Experimental benchmarking of in silico annotations at prostate cancer risk loci and hypothesis-generating cross-cancer prioritisation of ovarian risk variants

Abstract Genome-wide association studies (GWAS) have identified numerous non-coding risk variants for prostate and ovarian cancers, yet determining which are regulatorily active and which target genes they control remains challenging. We re-analysed a dinucleotide-barcode massively parallel reporter assay (DiR-seq) of 208 bi-allelic prostate cancer risk single-nucleotide polymorphisms (SNPs) in three prostate cancer cell lines under false-discovery-rate (FDR) control, identifying 50 regulatory SNPs at FDR < 0.10 (30 at FDR < 0.05), including 9 shared by at least two cell lines. Benchmarking standard in-silico functional annotations against this experimental ground truth revealed poor predictive performance (cross-validated ROC-AUC = 0.55–0.59), demonstrating that computational predictions cannot reliably substitute for measured allelic activity. We therefore treated the ovarian analysis as hypothesis-generating. Convergent annotation evidence prioritised 44 of 176 ovarian cancer GWAS leads (31 independent loci); a core of 21 SNPs (17 loci) was stable under alternative weights, leave-one-feature-out scoring and threshold changes, and candidates were enriched in ovarian and fallopian-tube regulatory chromatin (OR = 7.9, FDR = 5 × 10⁻5). Copy-number recurrence at candidate genes agreed between two tumour cohorts (Spearman ρ = 0.75, 95% CI 0.46–0.88) but no more than for chromosome-arm-matched control genes, so this observation is exploratory. Formal colocalisation with full summary statistics supported a shared causal variant for prostate cancer risk and TLCD3A expression at 17p13.3 (PP.H4 = 0.92), whereas the ovarian 17q21.31 signal could not be resolved to a single gene. Transcription-factor motifs disrupted in both cancers were over-represented for the interferon-regulatory factor (IRF) and STAT families (FDR = 3 × 10⁻6), a motif-level hypothesis. In an independent MPRA in primary ovarian epithelial cells, variants disrupting interferon-stimulated response elements (ISREs) were depleted, not enriched, for allele-specific activity (4.6% vs 9.2%; OR = 0.47, nominal p = 0.06, FDR = 0.10), so the IRF/STAT hypothesis was not supported; the same assay confirmed allele-specific activity at one prioritised ovarian candidate, rs2976384 (skew = − 0.67, FDR = 3 × 10⁻5). These results provide an audited prostate regulatory-SNP catalogue, quantify the limits of annotation-based fine-mapping, and offer testable cross-cancer hypotheses.