[Paper Review] Unsupervised learning of transcriptional regulatory networks via latent tree graphical models

Anthony Gitter, Furong Huang+3 more|arXiv (Cornell University)|Sep 20, 2016

Genomics and Chromatin Dynamics72 references3 citations

TL;DR

This paper proposes a latent tree graphical model to infer transcriptional regulatory networks from gene expression data without relying on transcription factor (TF) mRNA levels as proxies for regulator activity. By modeling hidden regulators as latent variables and using efficient unsupervised learning, the method identifies co-regulated gene groups, recovers known regulatory relationships in yeast stress response, and predicts novel condition-specific TF activities such as Msn4 binding under osmotic stress.

ABSTRACT

Gene expression is a readily-observed quantification of transcriptional activity and cellular state that enables the recovery of the relationships between regulators and their target genes. Reconstructing transcriptional regulatory networks from gene expression data is a problem that has attracted much attention, but previous work often makes the simplifying (but unrealistic) assumption that regulator activity is represented by mRNA levels. We use a latent tree graphical model to analyze gene expression without relying on transcription factor expression as a proxy for regulator activity. The latent tree model is a type of Markov random field that includes both observed gene variables and latent (hidden) variables, which factorize on a Markov tree. Through efficient unsupervised learning approaches, we determine which groups of genes are co-regulated by hidden regulators and the activity levels of those regulators. Post-processing annotates many of these discovered latent variables as specific transcription factors or groups of transcription factors. Other latent variables do not necessarily represent physical regulators but instead reveal hidden structure in the gene expression such as shared biological function. We apply the latent tree graphical model to a yeast stress response dataset. In addition to novel predictions, such as condition-specific binding of the transcription factor Msn4, our model recovers many known aspects of the yeast regulatory network. These include groups of co-regulated genes, condition-specific regulator activity, and combinatorial regulation among transcription factors. The latent tree graphical model is a general approach for analyzing gene expression data that requires no prior knowledge of which possible regulators exist, regulator activity, or where transcription factors physically bind.

Motivation & Objective

• To address the limitation of existing methods that assume TF mRNA levels reflect TF activity, which is biologically inaccurate due to post-transcriptional and post-translational regulation.
• To develop an unsupervised approach that infers hidden transcriptional regulators from gene co-expression data without prior knowledge of regulators, their activity, or binding sites.
• To identify co-regulated gene modules and assign biological meaning to latent variables through functional annotation and motif enrichment.
• To validate the method on a yeast stress response dataset, recovering known regulatory relationships and predicting novel ones such as condition-specific TF binding.

Proposed method

• Uses a latent tree graphical model where observed gene expression levels are conditionally dependent on hidden (latent) regulator variables, factorizing on a tree structure.
• Applies the guaranteed learning algorithm from Choi et al. [23] to recover the underlying tree structure and parameters from gene expression data without prior knowledge of the number or location of latent nodes.
• Employs a dynamic neighborhood selection method based on gene expression correlation distances to define the influence set of each latent node, using a tunable parameter λ = 0.15.
• Performs statistical testing (Fisher’s exact test and Bayesian FDR estimation via fdrtool) to assess overlap between latent node gene sets and known TF binding sites.
• Annotates latent nodes with Gene Ontology (GO) biological process terms and performs de novo motif discovery using WebMOTIFS to link latent regulators to known TF binding motifs.
• Uses conditional mean values of latent variables across osmotic stress samples to rank and identify active regulators, defining osmotic stress-specific TFs.

Experimental results

Research questions

• RQ1 Can a latent tree graphical model recover biologically meaningful transcriptional regulators from gene expression data without assuming TF mRNA levels reflect TF activity?
• RQ2 How well can the method identify known regulatory relationships and co-regulated gene modules in a well-characterized biological system like yeast stress response?
• RQ3 To what extent can latent variables be annotated as specific transcription factors or functional gene groups through motif enrichment and GO term analysis?
• RQ4 Can the method predict novel, condition-specific TF activities not detectable by standard mRNA-based approaches?
• RQ5 How does the performance of this latent tree model compare to established methods like ARACNE in terms of scalability and biological relevance?

Key findings

• The method successfully recovers known aspects of the yeast transcriptional regulatory network, including co-regulated gene modules and combinatorial regulation among transcription factors.
• It predicts novel, condition-specific TF activities, such as Msn4 binding under osmotic stress, which is not detectable via mRNA expression alone.
• The top 50% of latent nodes ranked by conditional mean in sorbitol-induced osmotic stress samples are significantly associated with known osmotic stress TFs, and this result is robust across different thresholds and sample subsets.
• A significant overlap is found between the extended neighborhood of influence of latent nodes and known TF binding sites, with statistical significance controlled via Bayesian FDR estimation.
• De novo motif discovery identifies known yeast TF binding motifs that match the sequence patterns in the gene sets associated with latent nodes, supporting their biological relevance.
• The latent tree model outperforms ARACNE in scalability, completing analysis on a 1035-gene subset in under 8 hours, whereas ARACNE failed to converge on the full dataset in reasonable time.