Computational Predictions Fail to Match Measured Activity at Prostate Cancer Risk Variants

Genome-wide association studies have linked thousands of non-coding DNA variants to prostate and ovarian cancer, but knowing which variants actually switch genes on or off has remained stubbornly difficult. A new study published in Genomics & Informatics on 28 October 2026 tackled this problem by re-analysing a massively parallel reporter assay of prostate cancer risk variants under strict statistical error control, and then asking a blunt question: can the computational annotations that researchers routinely rely on actually predict which variants are functionally active? The answer, according to the data, is largely no.

The research team, led by Wiaam Ahmed Al-Amili and colleagues, re-analysed a dinucleotide-barcode massively parallel reporter assay, or DiR-seq, that tested 208 bi-allelic prostate cancer risk single-nucleotide polymorphisms across three prostate cancer cell lines: 22Rv1, PC-3 and LNCaP. The original study had interpreted these data with uncorrected statistical thresholds, leaving room for a more rigorous treatment. Applying false-discovery-rate control with the MPRAnalyze framework, the team identified 50 regulatory SNPs at FDR below 0.10, of which 30 remained significant at the stricter threshold of FDR below 0.05. Nine SNPs were shared by at least two cell lines, forming a robust core, while 39 were cell-line-specific, including 33 unique to 22Rv1.

The flagship variant, rs684232, showed strong allele-specific activity in 22Rv1 cells, with a log2 fold-change of +0.57 for the T allele relative to the C allele and an FDR of approximately 1 × 10⁻⁶, meaning the C risk allele lowers reporter activity. Notably, when the team oriented each allelic effect by the reported GWAS risk allele, they found that risk alleles do not uniformly increase enhancer activity: among 33 regulatory SNPs with a matching reported risk allele, the skew was positive for 15 and negative for 18. A secondary cross-check using a limma empirical-Bayes model confirmed the direction of all 63 per-line calls.

With this experimentally grounded catalogue in hand, the researchers benchmarked standard in-silico annotations against it. They assessed transcription factor motif disruption with motifbreakR, regulatory feature overlap via Ensembl, and expression quantitative trait loci from GTEx, then trained elastic-net logistic regression and random forest classifiers to predict the experimental regulatory calls. The models performed barely better than chance: cross-validated ROC-AUC values ranged from 0.55 to 0.59. Adding cell-type-matched ENCODE epigenomic tracks, such as open chromatin and active histone marks, only marginally improved performance. The practical implication is stark: computational predictions cannot reliably substitute for measured allelic activity, and annotation-based fine-mapping should be treated with corresponding caution.

Because the benchmark showed annotations to be weak predictors, the team deliberately framed their extension to ovarian cancer as hypothesis-generating rather than definitive. Starting from 176 genome-wide-significant ovarian cancer GWAS lead variants, they scored each on convergent evidence including motif disruption, regulatory overlap and ovary eQTL status. This yielded 44 candidates across 31 independent loci, and a robust core of 21 SNPs across 17 loci that survived leave-one-feature-out scoring, alternative weightings, randomised weights and threshold changes. The candidates were strongly enriched in ovarian and fallopian-tube regulatory chromatin, with an odds ratio of 7.9 and FDR of 5 × 10⁻⁵, though the authors note this enrichment is partly non-independent of the features used for scoring.

Leading ovarian candidates included rs4808616 near ANKLE1, variants in the HOXD3/HAGLROS region, rs244353 near STXBP4, rs4743687 near SMC2, and rs1862626, which disrupts an ESR1 motif. Survival analysis of 43 candidate genes in the TCGA-OV tumour cohort produced no association that survived multiple-testing correction; the strongest nominal signal, HAGLROS with a log-rank p-value of 0.007, had an FDR of 0.30 and was not interpreted further. Copy-number alterations at candidate genes were recurrent across two independent ovarian tumour cohorts and correlated between them, with a Spearman correlation of 0.75, but the authors report this as exploratory because chromosome-arm-matched control genes showed similar cross-cohort agreement, reflecting the shared copy-number landscape of high-grade serous ovarian cancer rather than anything specific to the candidates.

