Mycobacterium tuberculosis Exploits a Molecular Off Switch of the Immune System for Intracellular Survival

Mycobacterium tuberculosis (M. tuberculosis) survives and multiplies inside human macrophages by subversion of immune mechanisms. Although these immune evasion strategies are well characterised functionally, the underlying molecular mechanisms are poorly understood. Here we show that during infection of human whole blood with M. tuberculosis, host gene transcriptional suppression, rather than activation, is the predominant response. Spatial, temporal and functional characterisation of repressed genes revealed their involvement in pathogen sensing and phagocytosis, degradation within the phagolysosome and antigen processing and presentation. To identify mechanisms underlying suppression of multiple immune genes we undertook epigenetic analyses. We identified significantly differentially expressed microRNAs with known targets in suppressed genes. In addition, after searching regions upstream of the start of transcription of suppressed genes for common sequence motifs, we discovered novel enriched composite sequence patterns, which corresponded to Alu repeat elements, transposable elements known to have wide ranging influences on gene expression. Our findings suggest that to survive within infected cells, mycobacteria exploit a complex immune “molecular off switch” controlled by both microRNAs and Alu regulatory elements.

Examination of the upstream 1500bp region of the significantly differentially expressed (SDE) downregulated genes compared to the SDE up-regulated genes from M. tuberculosis infection of whole blood over 96 h, identified a series of motifs shown to combine to form cassettes using the GSP algorithm. Node height is proportional to frequency. The "tree" arrangement shows which shorter cassettes fall within longer cassettes but does not imply any evolutionary or other arrangement.

Supplementary text 1 -Analysis of Microarray Time-course
Data see text script below S1 Text. Details of novel script used for time-course analysis Technical details for (i) the smoothing splines mixed-effects (SME) model used to fit the microarray time course data and (ii) the Wald type test applied to the fitted SME models to detect significantly differentially expressed probes.

Analysis of Microarray Time Course Data
We fit each probe on the microarray using a smoothing splines mixed-effects (SME) model (Berk et al. 2011). The SME model is a specific example of the functional mixed-effects model which has become popular in the analysis of replicated time course gene expression data (Storey et al. 2005;Liu and Yang 2009;Berk et al. 2010) due to its ability to handle missing observations, small sample sizes, few and irregularly spaced time points and subject heterogeneity that typifies such experiments. Under the SME model, it is assumed that the observations on subject i have arisen from some underlying smooth function of time, y i (t), which can be decomposed into the following components: where µ(t) is a mean function of time across all subjects, v i (t) is subject i's deviation from that mean function, also assumed to be a smooth function of time, and i (t) is an error process. Analogous to the standard mixed-effects model, µ(t) is assumed to be some fixed population parameter while the v i (t) are assumed to be randomly sampled from the population as a whole. Typically, the v i (t) are assumed to be independent realisations of a Gaussian Process with zero mean and covariance function γ(s, t) (Wu and Zhang 2006). In the SME model the functions µ(t) and v i (t), i = 1, · · · , n are represented as smoothing splines. To begin, (1) is written in matrix-vector format as: where y i is an N i length vector of all observations collected on subject i, X i is a known incidence matrix of dimension N i × p where p is the number of distinct design time points across all subjects, µ is a vector containing the values of µ(t) evaluated at the design time points and similarly for v i , and i is an N i length vector of all error terms corresponding to the observations on subject i. The incidence matrix X i maps the design time points onto the time points at which subject i was actually observed. This is in order to cater for two situations: (1) when data is missing, either due to the experimental design or due to errors in the measurement process such that subject i is not observed at all of the design time points and (2) when multiple observations on the same subject are available for a given time point due to technical replication. X i is constructed in the following way: the jth row, corresponding to the jth observation on subject i, taken at time t ij , contains zeroes in every column aside from the one corresponding to t ij which contains a one.
(2) is in the form of the standard linear mixed-effects model (Harville 1977). Standard practice is to assume that the randomly sampled v i are i.i.d. multivariate normal with zero mean and covariance matrix D. Independently, the error terms i are assumed to be i.i.d. multivariate normal with zero mean and covariance matrix R although typically this is simplified to R = σ 2 I. Under these assumptions, y i is itself multivariate normally distributed with mean vector X i µ and covariance matrix V i = X i DX T i + σ 2 I. The model parameters µ, D and σ 2 can then be estimated via maximum likelihood either by treating the random effects v i as missing data and employing the Expectation-Maximisation algorithm (Laird and Ware 1982), or through direct maximisation (Lindstrom and Bates 1988).
In the SME model, however, as the functions µ(t) and v i (t) are represented as smoothing splines, it is instead necessary to estimate the model parameters by maximising the penalised likelihood. This is the same as the standard likelihood with the addition of penalty terms for the roughness of the functions. Usually this roughness is quantified as the integral of the squared-second derivative (Ramsay and Silverman 2005).

1
The penalised complete log-likelihood across all subjects can then be written as: where y is an N = n i=1 N i length vector formed by concatenating all of the y i vectors and similarly for v, and λ * µ and λ * v are smoothing parameters controlling the roughness of the fit. These smoothing parameters are non-negative real values which allow for a full spectrum of non-linear behaviours to be considered. When zero, the penalty term is non-existant and the functions can interpolate the data points (assuming no technical replication). As the smoothing parameters tend to infinity, the penalty term dominates, leading to a linear fit as the second derivative will then be zero. Note that, as presented here, the same smoothing parameter λ * v is used for all subject specific functions. In principle, a separate parameter per subject specific function could be used, at the expense of computational cost. However, the idea of a common smoothing parameter is conceptually sound, given the assumption that the subject specific functions arise from the same underlying Gaussian Process (Wu and Zhang 2006).