Polygenic stratification of cancer risk inside autoimmune cohorts rests on one two-biobank demonstration

Question: How does polygenic risk score stratification perform for stratifying cancer risk in autoimmune disease populations using biobank cohorts such as All of Us and FinnGen?

Answer

Across the retrieved papers, the answer is that PRS stratification performs surprisingly well in this specific setting — but the performance depends on modelling the autoimmune-disease PRS as an effect modifier rather than as a nuisance covariate, and the entire quantitative case rests on a single study. In autoimmune hypothyroidism, the combination of a hypothyroidism PRS (H-PRS) and a skin cancer PRS (SC-PRS) produced hazard ratios from 2.32 to 3.37 across joint percentile strata and a ten-year skin cancer risk of 14.9% versus 2.8% between extreme joint quartiles in 55-year-old women, with the hypothyroidism and H-PRS associations replicating in All of Us after discovery in FinnGen [6]. The critical wrinkle is directional: the diagnosis of autoimmune hypothyroidism raises skin cancer hazard while genetic liability to hypothyroidism lowers it, and the interaction term is negative [6]. Taken together, these papers support PRS-guided stratification of cancer risk in at least one autoimmune population as a genuine, replicated signal — and simultaneously show that the field has not established whether that result generalises to any other autoimmune disease, any other cancer, or any ancestry stratum.

The autoimmune diagnosis and the autoimmune PRS point in opposite directions

In FinnGen, autoimmune hypothyroidism was associated with increased skin cancer hazard (HR 1.10, 95% CI 1.06–1.15), whereas the hypothyroidism PRS was inversely associated with risk (HR 0.97 per 1-SD increase, 95% CI 0.95–0.98) [6]. The hypothyroidism × H-PRS interaction was itself negative (HR 0.94, 95% CI 0.91–0.98), meaning hypothyroidism-associated risk was greatest in patients with the lowest genetic liability to hypothyroidism [6]. This is not an isolated curiosity: in a phenome-wide analysis of 13 cancer-specific PRSs in 560,287 Million Veteran Program participants, the thyroid cancer PRS was positively associated with a hypothyroidism phecode and the oral cancer PRS with type 1 diabetes, among five positive pleiotropic cardiometabolic associations [9]. Genetic correlations between thyroid cancer and autoimmune thyroid disease are substantial (r² = 0.21–0.97) in a 19-biobank meta-analysis [12], and cluster-specific polygenic scores derived from pleiotropy-based grouping were associated with thyroid disease, cancer and metabolic traits across ancestries [13]. The practical consequence is that a cancer PRS applied to an autoimmune cohort without the autoimmune PRS in the model is mis-specified by construction, because the two scores carry opposing and interacting information.

Stratification is delivered by joint clinical-genetic profiles, not by the PRS alone

Among hypothyroid patients, predicted ten-year risk rose with higher SC-PRS and lower H-PRS, and combined clinical-genetic profiles had HRs from 2.32 at the 20th H-PRS/80th SC-PRS percentiles to 3.37 at the 5th/95th percentiles — estimates the authors report as exceeding published point estimates for established dermatological risk factors [6]. In women aged 55, ten-year skin cancer risk was 14.9% (95% CI 8.4–20.9%) in the highest-SC-PRS/lowest-H-PRS quartiles versus 2.8% (95% CI 1.0–4.7%) in the opposite quartiles [6]. This magnitude stands out against the increments reported for single-axis PRS integration elsewhere: adding an endometrial cancer PRS to epidemiological risk factors moved AUC only from 0.73 to 0.74 [2], and combining a CRC PRS with FIT did not meaningfully improve prediction at first screening [4]. A microsimulation of PRS-stratified prostate cancer screening found realistic PRS implementations were dominated in every scenario [3]. The contrast suggests the gain in autoimmune populations comes from conditioning on the disease state, which is where the heterogeneity is concentrated, rather than from the cancer PRS being unusually strong.

