HIV Acquisition Meta-Analysis

[jaamarks]

Performing meta-analyses HIV Acquisition.

The parent issue is GitHub Issue 97.

Analysis Description

hiv-acquisition-gwas-meta-analysis-description.xlsx

[jaamarks]

We recently added WIHS3 to the latest iteration of metas. These can be found on S3 at:
s3://rti-hiv/meta_new/{023–027}

[jaamarks]

WIHS3_AA GWAS (N=2,009)

[jaamarks]

023: cross-ancestry (N=12,617)

meta015 + WIHS3_AA:

• UHS1–4_EA (N=3,013)
• VIDUS_EA (N=931)
• WIHS1_EA (N=720)
• UHS1–4_AA (N=4,015)
• WIHS1_AA (N=2,009)
• WIHS2_AA (N=844)
• WIHS3_AA (N=729)
• WIHS1_HA (N=356)

Previous	Current

[jaamarks]

024: AFR (N=7,597)

meta017 + WIHS3_AFR:

• UHS1–4_AFR (N=4,015)
• WIHS1_AFR (N=2,009)
• WIHS2_AFR (N=844)
• WIHS3_AFR (N=729)

Previous	Current

[jaamarks]

025: cross-ancestry (N=12,261)

meta018 + WIHS3_AA:

• UHS1–4_EA (N=3,013)
• VIDUS_EA (N=931)
• WIHS1_EA (N=720)
• UHS1–4_AA (N=4,015)
• WIHS1_AA (N=2,009)
• WIHS2_AA (N=844)
• WIHS3_AA (N=729)

Previous	Current

[jaamarks]

026: cross-ancestry (N=26,198)

meta021 + WIHS3_AA:

• McLaren_EA (N=13,581) (GC applied)
• UHS1–4_EA (N=3,013)
• VIDUS_EA (N=931)
• WIHS1_EA (N=720)
• UHS1–4_AA (N=4,015)
• WIHS1_AA (N=2,009)
• WIHS2_AA (N=844)
• WIHS3_AA (N=729)
• WIHS1_HA (N=356)

Previous	Current

[jaamarks]

027: cross-ancestry (N=25,842)

meta026 - WIHS1_HA:

• McLaren_EA (N=13,581) (GC applied)
• UHS1–4_EA (N=3,013)
• VIDUS_EA (N=931)
• WIHS1_EA (N=720)
• UHS1–4_AA (N=4,015)
• WIHS1_AA (N=2,009)
• WIHS2_AA (N=844)
• WIHS3_AA (N=729)

[jaamarks]

Rerun using cohort-level data

We reran metas 023–027 except we used the individual cohort-level GWAS summary stats instead of the previous meta-analyses.

023 AFR+AMR+EUR (N=12,617)

~ meta015 + WIHS3_AA
(see s3://rti-hiv/meta_new/015/results/figures/ for previous results)

• UHS1–4_EA (N=3,013)
• VIDUS_EA (N=931)
• WIHS1_EA (N=720)
• UHS1–4_AA (N=4,015)
• WIHS1_AA (N=2,009)
• WIHS2_AA (N=844)
• WIHS3_AA (N=729)
• WIHS1_HA (N=356)

024 AFR-specific (N=7,597)

~meta017 + WIHS3_AA
(see s3://rti-hiv/meta_new/017/results/figures/ for previous results)

• UHS1–4_AA (N=4,015)
• WIHS1_AA (N=2,009)
• WIHS2_AA (N=844)
• WIHS3_AA (N=729)

025 AFR+EUR (N=12,261)

~meta018 + WIHS3_AA
(see s3://rti-hiv/meta_new/018/results/figures for previous results)

• UHS1–4_EA (N=3,013)
• VIDUS_EA (N=931)
• WIHS1_EA (N=720)
• UHS1–4_AA (N=4,015)
• WIHS1_AA (N=2,009)
• WIHS2_AA (N=844)
• WIHS3_AA (N=729)

