Aggregating single nucleotide polymorphisms improves filtering for false-positive associations postimputation

Area of Science:

• Genetics and Genomics
• Statistical Bioinformatics
• Population Genetics

Background:

• Imputation in genome-wide association studies (GWAS) introduces bias in P-values, complicating the identification of true associations.
• Current imputation quality measures and filtering thresholds (e.g., IMPUTE info 0.3 and 0.8) present a trade-off between discarding true associations and retaining false ones, particularly for SNP array data.
• Existing post-imputation filtering methods often treat single nucleotide polymorphisms (SNPs) independently, neglecting the impact of linkage disequilibrium on association signals.

Purpose of the Study:

• To quantify imputation-introduced bias in P-values within GWAS.
• To evaluate the effectiveness of existing post-imputation filtering methods and thresholds on individual SNPs and aggregated signals across different genotype formats (best guess and dosage).
• To propose and validate a novel filtering approach, the Midrange Filter, designed to improve the accuracy of association signal identification.

Main Methods:

• Simulated 1536 small case-control studies on human chromosome 19 to assess imputation bias and filtering performance.
• Compared established IMPUTE info thresholds (0.3, 0.8) using individual SNPs and aggregated spikes in both 'best guess genotype' and 'dosage' formats.
• Applied two recently published filtering methods, Iam hiQ and MagicalRsq, and introduced the proposed Midrange Filter method.

Main Results:

• Differences in false signals and imputation quality were observed between genotype formats, particularly in the midrange imputation quality scores.
• In the midrange, 51% (best guess) and 60% (dosage) of associated SNPs represented true associations; aggregated spikes predominantly comprised true associations.
• The proposed Midrange Filter method demonstrated efficacy in classifying spikes based on thresholds and formats, preserving all true associations in most simulations.

Conclusions:

• The Midrange Filter offers a promising approach to mitigate imputation bias in GWAS by considering aggregated signals and genotype formats.
• This method effectively balances the removal of false signals with the preservation of true associations, addressing limitations of SNP-independent filtering.
• The Midrange Filter's performance was validated using real data from the PsyCourse study, highlighting its practical applicability in genetic research.