A Multivariate Regression Approach to Association Analysis of Quantitative Trait Network

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.

💡 Research Summary

The paper addresses a fundamental challenge in genetic association studies of complex diseases: many clinically relevant phenotypes are highly correlated, yet most existing methods treat each phenotype independently. To exploit the shared genetic architecture among correlated quantitative traits, the authors introduce a novel statistical framework called Graph‑guided Fused Lasso (GFlasso). The method first encodes the dependency structure among traits as a network (nodes represent traits, edges reflect pairwise correlations or biologically motivated connections). This network is then incorporated into a multivariate linear regression model that relates genotype data (single‑nucleotide polymorphisms, SNPs) to the set of traits.

Mathematically, the objective function consists of three components: (1) a least‑squares loss measuring the fit of the regression coefficients β to the observed trait values, (2) an L1‑lasso penalty (λ1‖β‖1) that enforces sparsity in the SNP selection, and (3) a graph‑guided fused‑lasso penalty (λ2∑(i,j)∈E wij‖βi−βj‖1) that penalizes differences between the coefficient vectors of traits that are linked in the network. The edge weights wij are proportional to the strength of the trait‑trait correlation, so that strongly connected traits are forced to share similar genetic effects. By jointly optimizing these terms, GFlasso can identify SNPs that simultaneously influence a subgroup of correlated traits, while still maintaining overall sparsity.

Optimization is performed using an efficient alternating‑direction method of multipliers (ADMM) or a coordinate‑descent scheme, allowing the method to scale to thousands of SNPs and dozens of traits. Hyper‑parameters λ1 and λ2 are tuned via cross‑validation or information‑theoretic criteria (e.g., BIC).

The authors evaluate GFlasso on two fronts. First, they generate realistic simulated data using HapMap genotypes, constructing trait networks with distinct structures (clustered, random, and noisy). They embed a set of causal SNPs that affect specific trait clusters. Compared with traditional single‑marker analysis, standard lasso, and ridge regression (which ignore trait structure), GFlasso achieves markedly higher true‑positive rates for causal SNPs, especially when the underlying trait network is well‑specified. False‑positive rates remain comparable, indicating that the added fused‑lasso penalty does not inflate spurious discoveries.

Second, they apply the method to an asthma cohort comprising several thousand individuals, genome‑wide SNP data, and about twenty quantitative clinical phenotypes (e.g., lung‑function measures, inflammatory markers). GFlasso reproduces known asthma‑associated loci such as IL33 and ORMDL3, and additionally uncovers novel SNP‑trait associations that are biologically plausible (e.g., SNPs linked to specific spirometry indices). Importantly, many of these novel signals are detected only when the trait network is leveraged, underscoring the power of borrowing information across correlated phenotypes.

The discussion highlights several strengths of GFlasso: (i) explicit modeling of trait inter‑relationships enables information sharing and improves statistical power; (ii) simultaneous analysis of all traits reduces the multiple‑testing burden inherent in separate univariate scans; (iii) the fused‑lasso penalty yields interpretable groups of traits driven by common genetic factors. Limitations include the need for a pre‑specified trait network (which may be noisy or incomplete) and potential over‑shrinkage when the network is overly dense. The authors suggest future extensions such as data‑driven network learning, incorporation of non‑linear effects via kernel methods, or integration with graph neural networks.

In conclusion, GFlasso provides a principled, scalable, and effective approach for multivariate GWAS when phenotypes form a correlated network. By marrying sparse regression with graph‑guided regularization, it uncovers genetic variants that would be missed by conventional univariate or structure‑agnostic multivariate methods, thereby advancing the analytical toolkit for precision medicine and complex‑trait genetics.