
	1	156135991C>T	Arg343Trp	rs61177390	8.81×10−6	1	0.0007	79.1 (4.9–1264.6)
	1	156136359C>T	Arg435Cys	rs150840924	1.77×10−5	2	0.0015	78.7 (11.1–559.2)
	1	156136419C>T	Arg455Cys	rs397517892	1.16×10−5	1	0.0007	60.3 (3.8–964.5)
MC4R	18	60372111T>C	Tyr80Cys	rs1368643838	0	1	0.0007	237.7 (9.7–5836.9)
	18	60371901G>A	Thr150Ile	rs766665118	8.79×10−6	1	0.0007	79.2 (5.0–1267.7)
NR0B2	1	26911981C>T	Arg213His	rs367827644	6.16×10−5	2	0.0015	22.7 (4.7–109.2)
NTRK2	9	84948490G>A	Arg598His	rs147652140	8.80×10−6	1	0.0007	79.2 (5.0–1267.2)
	9	84955380G>A	Gly679Ser	9–84955380-G-A	0	1	0.0007	237.6 (9.7–5834.6)
POMC	2	25161456G>C	His143Gln	rs201519174	3.48×10−4	2	0.0015	4.0 (1.0–16.6)
	2	25161281G>A	Gln202Ter	rs139849371	0	1	0.0007	222.2 (9.0–5456.9)
	2	25161179G>C	Arg236Gly	rs28932472	3.96×10−3	10	0.0074	1.8 (0.9–3.3)
SLC2A2	3	170999138C>T	Arg366His	rs755000812	8.83×10−6	1	0.0007	78.9 (4.9–1262.6)
	3	170997958_170997962del	Ala506fs	rs1560031127	0	1	0.0007	236.7 (9.6–5813.5)
TBC1D4	13	75299364C>T	Arg1041His	rs199560195	0	1	0.0007	235.8 (9.6–5791.9)

Abbreviations: NFE MAF, non-Finish European population minor allele frequency in gnomAD v2.1.1

^aPCOS frequency calculated out of total chromosomes sequenced in PCOS cases (n=1350)

Table 5.

Effects of variants in 15 highlighted PCOS genes on protein function.

