Detecting genome-wide directional effects of transcription factor binding on polygenic disease risk

Abstract

Biological interpretation of GWAS data frequently involves analyzing unsigned genomic annotations comprising SNPs involved in a biological process and assessing enrichment for disease signal. However, it is often possible to generate signed annotations quantifying whether each SNP allele promotes or hinders a biological process, e.g., binding of a transcription factor (TF). Directional effects of such annotations on disease risk enable stronger statements about causal mechanisms of disease than enrichments of corresponding unsigned annotations. Here we introduce a new method, signed LD profile regression, for detecting such directional effects using GWAS summary statistics, and we apply the method using 382 signed annotations reflecting predicted TF binding. We show via theory and simulations that our method is well-powered and is well-calibrated even when TF binding sites co-localize with other enriched regulatory elements, which can confound unsigned enrichment methods. We apply our method to 12 molecular traits and recover many known relationships including positive associations between gene expression and genome-wide binding of RNA polymerase II, NF-κB, and several ETS family members, as well as between known chromatin modifiers and their respective chromatin marks. Finally, we apply our method to 46 diseases and complex traits (average N = 289, 617) and identify 77 significant associations at per-trait FDR < 5%, representing 12 independent signals. Our results include a positive association between educational attainment and genome-wide binding of BCL11A, consistent with recent work linking BCL11A hemizygosity to intellectual disability; a negative association between lupus risk and genome-wide binding of CTCF, which has been shown to suppress myeloid differentiation; and a positive association between Crohn's disease (CD) risk and genome-wide binding of IRF1, an immune regulator that lies inside a CD GWAS locus and has eQTLs that increase CD risk. Our method provides a new way to leverage functional data to draw inferences about causal mechanisms of disease.

Sequential regulatory activity prediction across chromosomes with convolutional neural networks

“…experiment with a peak calling algorithm (Methods). We trained a Basset model using a set of hyperparameters studied in a recent application of the method (Reshef et al 2017). Basenji quantitative predictions achieved a greater AUPRC for all experiments, increasing the average from 0.435 to 0.577 and median from 0.449 to 0.591 (Supplemental Fig.…”

“…Linkage disequilibrium (LD) complicates the comparison to eQTL statistics; marginal associations and significance calls depend on correlated variants in addition to the measured variant, and association scans are better powered for variants that tag more genetic variation . To put SED on a level plane with the eQTL statistics, we distributed the SED scores according to variant correlations to form a signed LD profile of our SED scores, here denoted SED-LD (Methods) (Reshef et al 2017).…”

Sequential regulatory activity prediction across chromosomes with convolutional neural networks

“…To benchmark this framework against our previous Basset (Kelley et al 2016), we focused on the 949 DNase-seq experiments, divided the longer sequences into 1024 bp subsequences, and classified each subsequence in each experiment with a peak calling algorithm (Methods). We trained a Basset model using a set of hyper-parameters studied in a recent application of the method (Reshef et al 2017). Basenji quantitative predictions achieved a greater AUPRC for all experiments, increasing the average from 0.435 to 0.577 and median from 0.449 to 0.591 (Supplemental Fig S1).…”

“…For each dataset, we called peaks on the smoothed, normalized counts in the center 256 bp of the subsequences using a Poisson model parameterized by the maximum of a global and local null lambda similar to the MACS2 approach and applied a 0.01 FDR cutoff (Zhang et al 2008). We trained the Basset model using the most effective hyperparameters yet discovered, which were used in a recent application (Reshef et al 2017).…”

Transcriptome-wide association studies: opportunities and challenges

“…Crohn's Strong Genome-wide, variants that increase binding of IRF1 (a transcriptional activator of the innate immune response) tend to increase Crohn's risk, and vice versa 56 . High-density genotyping of the IRF1/SLC22A4/5 locus indicates that IRF1, but not SLC22A4/5, associates with Crohn's disease risk, and IRF1 expression, but not SLC22A4 expression, in GI biopsies was increased among Crohn's cases 57 .…”