Supervised Bayesian joint graphical model for simultaneous network estimation and subgroup identification

1. Problem Statement

Cancer is characterized by significant heterogeneity, necessitating the identification of distinct patient subgroups to understand biological mechanisms and improve clinical outcomes. Existing methods for subgroup identification fall into two categories:

• Unsupervised: Relies solely on molecular data distributions (means, variances, or network structures) without considering clinical outcomes. These often yield subgroups lacking clinical relevance.
• Supervised: Leverages associations between molecular predictors and clinical outcomes (e.g., survival). However, current supervised methods typically focus on heterogeneous regression coefficients while neglecting the interconnections (network structures) among the molecular predictors themselves.

The Gap: There is a lack of methods that can simultaneously identify patient subgroups and estimate heterogeneous biological networks in a supervised manner, where the network structures are informed by clinical outcomes.

2. Methodology: Supervised Bayesian Joint Graphical Model (SBJGM)

The authors propose a novel Bayesian framework that integrates a Finite Mixture of Regression Models with Gaussian Graphical Models (GGMs).

A. Statistical Model

• Data Structure: The model handles high-dimensional molecular predictors ($X$) and a survival outcome ($T, \Delta$).
• Subgroup Membership: Subjects belong to $K$ latent subgroups ($g_i \in \{1, \dots, K\}$).
• Joint Distribution: For each subgroup $k$, the model assumes:
  1. Predictor Distribution: $X | g_i=k \sim N(\mu_k, \Omega_k^{-1})$. Here, $\Omega_k$ is the precision matrix representing the network structure (conditional dependencies) specific to subgroup $k$.
  2. Outcome Distribution: An Accelerated Failure Time (AFT) model is used: $Z_i = \beta_{0k} + \beta_k^T X_i + \epsilon_i$, where $Z$ is the log-survival time. This captures the heterogeneous association between predictors and survival across subgroups.
• Reparameterization: To ensure scale invariance under the mixture model, regression coefficients are reparameterized based on the noise level ($\tau_k$).

B. Prior Specifications

The Bayesian framework employs specific priors to handle sparsity and similarity:

1. Spike-and-Slab Laplace Prior: Applied to the off-diagonal elements of the precision matrices ($\omega_{k,jl}$) to induce sparsity in the network edges. This allows for rigorous uncertainty quantification via posterior inclusion probabilities (PIPs).
2. Laplace Priors: Applied to regression coefficients ($\beta_{k,j}$) and mean vectors ($\mu_{k,j}$) to handle high-dimensional variable selection.
3. Novel Similarity Prior: A key innovation is a latent vector $\theta_{jl}$ shared across subgroups for the edge weights. This prior encourages sign consistency and similarity in the magnitude of connection effects across different subgroups. It is designed to borrow information effectively without forcing identical network structures, addressing the biological reality that subtypes often share core pathways while differing in specific regulatory mechanisms.

C. Inference Algorithm

• Estimation: Due to the complexity of the posterior, the authors develop a three-step strategy based on an Expectation-Maximization (EM) algorithm to compute the Maximum a Posteriori (MAP) estimates.
  • Step 1: EM algorithm to obtain MAP estimates of parameters ($\beta, \mu, \Omega, \tau$).
  • Step 2: Calculation of Marginal MAP Plug-in Posterior Inclusion Probabilities (PIPs) for edge selection.
  • Step 3: Thresholding PIPs (e.g., > 0.5) to determine the final network sparsity patterns.
• Optimization: Coordinate descent and ADMM algorithms are used within the M-step to ensure the positive definiteness of the precision matrices.

3. Key Contributions

1. First Supervised Joint Estimation: This is among the first methods to jointly estimate multiple heterogeneous networks and identify subgroups in a supervised setting, explicitly linking network structures to clinical outcomes.
2. Novel Similarity Prior: Unlike previous group-penalty methods that only enforce shared sparsity patterns, the proposed similarity prior captures the similarity in the values and signs of connection effects, facilitating more biologically plausible information borrowing.
3. Theoretical Guarantees: The authors rigorously establish the consistency of the estimators. They prove that the estimation error converges at a minimax optimal rate and that the method achieves variable selection consistency (correctly identifying non-zero edges and coefficients) even in high-dimensional settings with censored outcomes.
4. Computational Efficiency: By utilizing a MAP-based EM approach rather than traditional MCMC, the method achieves high computational efficiency while maintaining theoretical validity through PIPs.

4. Results

A. Simulation Studies

• Setup: Simulations were conducted with varying numbers of subgroups ($K=2,3$), sample sizes (balanced and imbalanced), and network structures (Power-law, Erdös-Rényi, Nearest-neighbor).
• Performance Metrics: Clustering Error (CE), Precision Matrix Squared Error (PME), True Positive Rate (TPR), and False Positive Rate (FPR).
• Findings:
  • SBJGM consistently outperformed unsupervised methods (SCAN, BCGS) and two-stage supervised methods (FMGemBag, FMJGL) in both subgroup identification and network recovery.
  • Even when compared to "ideal" methods that knew the true subgroup memberships, SBJGM performed competitively, demonstrating the effectiveness of the similarity prior.
  • The method was robust to imbalanced sample sizes and varying degrees of similarity between subgroups.

B. Real Data Application (TCGA SKCM)

• Dataset: Cutaneous melanoma (SKCM) data from The Cancer Genome Atlas (360 patients, 91 filtered genes).
• Findings:
  • The model identified two distinct subgroups with significantly different survival outcomes (Log-rank $P = 0.005$).
  • Network Analysis: The two subgroups shared 52 edges with consistent signs but also possessed unique edges.
  • Biological Validation: Gene Ontology (GO) enrichment analysis revealed that the unique edges in each subgroup were linked to distinct biological pathways, confirming the clinical relevance of the identified subgroups.
  • Prediction Accuracy: In resampling tests, SBJGM achieved the lowest negative log-likelihood (NLS), indicating superior predictive performance compared to alternative methods.

5. Significance

This paper addresses a critical gap in precision oncology by providing a statistical tool that bridges molecular network biology and clinical outcomes.

• Clinical Impact: By identifying subgroups based on how their specific gene networks influence survival, the method offers more actionable insights for personalized treatment than methods relying solely on gene expression means.
• Methodological Advancement: It advances the field of high-dimensional statistics by successfully integrating supervised learning, censored data handling, and complex network structure estimation within a unified Bayesian framework.
• Future Directions: The authors suggest extending the framework to other survival models (e.g., Weibull AFT) and incorporating multi-omics data (mutations, methylation) for a more comprehensive view of cancer heterogeneity.