Conditioning analysis

Multiple association signals could exist in one locus, especially when observing complex LD structures in the regional plot. Conditioning on one signal allows the separation of independent signals.

Statistical Model

In a standard GWAS association test, we model the phenotype \(Y\) as:

\[ Y = \mu + \beta_1 G_1 + \epsilon \]

where \(G_1\) is the genotype of the variant of interest, \(\beta_1\) is its effect size, and \(\epsilon\) is the error term.

When conditioning on a lead variant \(G_2\), we include it as a covariate in the regression model:

\[ Y = \mu + \beta_1 G_1 + \beta_2 G_2 + \epsilon \]

This allows us to test the association of \(G_1\) with \(Y\) while controlling for the effect of \(G_2\), effectively removing the effect of \(G_2\) and revealing independent association signals.

Several ways to perform the conditioning analysis:

Adding the lead variant to the covariates step-wisely and rerun the association test.
Conditional & joint association analysis using GWAS summary statistics (GCTA-COJO).

Adding the lead variant to the covariates

First, extract the individual genotype (dosage) to the text file. Then add it to covariates.

plink2 \
  --pfile chr1.dose.Rsq0.3 vzs \
  --extract chr1.list \
  --threads 1 \
  --export A \
  --out genotype/chr1

The exported format could be found in Export non-PLINK 2 fileset.

Note

Major allele dosage would be outputted. If adding ref-first, REF allele would be outputted. It does not matter as a covariate.

Then just paste it to the covariates table and run the association test.

Note

Some association test software will also provide options for condition analysis. For example, in PLINK, you can use --condition <variant ID> for condition analysis. You can simply provide a list of variant IDs to run the condition analysis.

GCTA-COJO

If raw genotypes and phenotypes are not available, GCTA-COJO performs conditioning analysis using sumstats and external LD reference.

cojo-top-SNPs 10 will perform a step-wise model selection to select 10 independently associated SNPs (including non-significant ones).

gcta \
  --bfile chr1 \
  --chr 1 \
  --maf 0.001 \
  --cojo-file chr1_cojo.input \
  --cojo-top-SNPs 10 \
  --extract-region-bp 1 152383617 5000 \
  --out chr1_cojo.output

Note

bfile is used to generate LD. A size of > 4000 unrelated samples is suggested. Estimation of LD in GCTA is based on the hard-call genotype.

Input file format less chr1_cojo.input:

ID      ALLELE1 ALLELE0 A1FREQ  BETA    SE      P       N
chr1:11171:CCTTG:C      C       CCTTG   0.0831407       -0.0459889      0.0710074       0.5172  180590
chr1:13024:G:A  A       G       1.63957e-05     -3.2714 3.26302 0.3161  180590

Here A1 is the effect allele.

Then --cojo-cond could be used to generate new sumstats conditioned on the above-selected variant(s).

Conditional Analysis vs. Fine-mapping

While both conditional analysis and fine-mapping are used to dissect association signals in a locus, they address different questions and use different approaches:

Aspect	Conditional Analysis	Fine-mapping
Primary Goal	Identify independent association signals in a locus	Identify causal variant(s) within a locus
Question Asked	"Are there multiple independent signals after controlling for the lead variant?"	"Which specific variant(s) are likely to be causal?"
Method	Regression-based: add variants as covariates and test remaining associations	Bayesian methods: estimate posterior inclusion probabilities (PIPs)
Output	P-values for independent associations	PIPs and credible sets for causal variants
Focus	Signal separation: Distinguishes independent signals from correlated ones	Causal inference: Distinguishes causal variants from tag variants (in LD)
When to Use	When you want to identify all independent association signals	When you want to pinpoint the most likely causal variant(s)

Complementary Approaches

Conditional analysis is often used as a first step to identify how many independent signals exist in a region
Fine-mapping is then applied to each independent signal to identify the causal variant(s) within that signal
Together, they provide a comprehensive understanding of the genetic architecture of a locus

References

https://www.cog-genomics.org/plink/1.9/assoc
Yang, J., Ferreira, T., Morris, A. P., Medland, S. E., Genetic Investigation of ANthropometric Traits (GIANT) Consortium, DIAbetes Genetics Replication And Meta-analysis (DIAGRAM) Consortium, ... & Visscher, P. M. (2012). Conditional and joint multiple-SNP analysis of GWAS summary statistics identifies additional variants influencing complex traits. Nature genetics, 44(4), 369-375.