026 AFR+AMR+EUR (N=26,198)

~meta021 + WIHS3_AA
(see s3://rti-hiv/meta_new/021/results/v03/figures for previous results)

• McLaren_EA (N=13,581) (GC applied)
• UHS1–4_EA (N=3,013)
• VIDUS_EA (N=931)
• WIHS1_EA (N=720)
• UHS1–4_AA (N=4,015)
• WIHS1_AA (N=2,009)
• WIHS2_AA (N=844)
• WIHS3_AA (N=729)
• WIHS1_HA (N=356)

027 AFR+EUR (N=25,842)

~meta022 + WIHS3_AA
(see s3://rti-hiv/meta_new/022/results/v05/figures/ for previous results)

• McLaren_EA (N=13,581) (GC applied)
• UHS1–4_EA (N=3,013)
• VIDUS_EA (N=931)
• WIHS1_EA (N=720)
• UHS1–4_AA (N=4,015)
• WIHS1_AA (N=2,009)
• WIHS2_AA (N=844)
• WIHS3_AA (N=729)

These results have been stored on S3 at:
s3://rti-hiv/meta_new/{023–027}/results/v02/

[jaamarks]

HIV Acquisition Meta-Analysis with TOPMed Imputed Data

0028 AFR+AMR+EUR (N=12,617)

• UHS1–4_EUR (N=3,013)
• UHS1–4_AFR (N=4,015)
• VIDUS_EUR (N=923)
• WIHS1_EUR (N=720)
• WIHS1_AMR (N=356)
• WIHS1_AFR (N=2,009)
• WIHS2_AFR (N=844)
• WIHS3_AFR (N=729)
  
  

1000g_p3 results (meta 0023)

TOPMed results (meta 0028)

no GC

GC applied to each

[jaamarks]

Troubleshooting

• Replicate previous 1000g_p3 results

details

• I reran the previous 1000g_p3 meta-023 and received consistent results. The only caveat is that I noticed that the previous set included WIHS2_EUR, which it shouldn't have. Therefore the results are nearly the same, but not exactly. However, the overall goal was to see if we could replicate the previous results to make sure that our pipeline was not defective. We showed that the pipeline is not defective and therefore we can continue our troubleshooting. To be most consistent, we should run each 1000g_p3 cohort through the new GWAS pipeline.
  

• Apply genomic control to all cohorts.

details

This did not tame the inflation.

• Confirm correct GWAS results were being used.

details

I redownloaded each set of TOPMed GWAS results from S3 to double check that we were using the correct results. This checks out.

Also, confirm GWAS results in the HIV bucket were correctly copied from the associated Cromwell output.When each GWAS is complete, the results are copied over from the Cromwell bucket to the appropriate project bucket in the S3 organizational structure. We should verify that the GWAS results in the HIV bucket were copied over correctly. This checks out.

• Verify meta-analysis pipeline is correct by replicating old results

details

We needed to verify that the meta-analysis pipeline (bash script) we have been using is correct. We attempted to replicate the previous 1000g_p3 HIV acquisition GWAS meta-analysis results. The results were consistent, there for the pipeline is correct.

• Try stricter thresholds (rsq>0.8 and maf>0.05)

details

During our lab meeting on Wednesday (9/23/2020) we discussed applying a stricter rsq threshold. It was suggested that since we haven't had much experience with TOPMed imputed data, the same filters we apply to 1000g_p3 data might not be directly applicable. We will therefore experiment with applying a stricter rsq threshold. In particular, we will use the GWAS results with rsq > 0.8 applied.

• rsq 0.3 & MAF 0.01, without GC: lambda=1.243

• rsq 0.3 & MAF 0.01, with GC: lambda=1.203

• rsq 0.8 & MAF 0.01, without GC: lambda=1.242

• rsq 0.8 & MAF 0.01, with GC: lambda=1.204

• rsq 0.8 & MAF 0.05, without GC: lambda=1.238

• rsq 0.8 & MAF 0.05, with GC: lambda=1.20

• Apply genomic control twice

details

We recommend applying genomic control correction to all input files that include genomewide data and, in addition, to the meta-analysis results. (METAL documentation).

