Estimating Gaussian graphical models of multi-study data with Multi-Study Factor Analysis

Technical Summary: Multi-study Factor-based Gaussian Graphical Models with Applications to the Metabolomics of Gestational Diabetes Mellitus

Problem Statement
Gestational diabetes mellitus (GDM) is a metabolic disorder associated with adverse maternal and neonatal outcomes. While metabolomic profiling has successfully linked specific metabolites to GDM status, emerging evidence suggests that GDM is not characterized by isolated metabolite dysfunction but by coordinated changes in groups of metabolites forming distinct patterns or signatures. Existing statistical approaches face a limitation: standard factor analysis (FA) identifies latent variables (factors) representing coordinated features but does not model the conditional dependencies (network structures) between them. Conversely, network models, such as Gaussian Graphical Models (GGMs), effectively represent conditional dependencies but typically do not deconvolute shared signals from condition-specific signals across multiple studies or groups. Current differential network analysis methods often estimate networks for each condition separately or directly estimate differences, failing to explicitly separate shared and study-specific contributions to these dependencies. This gap necessitates an integrated approach capable of identifying shared and divergent network patterns based on latent variable representations across multiple conditions.

Methodology: MSFA-X
The authors introduce MSFA-X (Extended Multi-Study Factor Analysis), a novel framework that combines Multi-Study Factor Analysis (MSFA) with network modeling to construct latent variable-based GGMs.

• Model Formulation: MSFA-X extends standard MSFA by modeling observed data $x_{is}$ (for subject $i$ in study $s$) as a sum of shared factors ($f_{is}$), study-specific factors ($l_{is}$), and decomposed noise components. Unlike standard MSFA where the residual covariance $\Psi_s$ is a single diagonal matrix, MSFA-X partitions the residual noise into a shared component ($\Gamma$) and a study-specific component ($H_s$). The covariance matrix for study $s$ is decomposed as:
  $$\Sigma_s = \Phi\Phi^\top + \Lambda_s\Lambda_s^\top + \Gamma + H_s$$
  where $\Phi$ represents shared loadings, $\Lambda_s$ represents study-specific loadings, and $\Gamma$ and $H_s$ are full-rank diagonal matrices.

• Precision Matrix Estimation: This decomposition ensures that the shared covariance component ($\Phi\Phi^\top + \Gamma$) and the study-specific covariance components ($\Lambda_s\Lambda_s^\top + H_s$) are invertible. This allows for the direct estimation of shared and study-specific precision matrices ($\Theta = \Sigma^{-1}$), which define the edges of the GGMs based on partial correlations. This approach avoids the need for tuning parameters typically required by penalized likelihood methods like the graphical lasso (glasso) when applied to singular matrices.

• Estimation and Identifiability: The model parameters are estimated using an Expectation-Conditional Maximization (ECM) algorithm. The authors address several identifiability issues:

  1. Rotation Indeterminacy: Resolved via block lower-triangular constraints on loading matrices.
  2. Label Switching: Addressed by the same constraints.
  3. Separation of Components: Ensured by rank constraints on the combined loading matrix.
  4. Noise Decomposition: The separation of $\Gamma$ and $H_s$ is not statistically identifiable (only their sum is). The authors propose a practical strategy of estimating $\Gamma$ at the center of the non-identifiability region ($\hat{\gamma}_p = 0.5 \times \min_s \hat{\psi}_{ps}$), arguing that while the individual noise estimates are not consistent, the resulting precision matrices remain practically meaningful.

• Significance Testing: Edge significance is assessed using an empirical null approach. Null data is generated under the assumption of no conditional associations, and empirical p-values are calculated for observed edge weights based on the distribution of null edge weights.

Key Results

• Simulation Studies: Extensive simulations across ten settings (varying sample sizes, number of studies, factors, and noise levels) demonstrated that MSFA-X significantly outperforms standard graphical lasso benchmarks in recovering both shared and study-specific network structures. The method achieved high modified-RV coefficients (often >0.99) between estimated and true adjacency matrices, whereas glasso-based approaches yielded substantially lower coefficients, particularly for study-specific networks.
• HAPO Study Application: Applied to metabolomic data from the Hyperglycemia and Adverse Pregnancy Outcomes (HAPO) study (3,463 participants: 576 with GDM, 2,887 without), MSFA-X identified:
  • Shared Networks: Two distinct subnetworks: one linking amino acids (AAs) with long-chain acylcarnitines, and another linking short- and medium-chain acylcarnitines.
  • Condition-Specific Differences: In women without GDM, glutamine/glutamate acted as a central hub connecting AAs to short- and medium-chain acylcarnitines. In women with GDM, this hub function was lost; glutamine/glutamate became unconnected to the rest of the network. Conversely, aromatic amino acids (histidine, phenylalanine, tyrosine) became strong hubs specifically in the GDM network.
  • Biological Validation: The latent factors derived from MSFA-X were associated with newborn adiposity (sum of skinfolds). Specifically, the GDM-specific factor (loading on aromatic AAs) and shared factors were significantly associated with newborn adiposity in women with GDM, even after adjusting for covariates. This association was not observed in women without GDM, suggesting that the metabolic dysregulation captured by these latent factors confers specific risk to offspring of mothers with GDM.

Significance and Claims
The paper claims that MSFA-X provides a distinctive and interpretable perspective for analyzing multi-condition data by explicitly deconvoluting shared and study-specific signals, a capability lacking in existing differential network analysis methods. The authors assert that the method successfully recovers graphical structures induced by latent variables and outperforms naive benchmarks.

The application to GDM reveals that the metabolic milieu of GDM involves a systemic disruption where key network nodes (like glutamine/glutamate) decouple, and new hubs (aromatic AAs) emerge. The authors conclude that these latent factors capture biologically meaningful variation in metabolomic profiles that relates to newborn adiposity, offering a potential mechanism for understanding how maternal GDM influences neonatal outcomes. The method is presented as a general tool applicable to any multivariate normal data exhibiting latent variable structures in multi-condition settings, not limited to metabolomics. The R package implementing MSFA-X is made available to the research community.