Genetic Stratification Based on Biological Networks in Autism Spectrum Disorders

Thursday, May 11, 2017: 12:00 PM-1:40 PM
Golden Gate Ballroom (Marriott Marquis Hotel)
Y. M. KIM1, T. Rolland1, R. Delorme1, G. Dumas1 and T. Bourgeron2, (1)Institut Pasteur, Paris, France, (2)Neuroscience, Institut Pasteur, Paris, France

The genetic architecture of Autism Spectrum Disorders (ASD) has proved to be complex and heterogeneous. For better prognosis and treatment, genetic stratification is suggested which aims to classify patients into more homogeneous subgroups. Cancer research has recently proposed Network Based Stratification (NBS) to stratify tumors into meaningful subtypes of similar molecular profiles by integrating protein interactions (Hofree et al 2013).


We aim to classify heterogeneous patients with genomic data into more homogeneous subgroups to associate them with clinical outcomes. NBS combines genomic information of each patient and Protein-Protein Interaction (PPI) networks. We developed an open source NBS toolkit in Python and propose to probe the impact of mathematical parameters on ASD stratification at both biological and phenotypical levels.


At first, we validate the NBS method with the original cancer data contributed by The Cancer Genome Atlas: 235 patients with uterine cancer. Once validated, we applied the NBS to an ASD cohort that involves 115 autistic patients, 354 related and 230 controls. We have the whole exome sequence of all participants, and clinical data.

After mutated gene’s score were diffused over neighbors, just like a thermal conduction, we reduced matrix dimension using Non-negative Matrix Factorization (NMF). This procedure aims to decompose a matrix into two lower rank matrices whose product can approximate the original matrix (Lee et al 1999). In addition, for the second application of graph topology, we introduce a Graph regularized NMF (GNMF) algorithm which respects the structure of the underlying gene interaction network and avoids the limitation of Euclidean space by incorporating a geometrically based regularization (Cai et al 2010). Finally, a consensus is calculated across 1,000 resampling iterations with hierarchical clustering. We then compare and analyze each subgroup characteristics at genomic, phenotypic and network topological level.


Our systematic and exhaustive investigation of all the parameters delivers the first guidelines to run NBS in ASD cohorts. The pilot study on cancer data already indicated that: 1) the diffusion step was essential but not the GNMF; 2) median quantile normalization could lead to even better results with just diffusion and NMF.

Our ASD study focused on rare deleterious germline mutations, whether inherited or de novo. Our preliminary results on ASD data reveal that: 1) in contrast to cancer, homogeneity of mutation profile across patients is preventing a drive of the clustering by mutation numbers; 2) modifying the diffusion factor allows to change the size of the mutated sub-network areas and potentially open up the ability to investigate local and global effects; and 3) NBS with a systematic PPI network (Rolland et al 2014) outperforms previous attempt with literature based networks. Moreover, several subnetworks stratifying patients with ASD were identified and will be presented.


Moving beyond the traditional dichotomy between monogenic and polygenic approaches, the NBS provides a versatile method to tackle heterogeneity of ASD and combined with proper clinical data has the possibility to uncover new relevant genotype-phenotype relationships, such as comorbidities and drugs response, for better prognosis and care of patients with ASD.

See more of: Genetics
See more of: Genetics