BINDER: computationally inferring a gene regulatory network for Mycobacterium abscessus

Abstract

Background: Although many of the genic features in Mycobacterium abscessus have been fully validated, a comprehensive understanding of the regulatory elements remains lacking. Moreover, there is little understanding of how the organism regulates its transcriptomic profile, enabling cells to survive in hostile environments. Here, to computationally infer the gene regulatory network for Mycobacterium abscessus we propose a novel statistical computational modelling approach: BayesIan gene regulatory Networks inferreD via gene coExpression and compaRative genomics (BINDER). In tandem with derived experimental coexpression data, the property of genomic conservation is exploited to probabilistically infer a gene regulatory network in Mycobacterium abscessus.Inference on regulatory interactions is conducted by combining 'primary' and 'auxiliary' data strata. The data forming the primary and auxiliary strata are derived from RNA-seq experiments and sequence information in the primary organism Mycobacterium abscessus as well as ChIP-seq data extracted from a related proxy organism Mycobacterium tuberculosis. The primary and auxiliary data are combined in a hierarchical Bayesian framework, informing the apposite bivariate likelihood function and prior distributions respectively. The inferred relationships provide insight to regulon groupings in Mycobacterium abscessus.

Results: We implement BINDER on data relating to a collection of 167,280 regulator-target pairs resulting in the identification of 54 regulator-target pairs, across 5 transcription factors, for which there is strong probability of regulatory interaction.

Conclusions: The inferred regulatory interactions provide insight to, and a valuable resource for further studies of, transcriptional control in Mycobacterium abscessus, and in the family of Mycobacteriaceae more generally. Further, the developed BINDER framework has broad applicability, useable in settings where computational inference of a gene regulatory network requires integration of data sources derived from both the primary organism of interest and from related proxy organisms.

Keywords: Bayesian inference; Data integration; Gene regulatory network; Mycobacterium abscessus.

Fig. 1

Number of target genes in the 34 orthologous M. tuberculosis regulons. Also illustrated is the the extent of orthology between M. tuberculosis and M. abscessus

Fig. 2

Simulation results illustrating the mean absolute deviation (MAD) between the true and estimated regulation interaction probabilities achieved by the deterministic, non-auxiliary and BINDER approaches across a range of dispersion parameter settings

Fig. 3

ROC analysis for ${\theta }_{r,t}^{50\%}$ posterior estimates for the BINDER, deterministic and non-auxiliary approaches and gene importance estimates for iRafNet for the r= fur and r= lexA regulons in E. coli and B. subtilis. BINDER (all) denotes results from analysis of BINDER applied to the complete coexpression data; BINDER relates to its application to the reduced data set

Fig. 4

Posterior estimates of ${\theta }_{r,t}^{50\%}$ for the BINDER, deterministic and non-auxiliary approaches for r= fur and r= lexA regulons in E. coli and B. subtilis, factored by established interaction status

Fig. 5

For the lexA regulon in B. subtilis and for targets where the auxiliary data ME=0 and PE=0, estimates of ${\theta }_{\text{lexA},t}^{50\%}$ for the BINDER, deterministic and non-auxiliary approaches, factored by known interaction status. The primary data values are CM and CP; points are jittered slightly for visibility

Fig. 6

Heat map illustrating the similarity between mean predicted ${\theta }_{r,t}^{50\%}$ values achieved by BINDER under three distinct prior distribution parameterisations (uninformative, informative, precise) on the set of N=167,280 regulator-target pairs

Fig. 7

Abacus plot illustrating interaction candidates achieving ${\theta }_{r,t}^{50\%}>0.9$ for the uninformative parameterisation; larger points are suggestive of less uncertainty; circles correspond to validated regulatory interactions in M. tuberculosis; shading corresponds to the posterior ${\theta }_{r,t}^{50\%}$ estimate. Regulators and targets are arranged by genomic position

Fig. 8

Central 95% of mass of the posterior distributions for ${\tau }_{{\text{ME}}_r}$, ${\tau }_{{\text{PE}}_r}$ and ζ_r under the uninformative parameterisation with posterior means indicated by dots for each of the R=34 regulators

Fig. 9

Central 95% of mass of posterior distributions for ϕ_r, ${\psi }_{{\text{CM}}_r}$ and ${\psi }_{{\text{CP}}_r}$ under the uninformative parameterisation with posterior mean values denoted by dots for each of the R=34 regulators

Fig. 10

Graphical representation of the hierarchical BINDER model; squares correspond to observed data, large discs correspond to random parameters and small discs correspond to fixed hyperparameters; the surrounding boxes denote observation-specific parameters and data