Gene	Variant (p.)	REVEL Score	Affected functional protein domain	Previous Reports of Functional Impact	Previous Reports of Pathogenicity	Classification
ABCC8	Pro69Thr	0.891			---	VUS
	Gly92Ser	0.887			CHI(84)	Pathogenic
	Leu135Val	0.689		Impairs KATP channel activity(85)	MODY,(86) CHI,(87) PAH(85)	Pathogenic
	Asp310Asn	0.833	ABC transmembrane type-1 domain 1	Decreases channel surface expression(88)	CHI(88–91)	Pathogenic
	Cys418Arg	0.657	ABC transmembrane type-1 domain 1		CHI,(92, 93) MODY,(94, 95) T2D(96)	Likely Pathogenic
	Leu511Pro	0.940	ABC transmembrane type-1 domain 1		CHI(97)	Pathogenic
	Ala513Thr	0.812	ABC transmembrane type-1 domain 1		T2D(98)	Pathogenic
	Ile681Thr	0.795	ABC transporter domain 1		---	Likely Pathogenic
	Gln808Lys	0.602	ABC transporter domain 1		PAH,(99) T1D(100)	Pathogenic
	Ala847Thr	0.723	ABC transporter domain 1		CHI(101)	Pathogenic
	Glu925fs	NA	ABC transporter domain 1		---	Likely Pathogenic
	Ala1536Val	0.951	ABC transporter domain 2		CHI,(102) MODY (Ala1536Thr)(103)	Pathogenic
ACACA	Arg393His	0.888	ATP-grasp domain		---	Likely Pathogenic
	Arg1860Gln	0.908	Carboxyl transferase domain		---	Likely Pathogenic
	Thr1918Ala	0.601	Carboxyl transferase domain		---	Likely Pathogenic
	Val2105Leu	0.882	Carboxyl transferase domain		---	Likely Pathogenic
BSCL2	Cys18Ter	NA			---	Likely Pathogenic
	Tyr120Cys	0.906			---	VUS
	Arg209His	0.837			---	VUS
	Ala282Val	0.784			CMT(104)	Pathogenic
GCK	Ala259Thr	0.901	Hexokinase large subdomain		MODY,(105–111) T2D(112)	Pathogenic
	Arg377Gly	0.991	Hexokinase large subdomain		MODY (Arg377His,(113, 114) Arg377Ser,(113) Arg377Leu,(115–117) Arg377Cys(118, 119))	Likely Pathogenic
GCKR	Val103Met	0.689	Sugar isomerase (SIS) domain 1	Severe loss of function in HeLa cells(120)	SHTG,(121–123) HLP1(124)	Pathogenic
	Arg301Ter	NA				Likely Pathogenic
	Arg358Ter	NA	Sugar isomerase (SIS) domain 2			VUS
	Thr379fs	NA	Sugar isomerase (SIS) domain 2	Severe loss of function in HeLa cells(120)		Pathogenic
INSR	Val1012Met	0.634			T2D(125)	Likely Pathogenic
	Met1269Leu	0.830	Protein tyrosine kinase domain		---	VUS
KCNJ11	Ala96Val	0.616			CHI(98, 126)	Pathogenic
	Lys222Gln	0.714			CHI(98)	Pathogenic
	Pro380_Lys381dup	Damaginga			T2D(98)	Pathogenic
LEPR	Ser389Asn	0.629	Ig-like domain		Severe obesity,(127) HGTP(128)	Pathogenic
	Leu662Ser	0.732	Fibronectin type-III domain 3		Hyperphagia and severe obesity(129)	Pathogenic
	Trp664Cys	0.867	Fibronectin type-III domain 3		Hyperphagia and severe obesity (Trp664Arg)(130)	Likely Pathogenic
LMNA	Arg25Leu	0.873			FPLD2(131, 132)	Pathogenic
	Arg343Trp	0.683	Intermediate filament rod domain		---	Likely Pathogenic
	Arg435Cys	0.667	Lamin tail domain		RD,(133, 134) HGPS(134, 135)	Pathogenic
	Arg455Cys	0.804	Lamin tail domain		PCOS,(72) EDMD (R455P)(136)	Pathogenic
MC4R	Tyr80Cys	0.722			---	VUS
	Thr150Ile	0.685		Impairs receptor signaling in response to a-MSH binding(137–139)	Hyperphagia and severe obesity(137, 139–143)	Pathogenic
NR0B2	Arg213His	0.887			Mild obesity (Arg213Cys)(144)	VUS
NTRK2	Arg598His	0.679	Protein kinase domain		---	Likely Pathogenic
	Gly679Ser	0.922	Protein kinase domain		---	Likely Pathogenic
POMC	His143Gln	0.822	a-MSH, corticotropin peptide region		Severe early-onset obesity(145)	Pathogenic
	Gln202Ter	NA			---	Likely Pathogenic
	Arg236Gly	0.816	Lipotropin beta, Met-enkephalin, Beta-endorphin peptide region	Prevents normal processing of POMC(146)	Severe early-onset obesity(145–151)	Pathogenic
SLC2A2	Arg366His	0.843			---	Likely Pathogenic
	Ala506fs	NA			---	Likely Pathogenic
TBC1D4	Arg1041His	0.602	Rab-GAP TBC domain		Severe IR(152)	Pathogenic

^aSmall insertion Pro380_Lys381dup in KCNJ11 was assessed with SIFT coding indel algorithm rather than REVEL score.