Replication in two biobanks is the strongest feature of the evidence — and its ceiling

The hypothyroidism and H-PRS associations replicated in All of Us, and similar H-PRS risk gradients appeared for melanoma, basal cell carcinoma and squamous cell carcinoma in both cohorts [6]. That two-cohort concordance across a Finnish isolate and a diverse US programme is the methodological high point of the retrieved literature on this question. It is also the whole of it: no retrieved paper tests an equivalent design in rheumatoid arthritis, IBD, lupus, psoriasis or type 1 diabetes, despite PRS analyses in UK Biobank showing both shared and distinct genetic architectures across autoimmune conditions and support for poly-autoimmunity [10]. Because autoimmune PRSs overlap only partially across diseases, a stratification scheme validated in autoimmune hypothyroidism cannot be assumed to transfer to another autoimmune diagnosis without re-derivation [10]. The nearest analogues stop short of cancer outcomes: a psoriasis PRS predicts psoriasis severity with an effect comparable to obesity and smoking in four European biobanks [15], and a colorectal cancer PRS used as a proxy recovered EHR-derived associations with breast cancer, prostate cancer, obesity, smoking and alcohol phenotypes in UK Biobank and eMERGE [14].

Relative-risk metrics understate or misstate performance; absolute risk changes the verdict

The authors of the hypothyroidism study frame their result in ten-year absolute risk within patient subgroups rather than as discrimination [6], which matters because AUC- and relative-risk-based evaluations do not capture background risk from contextual factors. A conditional average derivative estimator applied to breast and prostate cancer PRSs inside age-specific absolute risk models fit in All of Us with SEER inputs showed that despite the lowest discriminatory accuracy in African Americans, PRS utility for prostate cancer risk and mortality stratification was 2- and 7-fold higher in African Americans than European Americans [7]. In the same biobank, breast cancer PRS performance varied by self-reported household income, with SES-specific re-estimated effect sizes applied across White, Black and Hispanic participants [8]. For an autoimmune population — which is itself socially and clinically selected, with skewed comorbidity and surveillance patterns — the implication is that a single pooled AUC figure will not predict how the score behaves in the subgroup that actually receives it.

Translation is the weakest link, not discrimination

Twenty-seven randomised trials of PRS disclosure found no statistically significant effect on any of 22 outcomes tested in two or more trials, with relative risk 1.12 (0.77–1.61) for screening attendance and 0.95 (0.32–2.79) for disease incidence [5]. Instrument-level noise compounds this: a 313-variant breast cancer PRS measured across five genomic platforms classified 28% of individuals as high-risk on at least one platform but only 8% on all five, with pre-calibration agreement of κ = 0.552 improving to 0.650 after calibration [1]. A stratified surveillance pathway built on a PRS that changes classification with the assay platform, delivered to patients whose behavioural response to disclosure has not been demonstrated, has no demonstrated downstream benefit.

Conflicts and gaps

The central tension sits inside the one directly relevant paper: an increased risk conferred by the autoimmune diagnosis alongside a protective association of the autoimmune PRS, resolved by the authors as heterogeneity within the patient group rather than by any mechanistic account [6]. Whether the "diagnosis positive, PRS negative" pattern is a general property of autoimmune–cancer pairs or specific to hypothyroidism and keratinocyte cancers is untested; the retrieved literature offers no second autoimmune disease with a comparable design. No retrieved paper reports ancestry-stratified estimates within the autoimmune stratum, even though the All of Us replication cohort is ancestrally diverse and ancestry-dependent transferability is documented for cancer PRSs generally [7] [8]. Ascertainment is unaddressed: autoimmune and cancer phenotypes in FinnGen and All of Us are EHR/registry derived, and differential surveillance in patients with a chronic autoimmune diagnosis could generate detection-driven risk gradients indistinguishable from genetic effect modification. Direction of causation also remains inferential — Mendelian randomisation indicates multiple sclerosis increases gastric cancer risk through shared inflammatory pathways [11], which neither confirms nor excludes the effect-modification model used in the skin cancer work.

What would change the answer

A systematic screen applying the H-PRS/SC-PRS design to every autoimmune disease with adequate case counts in FinnGen and All of Us, reporting formal interaction tests and calibration rather than AUC alone, would establish whether the result generalises or is an outlier. Fitting the combined autoimmune-PRS plus cancer-PRS model inside absolute risk frameworks with registry-based incidence inputs, as done for breast and prostate cancer in All of Us with SEER [7], would convert relative hazard gradients into the ten-year risks that drive screening decisions and would allow ancestry- and SES-specific utility to be quantified in the autoimmune stratum. Bi-directional Mendelian randomisation on autoimmune liability and site-specific cancer, of the kind used to test MS–gastric cancer causality [11], would resolve whether the inverse PRS association reflects a causal protective pathway or index-event and surveillance artefacts. Finally, a decision-curve or microsimulation analysis of PRS-guided surveillance in autoimmune patients — the design that undercut PRS-stratified prostate screening [3] — would test whether the observed risk separation survives contact with uptake, false positives and cost.

