The activation of viral pattern sensors combined with cytokine signaling initiates a genetic program in DCs that is critical to controlling infection. We have found that the timing of this program is highly conserved. In order to deduce the underlying regulatory architecture, we developed an integrative method to reverse engineer the transcriptional network and its progression through time. Using genome-wide transcriptional profiling data covering the first 18 hours post-infection, our approach identified a single regulatory network composed of 24 transcription factors that accounts for the up-regulation of a majority of genes (). This network captures and connects through time many known elements of the antiviral response, including IRFs, STATs and NFkB. The cascading pattern of TF activity () and the interconnectedness of the network () combined with the tight conservation of target up-regulation times (), all indicate an integrated control as opposed to an underlying parallel architecture. It is important to carefully interpret the network structure and ordering of events defined by the predicted regulatory network since the overall genetic program results from the integration of multiple signals that may arrive at different times. The NDV infection was performed using an MOI of 0.5 so that only approximately half the cells are infected and thus many cells are simply responding to the interferon and other cytokines secreted by infected cells, rather than directly to the virus. While this implies that the gene expression measurements reflect the operation of two distinct signaling pathways, the inferred network structure suggests that both rapidly converge on an overlapping set of downstream transcriptional regulators. Nevertheless, this could explain why NFkB activity follows STAT activity in our network. We conclude that the uninhibited antiviral response is the result of a single transcriptional cascade with a convergent architecture, and is not implemented as a series of independent transcriptional events.
Six transcription factors whose involvement in the network was predicted by our method are not part of the general pathogen response signature (29
) or the DC core response (1
). These represent potentially novel antiviral TFs: ALX1, FOXC1, FOXO3, MAX, RUNX3 and ZEB1. In order to validate the transcriptional network, we attempted to experimentally validate the regulatory connections for three of these novel transcription factors (ALX1, FOXC1 and RUNX3). We considered both incoming links, testing regulation of the novel TF, and outgoing links, testing the ability of the novel TF to regulate other genes. Using EMSA, we validated 4 out of the 12 regulatory relationships tested indicating an antiviral role for all of the TFs tested. Moreover, this success rate likely represents a lower bound for the specificity of our method in determining regulatory relationships for two main reasons. First, we did not include the many well-known players in our experimental validation. Second, a negative EMSA result may simply reflect sub-optimal binding conditions or be indicative of a relatively unstable complex. Furthermore, although the temporal activation of PLA1A, IFNA14 and RUNX3 does not necessarily correlate to the binding kinetics of ALX1, FOXC1 and RUNX3 putative binding sites respectively, this is likely a reflection of their individual transcriptional potentials or ability to recruit DNA modifying enzymes. In the case of the putative FOXC1 binding site, this element represents the first documented example of a non-IRF controlling the transcription of this genetic cluster. Although this site was found to overlap the TATA-box, it has also been observed that FOX proteins can regulate genes through direct binding to TATA-boxes (35
). Further experiments will be needed to verify the true impact of this site and the cause for the FOXC1 binding site we have found.
ALX1 had the highest number of experimentally validated links. Based on the occupancy of the IRF7 upstream of the ALX1 TSS, it would appear that transcription is, in part, mediated directly by Type I interferon signaling (). In the network, ALX1 is predicted to cooperate with IRFs in the regulation of down-stream genes as part of a feed-forward loop (). ALX1 encodes for CART1, a paired-class homeodomain protein that is necessary for survival of the forebrain mesenchyme in rodents (36
). While the binding site preferences for CART1 have been characterized, (37
) the function of this protein in humans is not known. Although the precise role for CART1 in the antiviral response is not yet clear, a gene ontology analysis of ALX1 predicted targets suggests that this factor is functionally similar to STAT1 and FOXO3 (). ALX1 target genes are involved in a variety of functions, but the most enriched GO category predicts that this factor is most likely involved in the negative regulation of macromolecule biosynthetic processes (GO:0010558 – see ). Our network analysis places ALX1 within a well-known network motif (38
), supporting a critical role for this transcription factor in coordinating the antiviral response (). Our experimental results also appear to indicate the workings of such a feed forward control element of PLA1A by ALX1 and another virus activated transcription factor. Looking at our EMSA results we see that following the assembly of a single DNA binding complex at 4 hours, a second faster migrating complex was visible by 8 hours post-infection (). These results suggest the exciting possibility that ALX1, and other novel TFs identified in this study, may cooperate both with each other, or with known virus activated factors such as IRF3, IRF7, STAT1, and/or NFkB.
Network motif involving the putative antiviral transcription factor ALX1
Combinatorial regulation can play an important role in mammalian gene regulation, but is not specifically incorporated in our approach. However, multiple regulators are predicted for most target genes and the potential for combinatorial control can be investigated indirectly by analysis of TF target overlap (). Consistent with our idea of cascading control, we found that most of the implicated TFs do not exhibit significant pair-wise overlap of target genes. However, a cluster of high overlap, including ALX1 (mentioned above) and also FOXO3, FOXC1 and STAT, was identified and may also reflect cooperative activity (black box, ). Together these TFs account for approximately 70% of the genes controlled by the network. The cooperative nature of the network is further evidenced by the fact that only ~19% of genes controlled by the network are targeted by a single matrix. Thus, while TFs pass along the ‘baton’ in an orderly fashion, individual genes may be controlled by multiple TFs at various points in the cascade and the network as a whole is the minimal unit of control.
The network building approach developed here is generally applicable to transcriptional profiling time-series. It differs from most current analyses in several important ways. First, we grouped genes according to their initial time of up-regulation. Second, instead of simply identifying the controlling TFs, we connected these together mechanistically to generate a complete transcriptional cascade. In cases where the time resolution of microarray sampling may not be sufficiently dense to isolate genes with common cis-regulatory logic, we suggest that the model-based analysis might be generalized to estimate up-regulation times on a finer scale. Several assumptions underlying our network reconstruction method are important to keep in mind. The analysis is limited to transcription factors whose binding preferences are known and are included in the TRANSFAC database. Nevertheless, as we have demonstrated, it is possible to implicate factors with previously unknown functions such as ALX1. Our method is also restricted to TFs that are differentially-expressed at the mRNA level. Consequently, it will miss factors that are post-transcriptionally regulated, which may be particularly important in the earliest stages of the response. Furthermore, since only up-regulated genes are included, the analysis will also not identify transcriptional repressors. While it is conceptually possible to extend the network reconstruction procedure to down-regulated genes, the lack of correlation among down-regulation times (Supplementary Figure 1
) suggested that a transcriptional cascade may not be the best model for these data.
To study the highly dynamic processes involved in immune interaction with pathogens requires a methodology that incorporates time directly into the analysis rather than subsuming it. The method presented here is a first step in the construction of such a methodology. Using our time-centric promoter analysis methods we have elucidated the transcriptional network underlying an uninhibited antiviral response in human DCs. Our results indicate a robust convergent design and stepwise execution of the antiviral program. While inherently limited to discovering regulation by TFs whose binding preferences are known (and annotated in TRANSFAC), the inferred network nonetheless accounts for a majority of up-regulated genes and time points. The identification of key transcriptional players in the antiviral response, along with knowledge of their timing and regulatory architecture, provides a framework to identify the specific mechanisms used by human pathogens to subvert normal immune function.