Polygenic and Gut Metagenomic Risk Integration
Summary¶
Liu, Ritchie, Teo, et al. (2024) use the prospective, population-based FINRISK 2002 cohort (5,676 participants with linked genotyping and baseline shotgun gut metagenomic sequencing, median 17.8 years of electronic-health-record follow-up) to jointly evaluate polygenic risk scores (PRS), gut microbiome risk scores, and conventional risk factors for incident coronary artery disease (CAD), type 2 diabetes (T2D), Alzheimer disease (AD), and prostate cancer. PRSs consistently and significantly improved prediction over conventional risk factors alone for all four diseases; gut microbiome scores also significantly predicted all four diseases individually, but their incremental value on top of conventional risk factors was small and, after adjustment, was significant for T2D and prostate cancer only, not CAD or AD. Combined models integrating PRS, gut microbiome score, and conventional risk factors achieved the best predictive performance for every disease studied.
Cohort and Method¶
- Cohort: FINRISK 2002, a Finnish population-representative survey cohort (8,783 total participants); 5,676 had both genotyping and eligible gut shotgun metagenomic sequencing (excluding low-read-depth samples, baseline pregnancy, extreme BMI, and recent antibiotic use).
- Outcomes: 333 incident CAD, 579 incident T2D, 273 incident AD, and 141 incident prostate cancer cases, identified via linkage to Finnish national health registries (ICD/ATC codes) over a median 17.8-year follow-up.
- PRS: previously validated, PGS-Catalog-sourced scores for each disease (CAD: PGS000018; T2D: PGS000036; AD: PGS000334; prostate cancer: PGS000662).
- Gut microbiome score: a ridge logistic regression model trained on abundance of 235 species-level taxonomic groups (10× three-fold stratified cross-validation) to generate a predictive microbiome risk score per disease.
- Statistical models: sex-stratified (except prostate cancer) Cox proportional hazards models, comparing individual risk factors, PRS, gut microbiome score, and combined models by concordance statistic (C-statistic) and hazard ratio (HR) per standard deviation.
Individual Predictor Performance¶
- Baseline age dominates: for CAD, AD, and prostate cancer, baseline age was the single strongest individual predictor (C-statistics 0.719, 0.880, and 0.769 respectively); for T2D, BMI was strongest (C-statistic 0.745).
- PRS effect sizes: HR per SD was 1.68 for CAD (P=2.25×10⁻¹⁹), 1.42 for T2D (P=6.48×10⁻¹⁵), 1.92 for AD (P=4.27×10⁻³²), and 1.73 for prostate cancer (P=5.50×10⁻¹¹) — all significant after adjusting for disease-specific conventional risk factors, and independent of family history (implying PRS and family history are complementary, non-redundant predictors).
- PRS incremental gain (ΔC-statistic over conventional risk factors): +0.023 (CAD), +0.010 (T2D), +0.017 (AD), +0.027 (prostate cancer) — all significant. In an AD subanalysis restricted to participants ≥60 years, the PRS alone (C-statistic 0.667) outperformed any single conventional risk factor or their full combination, and improved C-statistic over conventional factors by 0.064.
- Gut microbiome effect sizes: HR per SD was 1.28 (CAD), 1.40 (T2D), 1.34 (AD), and 1.50 (prostate cancer) when modeled alone — all significant. After adjusting for conventional risk factors, the microbiome score's effect was attenuated and remained significant only for T2D (HR=1.20, P=9.13×10⁻⁶) and prostate cancer (HR=1.23, P=0.020); no significant independent effect remained for CAD or AD.
- Gut microbiome incremental gain: small — ΔC-statistic of only +0.004 (T2D) and +0.005 (prostate cancer, not statistically significant, 95% CI crossing zero) over conventional risk factors; species-level taxa-based cross-validated classifiers achieved modest standalone AUROC (0.564–0.613 across the four diseases).
- Diversity metrics: gut microbial α-diversity (Shannon/Chao-Shannon index) was negatively associated with incident T2D (HR≈0.89–0.90 per SD), consistent with prior reports linking lower microbial diversity to insulin resistance; species richness was positively associated with incident prostate cancer (HR=1.23).
Combined (Integrated) Model Performance¶
Models combining conventional risk factors + PRS + gut microbiome score achieved the highest C-statistics of any model configuration for all four diseases, with ΔC-statistic over conventional-risk-factors-alone of:
| Disease | ΔC-statistic (combined vs. conventional) |
|---|---|
| CAD | +0.024 (95% CI 0.013–0.035) |
| T2D | +0.014 (95% CI 0.007–0.021) |
| AD | +0.017 (95% CI 0.009–0.024) |
| Prostate cancer | +0.031 (95% CI 0.011–0.050) |
Adding PRS + gut microbiome score to baseline age alone produced larger gains — e.g., +0.076 C-statistic for T2D over age alone — comparable in magnitude to adding the entire panel of conventional risk factors, underscoring that age-only screening substantially underuses available predictive information.
Significance and Limitations¶
This is among the first studies to jointly model host genetics, gut metagenomics, and conventional clinical risk factors for prospective incident-disease prediction (rather than cross-sectional or retrospective case-control comparisons, which are more susceptible to selection bias). The authors conclude that while the gut microbiome is significantly associated with several incident diseases, its added predictive value on top of PRS and conventional risk factors is modest and disease-specific — the clearest evidence supports T2D and, to a lesser extent, prostate cancer, with limited support for CAD or AD. The study explicitly cautions that larger cohorts are needed to more precisely quantify the microbiome's incremental clinical utility, and that PRS provides the more consistent multiomic improvement across the diseases studied.
See Also¶
- Polygenic Risk Scores — the underlying PRS concept and methodological considerations this study builds on.
- PGS Catalog — source of the validated disease PRS used in this study.
- Multimodal Disease Risk Prediction — a related paradigm of combining multiple data modalities for risk prediction, though using deep sequence models rather than Cox regression of discrete risk scores.