The cross-cancer analysis produced one striking motif-level observation. Intersecting disrupted transcription factor motifs between the prostate regulatory SNPs and the ovarian candidates revealed that the shared fraction was dominated by the interferon-regulatory factor (IRF) and STAT families: IRF1 through IRF9 and STAT2 motifs were disrupted in both cancers, an over-representation with an FDR of 3 × 10⁻⁶. Eighteen risk SNPs disrupted these motifs, frequently hitting multi-factor interferon-stimulated response element cassettes. The researchers proposed this as a testable hypothesis of shared immune-regulatory disruption, while cautioning that IRF position weight matrices are highly similar, so a single SNP can disrupt several of them at once.

Formal colocalisation using full summary statistics provided firmer ground at the flagship prostate locus. At 17p13.3, the prostate GWAS signal fine-mapped to a three-variant credible set led by rs684232, and coloc.abf supported a shared causal variant for prostate cancer risk and TLCD3A expression with a posterior probability of 0.92, with more moderate support for GEMIN4 at 0.77. For VPS53, the dominant eQTL signal was driven by a different variant, but a secondary, independent VPS53 eQTL credible set containing rs684232 overlapped the GWAS credible set, consistent with earlier reports. At the ovarian 17q21.31 locus, by contrast, the signal lies within the extensive H1/H2 inversion haplotype, with a credible set of 1,898 variants, and could not be resolved to a single gene; the team explicitly withdrew a previously reported colocalisation with LINC02210 that had relied on a regional overlap test rather than formal colocalisation.

The experimental test of the interferon hypothesis delivered a clear negative result. In an independent MPRA dataset covering 5,095 cancer risk variants across 14 primary human cell types, the team re-analysed data from primary ovarian epithelial cells, yielding 497 allele-specific SNVs. Variants disrupting interferon-stimulated response elements were depleted, not enriched, for allele-specific activity: 4.6 percent were active versus 9.2 percent of non-disrupting variants, an odds ratio of 0.47 with a nominal p-value of 0.06 and FDR of 0.10. The same assay did confirm allele-specific activity at one prioritised ovarian candidate, rs2976384, with an allelic skew of −0.67 and FDR of 3 × 10⁻⁵. Cross-platform reproducibility between the two assays was otherwise low, with correlations near zero for the 16 shared SNPs, likely reflecting differences in cellular context and assay design, including 55-base-pair inserts with one barcode per allele in DiR-seq versus ten barcodes per allele in the independent dataset.

The study’s conclusions are deliberately measured. It delivers an audited, FDR-controlled catalogue of prostate cancer regulatory variants, quantifies the limits of annotation-based fine-mapping, and offers a set of explicitly hypothesis-generating cross-cancer candidates, one of which showed experimental support. The null result on the IRF/STAT hypothesis reinforces the paper’s central caution: motif-level predictions are often uncoupled from measured function. The authors point to the limitations of episomal reporter assays, which lack endogenous chromatin context and three-dimensional genome organisation, and call for tissue-matched allele-specific binding assays, CRISPR-based enhancer perturbations and in-vivo validation as the necessary next steps for translating non-coding cancer risk associations into biological and eventually clinical insight.

Subject of Research: Functional benchmarking of in silico annotations against measured regulatory activity at prostate and ovarian cancer risk variants

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

Article References: Al-Amili, W. A., Ayyed Najm, M., Jumaah, S. S., & Suleiman, A. A. (2026). Experimental benchmarking of in silico annotations at prostate cancer risk loci and hypothesis-generating cross-cancer prioritisation of ovarian risk variants. Genomics & Informatics, 24(1), Article 24. https://doi.org/10.1186/s44342-026-00087-2

Image Credits: AI Generated

DOI: 10.1186/s44342-026-00087-2

Keywords: prostate cancer, ovarian cancer, GWAS, massively parallel reporter assay, regulatory variants, non-coding DNA, colocalisation, transcription factor motifs, fine-mapping, functional genomics, in silico annotation, eQTL