The Impact of Stability Considerations on Genetic Fine-Mapping

Reviewer #1 (Public Review):

Aw et al. have proposed that utilizing stability analysis can be useful for fine-mapping of cross populations. In addition, the authors have performed extensive analyses to understand the cases where the top eQTL and stable eQTL are the same or different via functional data.

Major comments:

1. It would be interesting to see how much fine-mapping stability can improve the fine-mapping results in cross-population. One can simulate data using true genotype data and quantify the amount the fine-mapping methods improve utilizing stability idea.

2. I would be very interested to see how other fine-mapping methods (FINEMAPE, SusiE, and CAVIAR) perform via the stability idea.

3. I am a little bit concerned about the PICS's assumption about one causal variant. The authors mentioned this assumption as one of their method limitations. However, given the utility of existing fine-mapping methods (FINEMAPE and SusiE), it is worth exploring this domain.

Reviewer #2 (Public Review):

Aw et al presents a new stability-guided fine-mapping method by extending the previously proposed PICS method. They applied their stability-based method to fine-map cis-eQTLs in the GEUVADIS dataset and compared it against what they call residualization-based method. They evaluated the performance of the proposed method using publicly available functional annotations and claimed the variants identified by their proposed stability-based method are more enriched for these functional annotations.

While the reviewer acknowledges the contribution of the present work, there are a couple of major concerns as described below.

Major:

1. It is critical to evaluate the proposed method in simulation settings, where we know which variants are truly causal. While I acknowledge their empirical approach using the functional annotations, a more unbiased, comprehensive evaluation in simulations would be necessary to assess its performance against the existing methods.

2. Also, simulations would be required to assess how the method is sensitive to different parameters, e.g., LD threshold, resampling number, or number of potential sets.

3. Given the previous studies have identified multiple putative causal variants in both GWAS and eQTL, I think it's better to model multiple causal variants in any modern fine-mapping methods. At least, a simulation to assess its impact would be appreciated.

4. Relatedly, I wonder what fraction of non-matching variants are due to the lack of multiple causal variant modeling.

5. I wonder if you can combine the stability-based and the residualization-based approach, i.e., using the residualized phenotypes for the stability-based approach. Would that further improve the accuracy or not?

6. The authors state that confounding in cohorts with diverse ancestries poses potential difficulties in identifying the correct causal variants. However, I don't see that they directly address whether the stability approach is mitigating this. It is hard to say whether the stability approach is helping beyond what simpler post-hoc QC (e.g., thresholding) can do.

7. For non-matching variants, I wonder what the difference of posterior probabilities is between the stable and top variants in each method. If the difference is small, maybe it is due to noise rather than signal.

8. It's a bit surprising that you observed matching variants with (stable) posterior probability ~ 0 (SFig. 1). What are the interpretations for these variants? Do you observe functional enrichment even for low posterior probability matching variants?