Abbreviations: REVEL, rare exome variant ensemble learner; VUS, variant of uncertain significance; CHI, congenital hyperinsulinemia; PAH, pulmonary arterial hypertension; T1D, type 1 diabetes; CMT, Charcot-Marie-Tooth disease; SHTG, severe hypertriglyceridemia; HLP1, Type I hyperlipoproteinemia; HGPT, hypertriglyceridemic pancreatitis; FPLD2, familial partial lipodystrophy type 2; RD, restrictive dermopathy; HGPS, Hutchinson-Gilford progeria syndrome; EDMD, Emery-Dreyfuss muscular dystrophy.

Discussion

PCOS is a complex, multisystem genetic trait that is strongly associated with IR. To test the hypothesis that genetic forms of IR contribute to the etiology of PCOS, we evaluated genes in three categories of genetic IR for protein-altering variants in PCOS patients. From a list of 310 genes, we identified 153 genes with significant a priori evidence to warrant investigation. Out of these, 62 genes were significantly enriched for protein-altering variation in WES data from a cohort of 675 PCOS cases compared to population controls. Although individual variants are rare, 51% of PCOS cases were carriers of protein-altering variants in genes significantly enriched for variants compared to population controls, thus representing an incredibly common genetic feature of PCOS. These data provide support that for a significant subset of PCOS cases, IR may be a primary driver of the PCOS phenotype rather than a consequence of elevated testosterone levels.

To test the role of IR in PCOS in an unbiased and comprehensive manner but also maximize interpretability, we developed a gene curation and prioritization pipeline. The gene scoring system (Fig 2) was developed to optimize the identification of variants in genes amenable to clinical interpretation. We started with 310 genes in total and tested the genes with the strongest molecular and clinical evidence for a role in IR for further genetic evaluation. We evaluated 153 genes that survived our gene scoring system from the initial list of 310 genes, representing 49% of the genes initially considered. In contrast, out of 100 randomly selected genes scored alongside the IR gene sets, only 16 (16%) survived our prioritization screen. When we later examined the narrower set of 15 selected genes (Box 1) at the variant level, we did not identify any benign or likely benign variants, and only seven of 48 variants remained VUS (Table 5). By selecting genes with substantial evidence for disease association and high genomic constraint, and by using a protein functional prediction filter to select variants in those genes, our pipeline strongly enriched our results towards pathogenic variants. Our stringent gene prioritization and variant filtering system allowed us to clearly interpret the significance of variants in the final dataset. Although this may have resulted in a degree of type 2 error due to the exclusion of genes without strong of a priori evidence for a role in IR, this is preferable to type 1 error and the inclusion of false positive genes.

Box 1.

ABCC8 encodes the SUR1 protein, part of an ATP-sensitive potassium channel that regulates insulin release from the pancreas. It has been implicated in T2D through GWAS, and rare variants also cause familial hyperinsulinemic hypoglycemia. Common variants associated with T2D have been shown to lack association with PCOS (45), but the coding region has not been assessed in the PCOS context.

ACACA encodes a member of the insulin signaling pathway that regulates de novo fatty acid synthesis. Homozygous variants in this gene cause a rare metabolic syndrome called Acetyl-CoA carboxylase deficiency. It has not previously been a candidate gene for PCOS, but has known roles in oocyte maturation (46) and ovarian steroidogenesis (47).

BSCL2, also known as seipin, encodes an essential regulator of lipid droplet formation and adipocyte development. Variants in this gene cause CGL2, a recessive lipodystrophy syndrome. A relationship between CGL2 and PCOS was first proposed nearly 50 years ago (48), but only sporadic case reports document symptoms of PCOS in CGL2 patients,(49) and a previous sequencing effort in a PCOS cohort did not identify variants in BSCL2 (15).

GCK encodes the glucokinase enzyme, a member of the insulin signaling network whose dysfunction causes MODY and other forms of diabetes mellitus. Although at least one group has included GCK on a candidate sequencing panel (15), no variants have been associated with PCOS before.

