Fine-mapping inflammatory bowel disease loci to single-variant resolution

Huang et al 2017 Nature

Overview

Inflammatory bowel disease (IBD)

{ width=60% }

• Ulcerative colitis (UC): colon inflammation & ulcers
• Crohn’s disease (CD): colon + ileum inflammation

Contributions of this study

The pace of identifying associated loci outstrips that of defining specific molecular mechanisms and extracting biological insight from associations

• 200 IBD loci identified, few resolved to functional variants
• This study: 94 loci evaluated, 45 variants resolved
• Approach: genetic «fine-mapping» + detailed annotations
  • Dense genotyping array of Europeans
  • 3 fine-mapping methods
  • Enrichment in regulatory regions
  • 3 eQTL colocalization methods

###

{ width=90% }

Fine-mapping: methods and results

1. Model selection with BIC
2. Bayesian single effect model & conditional regression
3. Bayesian LASSO

Method 1

Step 1: model selection of independent signals

• BIC-based selection
• Penalized likelihood across traits and variants
• Greedy search algorithm (steepest descent)
• A ``swap-out repositioning” adjustment

Method 1

Step 2: construction of confident sets (CS)

• 95% CS that contains exactly one signal
• Conditional CS in the presence of multiple signals
• Potentially multiple CS generated per region

Not accounting for uncertainty of alternative signals

Method 2 (Based on Jostins & McVean 2016)

Step 1: multivariate regression with correlated normal prior

• Assuming all loci are shared between traits
• Model and inference: a special case of eQTLBMA
  • No model averaging
• BF and CS assuming single effect

Method 2 (Based on conditional regression)

Step 2: mapping secondary signals

• Conditional regression followed by Step 1
• Stopping criteria: number of signals by method 1

Method 3 (Fang & Georges 2016 biorxiv)

Step 1: univariate, multiple regression model

• Multiple causal variants jointly using a LASSO prior
• Sample from posterior rather than computing it
  • Sample $y$ from liability threshold model
  • Sample configuration restricted to clusters (MH)
  • Sample parameters and hyperparameters (Gibbs)
• Get CS from posterior samples for each cluster

SNP clusters are created beforehand via hierachical clustering on LD. A correlation threshold has to be specified to define clusters.

Method 3

Step 2: cross trait CS construction

• Simply merge overlapping clusters

Consolidating 3 methods

Fine-mapping summary

1. 97 loci analyzed, 94 have «concensus» CS
2. Size of CS: from 1 (fine) to >400 (not so fine)
   • 68/94 regions have single variant
3. 18 single causal variants with >95% certainty
   • 13 are novel
4. 27 single causal variants with >50% certainty
5. Most associated with both UC and CD
6. Number of candidate genes reduced from 669 to 233

Fine-mapping summary

Proportion of variance explained (PVE)

• CD: 25% PVE by fine-mapping out of 28% by all loci
• UC: 17% out of 22%

Importance of secondary and tertiary associations

{ height=60% }

Enrichment and colocalization: methods and results

Protein coding variants: an «assurance»

• 18-fold enrichment in non-synonymous
• Only one variant with >50% certainty is synonymous
• All previously reported IBD risk coding variants are included with >95% certainty

Enrichment in TF families

• CD signals overlap more with immune cell peaks
• UC signals overlap more with gut peaks

Co-localization method 1: naive overlap

Background overlap rate: $$\rho_i = 1 - (1 - \frac{1}{N_i})^{C_i}$$

where $N_i$ is number of variants in region $i$, $C_i$ of which are in IDB CS. Expected background overlap is $\sum_i \rho_i$.

Result of method 1

• Non-significant overlap of top PBMC eQTLs in CS identified
  • in neither of the 2 different studies investigated
• Overlap of eQTLs for CD4+ cells and ileum cells observed

Co-localization method 2: conditional p-value

Nica et al 2010 PLoS Genet

$$\delta = -\log(P{eQTL}) - [-\log(P{eQTL|GWAS})]$$

• Empirical p-value for $\delta$ by comparing to dropping other variants

Result of Method 2

Co-localization method 3: Model comparisons

Uses summary statistics and Baysian «configuration» models

{ height=50% }

Result of Method 3

• Additionally found moderate enrichment in colon
  • p-value 0.04 compared to 0.216 via method 2

Final remarks

Conclusions

• Significant enrichment in coding region and strong immune signature
• Moderate, cell/tissue specific eQTL colocalization
• 21 non-coding variants in IBD CS not interpreted

Discussions

• Direct genotyping of (common) causal variant is essential
• Improve fine-mapping resolution in other ethnicities
  • but not enough data to pursue
    

Criticisms??