A method for scoring the cell-type specific impacts of non-coding variants in personal genomes

To train a sequence-based regression model, we collected paired RNA-seq and ATAC-seq data for 42 samples of 18 tissues from the ENCODE project. The chromatin accessibility scores are calculated from ATAC-seq data. Then, we use the model trained based on ENCODE samples to predict open scores for GTEx samples. The training and predicting process is implemented by running

matlab predict.m

Based on the Ropen model, we can predict the chromatin accessibility score of a given region using the TF expression and genomic sequence information as input, where the TF binding sites can be derived from the sequences of the reference genome, or alternatively, it can be learned from whole-genome sequencing data. We use the change of chromatin accessibility scores before and after SNP mutation to measure the influence of a variant on an RE. To quantify this influence, we define the absolute value of log fold change between chromatin accessibility scores calculated based on the reference genome (REF) and that calculated based on whole-genome sequencing (WGS) data as the causal score for the given region.

matlab causal_delta_score.m

For a given risk SNP identified from GWAS summary data, we define the 200kb region centering at this SNP as a risk loci. Then, we define the variant causality score (VCS) for a variant in the risk loci as the average (across donors) of its causal scores ΔOs calculated by OpenCausal. Finally, by ranking the variants in this risk loci according to their VCS, we can prioritize putative causal variants for the GWAS trait. The prioritization is implemented by running

python prioritize.py &

Download OpenCausal by

git clone https://github.com/liwenran/OpenCausal

This project is licensed under the MIT License - see the LICENSE.md file for details