GCKR encodes glucokinase regulatory protein. It qualified for inclusion in this study due to a GWAS association with triglyceride level but also has an OMIM association with fasting plasma glucose levels, suggesting a multifaceted role in glucose homeostasis and IR. One previous study examined common copy-number variation in GCKR in PCOS cases with little success (50), and no studies have sequenced GCKR for coding variants in PCOS until now.

INSR, which encodes the insulin receptor, has long been a candidate gene for PCOS (8, 51–62). However, its candidacy has been based on associations with common, noncoding variants. Since its most penetrant related diseases Donohue syndrome and Rabson-Mendenhall syndrome are recessive disorders, INSR may not be an ideal candidate gene for rare causal variants in the heterozygous state.

KCNJ11 encodes an ATP-sensitive potassium channel essential for regulating insulin release from pancreatic beta cells. It has long been a GWAS candidate for T2D, and rare variants in this gene cause familial hyperinsulinemic hypoglycemia and MODY. Multiple studies have tested for association between common T2D polymorphisms and PCOS and found no relationship (25, 26, 63), but rare variation has not been investigated until the present work.

LEPR encodes the leptin receptor, a key regulator of appetite and fat metabolism expressed in the liver. Common variants in LEPR have been associated with BMI in GWAS, and rare variants have been shown to cause monogenic morbid obesity. Common variants have been associated with PCOS in several populations (27, 28, 64–69), but a previous attempt to sequence the gene for rare variation in a PCOS cohort did not yield any variants (15).

LMNA encodes the lamin A/C intermediate filament proteins expressed in all differentiated cell types. Because of variants in this gene that cause FPLD2, a syndrome with overlapping features of hyperandrogenism and menstrual dysfunction, LMNA has been a prime candidate gene for PCOS. Several studies have identified rare pathogenic variants in LMNA in women with PCOS (15, 16, 20, 70–72), cementing it as a reproducible candidate gene (73).

MC4R, which encodes a melanocortin receptor primarily expressed in the brain, is a preeminent obesity gene due to GWAS associations and high effect, monogenic cases alike. Common MC4R variants have been tested for association with PCOS in multiple studies, but have found no associations with PCOS itself, only with BMI/obesity (74–77).

NR0B2, also known as SHP, encodes an orphan receptor expressed in the liver and GI tract with proposed roles in adipogenesis and cholesterol metabolism. Variants in this gene cause mild early onset obesity and have never been reported in the context of PCOS. In one study of gene expression in adipocytes differentiated from human embryonic stem cells derived from PCOS patients and controls, NR0B2 expression was found to be higher in PCOS-derived adipocytes than those from controls (78).

NTRK2, also named TRKB, encodes a tyrosine kinase receptor (TrkB) that binds its ligand brain-derived neurotrophic factor (BDNF). It is expressed in the brain and variants are associated with hyperphagic obesity. Although it hasn’t been considered a genetic target in PCOS before, several functional studies have established a role for BDNF-TrkB signaling in germ cell nest breakdown (79), oocyte maturation (80), and follicle development (81).

POMC encodes a precursor peptide expressed in the pituitary that gets processed into several peptide hormones, including the melanocortins. Its dysfunction causes hyperphagic obesity and has not yet been shown to have a direct role in PCOS, although one study did include it in their gene panel but did not identify any variants (15).

SLC2A2 encodes the GLUT-2 glucose transporter expressed primarily in the liver. Variants in this gene cause Fanconi-Bickel syndrome, a glycogen storage disease, as well as susceptibility to T2D. Two studies have considered the role of SLC2A2 in PCOS: one found a strong association between PCOS and a common variant (29), and the other sought to associate noncoding SNPs with metformin response (82).

TBC1D4, also known as AS160, encodes a substrate of AKT phosphorylation downstream of insulin receptor activation. Common variation in TBC1D4 is associated with T2D. One study has shown a lesser presence of AS160 phosphorylation in muscle biopsies of PCOS patients compared to controls (83), but no studies have sequenced TBC1D4 in PCOS cases until now.

