Which Inherited Pathways Thin the Fibrous Cap? Pathway-Partitioned Polygenic Scores Against Coronary OCT

A polygenic risk score adds up thousands of small inherited DNA differences into one number for a person's lifetime liability to coronary artery disease. It predicts heart attacks reasonably well, but it says nothing about what is going wrong inside the artery wall. Intravascular optical coherence tomography (OCT), a catheter that images the vessel wall with near-infrared light at 10–15 µm resolution, shows exactly that. It can see whether a fatty plaque is covered by a dangerously thin fibrous cap (a "thin-cap fibroatheroma," cap under 65 µm), whether the plaque is mostly lipid or mostly calcium, and whether macrophages have moved in. No one has connected the two. We propose to split the coronary polygenic score into biologically defined parts: lipid handling, inflammation and macrophage traffic, extracellular matrix and collagen turnover, and vascular tone. We then test whether each part predicts a different OCT-measured plaque feature in patients who already underwent OCT.

The claim is falsifiable and specific. If inherited risk acts through distinct mechanisms, the score-by-feature association matrix should be block-structured: lipid-pathway liability tracks lipid arc and cap thinning, matrix-pathway liability tracks cap thickness and rupture, and inflammatory liability tracks macrophage signal. Size- and heritability-matched random gene sets should show no such structure. If the matrix is flat, then the polygenic score is clinically useful but mechanistically uninformative at the vessel wall. That is also worth knowing before anyone proposes "genetics-guided" plaque therapy.

What Exists and What Is Missing

Three bodies of work motivate the design. Effect of clonal hematopoiesis on plaque morphology and prognosis linked acquired blood-cell mutations to OCT plaque vulnerability in 355 Fuwai Hospital STEMI patients with culprit-lesion OCT. Carriers of TET2/ASXL1 mutations had more thin-cap fibroatheromas (74.9% vs. 51.9%) and ruptured caps (70.2% vs. 47.1%). That study profiled only somatic mutations, not inherited common variants. The large UK Biobank studies, Polygenic risk score improves the accuracy of a clinical risk score for coronary artery disease and Highly elevated polygenic risk scores are better predictors of myocardial infarction risk early in life than later, showed that an ensembled score is nearly uncorrelated with the Pooled Cohort Equations (r = 0.01) and adds discrimination. Their endpoints, however, are billing-code heart attacks, with no vessel-wall phenotype. Pathway-restricted scores have isolated specific imaging signatures in the brain. NETRIN1-pathway PRS and white matter microstructure found white-matter changes that the genome-wide score obscured, and glutamatergic gene scores tracked brain glutamate. We borrow this partitioning logic and apply it to a vascular imaging phenotype for the first time.

Two lines of work supply method. FiAt-Net segments fibrous caps in 3D OCT pullbacks with 81.93% F1 on PREDICT-trial data. Attenuation-coefficient A-line classification and volumetric SVM plaque classification give per-A-line tissue labels, where an A-line is a single radial depth scan through the wall. Together these let us replace binary "TCFA yes/no" calls with continuous measurements: minimum cap thickness, lipid arc in degrees, calcium arc, and macrophage-signal extent.

One paper constrains the design. Evaluation of Phenotyping Errors on Polygenic Risk Score Predictions shows that label noise, especially noise correlated with covariates, drags polygenic prediction toward chance. With only a few hundred patients we cannot afford noisy binary phenotypes. The continuous, reader-validated measurements are therefore central to the study's statistical power, not a convenience.

Core Claim

Coronary polygenic liability decomposes into pathway components that map onto distinguishable OCT plaque phenotypes. This mapping is not explained by measured LDL cholesterol, by high-sensitivity C-reactive protein, or by clonal hematopoiesis status.

Research Program

Cohorts and genotyping. The base cohort is the 355 Fuwai STEMI patients with OCT. Their blood-cell DNA already exists because it was deep-sequenced for clonal hematopoiesis. The panel is targeted, so we would add genome-wide array genotyping and imputation. For replication and to cover non-acute disease, we would genotype the Beijing Anzhen OCT cohort, which includes patients undergoing angiography for broader indications. Both cohorts are Han Chinese. The UK Biobank result that score gains were negligible in African-ancestry participants warns us not to use European-only weights. We therefore derive effect sizes from GWAS that include Japan Biobank alongside CARDIoGRAMplusC4D, as the integrated score did.

Partitioned scores. We compute PRS-CS posterior SNP weights genome-wide once, then assign each SNP to a pathway by gene annotation. PleioPred showed that functional annotation improves weight estimation, and we follow it in using annotation for the assignment. Pathway sets are fixed before any OCT data are unblinded. Unassigned SNPs form a residual score, so the parts sum to the genome-wide score. This guarantees the decomposition loses nothing relative to the global score.

Phenotypes. We fine-tune a FiAt-Net-style cap segmenter and an attenuation-based tissue classifier on a manually annotated subset of Fuwai pullbacks. Agreement with two blinded expert readers must reach an intraclass correlation of at least 0.8 before the outputs are used genetically.

