Functional analysis of low grade glioma genetic variants using statistics and physics-inspired deep learning methods

Large-scale genome-wide association studies (GWAS) have implicated thousands of germline variants in modulating individual's risk of diseases, including cancer. For low grade gliomas (LGGs), at least 25 risk loci have been identified, whose molecular functions, however, remain largely unknown. Understanding how the risk loci function in tumorigenesis poses a major challenge in the field, owing to potential confounding factors and the lack of relevant types of experimental data in the brain. Based on statistical methods and physics-inspired deep learning methods, this work presents a comprehensive computational framework for performing functional analysis of LGG GWAS loci. We hypothesized that GWAS loci contain causal single nucleotide polymorphisms (SNPs) which reside in accessible open chromatin regions and modulate the expression of target genes by perturbing the binding affinity of transcription factors (TFs). We performed an integrative analysis using genomic, epigenomic and transcriptomic data from public repositories and identified the candidate (causal SNP, target gene, TF) triplets that might contribute to oncogenesis. We assessed a candidate causal SNP's potential regulatory role via convolutional neural network (CNN) and simulated-annealing-based interpretation methods. Finally, we applied tensor train decomposition (TT-decomposition) to neural network parameter reduction and demonstrated that the reduced convolutional neural network performed well. This work helps understand the molecular mechanisms underlying genetic risk factors of low grade glioma. The CNN and TT-decomposition-based deep learning approach may benefit future functional genomic studies, where TF chromatin immunoprecipitation followed by sequencing (ChIP-seq) data are not readily available in the brain