Both dominant and recessive forms of highly penetrant, monogenic diabetes, obesity, and dyslipidemia have been identified previously, but it was unclear whether the same genes or causal variants are also implicated in PCOS, a common, complex trait. Our gene curation strategy included genes with OMIM associations of MODY, monogenic obesity, lipodystrophy, or other syndromic pathologies with obesity or IR as a feature. We assayed 41 genes related to known monogenic manifestations of IR from the WES data collected for our PCOS cohort. Of these, 29 genes contained at least one variant surviving our functional filter in least one PCOS case and 21 of these genes showed significant association with PCOS. These include the MODY genes KCNJ11, GCK, HNF4A and PAX4, familial hyperinsulinemic hypoglycemia gene ABCC8, monogenic obesity genes LEPR, NR0B2, POMC, MC4R, and NTRK2, Bardet-Biedl syndrome genes BBS2 and BBS4, partial lipodystrophy genes LMNA, LIPE and MFN2, and generalized lipodystrophy genes BSCL2, AGPAT2, CAV1, and PCYT1A. Along with our previous study of PCOS patients with pathogenic variants in LMNA (16), this further underscores that within the heterogeneous group of individuals diagnosed with PCOS, some may have distinct monogenic etiologies. This is essential to understanding the pathogenesis of PCOS, especially in the context of genetic causes of obesity and IR. Accurate and effective treatment of PCOS depends on recognizing obesity and IR as potential primary drivers of the condition and thus targets for treatment.

Our results add to a growing body of studies that challenge classical inheritance models especially in the context of common, complex traits (44). It is possible that the features of PCOS include symptomatic heterozygous forms of recessive traits like generalized lipodystrophy or severe IR. If variants can only exert their effects in the recessive state, we would expect their prevalence in population databases to be too common to result in statistical enrichment in our cohort. In contrast, numerous genes significantly enriched for variation in PCOS cases in our study are in fact related to recessive disorders of IR. Examples include ABCC8 (familial hyperinsulinemic hypoglycemia), BSCL2, PCYT1A, CAV1, and AGPAT2 (generalized lipodystrophy), BBS2 and BBS4 (Bardet-Biedl syndrome), SLC2A2 (Fanconi-Bickel syndrome), ACACA (Acetyl-CoA carboxylase deficiency), NPC1 (Niemann-Pick disease), and G6PC and GYS1 (glycogen storage disorders) (Table S2). The prevalence of heterozygous protein-altering variants in these genes in our PCOS cohort supports possible additive inheritance for these genes, or an otherwise symptomatic heterozygote model.

Although the insulin signaling, obesity, and dyslipidemia gene sets were curated and stratified using the same strategy, there were clear differences in how each gene set contributes to the heritability of PCOS. The obesity and dyslipidemia sets were primarily defined by genes already demonstrated to be associated with disease, while the insulin signaling gene set was largely selected by KEGG molecular pathway designations. We expected that our prioritization pipeline would differentially weight genes in the disease-defined gene sets higher; in fact, 57% of insulin signaling genes were eliminated before assaying from the sequencing dataset, whereas only 41% of obesity genes and 15% of dyslipidemia genes were eliminated (Fig 1). This was also reflected in the distributions of genes receiving each level of supporting evidence (Fig 3), where the insulin signaling gene set has a right-skewed distribution (most genes receiving low scores), the obesity gene set is nearly normally distributed, and the dyslipidemia gene set has a left-skewed distribution of evidence (most genes receiving high scores). In contrast, however, 84% of the 100 randomly selected genes were eliminated before assaying from the sequencing dataset, and the distribution of genes receiving each level of supporting evidence was extremely skewed towards genes receiving low scores (Fig 3). This underscores that while the insulin signaling gene set appears the weakest out of the IR gene sets, it is still starkly different from unselected, arbitrary genes. It is possible that variants in genes in the insulin signaling gene set have a smaller effect size than variants in dyslipidemia genes, with variants in obesity genes falling in the middle. The full effects of these variants may require polygenic inheritance or epistatic interactions within the pathway.

