Modeling multipartite networks integrating high-throughput phospho-proteomics (ModPhosphoNet)

Scientific background

Network models are a common and powerful formalism for studying cell functioning and its deregulations during cancer. They allow the integration of a large number of entities into a single complex network. However, the analysis of these large networks remains a challenge. Notably, not many integrative studies include phosphoproteomics data, an important layer of information for understanding signaling pathways, due to the challenges in analyzing phosphoproteome data. In Medulloblastoma (MB), the most common malignant pediatric brain tumor, we have revealed that key signaling pathways are controlled at the post-transcriptional level in two out of four disease subgroups [1]. Thus, we propose here a novel integrative and modeling strategy to accurately integrate information about phosphoproteome and its potential functional consequences.

Results

Two main computational approaches have been explored: (i) stochastic blockmodeling (SBM) of multipartite networks and (ii) dynamical simulation of multi-layer networks. Blockmodeling is an increasingly popular class of models in statistical analysis of graphs that can be used to discover or understand the latent structure of a network, as well as for clustering purposes [2].

Each sample and each gene have as an output a certain probability to belong to a given block. This can be interpreted as a classification of patients into molecular similar clusters (subgroups) on the basis of similarly expressed genes (signatures). The output of a stochastic block modeling algorithm is a probability distribution of membership, i.e., the probability of a given sample to be associated to a subgroup and, at the same time, the probability of a gene to characterize a given subgroup. This mixed probabilistic membership is a major advantage of stochastic block modeling with respect to standard clustering approaches.

First, we proposed to use SBM algorithms developed for multipartite graphs to identify clusters of patients based on their connection behavior in multiple molecular networks (proteomics, phosphoproteomics and transcriptomics). The output of the SBMs analyses have been interpreted as weighted signatures per MB subgroups encompassing transcription factors, proteins and phosphorylated states of kinases. Second, we used this information to build network models specific for each disease subgroup. These networks can be used as basis for mathematical modeling and dynamical simulations to predict possible points of intervention.

[1] Forget A et al, Aberrant ERBB4-SRC Signaling as a Hallmark of Group 4 Medulloblastoma Revealed by Integrative Phosphoproteomic Profiling. Cancer Cell 10;34(3):379-395.e7 (2018)

[2] Lee C et al, A review of stochastic block models and extensions for graph clustering. Applied Network Science 4(1)(2019)