There is a critical design detail. Culprit lesions in STEMI have usually already ruptured, so their cap thickness cannot be measured and their morphology is a post-event state. Our primary phenotypes therefore come from non-culprit plaques within the same pullback. Culprit-lesion rupture versus intact-cap erosion is a secondary, binary outcome.

Power. For a continuous phenotype, the sample needed to detect a score explaining a fraction $R^2$ of variance is roughly $n \approx (z_{1-\alpha/2}+z_{1-\beta})^2 / R^2$. This follows from the large-sample test of a regression slope. At 80% power and $\alpha=0.05$ that gives $n \approx 7.85/R^2$, about 390 patients for $R^2 = 0.02$. With four pathways, four phenotypes and Bonferroni correction, Fuwai alone is underpowered for small effects. We therefore treat Fuwai as discovery of the matrix structure, judged as a whole-matrix test rather than cell by cell, and Anzhen as replication. We will state this limit openly instead of reporting isolated significant cells.

Index-event bias. Every Fuwai patient already had a heart attack. Among cases, people with high inherited risk may need fewer vulnerable plaque features to have had an event. That can produce null or even inverse associations, an artifact of selecting on the outcome. The Anzhen cohort, which is not restricted to acute events, is our check: a sign flip between cohorts would point to selection rather than biology.

Baselines

The naive reference is a regression of each OCT phenotype on age, sex, smoking, diabetes, blood pressure, measured LDL and hsCRP. It sets how much plaque variance bedside data already explain. On the same cohort we also compute the Pooled Cohort Equations risk as a single covariate. No reported result found on Lacuna for clinical scores predicting continuous OCT cap thickness.

The strongest existing competitor is a single genome-wide coronary score built as in the integrated PRS study. That study reports a C-statistic of 0.640 for the score alone, 0.718 for the Pooled Cohort Equations, and 0.753 combined, all for incident CAD in UK Biobank. No reported result found on Lacuna for any polygenic score against OCT plaque features. We build it from the same GWAS, SNPs and PRS-CS weights as the partitions, so the only difference is aggregation. Beating it means the partitioned model explains significantly more phenotype variance across the four features than the global score. The test is a likelihood-ratio test with three extra degrees of freedom, cross-validated within cohort and replicated in Anzhen.

The somatic baseline is TET2/ASXL1 clonal hematopoiesis status at the 0.5% variant allele frequency threshold, the setting where the Fuwai study found 74.9% vs. 51.9% TCFA prevalence. We fit it in the same patients. The partitions must add explained variance beyond it, and we test the interaction between inherited risk and clonal hematopoiesis explicitly.

Two ablations isolate the key ingredients. The first replaces each biological partition with 1,000 random SNP sets matched for SNP count and heritability contribution. The observed block structure, measured as the share of explained variance on prespecified diagonal cells, must exceed the 95th percentile of the random-set distribution. The second reruns the primary analysis with expert binary TCFA and lipid-rich labels instead of automated continuous measurements. It tests whether quantitative phenotyping actually buys power, as the phenotyping-error paper predicts.

Why not just…?

Why not just correlate the genome-wide coronary score with TCFA? This is the obvious study and our main baseline. Even a positive result would only show that genetically high-risk people have worse plaques, which the incident-MI literature already implies. It cannot say through which mechanism. The random-set ablation directly tests whether partitioning carries information beyond the total. If it does not, we report that the global score is sufficient.

Why not just use measured LDL and hsCRP? This is a real threat for the lipid partition. Most lipid-pathway genetic liability may act through circulating LDL, and then adding the lipid score after adjusting for LDL will show nothing. That outcome is informative in itself. The matrix and inflammatory partitions are where inherited signal should survive adjustment, because there is no routine blood measurement of collagen turnover. We report every partition with and without biomarker adjustment. hsCRP also fluctuates acutely, especially after STEMI, whereas germline scores do not.

Why not just do this in UK Biobank with retinal OCT or with CT angiography, where samples are huge? Retinal OCT predicts events, with AUC 0.75 in García et al., but it images the choroid, not coronary caps. CT angiography cannot resolve sub-65 µm caps: O2CTA found clinician agreement with OCT tissue labels of only κ = 0.113. Under the phenotyping-error results, that level of noise would swamp the small effects we seek. Scale is the honest counterargument. If our OCT-derived partition map replicates, a natural follow-up is to test it on OCT-supervised CT labels of the O2CTA type at biobank scale. We do not start there, because we would be unable to tell a null from measurement noise.

Decisive Checks

The study succeeds if three conditions hold together. The prespecified diagonal share of the partition-by-phenotype matrix beats the matched random-set null in Fuwai. The same blocks replicate with the same sign in Anzhen. The matrix and inflammatory partitions retain association after adjustment for LDL, hsCRP and clonal hematopoiesis. A flat matrix with a significant global score would mean that inherited risk worsens plaques diffusely. A sign reversal between the acute and non-acute cohorts would mean index-event bias dominates the STEMI data. Either result is publishable and redirects the field. Only the block-structured outcome would justify pathway-guided genetic stratification of plaque risk.