Source-linked AI summary
A Multivariate Regression Approach to Association Analysis of Quantitative Trait Network
Seyoung Kim, Kyung-Ah Sohn, Eric P. Xing
TL;DR
Association mapping for many correlated phenotypes is difficult because conventional analyses treat traits separately or use two-stage procedures. The paper proposes GFlasso, a graph-guided multivariate regression method whose fused penalties share information across connected traits. The authors report improved detection of true causal SNPs when phenotype correlation is incorporated, while noting limitations in graph learning and marker dependence.
Problem
Association methods often analyze correlated phenotypes separately or use limited two-stage summaries instead of modeling their full dependency structure jointly.
Method
GFlasso uses a phenotype correlation graph with lasso and fusion penalties in one multivariate regression framework to identify SNPs affecting correlated trait subgroups.
Results
The authors report a significant advantage in detecting true causal SNPs when phenotype correlation is incorporated through the proposed methods.
Takeaways & Limitations
GFlasso can detect associations with phenotype-network subgraphs while retaining sparse estimates and borrowing information across correlated traits.
Takeaways & Limitations
The study uses a simple thresholded correlation graph and models dependencies among phenotypes but not among genetic markers.
Abstract
from arXiv · showhide
Many complex disease syndromes such as asthma consist of a large number of highly related, rather than independent, clinical phenotypes, raising a new technical challenge in identifying genetic variations associated simultaneously with correlated traits. In this study, we propose a new statistical framework called graph-guided fused lasso (GFlasso) to address this issue in a principled way. Our approach explicitly represents the dependency structure among the quantitative traits as a network, and leverages this trait network to encode structured regularizations in a multivariate regression model over the genotypes and traits, so that the genetic markers that jointly influence subgroups of highly correlated traits can be detected with high sensitivity and specificity. While most of the traditional methods examined each phenotype independently and combined the results afterwards, our approach analyzes all of the traits jointly in a single statistical method, and borrow information across correlated phenotypes to discover the genetic markers that perturbe a subset of correlated triats jointly rather than a single trait. Using simulated datasets based on the HapMap consortium data and an asthma dataset, we compare the performance of our method with the single-marker analysis, and other sparse regression methods such as the ridge regression and the lasso that do not use any structural information in the traits. Our results show that there is a significant advantage in detecting the true causal SNPs when we incorporate the correlation pattern in traits using our proposed methods.
1 Introduction
The paper addresses association mapping when many correlated phenotypes are measured together, a setting inadequately handled by independent or two-stage analyses. It introduces GFlasso, which uses the phenotype correlation graph and fused penalties to jointly identify markers affecting correlated trait subgroups.
- Motivation: High-throughput studies measure many correlated phenotypes, while much association methodology remains focused on univariate traits.This creates a need for methods that analyze genetic markers and multiple complex traits together.
- Motivation: Earlier multivariate approaches sought loci influencing phenotypes jointly but did not explicitly model complex interaction patterns among traits.
- Motivation: Network-based approaches can represent phenotype correlation, but prior methods often clustered traits and combined single-phenotype tests afterward.The paper identifies these steps as losing information from the full network and increasing multiple-testing demands.
- Proposed framework: GFlasso directly incorporates the quantitative-trait network into a single multivariate regression framework without additional phenotype clustering.It targets SNPs perturbing subsets of tightly correlated phenotypes rather than only individual traits.
- Proposed framework: The lasso provides sparsity by setting coefficients for irrelevant markers to zero, but its multiple-trait use amounts to separate regressions for each phenotype.Thus, the baseline lasso does not combine information across correlated traits.
- Proposed framework: GFlasso adds a fusion penalty that encourages shared predictors across connected traits, with graph-constrained and graph-weighted variants.The graph-constrained variant uses topology, whereas the graph-weighted variant adjusts constraint strength through edge weights.
2 Material and Methods
The paper formulates multiple-trait association mapping as multivariate regression and extends lasso with a phenotype correlation graph. GFlasso methods fuse coefficients for connected traits, using graph topology or correlation strength to identify SNPs affecting correlated trait subgroups.
- Regression setup: Genotypes are represented by an N × J matrix X, while Y contains K quantitative traits measured for the same individuals.Each genotype entry is coded 0, 1, or 2 according to minor-allele count.
- Regression setup: With many SNPs, ordinary regression can produce unstable, poorly interpretable estimates, motivating sparse methods such as ridge regression and lasso.The lasso controls sparsity through an L1 penalty governed by λ.
- Regression setup: Multiple-trait lasso fits K independent regressions and therefore does not combine information from correlated traits in its coefficient estimates.This baseline treats each trait independently despite shared marker effects or trait relatedness.
- Graph-guided fused lasso: GFlasso represents trait correlations as an edge-weighted graph learned from pairwise Pearson correlations above threshold ρ, then guides multivariate lasso estimation with structured penalties.The graph contains one node per trait, and each edge weight is the absolute correlation between connected traits.