References

[1] Ho, P. J., Khng, A. J., Tan, J. H. J., Goy, P.-A. V., Kamila, K. A., Li, Z., Ho, W. K., Tan, I. B. H., Chong, D. Q., Lo, E., Goh, L. L., Wee, H. L., Hartman, M., Dorajoo, R., Bertin, N., Li, J.. "Polygenic Risk Scores Across Genomic Platforms for Reliable Breast Cancer Risk Stratification". bioRxiv. 2025-04-16T00:00:00.000Z. URL: https://doi.org/10.1101/2025.04.10.648141

[2] Wang, X., Dossus, L., Gunter, M. J., Crosbie, E. J., Ong, J.-S., Glubb, D. M., O'Mara, T. A.. "Risk Stratification for Endometrial Cancer Reveals Independent Contributions of Polygenic Risk and Body Mass Index". medRxiv. 2025-02-21T00:00:00.000Z. URL: https://doi.org/10.1101/2025.02.19.25322538

[3] Dixon, P., Aning, J., Martin, R., Clements, M.. "Modelling the potential contributions of polygenic risk stratification to cost-effective screening for prostate cancer". medRxiv. 2025-05-22T00:00:00.000Z. URL: https://doi.org/10.1101/2025.05.19.25327897

[4] Krogh Nohr, A., Giehm Overby, M., Munk Nielsen, M., Hedegaard Jensen, R., Ostergaard Poulsen, L., Thorlacius-Ussing, O., Ladefoged Rasmussen, S., Andersen, B., Göogenur, I., Sorensen, E., Birger Vesterager Pedersen, O., Erikstrup, C., Brons, N., Schwinn, M., Mikkelsen, C., Topholm Bruun, M., Moller Jorgensen, M., Anders Bertelsen, C., Georg Hillingso, J., Hogdall, E., Rye Ostrowski, S., DBDS Genomic Consortium,, DCB Research Consortium,, Bogsted, M., Brondum, R. F.. "Combining genome-wide polygenic scores with registry data for colorectal cancer risk-based screening". medRxiv. 2025-12-04T00:00:00.000Z. URL: https://doi.org/10.64898/2025.12.04.25341629

[5] Russo, L., Lentini, N., Farina, S., Cristiano, A., Adduci, A., Perilli, A., Cao, C., Pastorino, R., Boccia, S., Ioannidis, J.. "Effects of polygenic risk score communication on health outcomes: systematic review and meta-analysis". medRxiv. 2025-11-06T00:00:00.000Z. URL: https://doi.org/10.1101/2025.11.04.25339466

[6] Molotkov, I., Reeve, M. P., Kolosov, N., FinnGen,, Tsao, H., Palotie, A., Ripatti, S., Daly, M. J., Gupta, S., Artomov, M.. "Genetic background shapes skin cancer risk in patients with autoimmune hypothyroidism". medRxiv. 2026-10-19T00:00:00.000Z. URL: https://doi.org/10.64898/2026.10.15.26365654

[7] Martina, F., Kachuri, L., Witte, J., Huo, D., Natarajan, P., Chatterjee, N.. "Contextualizing the Utility of Polygenic Risk Scores using Absolute Risk Models in Diverse Ancestry Populations". medRxiv. 2026-06-04T00:00:00.000Z. URL: https://doi.org/10.64898/2026.06.03.26354842

[8] Domian, H. I., Tian, X., Ong, D., Hamilton, L., Shieh, Y., Musharoff, S. A.. "Breast cancer polygenic risk score performance varies by socioeconomic status". medRxiv. 2026-06-04T00:00:00.000Z. URL: https://doi.org/10.64898/2026.06.03.26354819

[9] Wang, L., Desai, H., Teerlink, C., Clark, O., Pridgen, K. M., VA Million Veteran Program,, Lee, J., Chang, K.-M., Kember, R. L., Vujkovic, M., Lynch, J. A., Maxwell, K. N.. "Cross-ancestry and phenome-wide associations of cancer-specific polygenic risk scores". medRxiv. 2025-05-16T00:00:00.000Z. URL: https://doi.org/10.1101/2025.05.15.25327704

[10] Saurabh, R., Wohlers, I., Moeller, M., Busch, H.. "Genetic Profiling of Autoimmune Diseases and Exploring Clusters Through Polygenic Risk Score Analysis Using Cohort Data from the UK Biobank". medRxiv. 2026-05-13T00:00:00.000Z. URL: https://doi.org/10.64898/2026.05.09.26352677

