Robust knowledge-guided biclustering for multi-omics data

Qiyiwen Zhang, Changgee Chang, Qi Long

TL;DR

This paper introduces Bayesian Graph-Guided Biclustering (BGB), a novel method that integrates biological knowledge from graphs with multi-modal omics data for more robust and interpretable biclustering. BGB outperforms existing methods by effectively leveraging graph information, even when it's noisy, and can handle diverse data types simultaneously.

What Problem Does This Paper Address?

Biclustering is a powerful technique for identifying groups of similar samples and features within large datasets, particularly useful in high-dimensional biological data. However, existing methods often struggle with:

  • Integrating Multi-Modal Data: Many methods are designed for a single type of data (e.g., gene expression) and cannot effectively combine diverse omics data like genomics, transcriptomics, and epigenomics.
  • Leveraging Biological Knowledge: While biological knowledge, often represented as graphs (e.g., gene regulatory networks), has proven beneficial in other statistical tasks, its integration into biclustering methods has been limited. Existing graph-guided approaches have restrictive assumptions, such as requiring diagonal dominance in covariance matrices, which hinders their application to larger or more complex biological networks.
  • Handling Overlapping Biclusters: Biological phenomena often involve features participating in multiple biological processes or subtypes. Many biclustering methods assume non-overlapping clusters, which can be biologically unrealistic.

This paper aims to bridge these gaps by proposing a flexible and robust biclustering framework that can handle multi-modal data and effectively incorporate biological graph information.

Methods

The core of the proposed method, Bayesian Graph-Guided Biclustering (BGB), lies in its novel Bayesian hierarchical model and efficient computational framework:

  • Unified Multi-Modal Model: BGB uses a data augmentation technique (Pólya Gamma latent variables) to create a unified likelihood function that can accommodate various data types, including Gaussian, binomial, and negative binomial distributions. This allows for the simultaneous analysis of different omics data.
  • Knowledge-Guided Priors: The key innovation is the incorporation of biological graph information through novel hierarchical sparsity-inducing priors for the factor loading matrix (W). These priors allow for:
  • Flexible Graph Integration: Unlike previous methods, BGB's priors do not require diagonal dominance, making them applicable to larger and more complex biological networks.
  • Adaptive Shrinkage: The priors induce sparsity in the factor loadings, meaning only relevant features are selected for each bicluster. This sparsity is guided by the graph structure, encouraging features connected in the graph to be selected or deselected together.
  • Robustness to Noise: The model is designed to be robust to noisy graph information, as demonstrated by its performance even with imperfect graph structures.
  • Efficient MCMC Inference: An efficient Markov Chain Monte Carlo (MCMC) algorithm is developed for posterior sampling and inference, utilizing data augmentation and Gibbs sampling techniques.
  • Model Selection: A variant of the Deviance Information Criterion (DIC) is used to select optimal tuning parameters and the number of biclusters (L).

Datasets

The paper evaluates BGB using:

  • Simulated Data: Two simulation settings (non-overlapping and overlapping biclusters) were created with varying graph structures (null, within-pathway, mixed, and noisy across-pathway) and data types (Gaussian and binomial).
  • Real Data: The Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset, integrating SNP and gene expression data from human subjects, was used for real-world validation.

Key Insights and Ideas

  • Biological Graphs are Powerful for Biclustering: The study strongly suggests that incorporating biological network information significantly enhances biclustering performance, leading to more biologically meaningful results.
  • Flexibility is Key: The novel priors in BGB overcome limitations of previous graph-guided methods, allowing for more adaptable integration of biological knowledge, especially for complex networks.
  • Robustness to Imperfect Knowledge: BGB's ability to perform well even with noisy or incomplete graph information is a critical advantage for real-world applications where perfect biological networks are rare.
  • Unified Handling of Diverse Data: The ability to seamlessly integrate different omics data types within a single framework is a major step towards a more holistic understanding of biological systems.
  • Overlapping Biclusters are Biologically Relevant: The method's capability to identify overlapping biclusters aligns better with the complex, interconnected nature of biological processes.

Key Results

  • Superior Performance: BGB consistently outperformed existing biclustering methods (GBC, SSLB, iSSVD, ISA) in both simulation studies and real data analysis across various metrics (Relevance, Recovery, Census Score, Clustering Accuracy).
  • Graph Incorporation Benefits: BGB with graph incorporation showed significant improvements over BGB without graph information, highlighting the value of the proposed priors.
  • Robustness Demonstrated: BGB's performance remained strong even with noisy graph structures (G3 and G4), surpassing other methods.
  • Real Data Validation: On the ADNI dataset, BGB identified biologically relevant pathways (e.g., sensory perception, olfactory transduction) related to Alzheimer's disease, providing interpretable insights.
  • Overestimation of L: While most methods tended to overestimate the number of biclusters (L), BGB's overestimation was generally within a narrower margin.

Limitations and Future Work

The authors acknowledge several areas for future research:

  • Fully Bayesian Model Selection: Extending the model to include a fully Bayesian approach for model selection (e.g., determining model complexity and optimal parameters) could further improve interpretability.
  • Algorithmic Optimizations: Exploring algorithmic optimizations and parallel computing techniques could enhance the computational efficiency of BGB, especially for very large datasets.
  • Broader Applications: Expanding the model to handle more complex scenarios, such as incomplete multi-omics data, different types of biological data (e.g., imaging, fMRI), and more intricate graph structures, is a promising direction.
  • Integration with Deep Learning: Integrating the graph-incorporated technique into deep learning architectures could lead to more interpretable deep learning models for biological data analysis.

Summary updated

Abstract

Biclustering is a useful method for simultaneously grouping samples and features and has been applied across various biomedical data types. However, most existing biclustering methods lack the ability to integratively analyze multi-modal data such as multi-omics data such as genome, transcriptome and epigenome. Moreover, the potential of leveraging biological knowledge represented by graphs, which has been demonstrated to be beneficial in various statistical tasks such as variable selection and prediction, remains largely untapped in the context of biclustering. To address both, we propose a novel Bayesian biclustering method called Bayesian graph-guided biclustering (BGB). Specifically, we introduce a new hierarchical sparsity-inducing prior to effectively incorporate biological graph information and establish a unified framework to model multi-view data. We develop an efficient Markov chain Monte Carlo algorithm to conduct posterior sampling and inference. Extensive simulations and real data analysis show that BGB outperforms other popular biclustering methods. Notably, BGB is robust in terms of utilizing biological knowledge and has the capability to reveal biologically meaningful information from heterogeneous multi-modal data.
BibTeX
@article{zhang2024robust,
  author = {Qiyiwen Zhang and Changgee Chang and Qi Long},
  title = {{Robust knowledge-guided biclustering for multi-omics data}},
  journal = {Briefings in Bioinformatics},
  year = {2024},
  doi = {10.1093/bib/bbad446},
}
Figure