This repo contains the code for the Master's thesis 'Predicting the molecular mechanisms of genetic variants'.

It implements a Nextflow workflow, which retrieves all necessary genomic features and runs inference with several neural networks to obtain features for the mode of action prediction model - classifying whether the (fine-mapped) variant acts as a spicing (s)QTL or gene expression affected by chromatin accessibility (ce)QTL.

Genomic features

  1. Binary variable indicating whether the variant is located within the gene body
  2. Distance from the variant to the closest annotated splice junction (GENCODE v39 annotation)
  3. Number of overlaps with open chromatin regions in 5 cell types. We used the same ENCODE DNASE/ATAC-seq experiments on which ChromBPNet models were trained.
  4. Number of overlaps with binding sites of RNA binding proteins. We took the binding sites of 211 RBPs, identified by Nostrand et al.

Neural features

  1. Splicing scores from SpliceAI and Pangolin. Each model produces two scores: maximum increase and decrease in the probability of a site being a splice junction in a 1000bp window around the variant.
  2. Enformer SAD scores for five CAGE tracks (gene expression) and five DNASE tracks. SAD score is a difference between Enformer predictions for reference and alternative alleles, averaged over the eight flanking bins representing 1000bp window.
  3. ChromBPNet difference scores for five cell types.