Background: Due to the interplay between many factors, systems medicine of complex diseases is most often centered either on individual genes, or on statistical associations. Complex network analysis could potentially explain systems-wide differences between patients and controls. In order to achieve this understanding using incomplete existing data, we need these powerful coarse-graining procedures. First, disease genes have been shown to co-localise for many complex diseases in protein-interaction networks, and thereby forming disease modules, however the identification of such modules is not well-defined (Gustafsson Genome Medicine 2014). We have developed ModifieR, an R-package which combines 10 existing and new methods using statistical overlaps between multiple omics as a guiding principle. Second, upstream hub transcription factors (TFs) regulating those modules have shown to be good candidates for early predictive medicine (Gustafsson Science Transl Med 2015). Third, those TFs tend to be interconnected and form core circuits, which have a great genome-wide impact, and could be modelled by our recent dynamic ordinary differential equation based modelling tool, LASSIM (Magnusson, PLoS Comput Biol 2017, https://gitlab.com/Gustafsson-lab/ ). Each of these concepts could synergistically increase the analysis resolution. A critical problem is that parts of the models are patient-specific while parts are highly conserved across patients, which we work to resolve.
Methods: A key feature to estimate patient specific disease-relevant networks is multiple time-points omics in cells reflecting different disease-relevant stages. We hypothesise that by connecting patients at multiple stages we can infer the disease-relevant patient specific gene regulatory networks, from which we can draw statistical conclusions (extension of the work in Hellberg Cell Reports 2016). For this purpose, we are performing modelling of CD4+ T-cells in multiple sclerosis using generally applicable principles in open access pipelines. Briefly, we apply a linear transcription factor-target model based on eQTL-targeted sequencing and patients in two time-points. The model uses L1 constraints, similar to our previous models (Gustafsson Science Transl Medicine 2015, Gustafsson IEEE 2005), but is allowed to infer transcription factor activity based on target gene expression, using the CVX optimisation toolbox.
Results: We found that the inferred patient specific GRN models with L1 constraint predicted 50% of the transcriptomic changes using cross-validation. Generally, we found that the TF activity poorly correlated with its mRNA level. Moreover, we found TFs with a significantly higher effect in patients than controls.
Discussion: My research focus on three recent network medicine concepts that enable a gradually increased resolution of modelling for complex diseases, which I ultimately aim at integrating in a unifying precisions medicine toolbox.