Integration of Signaling Networks with Differential Gene Regulation Data Using a Probabilistic Pathway Impact Analysis Approach

The objective of the present study is to integrate the information on intracellular signaling network structures with large-scale gene expression data sets to predict functional changes in the cellular pathways that are operational under those conditions. Conventional approaches do not consider the network structure and transform the signaling pathway structures into lists of proteins, map the corresponding gene expression to these proteins and compute statistical confidence for observing these changes against a random chance. As the network structure is ignored in such ‘enrichment'-based methods, the results do not reflect the biological constraints and are likely to be erroneous in typical use. For example, consider the case when only a single component of a pathway with 10 total nodes is significantly down-regulated. The biological significance is likely to be dependent on whether the down-regulated component is an activator or an inhibitor and its location in the pathway structure. However, conventional methods ignore this important connectivity information. Recently, a method named SPIA was proposed to overcome this issue by incorporating the signaling pathway structure explicitly in assessing the impact of gene expression changes on cellular function (Traca et al., Bioinformatics 2009). A limitation of SPIA is that it considers the canonical signaling pathways in its assessment and does not reflect the particular biological context from which the gene expression data was obtained. We developed an alternative approach to SPIA that identifies context-specific signaling network structures that are consistent with the observed gene expression data. Our approach starts with a user-specified pathway structure, generates thousands of alternative pathway structures and evaluates each network for consistency with the gene expression data based on the probability of perturbation computed by the SPIA algorithm. This approach allows us to remove those interactions in the initial signaling pathway that are not consistent with the changes at the gene expression level. The probability of perturbation of each network and the enrichment-based statistics are combined into a significance metric for statistical evaluation of each network. We benchmarked our approach using three biologically motivated synthetic networks. In the benchmark studies, we considered 5-node, 7-node and 11-node synthetic networks with multiple activation and inhibitory interactions and feedback connections. In parallel, we generated nominal gene expression data sets with up-regulated activators and down-regulated inhibitors (data sampled from a normal distribution with appropriately signed mean of +/-1 and variance 0.09). For each of the benchmark cases, we searched for networks that are consistent with the simulated gene expression data sets by removing one or more connections in the pathways. Our results indicate that (1) networks with high consistency are likely to arise by differential expression of activators and inhibitors in opposing directions, (2) quantitative changes in the nodes with multiple upstream activation and inhibitory connections significantly affect the consistency, and (3) nodes with feedback interactions have large influence on whether the network is consistent with the gene expression data. In the biological case study, we applied the validated method to evaluate the biologically relevant interactions in the Angiotensin II receptor type 1 (AT1R) mediated pathway operational in the rat brainstem neurons dysregulated under high blood pressure conditions. We considered microarray gene expression data obtained from hypertensive animals and normotensive controls. The initial AT1R network was based on literature-based pathway interactions. We explored sixteen connections in the pathway as candidates for removal from the network structure and examined each of the 2^16 (=65536) networks for consistency with the microarray gene expression data using the probability of perturbation determined by SPIA. The networks with highly significant consistency showed remarkable similarity within the treatment groups but are distinct for hypertensive vs normotensive samples. Based on these results, we propose that the AT1R pathway interactions in the brainstem neurons are conditionally dependent on the hypertensive state of the animal. In summary, we present a novel approach to integrate signaling pathway structure information with the differential gene regulation and derive context-specific network structures based on the available gene expression data. Research Support: NIH R01GM083108 and R33HL088283.