- Graph-guided fused lasso: GcFlasso fuses coefficients for connected traits, while GwFlasso additionally weights fusion by correlation strength; negative correlations imply opposite coefficient directions.Fusion can propagate through densely connected subgraphs, producing block-structured coefficients for pleiotropic markers.
- Simulation design: The simulations use HapMap-based genotypes for 250 individuals, 10 phenotypes, and three correlated phenotype subnetworks of sizes 3, 3, and 4.Three SNPs affect the first subnetwork and four SNPs influence each of the other subnetworks.
3 Results
Simulation and asthma analyses compare single-marker, ridge, lasso, and graph-guided fused-lasso methods using association detection, prediction error, and estimated coefficient structure. Incorporating phenotype-network structure improves detection of causal SNPs and identifies SNP effects spanning correlated trait blocks.
- Simulation Study: The evaluation compares association methods using sensitivity/specificity and phenotype prediction error.Sensitivity and specificity assess detection of true association SNPs with few false positives; prediction error assesses phenotype prediction for new individuals.
- Simulation Study: The weighted fusion penalty significantly improves performance over the lasso in varying-sample-size simulations.These experiments use threshold ρ=0.3 and effect size 0.5 while varying sample size.
- Simulation Study: The weighted graph-guided fused-lasso method outperforms the other methods across all tested effect sizes.Figure 4 varies effect size from 0.3 to 1.0 with sample size 100 and graph threshold ρ=0.1.
- Simulation Study: Across graph thresholds, the weighted graph-guided method outperforms all other methods or performs at least as well as the lasso.At high threshold ρ=0.7, the graph contains nearly no edges, effectively removing the fusion penalty and bringing graph-guided performance close to the lasso.
- Simulation Study: Graph-guided methods remove most spurious SNP associations and produce coefficient blocks corresponding to subgroups of correlated phenotypes.Using edge weights avoids fusing effects across distinct true subgraphs connected only by weak correlations.
- Case Study Using Asthma Dataset: In the asthma dataset, graph-guided estimates align SNP effects with correlated clinical-trait blocks and suggest that rs3024660 and rs3024622 share a trait block with Q551R.The Q551R association concerns lung-physiology traits and had been previously associated with severe asthma.
4 Discussion
The discussion emphasizes jointly modeling correlated phenotypes with graph-guided fused lasso rather than relying on separate analyses or clustering. It also identifies graph-learning and marker-dependence assumptions as directions for extension.
- Method and motivation: GFlasso combines phenotype information in one framework and uses the full trait graph instead of clustering or a two-stage analysis.The approach targets SNPs perturbing subsets of tightly correlated phenotypes through graph-guided regularization.
- Method and motivation: The fusion penalty encourages coefficients for correlated phenotypes to align, with effects propagating across graph-connected subgraphs.The lasso component promotes sparsity, while fusion supports detection of phenotype submodules.
- Method and motivation: GwFlasso adds edge-weight information, while GcFlasso uses an unweighted graph to identify covariates jointly affecting correlated outputs.Both variants extend the graph-guided framework in different ways.
- Findings: Simulated and asthma datasets showed improved accuracy in detecting true associations when richer phenotype structure was incorporated.The reported improvement applies to the GwFlasso and GcFlasso methods.
- Limitations and extensions: The phenotype graph was learned using a simple thresholded-correlation scheme rather than more sophisticated network-learning methods.The authors suggest evaluating graphical Gaussian models and sparse variants as alternatives.
- Limitations and extensions: The framework assumes the graph is available before regression and models dependencies among phenotypes, but not among genetic markers.Jointly learning graph structure and regression coefficients, and incorporating marker dependencies, are proposed extensions.
Appendix: Parameter Estimation
The appendix describes estimating GwFlasso coefficients through a convex quadratic program for fixed tuning parameters, followed by cross-validation and gradient-based tuning updates.
- Quadratic-program estimation: GwFlasso coefficient estimates are obtained by solving a quadratic program because the objective and constraints are convex.The coefficients are concatenated into a single vector for the QP formulation.
- Quadratic-program estimation: For each graph edge, the QP encodes signed differences between phenotype-specific regression coefficients using a block matrix construction.Identity, negative-identity, and zero blocks represent the edge relationships.
- Parameter selection: The QP finds optimal coefficients for fixed s1 and s2, which correspond to λ and γ in Equation (5).These parameters govern the regularization terms in the optimization.
- Parameter selection: Cross-validation selects s1 and s2 by evaluating validation error over a parameter grid and choosing the pair with the lowest C(s1, s2).This grid search is followed by iterative updates using a finite-difference approximation to the validation-error gradient.
- Parameter selection: The initialization procedure first searches for s1 with s2 = ∞, then searches along s2 after fixing the selected s1.The reported procedure produced starting values close to the global optimum and converged within relatively few iterations.
Web Resources
The appendix lists the HapMap project as a web resource.
- Web resource: The HapMap project is provided as a web resource at hapmap.org.The resource is listed without additional description.