Skip to content

Pleiotropic Decomposition Regression (PDR)

Summary

Pleiotropic Decomposition Regression (PDR) decomposes each variant's marginal effect-size vector across a cluster of genetically correlated traits into a sum of a small number of independent components, each with its own "pattern" describing which traits it affects and how strongly. Applied to trait clusters built around coronary artery disease, asthma, and type 2 diabetes, it recovered mechanistically distinct components, such as separate BMI-driven and lipid-driven components within the CAD cluster, and produced posterior-mean effect-size estimates that substantially improved out-of-sample prediction accuracy relative to the original GWAS summary statistics.

Method

PDR models each component as a scale mixture of multivariate normal distributions governed by a trait-by-component "pattern matrix," and fits the pattern matrices and per-variant mixing weights by matching the empirical characteristic function of the observed GWAS summary statistics against the model's expected characteristic function at randomly sampled points (a time-domain moment-matching procedure), using iterative estimation and gradient descent.[1] The output — posterior-mean, denoised effect sizes for each variant, decomposed by component — pools information across genetically correlated traits in a way that is interpretable per component, unlike black-box multi-trait shrinkage.

Results

Applied to three trait clusters (5–6 traits each, selected by genetic correlation with a focal disease):[1]

  • Coronary artery disease (+ BMI, hypertension, total cholesterol, triglycerides, HDL): 3 components — a BMI-related component explaining ~18% of CAD heritability with brain-tissue enrichment; a hypertension-related component explaining ~50% of hypertension and CAD heritability with minimal BMI contribution; and a cholesterol-related component explaining ~50% of high-cholesterol/triglyceride/HDL heritability and ~14% of CAD heritability, enriched in visceral adipose tissue.
  • Asthma (+ BMI, eczema/allergies, eosinophil count, FEV1/FVC): 3 components — BMI-related (brain-tissue enrichment), inflammatory (eosinophils/eczema/asthma, spleen and small-intestine enrichment), and pulmonary-function-related (strongest FEV1/FVC association, esophagus/tibial-artery enrichment).
  • Type 2 diabetes (+ BMI, fasting glucose, HbA1c, total cholesterol, waist-hip ratio adjusted for BMI): 3 components corresponding to BMI, waist-hip ratio, and glucose/HbA1c-driven genetic effects.

Posterior-mean effect sizes from PDR improved out-of-sample replication r² over the original GWAS summary statistics by 94% for asthma (r² = 0.66 vs. 0.34) and 70% for type 2 diabetes (r² = 0.46 vs. 0.27); a within-sample CAD improvement of roughly 300% was also reported, though external replication data were unavailable for CAD at the time.[1]

Relationship to Disease Subtyping

PDR was developed as a general method for improving prediction and dissecting pleiotropy rather than as a clinical disease-subtyping tool, but its components are directly analogous to the biologically interpretable clusters produced by mechanism-first disease-subtyping methods such as Udler et al.'s type 2 diabetes clustering — see Mechanism-First Polygenic Risk Score Clustering.

Citations

[1] Ballard, J. L., & O'Connor, L. J. (2022). Shared components of heritability across genetically correlated traits. American Journal of Human Genetics, 109(4), 989–1006. DOI: 10.1016/j.ajhg.2022.04.003. Source: ballard2022-pleiotropic-decomposition-regression.md. Supports: method description and all component/prediction results above. Location: Full text — Methods and Results.