• Inspect gc parameter for each chromosome across each cohort

details

	chr1	chr2	chr3	chr4	chr5	chr6	chr7	chr8	chr9	chr10	chr11	chr12	chr13	chr14	chr15	chr16	chr17	chr18	chr19	chr20	chr21	chr22	chr23
UHS1–4 AFR	1.071	1.023	1.048	1.023	0.995	1.002	1.041	1.047	1.043	1.071	1.045	0.974	0.97	1.002	1.033	1.025	1.06	1.054	1.173	1.031	1.012	1.019	1.085
UHS1–4 EUR	1.038	1.108	1.039	1.021	1.094	1.165	1.089	1.04	1.084	1.054	1.09	1.033	1.072	1.056	1.093	1.017	1.011	1.099	1.151	1.186	1.034	1.108	1.02
VIDUS EUR	1.068	0.97	1.028	1.026	0.962	1.027	1.028	1.009	0.963	0.983	1.083	1.133	1.054	1.048	1.126	1.026	1.107	1.001	0.862	1.023	1.14	0.938	0.952
WIHS1 AFR	1.002	0.943	1.032	0.987	0.958	0.988	0.964	1.007	0.981	1.009	0.992	0.975	1.078	0.975	1.078	0.948	0.979	0.933	0.985	0.98	1	0.973	1.006
WIHS1 AMR	1.039	1.065	1.042	1.021	1.026	1.038	1.014	0.994	1.005	0.986	1.049	1.029	1.046	1.015	0.989	1.078	1.073	1.024	0.981	0.998	0.976	1.112	1.201
WIHS1 EUR	0.975	1.047	1.042	1.081	0.963	1.03	1.071	0.95	0.986	0.974	0.989	1.014	1.127	0.962	1.032	0.973	1.039	0.999	0.943	1.029	0.881	1.036	1
WIHS2 AFR	1.017	1.033	1.031	1.001	1.045	0.965	0.99	1.031	0.977	1.038	1.028	1.072	1.116	1.041	1.006	0.983	0.978	1.026	0.985	0.978	0.924	1.022	1.115
WIHS3 AFR	1.012	1.05	0.981	0.996	1.02	0.97	0.969	1.024	0.969	1.032	1.066	1.04	1.116	1.001	0.979	0.98	0.978	1.007	0.999	1.005	0.983	0.999	1.117

After GC was applied to each cohort/chromosome.

	chr1	chr2	chr3	chr4	chr5	chr6	chr7	chr8	chr9	chr10	chr11	chr12	chr13	chr14	chr15	chr16	chr17	chr18	chr19	chr20	chr21	chr22	chr23
GC	1.193	1.215	1.22	1.224	1.187	1.171	1.198	1.194	1.176	1.167	1.131	1.204	1.183	1.193	1.232	1.189	1.234	1.19	1.329	1.21	1.16	1.214	1.272

• QQ plot for each chromosome

details

No GC

	chr1	chr2	chr3	chr4	chr5	chr6	chr7	chr8	chr9	chr10	chr11	chr12	chr13	chr14	chr15	chr16	chr17	chr18	chr19	chr20	chr21	chr22	chr23
lambda	1.229	1.259	1.255	1.238	1.217	1.198	1.227	1.229	1.198	1.211	1.18	1.249	1.277	1.216	1.278	1.201	1.269	1.225	1.392	1.248	1.182	1.243	1.373

qq plots

GC applied to each cohort

	chr1	chr2	chr3	chr4	chr5	chr6	chr7	chr8	chr9	chr10	chr11	chr12	chr13	chr14	chr15	chr16	chr17	chr18	chr19	chr20	chr21	chr22	chr23
lambda	1.194	1.214	1.22	1.223	1.187	1.171	1.198	1.195	1.176	1.167	1.131	1.204	1.183	1.194	1.232	1.189	1.235	1.189	1.329	1.21	1.16	1.214	1.272

qq plots

GC applied twice

GC applied to each cohort and then again to the meta-analysis results.

qq plots

• Perform meta on subset of cohorts

details

Investigate whether any one cohort is disproportionately contributing to the inflation.

