Probabilistic Graphical Modeling under Heterogeneity

Liying Chen, Satwik Acharyya, Chunyu Luo, Yang Ni, Veerabhadran BaladandayuthapanibioRxiv 2023 · DOI

TL;DR

This paper introduces GraphR, a novel Bayesian framework for building biological networks that accounts for sample-specific heterogeneity. By incorporating intrinsic factors like cancer subtypes or spatial locations, GraphR provides more accurate and robust network inference, uncovering biological insights missed by traditional methods that assume homogeneity.

What Problem Does This Paper Address?

Biological systems, especially in areas like cancer research, are inherently complex and heterogeneous. Traditional network inference methods, often based on probabilistic graphical models (PGMs), typically assume that all samples are drawn from the same underlying distribution (homogeneity). This assumption breaks down when dealing with data exhibiting significant variations between samples, such as different tumor subtypes, stages of development, or spatial locations within a tumor. Ignoring this heterogeneity can lead to misleading conclusions, a phenomenon illustrated by Simpson's paradox, where trends observed in aggregated data disappear or reverse when data is disaggregated. The authors aim to address this limitation by developing a method that can explicitly model and leverage sample-specific heterogeneity to build more accurate biological networks.

Methods

The core of the proposed method is GraphR (Graphical Regression), a flexible Bayesian approach. GraphR extends Gaussian Graphical Models (GGMs) by making the precision matrix (which encodes conditional independence relationships) a function of intrinsic factors (covariates) that represent sample heterogeneity.

Key aspects of the GraphR methodology include:

  • Regression-based formulation: The partial correlations between nodes are modeled as linear functions of the intrinsic factors. This allows for direct incorporation of heterogeneity at various scales.
  • Sparse network estimation: To handle the high dimensionality of biological data, GraphR employs spike-and-slab priors, a Bayesian technique for variable selection, to encourage sparsity in the network structure.
  • Computational efficiency: Instead of computationally intensive MCMC methods, GraphR utilizes variational Bayes (VB) algorithms with a mean-field assumption. This significantly speeds up the inference process, making it scalable for large datasets.
  • Uncertainty quantification: The Bayesian framework allows for the estimation of posterior probabilities, providing uncertainty measures for identified edges and the influence of heterogeneity factors.

GraphR can handle various types of intrinsic factors, including binary (e.g., disease vs. control), categorical (e.g., cancer subtypes), and continuous (e.g., dedifferentiation levels).

Datasets

The authors evaluated GraphR using:

  • Realistic simulation settings: Multiple scenarios were designed to test graph structure recovery under different types and combinations of intrinsic factors (discrete, continuous, homogeneous).
  • Four diverse multi-omics and spatial transcriptomics datasets from breast cancer (BRCA):
    • Proteogenomics data for intrinsic subtypes of BRCA.
  • Proteomics data to study stemness-induced heterogeneity.
  • Pan-gynecological cancer data to investigate inter-tumor heterogeneity.
  • Spatial transcriptomics data to analyze gene expression networks with spatial heterogeneity.

Key Insights and Ideas

  • Simpson's Paradox in Graphical Models: The paper effectively demonstrates how ignoring sample heterogeneity can lead to spurious associations and mask true biological signals, highlighting the necessity for methods like GraphR.
  • Heterogeneity as a Feature, Not a Bug: GraphR treats sample-specific variations not as noise, but as crucial information that can refine network inference and reveal biologically meaningful differences.
  • Scalability for High-Dimensional Data: The use of variational Bayes makes GraphR computationally efficient, enabling its application to large-scale genomics and proteomics datasets where traditional Bayesian methods might be intractable.
  • Uncovering Novel Biological Discoveries: By accounting for heterogeneity, GraphR identified biological relationships and patterns that were not apparent with existing homogeneous methods, particularly in cancer subtypes and spatial gene expression.

Key Results

  • Simulation Performance: GraphR consistently outperformed six state-of-the-art methods in terms of graph structure recovery (measured by MCC, AUC, TPR, FDR, FPR) across various simulation settings, especially in heterogeneous scenarios. It achieved significantly lower FDR and FPR while maintaining high TPR.
  • Computational Efficiency: GraphR demonstrated substantial speed improvements over MCMC-based Bayesian approaches, with computation times for large datasets being orders of magnitude faster. For example, a large network estimation took seconds with GraphR compared to hours with MCMC.
  • Biological Applications:
    • In breast cancer subtypes, GraphR identified specific protein-protein interactions and highlighted the role of phosphorylated EGFR.
  • Analysis of stemness indices revealed interactions that change with dedifferentiation levels, linking genes like GAPDH and FOXM1 to tumor progression.
  • Pan-cancer analysis revealed shared and distinct proteomic networks across different gynecological cancers.
  • Spatial transcriptomics data showed spatially varying gene correlations, identifying genes like XBP1 and SPINT2 with distinct roles in tumor vs. normal regions.

Limitations and Future Work

  • Non-Gaussian Distributions: The current GraphR framework is based on Gaussian distributions. The authors suggest extending it to model graphical structures for discrete data.
  • Non-Linear Relationships: The current model assumes linear relationships between covariates and precision matrices. Future work could explore non-linear function mappings.
  • Spatial Genomic Networks: The authors envision further expansion of GraphR for analyzing spatial genomic and proteomic networks across multiple tissue samples.
  • Prior Knowledge Integration: While GraphR is designed to be flexible, further exploration of how to best integrate prior biological knowledge could enhance its applicability.

Summary updated

Abstract

Abstract Probabilistic graphical models are powerful and widely used tools to quantify, visualize and interpret dependencies in complex biological systems such as highthroughput genomics and proteomics. However, most existing graphical modeling methods assume homogeneity within and across samples which restricts their broad applicability to cases where sample-specific heterogeneity exists e.g. tumor heterogeneity. We propose a flexible Bayesian approach called Graph ical R egression (GraphR) which (a) allows direct incorporation of intrinsic factors of sample heterogeneity at different scales through a regression-based formulation, (b) enables sparse network estimation at a sample-specific level, (c) allows identification and uncertainty quantification of potential effects of heterogeneity on network structures, and (d) is computationally efficient through the use of variational Bayes algorithms. We illustrate the comparative efficiency of GraphR against existing methods in terms of graph structure recovery and computational cost across multiple realistic simulation settings. We use GraphR to analyze four diverse multi-omics and spatial transcriptomics datasets to study inter- and intra-sample molecular networks and delineate biological discoveries that otherwise cannot be revealed by existing approaches. We have developed a GraphR R-package along with an accompanying Shiny App that provides comprehensive analysis and dynamic visualization functions.
BibTeX
@misc{chen2023probabilistic,
  author = {Liying Chen and Satwik Acharyya and Chunyu Luo and Yang Ni and Veerabhadran Baladandayuthapani},
  title = {{Probabilistic Graphical Modeling under Heterogeneity}},
  year = {2023},
  doi = {10.1101/2023.10.13.562136},
  howpublished = {bioRxiv},
}
Figure