[11] Liao, H., Wang, X., Zhang, Y., Zhang, Z., Liao, Y.. "Multiple Sclerosis Drives Gastric Cancer via Shared Inflammatory Pathways: A Mendelian Randomization Study". medRxiv. 2025-09-03T00:00:00.000Z. URL: https://doi.org/10.1101/2025.09.01.25334880

[12] White, S. L., Brasher, M. S., Pattee, J., Zhou, W., Chapman, S., Jee, Y. H., Bell, C. C., Jamil, T. L., Barrio, M., Hirbo, J., Cox, N. J., Straub, P., Namba, S., Bertucci-Richter, E., Guare, L., EdrisMohammed, A., Morris, S., Mulford, A. J., Zhang, H., Fennessy, B., Tobin, M. D., Chen, J., Williams, A. T., John, C., van Heel, D. A., Mathur, R., Finer, S., Moksnes, M. R., Brumpton, B., Asvold, B. O., Peculis, R., Rovite, V., Konrade, I., Wang, Y., Crooks, K., Chavan, S., Fisher, M. J., Rafaels, N., Lin, M., Shortt, J., Sanders, A. R., Whiteman, D., MacGregor, S., Medland, S., Thorsteinsdottir,. "Global multi-ancestry genetic study elucidates genes and biological pathways associated with thyroid cancer and benign thyroid diseases". medRxiv. 2025-12-02T00:00:00.000Z. URL: https://doi.org/10.1101/2025.05.15.25327513

[13] Jee, Y. H., Pozdeyev, N., Gignoux, C., White, S., Haugen, B. R., Verma, A., Peculis, R., Rovite, V., Mulford, A. J., Sanders, A. R., Kitahara, C. M., Reeve, M. P., Kraft, P., Martin, A.. "Genetic drivers of etiologic heterogeneity in thyroid cancer". medRxiv. 2025-05-16T00:00:00.000Z. URL: https://doi.org/10.1101/2025.05.15.25327708

[14] Rosenthal, E. A., Wei, W.-Q., Luo, Y., Namjou-Khales, B., Schaid, D. J., Esplin, E. D., Lape, M., Kottyan, L., Pacheco, J. A. P. A., Weng, C., Gordon, A. S., Kullo, I. J., Crosslin, D. R., Grady, W. M., Hsu, L., Peters, U., Jarvik, G. P.. "Mixed-effects polygenic risk score Phenome-wide association study detects genetic correlation between colorectal cancer risk and phenotype data extracted from the electronic health record". medRxiv. 2025-02-28T00:00:00.000Z. URL: https://doi.org/10.1101/2025.02.26.25322864

[15] Saklatvala, J. R., Lessard, S., Teder-Laving, M., Thomas, L. F., Ramessur, R., Asvold, B. O., Barton, A., Baudry, D., Bowes, J., Brumpton, B., Chandran, V., Chatelain, C., de Rinaldis, E., Elder, J. T., Ellinghaus, D., Foerster, J., Franke, A., Gladman, D. D., Gulliver, W., Hüffmeier, U., Huilaja, L., Hveem, K., Khader, S., Kingo, K., Klinger, K., Koks, S., Liao, W., Nair, R. P., Nititham, J., Rahman, P., Reis, A., Stuart, P. E., Tasanen, K., Traks, T., Tsoi, L. C., Uebe, S., Watts, K., BSTOP study group,, Barker, J. N., Mahil, S. K., Langan, S. M., FinnGen,, Estonian Biobank research tea. "Genetic liability to psoriasis predicts severe disease outcomes". medRxiv. 2025-03-04T00:00:00.000Z. URL: https://doi.org/10.1101/2025.03.04.25323079

How this survey was made

This survey answers one question: “How does polygenic risk score stratification perform for stratifying cancer risk in autoimmune disease populations using biobank cohorts such as All of Us and FinnGen?”

We keep our own collection of recent preprints and refresh it every day from bioRxiv, medRxiv, arXiv, PsyArXiv, SocArXiv and ChemRxiv. For each survey, we search that collection for recent work and keep the strongest matches: papers that use the same terms as the question, and papers on the same topic in different words.

The text was written from the titles and abstracts of those papers; the full texts were not read. Statements that draw on a particular paper are marked with a number in brackets. The reference list just above is the 15 papers whose work this text draws on, chosen from the 128 papers those searches returned; the rest of the shortlist is not printed. It is not a complete review of everything published on the question.

Treat this as a starting point and a reading list, not a peer-reviewed review. The searches behind this survey ran on October 22, 2026 and covered work from the previous two years.