Source-linked AI summary
A sparse conditional Gaussian graphical model for analysis of genetical genomics data
Jianxin Yin, Hongzhe Li
TL;DR
The paper addresses gene-network inference when genetic variants affect gene-expression means and may confound dependencies among genes. It introduces a sparse conditional Gaussian graphical model with penalized estimation, then supports it through simulations, theory, and yeast eQTL analysis. The resulting networks are more accurate and less dense than standard expression-only Gaussian graphical models in the reported settings.
Problem
Genetic effects and complex genetic architecture can compromise gene-dependency inference when standard Gaussian graphical models analyze gene-expression data without adjustment.
Method
A sparse conditional Gaussian graphical model jointly estimates marker effects in gene-expression means and a sparse concentration matrix using seemingly unrelated regressions and penalized coordinate descent.
Results
Simulations found cGGM estimates and graph-structure recovery better than standard GGM procedures, while yeast analysis produced 94 cGGM links versus 341 glasso edges among 54 MAPK genes.
Takeaways & Limitations
Adjusting for shared genetic effects yields a more interpretable gene-expression network than using gene-expression data alone.
Takeaways & Limitations
Gene-expression data cannot be expected to recover complete signaling networks because many dependencies occur at protein or metabolite levels.
Abstract
from arXiv · showhide
Genetical genomics experiments have now been routinely conducted to measure both the genetic markers and gene expression data on the same subjects. The gene expression levels are often treated as quantitative traits and are subject to standard genetic analysis in order to identify the gene expression quantitative loci (eQTL). However, the genetic architecture for many gene expressions may be complex, and poorly estimated genetic architecture may compromise the inferences of the dependency structures of the genes at the transcriptional level. In this paper we introduce a sparse conditional Gaussian graphical model for studying the conditional independent relationships among a set of gene expressions adjusting for possible genetic effects where the gene expressions are modeled with seemingly unrelated regressions. We present an efficient coordinate descent algorithm to obtain the penalized estimation of both the regression coefficients and the sparse concentration matrix. The corresponding graph can be used to determine the conditional independence among a group of genes while adjusting for shared genetic effects. Simulation experiments and asymptotic convergence rates and sparsistency are used to justify our proposed methods. By sparsistency, we mean the property that all parameters that are zero are actually estimated as zero with probability tending to one. We apply our methods to the analysis of a yeast eQTL data set and demonstrate that the conditional Gaussian graphical model leads to a more interpretable gene network than a standard Gaussian graphical model based on gene expression data alone.
1. Introduction.
The paper addresses confounding in gene-network inference from eQTL data by introducing a sparse conditional Gaussian graphical model that adjusts for genetic effects while estimating gene dependencies.
- Motivation: Complex genetic architecture can compromise transcriptional dependency inference when gene-expression traits are analyzed without adequately modeling genetic effects.eQTL analyses identify loci associated with expression, but multiple genetic effects and gene–gene interactions can complicate the architecture.
- Motivation: Standard Gaussian graphical models use gene-expression data to infer regulatory networks but ignore genetic-variant effects on expression means.The paper frames graphical models as representations of dependencies among stochastic variables.
- Motivation: Shared genetic regulators can create spurious gene–gene links, producing false-positive and false-negative edge identifications.The proposed analysis targets conditional dependencies after shared genetic effects are removed.
- Contribution: The sparse cGGM simultaneously identifies genetic variants associated with expressions and constructs a sparse Gaussian graph from eQTL data.It models gene-expression means through covariates or markers and estimates sparse regression coefficients together with a sparse concentration matrix.
2. The sparse cGGM and penalized likelihood estimation.
The sparse cGGM models gene-expression means through genetic markers and estimates conditional gene dependencies through a sparse concentration matrix, using penalized likelihood and coordinate descent.
- Model formulation: Conditioning on SNP genotypes, the model represents gene expressions with a multivariate normal distribution whose mean depends on genetic markers.The coefficient matrix Γ models marker effects on means, while the covariance matrix Σ is assumed not to depend on the markers.
- Model formulation: The concentration matrix Θ=Σ^-1 encodes conditional gene dependence after shared genetic regulators are adjusted.Genes are graph nodes, and an edge exists exactly when the corresponding off-diagonal element of Θ is nonzero.
- Model formulation: The method jointly estimates sparse regression coefficients Γ and a sparse concentration matrix Θ because genetic effects and gene networks are both expected to be sparse.The approach is related to seemingly unrelated regressions because residual correlations across genes improve estimation of genetic effects.
- Optimization: For fixed Θ or Γ, the negative log-likelihood is convex in the other parameter block, enabling alternating coordinate-descent updates.The algorithm iteratively updates the concentration-matrix and regression-coefficient estimates until convergence.
- Penalized estimation: Penalized likelihood uses Lasso or adaptive Lasso penalties, with tuning parameters ρ and λ controlling sparsity in the cGGM.The penalties apply to regression and concentration parameters, respectively, and BIC selects the tuning parameters.
- Optimization: The iterative algorithm converges to a stationary point, decreases the penalized negative log-likelihood monotonically, and has complexity O(pq) plus the glasso complexity.Because the joint objective is not always convex, convergence to the global minimum is not guaranteed.
- Theory: Theoretical results cover asymptotic distributions, oracle properties, convergence rates, and sparsistency as dimensions may diverge with sample size.Sparsistency means zero parameters are estimated as zero with probability tending to one.
