In-silico validation

Disease knowledge assembly models were generated in order to capture the vast knowledge around AD and PD. The language used to build the underlying models is the open source version of the Biological Expression Language (BEL). BEL encodes knowledge-based (mostly literature-derived) “cause and effect” relationships into network models, which can be subjected to causal analysis using quantitative data such as gene expression. The models developed here not only represent a comprehensive view on the core established pathways involved in amyloid processing, but also cover a broad spectrum of events that lead to clinical readouts often seen in AD and PD patients, such as neuro-inflammatory cascades. These models constitute the core of the candidate mechanism graphs that NeuroMMSig is based on.

We do have an inventory of mechanisms NeuroMMSigDB, which can be queried and “run against data” – an essential feature required for any validation. NeuroMMSig entries are encoded in the Biological Expression Language (BEL), the “lingua franca” of mechanistic modelling. The BEL models can be used for in-silico validation through a variety of procedure:

  • Mapping of omics data and subsequent multivariate clustering

  • Algorithms such as “reverse causal reasoning” (RCR) that determine the concordance between a pattern observed in gene expression data and the encoded regulatory relationships in BEL

  • Algorithms such as “network perturbation amplitude” (NPA) that score patterns in gene expression data according to the distortion of physiological states under pathophysiology conditions

Any validation of candidate mechanisms has to take into account, that “signals” that provide supportive evidence for a mechanism-defined patient subgroup need to be studied in a longitudinal fashion. AETIONOMY has specifically paid attention to this aspect by

  • Visualization of biomarker levels change over states, which is facilitated through the “longitudinal study viewer” that partner FRAUNHOFER has developed in the course of IMI-project EPAD,

  • analyzing the role of disease mechanisms in the context of disease risk, specifically AD, and

  • “reverse reasoning” of complex dependency patterns in data (via Bayesian Network modelling).

Using prior knowledge and dimensionality reduction techniques (e.g. autoencoders) we can turn entire clinical trials into representations that can be used for mechanism validation.

The above outlined strategies allow for a rapid and systematic validation of all NeuroMMSig candidate mechanisms.