Confusion: "extract_outcome_data" function without without a filter for significant SNPs (whose p_value<5*10e-8)

[kamisama0101]

I got one confusion involving "extract_outcome_data" function in "TwoSampleMR" package: If is invalid to use non-genome wide significant SNPs from GWAS (whose p_value>5×10e-8) without the integration of a polygenic risk score for each individual, why was the"extract_outcome_data" function designed without a filter for significant SNPs? (whose p_value<5×10e-8). Will this design also cause any problem in analysis from analysis workflow on the TwoSampleMR official website？
https://mrcieu.github.io/TwoSampleMR/articles/perform_mr.html
Thank you for your help!

[nahadian]

Hi.

From what I understand, extract_outcome_data() is a data extraction function, not a SNP selection/filtering function. It assumes you've already constructed a valid SNP set to look up e.g., during instrument selection using extract_instruments(). The data extraction can still be narrowed down to any desired SNP set (e.g., instruments at p=5e-8 level) by specifying the "snps" parameter e.g.,:

instruments <- extract_instruments(..., p1 = 5e-8, ...) 

extract_outcome_data(snps = instruments$SNP, ...)

[remlapmot]

To add a little: extract_outcome_data() is purely for looking up the outcome GWAS associations for a SNP set you've already chosen. The genome-wide significance filter (p < 5×10⁻⁸) belongs at the instrument-selection / exposure stage — e.g. extract_instruments(..., p1 = 5e-8) or your own clumping/filtering — and those SNPs are then passed to extract_outcome_data(snps = ...).

Importantly, you should not filter the outcome SNPs on the outcome p-value. In two-sample MR you want the outcome association for every instrument, regardless of whether it reaches significance in the outcome GWAS — the instruments are selected on the exposure, and filtering them on the outcome would introduce selection bias. So the absence of a significance filter in extract_outcome_data() is by design, not a flaw, and the workflow on the website (perform_mr.html) is correct.