While our unbiased gene curation and filtering strategy was designed to reduce noise and increase interpretability, it may have introduced biases towards highly studied genes and rare variants. To mitigate the sequence variant interpretation bottleneck, we implemented rigorous pre-processing of genes, largely informed by the pieces of evidence that are incorporated in the ACMG variant curation (43). These criteria heavily favor genes associated with rare, monogenic disease and may not encompass genes contributing to PCOS heritability through more common, polygenic inheritance. Additionally, at the variant level, we applied a relatively stringent functional filter (REVEL ≥ 0.6). We chose this approach over a population frequency filter for two reasons: 1) to limit the inclusion of de novo variants that while stochastically rare, are not necessarily selected against and may not have clear functional impacts; and 2) to prevent the possible erroneous exclusion of relatively common variants with predicted functional effects. Our approach did, however, still yield almost entirely rare variation; the highest NFE MAF of a single variant in the REVEL-filtered dataset was 0.0633. These biases towards well-known genes and rare variants served the important purpose of increasing interpretability, but future efforts will need to expand and begin identifying the “medium-hanging fruit” type of variation that contributes to pathogenesis in ways other than rare, high-effect causes. These types of variants may require molecular evaluation for their effects to be clear.

We have demonstrated that genetic causes of IR may contribute to the heritability of PCOS in a large proportion of women, and that a candidate pathway-driven approach to focus WES data can yield manageable, interpretable data. The selected gene sets of insulin signaling, obesity, and dyslipidemia may contribute to PCOS etiology through distinct mechanisms, but all represent important areas of further investigation into disease pathogenesis. For example, the genes identified as significantly enriched for variation in PCOS cases in this study and the molecular pathways they affect are prime candidates for functional screens or mechanistic studies to clarify their causal roles. These genes and pathways may merit follow-up through whole-genome sequencing to investigate variation in regulatory regions or deep intronic sites not captured by WES. Noncoding variation has been proposed as a likely causal mechanism for common, complex traits like PCOS, but whole-genome sequencing analysis requires even greater focusing of the data to reach meaningful results. Our gene discovery strategy may be useful to identify novel genes that can subsequently be investigated in greater detail, such as for regulatory region or splice site variants. More broadly, our approach can be applied to other functional pathways that may underlie PCOS pathogenesis and other complex genetic traits. Ultimately, this work advances our understanding of the complex etiology of PCOS and continues to move the field towards personalized medicine and treatments targeted at the precise underlying causes of disease.

Table 3.

Candidate gene evidence and strength score for 15 highlighted PCOS genes.

