Information Enhanced Model Selection for Gaussian Graphical Model with Application to Metabolomic Data

Jie Zhou, Anne G. Hoen, Susan McRitchie, Wimal Pathmasiri, Weston D. Viles, Quang P. Nguyen, Juliette C. Madan, Erika Dade, Margaret R. Karagas, Jiang GuibioRxiv 2019 · DOI

TL;DR

This paper introduces a new method, Structural Bayesian Information Criterion (SBIC), to improve the selection of association networks in complex biological data, like metabolomics. By incorporating prior biological knowledge (e.g., metabolic pathways) into a model selection framework, SBIC and a novel two-step algorithm can uncover more accurate and biologically relevant relationships than existing methods.

What Problem Does This Paper Address?

Modern biological datasets, especially from "omics" technologies, are often characterized by a high signal-to-noise ratio, making it challenging to identify true associations between variables. This is particularly true in metabolomics, where the sheer number of metabolites and their complex interactions can obscure meaningful patterns. Traditional methods for inferring these association networks often struggle with this noise and may not effectively leverage existing biological knowledge, such as known metabolic pathways. The authors aim to bridge this gap by developing a method that explicitly integrates prior biological information into the network structure learning process, leading to more robust and interpretable results.

Methods

The core of the paper lies in two main contributions:

  1. Structural Bayesian Information Criterion (SBIC): The authors propose SBIC as a novel model selection criterion. It's an extension of the widely used Bayesian Information Criterion (BIC) and the Extended BIC (EBIC). SBIC explicitly incorporates prior structural information about the network into the BIC framework. This is achieved by defining a "difference graph" that quantifies the deviation of a candidate model from the prior structure. The paper shows that EBIC is a special case of SBIC, and theoretically demonstrates that SBIC is a consistent model selection criterion for high-dimensional Gaussian graphical models under certain conditions.
  2. Two-Step Algorithm for Candidate Model Pool Construction: To make the model selection feasible in high-dimensional settings, the authors propose a data-driven, two-step algorithm to build a "candidate model pool." This algorithm intelligently incorporates the prior structure:
  3. Step 1 (Edge Enrichment): This step aims to add edges that might have been missed by the prior structure but are supported by the data.
  4. Step 2 (Edge Pruning): This step removes potentially spurious edges that might have been included in the prior structure but are not supported by the data. This approach is implemented using existing tools like graphical lasso or glmnet, making it computationally practical.

Datasets

The paper utilizes the following dataset:

  • New Hampshire Birth Cohort Study (NHBCS) metabolomic data: This dataset consists of relative concentration data of metabolites from infant feces. The prior network structure for these metabolites was constructed using pathway information from the Kyoto Encyclopedia of Genes and Genomes (KEGG) database.

Key Insights and Ideas

  • Prior Knowledge is Crucial for Noisy Data: The paper strongly advocates for the integration of prior biological knowledge, particularly in noisy biological datasets. It argues that this prior information can significantly enhance the accuracy and interpretability of network inference.
  • SBIC as a Flexible Generalization: SBIC offers a more flexible approach to model selection than EBIC by allowing explicit incorporation of prior structure. This flexibility is key to leveraging domain-specific knowledge.
  • Two-Step Algorithm for Practicality: The proposed two-step algorithm is a practical solution for constructing a relevant set of candidate models in high-dimensional settings, avoiding exhaustive searches.
  • Metabolic Pathways as Significant Predictors: The application to metabolomic data demonstrates that metabolic pathways are indeed statistically significant factors influencing the conditional dependence between metabolites.
  • Discovering Novel Relationships: The method successfully identifies new relationships between metabolites that are not apparent through conventional pathway analysis alone, some of which are later validated in biological literature.

Key Results

  • Theoretical Consistency: The paper provides theoretical guarantees that SBIC is a consistent model selection criterion for high-dimensional Gaussian graphical models under mild conditions.
  • Superior Performance in Simulations: Simulation studies show that the proposed SBIC with the two-step algorithm outperforms existing methods in terms of true positive rate (TPR) and false positive rate (FPR), especially when the association strength is low.
  • Robustness to Misspecification: The method demonstrates robustness to the misspecification of the temperature parameter (which controls the influence of prior information), suggesting practical utility even when prior knowledge is not perfectly accurate.
  • Biological Significance in Metabolomics: The application to NHBCS data revealed that metabolic pathway involvement is a significant factor in metabolite associations. The method identified novel metabolite-metabolite relationships, some of which have been recognized in biological literature, highlighting its ability to uncover biologically meaningful connections.

Limitations and Future Work

  • Decomposability Assumption (for theoretical proofs): While the algorithm itself does not require decomposability, the theoretical proofs for consistency (Theorems 1 and 2) rely on strong assumptions, including the decomposability of the underlying graph. The authors acknowledge that results might still hold if these assumptions are relaxed.
  • Prior Structure Quality: The performance of SBIC is dependent on the quality and accuracy of the prior structure. While the method shows robustness to some misspecification, a significantly flawed prior could still negatively impact results.
  • Statistical Inference: The paper focuses on model selection and does not delve into the statistical inference of the network parameters after selection. The authors note that while inference is implicitly assumed to be reasonable after model selection, this is not always guaranteed in practice.
  • Generalization to Other Data Types: The current work focuses on Gaussian graphical models. The authors suggest that the algorithm can be adapted for other types of undirected graphical models, such as Ising models, which would be a valuable area for future research.
  • Exploring Other Prior Network Types: The paper mentions that other types of prior networks, such as those based on structural similarity, could be used, opening avenues for further investigation.

Summary updated

Abstract

Abstract In light of the low signal-to-noise nature of many large biological data sets, we propose a novel method to learn the structure of association networks using Gaussian graphical models combined with prior knowledge. Our strategy includes two parts. In the first part, we propose a model selection criterion called structural Bayesian information criterion (SBIC), in which the prior structure is modeled and incorporated into Bayesian information criterion (BIC). It is shown that the popular extended BIC (EBIC) is a special case of SBIC. In the second part, we propose a two-step algorithm to construct the candidate model pool. The algorithm is data-driven and the prior structure is embedded into the candidate model automatically. Theoretical investigation shows that under some mild conditions SBIC is a consistent model selection criterion for high-dimensional Gaussian graphical model. Simulation studies validate the superiority of the proposed algorithm over the existing ones and show the robustness to the model misspecification. Application to relative concentration data from infant feces collected from subjects enrolled in a large molecular epidemiological cohort study validates that metabolic pathway involvement is a statistically significant factor for the conditional dependence between metabolites. Furthermore, new relationships among metabolites are discovered which can not be identified by the conventional methods of pathway analysis. Some of them have been widely recognized in biological literature.
BibTeX
@misc{zhou2019information,
  author = {Jie Zhou and Anne G. Hoen and Susan McRitchie and Wimal Pathmasiri and Weston D. Viles and Quang P. Nguyen and Juliette C. Madan and Erika Dade and Margaret R. Karagas and Jiang Gui},
  title = {{Information Enhanced Model Selection for Gaussian Graphical Model with Application to Metabolomic Data}},
  year = {2019},
  doi = {10.1101/815423},
  howpublished = {bioRxiv},
}
Figure