TARGET DISCOVERY WITH
MULTI-OMICS MULTIVARIATE ANALYSES

The goal of our research is the identification of novel molecular therapeutic targets in solid tumors in a multivariate multi-omic setting.

Identifying active modules on a PPI network
We analyze differential entity expression between tumor and matching normal tissue at the DNA, mRNA, protein, and phospho-protein levels, which are mapped to a protein-protein interaction network (PPI). From this structure we identify active modules – small subsets in which differentially expressed entities are enriched which allow for biological and functional validation of contained potential drug targets.

Augmenting the established methods
There are many methods to identify active modules using single-omic data, recently using PPI embedding or random-field Ising models. We will discuss how to extend these methods to our multi-omic setting. Measuring the performance of active module identification methods is challenging due to the lack of a ground truth. We tackle this with a new performance metric, using the given differential expression results and PPI topology.

Active Modules  
In the multivariate multi-omic biomolecule analysis-based detection of novel pharmacological cancer treatment targets in cancer and matching normal tissues, the identification of active modules has been an evolving trend. Such modules are models of sets of proteins which physically interact to perform molecular functions and participate in cellular processes driving cancer phenotypes.

Detection
Active modules can be detected from T/N tissues by partitioning PPIs according to the degree of differential T/N biomolecule expression. To determine which modules are the most relevant for downstream validation and targeting steps, we show an approach to utilize the progression free survival (PFS) data of the patients from whom the tissues were obtained to rank the modules.

Representation
Active modules are of varying size but contain up to 60 biomolecules of three to four omics-analysis modalities which prevents a classical Cox survival regression. We therefore use a three-layer GCN with a linear readout layer on top to embed the molecule variables of the modules into fixed-size, low-dimensional representations.

Output
The output of the GCN is a module-specific embedding of the tissue data. The resulting low-dimensional embedding is used as the input for the Cox proportional hazards model. From it we obtain χ2 statistics for each module embedding along with coefficients for each predictor which enable a survival-based ordering of the modules.