Regulatory risk loci link disrupted androgen response to pathophysiology of Polycystic Ovary Syndrome

Abstract

A major challenge in deciphering the complex genetic landscape of Polycystic Ovary Syndrome (PCOS) lies in the limited understanding of how susceptibility loci drive molecular mechanisms across diverse phenotypes. To address this, we integrated molecular and epigenomic annotations from proposed causal cell-types and employed a deep learning (DL) framework to predict cell-type-specific regulatory effects of PCOS risk variants. Our analysis revealed that these variants affect key transcription factor (TF) binding sites, including NR4A1/2, NHLH2, FOXA1, and WT1, which regulate gonadotropin signaling, folliculogenesis, and steroidogenesis across brain and endocrine cell-types. The DL model, which showed strong concordance with reporter assay data, identified enhancer-disrupting activity in approximately 20% of risk variants. Notably, many of these variants disrupt TFs involved in androgen-mediated signaling, providing molecular insights into hyperandrogenemia in PCOS. Variants prioritized by the model were more pleiotropic and exerted stronger downregulatory effects on gene expression compared to other risk variants. Using the IRX3-FTO locus as a case study, we demonstrate how regulatory disruptions in tissues such as the fetal brain, pancreas, adipocytes, and endothelial cells may link obesity-associated mechanisms to PCOS pathogenesis via neuronal development, metabolic dysfunction, and impaired folliculogenesis. Collectively, our findings highlight the utility of integrating DL models with epigenomic data to uncover disease-relevant variants, reveal cross-tissue regulatory effects, and refine mechanistic understanding of PCOS.

Keywords: Polycystic Ovary Syndrome (PCOS); artificial intelligence; deep learning; disease-causal noncoding variants; enhancer variants; regulatory genomics.

Figure 1:

(A) PCOS susceptibility loci and their distribution in non-coding regions, (B) Gene Ontology annotations of target genes, (C) Fold enrichment of PCOS eVariants in GTEx cell-types (reported eVariants with enrichment binomial p-value < 0.01).

Figure 2:

(A) ROC and PRC curves of eleven cell-type specific TREDNet models, (B) A comparison of fold change (alternate / reference allele) in TREDNet scores between all variants and those exhibiting significant change in enhancer activity in MPRA, using Wilcoxon test (C) Fraction of SNVs overlapping with phastCons elements conserved across 30 primates, (D) TFs enriched among reSNVs compared with control SNVs (hypergeometric p-value < 0.01). (ns: p > 0.05, *: p <= 0.05,**: p <= 0.01,***: p <= 0.001)

Figure 3:

Regulatory impact of reSNVs prioritized by TREDNet. (A) Fold enrichment of reSNVs compared to pcosSNVs across cell-types (binomial p < 0.01). (B) Comparison of the number of GTEx cell-types impacted by reSNVs versus otherSNVs, (C) Normalized effect size of reSNVs versus otherSNVs. Left and right panels show differences for downregulating (NES ≤ −0.5) and upregulating (NES ≥ 0.5) variants, respectively. (D) Genomic overlap of an intronic reSNV (rs1784692) at the ZBTB16 locus with epigenomic features from cell types where it exhibits predicted allele-specific activity. The affected Androgen Receptor (AR) motif is shown below. (E) Functional enrichment of biological processes in the ZBTB16 protein interaction network (STRING database). The plot shows the top 10 terms (FDR < 0.001), with enrichment strength calculated as log₁₀(observed/expected). (ns: p > 0.05, *: p <= 0.05,**: p <= 0.01,***: p <= 0.001)

Figure 4:

reSNVs in FTO locus exhibiting significant fold change in TREDNet predicted enhancer activity. (A) Overlap of reSNVs with active regulatory regions of pathogenic cell-types (B) Intact Hi-C map of chromatin interactions from reSNVs in FTO locus in HUVEC (doi:10.17989/ENCSR788FBI)