Traditional statistical approaches for characterizing interactions have only a limited ability to capture the heterogeneity and complexity of pathway crosstalk and consequently a number of novel informatics approaches have been proposed (Myers et al., 2005; Li et al., 2008; Liu et al., 2010; Xu et al., 2011; Hsu and Yang, 2012; Atluri et al., 2013; Diao et al., 2013; Wang et al., 2013; Tegge et al., 2015). One could quantify via scores the likelihood that a pair of pathways will crosstalk based on existing biological datasets that provide evidence for possible crosstalks (including physical interaction, genetic interaction, and transcription factors). To have a more robust pathway crosstalk map, one could incorporate a wide array of evidences. Scores from each of these evidences can then be combined to build a generic pathway crosstalk reference map analogous to the “Kyoto Encyclopedia of Genes and Genomes” (KEGG) pathway reference map. The likelihood of two pathways crosstalking can be scored utilizing one of several different methods. One method could be based on the presence of “common elements,” such as shared enzymes and metabolites. For example, the regulatory signaling protein kinase A (PKA) and protein kinase C (PKC) pathways converge onto the MAPK kinase system using the small G protein, Ras, a common element of both the PKA and PKC pathway (Franco et al., 2017). Another method could rely on the presence of “interacting elements,” such as physical protein-protein interactions. An example of such an interaction in humans is the ANP32A protein binding to Axin-1, another protein, to enhance its suppression of the Wnt pathway, a key pathway in adult tumor formation (Stelzl et al., 2005). Using such methods, one could build a network from a generic pathway crosstalk reference map where the nodes represent pathways and the edges represent a statistically significant p-value for crosstalk likelihood between a pathway pair.

To determine which of the pathway crosstalks in the generic reference map may be utilized as a biomarker for AD, one could then go further to characterize patient-specific pathway crosstalks. For this purpose, one could make use of single-nucleotide polymorphism (SNP) and gene expression (transcriptomics) data collected in large naturalistic studies such as the Alzheimer's Disease Neuroimaging Initiative (ADNI) and the Dominant Inherited Alzheimer's Network (DIAN) or large clinical trials such as the Alzheimer's Disease Genetic Consortium (ADGC). Using SNP-data as an illustrative example, characterization of patient-specific pathway crosstalks could be broken down into four steps schematically as shown in Figure 1B:

Obtain a mapping of SNPs to pathways using genetic information.

If a different “-omics” data source is used such as proteomics or metabolomics, then a simple substitution of SNPs for the desired data source in the above schema is needed to obtain “patient specific data-enriched” pathways.

A number of other informatics approaches that attempt to predict crosstalk are possible and have been reviewed (Myers et al., 2005; Dotan-Cohen et al., 2009; Liu et al., 2010; Xu et al., 2011; Atluri et al., 2013). These approaches are limited to physical protein interaction alone and do not capitalize upon additional physical and functional crosstalk evidences available (Li et al., 2008). Moreover, personalization of pathway crosstalks and subsequent utilization of such crosstalks as prognostic biomarkers has not been broached.

A recent systems biology analysis conducted by Hu et al. (2017) constructed an AD pathway crosstalk map from 430 human genes associated with AD. A total of 68 biological pathways were found to be enriched by these AD-related genes. Pathway crosstalks were determined by the proportion of overlapping genes with a minimum of two shared genes between pathways as a requirement for inclusion in the pathway crosstalk map. Their crosstalk network revealed three core interacting modules, consisting of immune modulation-related pathways, cell growth/survival and neuroendocrine-related pathways, and neuronal and drug-metabolism pathways (Figure 1C). Such an AD immune-endocrine-neuronal regulatory network is consistent with our current understanding of AD pathogenesis. These promising findings highlight the need for moving beyond the traditional single-gene based studies to network and pathway-based methodologies. Furthermore, we believe enriching sophisticated pathway crosstalk analyses with patient-specific data could yield powerful personalized biomarkers that could advance our understanding of disease mechanisms and potential therapeutic targets.

Emerging robust bioinformatics models will ultimately need to be validated using existing datasets and biological models of disease. Some recent studies have begun to utilize such methods. Mukherjee et al. (2014) performed a network analyses incorporating a human protein-protein interaction database mined from 12 different sites including BIND, BioGRID, Intct to the HapMap2-imputed combined ADGC data set from 15 studies. They identified a set of significant modules and candidate genes and then demonstrated an initial functional validation of some of these candidate markers as modifiers of amyloid-beta toxicity in vivo using a transgenic C. elegans model. Liu et al. (2016) analyzed 528 biomarkers in ADNI data (proteomics, MRI, cognitive tests) to examine the sequence of network changes related to the risk for progression from mild cognitive impairment (MCI) to AD. A semi-mechanism-based Bayesian network with 26 nodes and 43 arcs was generated, which achieved a high 10-fold cross-validated prediction performance with 95% sensitivity and 65% specificity. The network analyses identified several markers of relevance such as fibrin clot formation and hyperinsulinemia. This is consistent with our own preliminary work using longitudinal ADNI data from 91 MCI subjects which suggests that constructing patient-specific SNP crosstalk maps may enhance the predictive accuracy above and beyond the traditional approach of combining discrete MRI and cognitive test markers. Clearly, further experimental validation will be critical and readers are referred elsewhere for a more comprehensive review of validation methods (Chen et al., 2015).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.