3. Theoretical properties.
The paper establishes theoretical error rates and evaluates the proposed estimates through Monte Carlo comparisons with glasso and mLasso.
- Theoretical analysis establishes convergence rates and sparsistency for the penalized sparse cGGM estimators.Sparsistency means zero parameters are estimated as zero with probability tending to one.
- The estimated concentration matrix has asymptotic bias of order (p_n + s_n)log p_n/n, which vanishes under a stated logarithmic sparsity condition.
- Regression-coefficient estimation has total squared error at least k_n/n and incurs a logarithmic high-dimensionality factor log(p_nq_n).
- Monte Carlo simulations compare the proposed estimates with glasso and a modified neighborhood-selection procedure called mLasso.The simulations select tuning parameters ρ and λ using BIC.
4. Monte Carlo simulations.
The simulations compare cGGM with glasso, adaptive variants, and modified neighborhood selection across settings where p<n and p>n. cGGM generally improves concentration-matrix estimation and graph-structure recovery, while high-dimensional sensitivity remains difficult.
- Models with p<n: For p<n, models 1–3 use n=250 with gene and marker dimensions ranging from (p,q)=(100,100) to (25,10).The compared settings assign sparse nonzero probabilities to concentration-matrix and regression coefficients.
- Results with p<n: For p<n, cGGM estimates the concentration matrix better than glasso across all three models and all measurements.Adaptive cGGM and adaptive glasso also improve estimates, but their gains are minimal.
- Results with p<n: For p<n, cGGM and adaptive cGGM achieve higher sensitivity, specificity, and MCC than glasso and adaptive glasso.Glasso often produces denser graphs and a larger Hamming distance because shared genetic variants can explain some apparent links.
- Results with p<n: Modified neighborhood selection performs better than glasso or adaptive glasso but worse than cGGM or adaptive cGGM in graph-structure selection.Unlike the graphical-model estimators, this procedure does not provide a concentration-matrix estimate.
- Results with p>n: For p>n, glasso has lower sensitivity and MCC than cGGM, whereas mLasso can have higher sensitivity but lower specificity in models 5 and 6.The mLasso procedure does not estimate the concentration matrix.
5. Analysis of yeast eQTL data.
The yeast eQTL analysis applied cGGM to MAPK-pathway and PPI-network genes while adjusting expression dependencies for genetic-marker effects. Compared with glasso and mLasso, cGGM produced sparser networks and recovered biologically interpretable MAPK links, although PPI recovery was limited.
- Data and analysis scope: Small sample size and limited biological perturbation restricted analysis to 54 MAPK-pathway genes and 1,207 PPI-network genes, rather than all 6,216 yeast genes.The observational data were used to construct conditional-independence networks, not directed graphs.
- Data and analysis scope: Among 585 markers tested against 54 genes, 839 significant associations were observed, including 702 associations involving 188 markers linked to at least two genes.Many genes were associated with multiple markers, indicating shared genetic regulation that could create false expression-level links if ignored.
- cGGM results: Using BIC-selected λ = 0.28 and ρ = 0.54, cGGM identified 94 links among 54 MAPK genes and nonzero genetic effects for 677 regression coefficients.The estimated concentration matrix contained 188 nonzero elements, and 43 genes were linked in the displayed graph.
- cGGM results: The cGGM graph recovered biologically related MAPK connections, including Fus3 with Dig1, Ste12 and Fus1, and Slt2/Mpk1 with Swi4 and Rlm1.Most links in the upper pheromone-response portion of the pathway were recovered, along with links associated with cell-wall repair and hypotonic-shock response.
- Method comparisons: Glasso selected 341 MAPK edges versus 94 for cGGM, while mLasso selected 45; 36 mLasso links overlapped cGGM and all 45 overlapped glasso.The authors attribute glasso-only edges partly to shared genetic variants and its constant-mean assumption.
- Method comparisons: For the 1,207-gene PPI analysis, cGGM selected 12,036 of 727,821 possible links, while glasso produced a much denser graph and cGGM achieved MCC 0.041 versus 0.030 for glasso.Low true-positive rates were attributed partly to PPI links not being reflected at the gene-expression level and to marginal-versus-conditional interaction differences.
6. Conclusions and discussion.
The proposed cGGM is useful for learning gene-network projections from eQTL data, but gene-expression data cannot fully recover signaling networks. The paper also identifies causal interpretation and genetic-variant identification as important boundaries for future work.
- Limitations: Gene-expression data cannot completely recover true signaling networks because some dependencies appear only at protein or metabolite levels.The resulting gene network should therefore be interpreted as a projection of broader signaling or protein-interaction networks.
- Conclusions: The authors nonetheless report that cGGM-derived yeast eQTL networks can explain the data and provide biological insights into gene interactions.
- Conclusions: The paper focuses on estimating sparse conditional graphical structure while adjusting for genetic effects on gene expressions.
- Future work: Explicitly modeling covariance among gene expressions may improve identification of associated genetic variants, but this possibility remains a theoretical question for future study.The authors connect this prospect to the original motivation for seemingly unrelated regression models.
- Interpretation: The undirected graphs characterize conditional independence rather than causal dependency.Directed causal gene networks require approaches such as structural equation modeling or genetic algorithms.