Mendelian randomisation (MR)

[edm1]

We have not planned exactly how this will be done.

One option is to take results straight from the Bristol IEU's MR Everything-vs-Everything (MR-EvE). Their methods preprint is here.

We can access the neo4j database from: http://eve-neo4j.mrbase.org/browser/ (Ed has username/password).

To think about:

• [ ] How do diseases/phenotypes map to EFO
• [ ] Can we map targets to genes (e.g. the metabolites)
• [ ] What parameters are available. How can these be used to assess (1) strength of evidence, (2) risk of bias
• [ ] What comparisons are available
• [ ] How would we represent these data in the portal?
• [ ] Taking into accounts the caveats, the main thing to ask is whether an association detected through this process could be used as 'evidence' in our system, bearing in mind that we can score lower to reflect caveats. What would this look like e.g. is the protein measurement a causal biomarker for the trait?

[edm1]

Ed's notes on the method (from November 2017):

Phenotypes used:

• 2407 phenotypes - can be viewed in the paper’s supplementary table 1:
  • 149 complex traits and diseases
  • 575 metabolites
  • 1683 plasma protein levels: http://biorxiv.org/content/early/2017/05/05/134551
  • Total of 715,681 interactions (only 61,029 interactions using MR-MoE)
• Includes EFO for most of the diseases and risk factors. But not for the metabolite and protein levels.

Instrument selection:

1. Top hits - Selecting SNPs association with p<5e-8 and clumping to obtain independent SNPs, then replicating in an independent sample.
2. Steiger filtering - In GWAS with high statistical power, it is difficult to know which trait a SNP (g) has its primary effect on. This could lead to the incorrectly concluding that B->A when in fact A->B. If trait A causes trait B, then g will be kept if the r2 is greater with A than B. (Currently not clear to me how this is done as they are using summary statistics, and don’t use the raw phenotypes.)

MR methods used (14 total):

• Mean-based methods (4 methods)
  • Inverse variance weighted (IVW). Fixed effects meta-analysis. Variants are weighted by the SE in the SNP-outcome associations. Assumes no horizontal pleiotropy.
  • IVW random effects. Relaxes horizontal pleiotropy assuming; allows it to be present but must be balanced (so that it only leads to increased heterogeneity around the regression line without affecting the slope).
  • Fixed effects Egger analysis. Further relaxes assumptions by allowing a non-zero intercept. This allows the overall horizontal pleiotropy to be directional, where its total effect influences the outcome in a specific direction.
  • Random effects Egger regression. Further allows heterogeneity around the slope, having account for overall directional horizontal pleiotropy. This assumes that the horizontal pleiotropy effects are not correlated with the SNP-exposure effects (the INSIDE assumption).
• Rucker framework adapted to MR. Uses estimates of heterogeneity in the IVW and Egger to navigate between these nested models. A jackknife approach (random selection with replacement of the complete set of instruments) is used to obtain a sampling distribution for the model estimates amongst the 4 above mean-based approaches. Where >15 instruments are available, they use 1000 rounds of jackknife estimations to calculate the mean and medians of the distributions. (3 estimates)
  • Point estimate
  • Mean of the jackknife
  • Median of the jackknife
• Median-based methods - advantageous as only half of the instruments need to be valid for the estimation to remain valid. (3 methods)
  • Simple median estimate
  • Weighted median estimate. Allows stronger instruments to contribute more towards the estimate by weighting each instrument by the inverse of its variance.
  • Penalised weighted median estimate. Introduces further weighting to penalise any instruments that contribute substantially towards the heterogeneity statistic (note: what statistic is used?)
• Mode-based estimators. Clusters the instruments into groups based in the similarity of causal effects, then returns the causal effect for the cluster that contains the largest number of instruments. (4 methods).
  • Simple mode. The unweighted mode of the empirical density function of causal estimates. With and without assumption that there is no measurement error in the exposure estimates (NOME).
  • Weighted mode. Weighted by the inverse variance of the outcome effect. With and without NOME.

Mixture of experts (MoE):

• They simulated 100,000 summary datasets
• And calculated 28 metrics on these datasets (e.g. num SNPs, sample sizes, F stats, outliers, etc.) full list in supplementary sheet 3.
• They then used the 14 MR methods above - each with and without Steiger filtering (28 methods in total), full list supplementary sheet 2.
• And used a random forest to predict which MR method gave the most accurate estimate based on the datasets metrics.

Which method gets used after Steiger filtering?

• If >5 instruments remain, use MoE to get best method
• If 2-5, use IVW random-effects approach
• If =1, Use Wald ratio on that variant
• If no variants, declare no estimate

Limitations:

• They advise that associations need additional interpretation.
• There was evidence for a large degree of horizontal pleiotropy.
• For a large number of traits there were <=5 variants after Steiger filtering, meaning that MR-MoE could not be used (less adjustment for horizontal pleiotropy).

• Some of the reported associations are biologically impossible. E.g. the outcome occurs earlier in life than the exposure (e.g. college completion -> childhood intelligence). This can partially be explained by “liability scale”. That is, genetic instruments are a proxy for the liability to achieve some level of educational attainment, rather than having completed some level of education.

  "In light of these issues, we emphasise that we do not consider the reporting of MR-MoE (or any othersingle MR method) to be seen as performing a Mendelian randomization study. Rather, it can be used to motivate further detailed follow up which should include sensitivity analyses (12), incorporating biological knowledge regarding instruments, and triangulation with other sources of evidence or experimental designs(14). Demanding a rigorous follow up of putative associations is also essential for avoiding issues that mayarise due to cherry picking of MR results. This is especially important when a large repository of putative associations, such as MR-EvE, is made publicly available."

[edm1]

Update:

We have a project underway to incorporate MR into the portal. It looks like, we’ll initially focus on generating high quality genetic instruments from protein QTL data following a similar methodology to this study: https://www.biorxiv.org/content/10.1101/627398v2. We can then expand to other molecular phenotypes such as metabolite QTLs.

I don’t feel we’d see much benefit, on top of coloc, from generating instruments from our expression-QTL (eQTL) DB as: (1) there is a high degree of pleiotropy (~25% of loci have strong eQTl coloc evidence with >1 gene), which would invalidate the assumption that the variant’s effect on the outcome (disease) is mediated entirely via its effect on the exposure (eQTL in this case); (2) we only have cis-eQTLs, which often only have a single causal variant. Multiple independent variants are required for most sensitivity analyses, without which we would be less confident in the MR estimates.

[d0choa]

This ticket has been superseded by Mohd's work