Gene	Encoded protein	OMIMa associated disease	MIM #	mRNA expressionb	Genomic constraintc	Score	OR (95% CI)	Count Cases
LMNA	Lamin A/C	Cardiomyopathy, dilated, 1A; Charcot-Marie-Tooth disease, type 2B1; Emery-Dreifuss muscular dystrophy 2, autosomal dominant; Emery-Dreifuss muscular dystrophy 3, autosomal recessive; Heart-hand syndrome, Slovenian type; Hutchinson-Gilford progeria; Lipodystrophy, familial partial, type 2; Malouf syndrome; Mandibuloacral dysplasia; Muscular dystrophy, congenital; Restrictive dermopathy 2	115200 605588 181350 616516 610140 176670 151660 212112 248370 613205 619793	low specificity	3.1	Strong	63.3 (19.3–207.5)	5
LEPR	Leptin receptor	Obesity, morbid, due to leptin receptor deficiency	614963	liver	2.07	Strong	585.1 (30.2–11332.3)	3
KCNJ11	Potassium inwardly rectifying channel subfamily J member 11	Diabetes mellitus, transient neonatal; Diabetes, permanent neonatal 2; Hyperinsulinemic hypoglycemia, familial; MODY, type 13	610582 618856 601820 616329	skeletal muscle, tongue	2.86	Strong	125.6 (21.0–752.4)	3
BSCL2	Lipid droplet biogenesis associated, seipin	Encephalopathy, progressive; Lipodystrophy, congenital generalized, type 2; Neuronopathy, distal hereditary motor, dominant 13; Silver spastic paraplegia syndrome	615924 269700 619112 270685	brain, pituitary gland	1.04	Strong	13.1 (5.7–31.6)	6
ACACA	Acetyl-CoA carboxylase alpha	Acetyl-CoA carboxylase deficiency	613933	low specificity	10.2	Strong	15.6 (6.4–40.3)	5
NTRK2	Neurotrophic receptor tyrosine kinase 2	Developmental and epileptic encephalopathy 58; Obesity, hyperphagia, and developmental delay	617830 613886	brain, thyroid gland	5.29	Robust	167.6 (15.2–1849.6)	2
GCK	Glucokinase	Diabetes mellitus, noninsulin-dependent, late onset; Diabetes mellitus, permanent neonatal 1; Hyperinsulinemic hypoglycemia, familial, 3; MODY, type II	125853 606176 602485 125851	brain, liver, pituitary gland	2.79	Strong	404.1 (19.4–8421.5)	2
ABCC8	ATP binding cassette subfamily C member 8	Diabetes mellitus, noninsulin-dependent; Diabetes mellitus, permanent neonatal 3; Diabetes mellitus, transient neonatal 2; Hyperinsulinemic hypoglycemia, familial, 1; Hypoglycemia of infancy, leucine-sensitive	125853 618857 610374 256450 240800	brain, pancreas, pituitary gland	2.85	Strong	5.7 (3.1–10.5)	11
SLC2A2	Solute carrier family 2 member 2	{Diabetes mellitus, noninsulin-dependent}; Fanconi-Bickel syndrome	125853 227810	liver	1.58	Strong	166.9 (19.4–1841.3)	2
POMC	Proopiomelanocort in	{Obesity, early-onset, susceptibility to}; Obesity, adrenal insufficiency, and red hair due to POMC deficiency	601665 609734	pituitary gland	−1.61	Strong	2.2 (1.3–3.8)	13
MC4R	Melanocortin 4 receptor	{Obesity, resistance to (BMIQ20)}; Obesity (BMIQ20)	618406 618406	brain, fallopian tube, retina	0.311	Strong	47.7 (7.5–304.7)	2
TBC1D4	TBC1 domain family member 4	{Diabetes mellitus, noninsulin-dependent, 5}	616087	skeletal muscle	1.8	Strong	249.8 (10.2–6133.5)	1
INSR	Insulin receptor	Diabetes mellitus, insulin-resistant, with acanthosis nigricans; Donohue syndrome; Hyperinsulinemic hypoglycemia, familial, 5; Rabson-Mendenhall syndrome	610549 246200 609968 262190	low specificity	5.47	Strong	1.3 (0.8–2.2)	15
NR0B2	Nuclear receptor subfamily 0 group B member 2	Obesity, mild, early-onset	601665	intestine, liver	0.578	Strong	24.0 (5.0–115.6)	2
GCKR	Glucokinase regulator	[Fasting plasma glucose level QTL 5]	613463	liver	0.167	Strong	1.8 (0.8–4.0)	6

^aOMIM associated disease designations (8/9/2024):

[ ] genetic variation leads to an abnormal laboratory test value but not a disease.

{ }genetic variation contributes to susceptibility of a trait rather than a Mendelian relationship.

^bExpression: Tissue RNA expression, tissue specificity from Human Protein Atlas (proteinatlas.org) (8/12/2024)

^cConstraint: Constraint on missense Z score computed in gnomAD v4 (8/9/2024), Z > ±3 considered significant. Abbreviations: QTL, quantitative trait locus.