Detect Spurious eQTL Associations (DSEA)

Many trans-eQTLs can be 'phantom cis-eQTL effects' driven by multimapping reads due to gene families, paralogues and regions of high sequence similarity across the genome (CBWD1 below). As multi-mapping is inevitable across RNA-seq experiments, these false trans-eQTLs are replicable across studies. Even when considering only uniquely mapping reads, sequence variation at the population level or technical sequencing error can lead to reads mapping in the wrong genomic location.

We have devised a simple post eQTL-analysis filter pipeline to mark / remove eQTLs that are spurious and/or non-interpretable.

Extract SNPs from dosage files that have apparent eQTL effects.

python get_dosages.py dir/to/dosages/ dir/to/trans/eqtls/eQTLs.sorted output.dosages

Run DSEA to annotate results with paralogues, pseudogene enrichment and poor read coverage (stratified by genotype)

R CMD BATCH FalseTransFilter.R gencode.v19.annotation.gtf trans.eQTLs.txt.sorted /path/to/BAMs/ output.dosages snplocations.txt listOfQuantifiedGenes.txt

Required files:

1. trans-eQTL results (MatrixQTL output format) - best SNP per gene.
2. GENCODE annotation used in study (e.g. v19) - not filtered.
3. list of ENSEMBL ids expressed in study
4. directory to BAM alignments
5. SNP dosages that are trans-eQTLs

Functionality:

1. Overall test for psuedogene enrichment given expression background.

2. Mark trans-eQTLs that are pseudogenes

3. eQTLs that are members of gene families are also marked. If gene families are all in cis, multi-gene families can also create multiple spurious cis-eQTL signals in which no one 'causal' gene can be selected.

4. Mark phantom cis-eQTLs. These are trans-eQTLs driven by multimapping. We have observed many false trans-eQTL signals that are driven by shared sequence homology with regions in cis. Whilst the QTL effect is real (differential expression due to genotype) - It should act in cis rather than trans (CBWD1 below). We automate the process of checking mean nucleotide % read coverage across all individuals in a study. If many reads align to very few basepairs, the overall coverage across the gene will be non-uniformly distributed. In addition we provide diagnostic plots of each spurious association and mark it for filtering.

Example: CBWD1

An example of this effect is the following trans-eQTL rs2034278-CBWD1 (P = 1.96 x 10^-19) Gtex replication ( P = 0.0078, Subcutaneous Adipose tissue), demonstrating these spurious signals are replicable across cohorts. When scanning for paralogous genes in cis of rs2034278, it becomes apparent that there is 1 gene, CBWD2. Mean basepair coverage calculated across 720 individuals clearly demonstrates that the signal is spurious, with only 3.27% of the gene having >= 1 read per basepair, despite being expressed at 1 CPM > 90% of individuals. It is likely that this trans-eQTL is actually a phantom cis-eQTL effect on CBWD2, and due to their high sequence similarity (98%, ENSEMBL) the reads are erronously assigned to a region of CBWD1.

This spurious signal can be compared to a trans-eQTL for NME3 that did not have a paralogue within 1MB of its associated SNP and was not found to have poor/non-uniform read coverage (83% of gene > 1 read per bp) (below).