number of associated SNPs in tsim Cohorts GWAS than meta-analysis of GWASs

[Manonanthini]

Hi,
The two-stage imputation strategy is great. Thank you for the method!
I have tried tsim in two differently genotyped independent Cohorts after the first imputation using the same reference panel and GWAS after merging them using hq SNPs followed by the second stage imputation using the Topmed r3 reference panel. I also performed GWAS on them independently and combined them using meta-analysis. I found less number of SNPs in the tsim merged GWAS compared to the meta-analysis of GWAS results of those cohorts. I found almost twice the number of SNPs in meta-analysis than merged GWAS. What should be the reason for that? How can I rectify it?

Regards
Mano

[Dongwon-Lee]

Hi Mano, thank you for your interest in our TSIM method! For the r3 TOPMED reference panel, you will need to use a lower r^2 cutoff for hqSNP to have sufficient coverage. As noted in the README, we have found that 0.98 should be used instead of 0.99. In fact, you can still get a comparable accuracy with an r^2 cutoff as low as 0.97 based on our new calibration analysis. We are in the process of revising the manuscript to reflect this. Stay tuned!

Dongwon

[Manonanthini]

Hi, Thanks for the reply. I need a clarification regarding ER2 value. Do we have to keep the threshold as 0.9 or can it be reduced further down? If so, what should be the minimum value to consider for ER2 ? Also for hq SNPs, is ER2 >0.9 & R2>0.97 or ER2 >0.9| R2>0.97 is used in the code? Regards Mano

…

On Thu, 13 Jun, 2024, 22:24 Dongwon Lee, ***@***.***> wrote: Hi Mano, thank you for your interest in our TSIM method! For the r3 TOPMED reference panel, you will need to use a lower r^2 cutoff for hqSNP to have sufficient coverage. As noted in the README, we have found that 0.98 should be used instead of 0.99. In fact, you can still get a comparable accuracy with an r^2 cutoff as low as 0.97 based on our new calibration analysis. We are in the process of revising the manuscript to reflect this. Stay tuned! Dongwon — Reply to this email directly, view it on GitHub <#1 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACPEQJEO6GNSTI6LE5QXPYDZHH5WLAVCNFSM6AAAAABJJCQDY2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCNRWG4YDAMZZGA> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[Dongwon-Lee]

Currently we use "AND". Since ER2 is calculated based on the directly genotyped data, it is not affected by the changes in the reference panel. Therefore, the ER2 threshold should remain the same.

[Manonanthini]

Hi,Thank you for the clarification. As the ER2 threshold 0.9 is very high and you have used 'AND' with R2>0.97, a lot of directly genotyped SNPs will be deleted from the dataset if I understand correctly. Shall i use 0.7 or 0.8 as threshold for ER2 to include more directly genotyped SNPs into the dataset? Regards Manonanthini

…

On Fri, 14 Jun, 2024, 17:19 Dongwon Lee, ***@***.***> wrote: Currently we use "AND". Since ER2 is calculated based on the directly genotyped data, it is not affected by the changes in the reference panel. Therefore, the ER2 threshold should remain the same. — Reply to this email directly, view it on GitHub <#1 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACPEQJDOREQ6PAVDP52G42TZHMCYFAVCNFSM6AAAAABJJCQDY2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCNRYGI2TQMJZGM> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[EasperOne]

Hi, I just read the fantastic paper and try to implement it. I got the same question as @Manonanthini here regarding relaxed thresholds. Is there any update on this matter?

Best,
Khoa

[Dongwon-Lee]

Sorry for the late reply. It took much longer than expected due to the update of the TOPMED reference panel. First, I would like to clarify that TSIM uses imputed genotypes even for the direct genotyped SNPs. We found that most of the imputed genotypes for these SNPs have very high concordance (R2>0.98 for imputed genotypes vs. WES-derived genotypes), regardless of their ER2 values. This strongly suggests that a lower ER2 is indicative of genotyping error and that the imputation server can effectively correct for genotyping error when using a stringent R2 threshold. Although we found slightly lower concordance for the SNPs with ER2<0.8, our additional analysis suggests that these SNPs are still comparable to other imputed SNPs. Therefore, we concluded that an ER2 threshold was not necessary with the stringent R2 threshold.