• Leave-one-out approach that @danahancock suggested during our lab meeting.
  

	uhs_afr	uhs_eur	vidus_eur	wihs1_afr	wihs1_amr	wihs1_eur	wihs2_afr	wihs3_afr
lambda	1.32	1.244	1.25	1.235	1.297	1.239	1.053	1.056

UHS_AFR gwas plots

Leave UHS1–4 AFR out.

UHS_EUR qq-plot

Leave UHS1–4 EUR out.

VIDUS_EUR qq-plot

Leave VIDUS EUR out.

WIHS1_AFR qq-plot

Leave WIHS1 AFR out.

WIHS1_AMR qq-plot

Leave WIHS1 AMR out.

WIHS1_EUR qq-plot

Leave WIHS1 EUR out.

WIHS2_AFR qq-plot

Leave WIHS2 AFR out.

WIHS3_AFR qq-plot

Leave WIHS3 AFR out.

• WIHS2&3 results check

details

Recreate QQ and Manhattan plots to see if we can replicate what came out of the workflow. This checks out.

	workflow	meta
WIHS2
WIHS3

[jaamarks]

TL;DR

(1) For the 1000g_p3 GWAS, all the WIHS3 subjects were also included within WIHS2.

(2) Including WIHS3 in the 1000g_p3 meta did inflate the results some, but it was not as pronounced as this current TOPMed meta.

(3) ProbABEL was used to run the 1000g_p3 GWAS for both WIHS2 and WIHS3 (as opposed to RVTEST which we used for the TOPMed GWAS). The individual GWAS results were deflated for both when using ProbABEL. This could explain why adding WIHS3 to the 1000g_p3 meta did not inflate the results as much as it did in the TOPMed meta. The inflation is most evident in the AFR-specific results (go figure).

All of WIHS3 was in WIHS2 for 1000g_p3

When comparing the phenotype files for WIHS2 and WIHS3 from the original 1000g_p3 results, we see that every subject in WIHS3 was within WIHS2.

• WIHS3 1000g_p3 phenotype: s3://rti-hiv/gwas/wihs3/data/acquisition/0001/archive/phenotype/0001/final/wihs3_aa_hiv_age_sex_PC3+PC8+PC2.txt
• WIHS2 1000g_p3 phenotype: s3://rti-hiv/gwas/wihs2/data/acquisition/archive/pheno/WIHS2 HIV status GWAS baseline.csv

command line

aws s3 cp s3://rti-hiv/gwas/wihs3/data/acquisition/0001/archive/phenotype/0001/final/wihs3_aa_hiv_age_sex_PC3+PC8+PC2.txt .
aws s3 cp s3://rti-hiv/gwas/wihs2/data/acquisition/archive/pheno/WIHS2 HIV status GWAS baseline.csv .

cut -d" " -f1 wihs3_aa_hiv_age_sex_PC3+PC8+PC2.txt   | tail -n +2 |\
  awk -F"_" '{print $2}' | xargs -I{} grep {} WIHS2\ HIV\ status\ GWAS\ baseline.csv

Inflation observed when adding WIHS3 to 1000g_p3 meta

We do observed some inflation when adding WIHS3 to the 1000g_p3 results. The lambda values do not increase as markedly as with the TOPMed imputed results, but one can observed, from visual inspection, a clear shift above the line y=x. The most appreciable inflation is observed within the AFR-specific meta-analysis. Click below to expand the 1000g_p3 vs TOPMed plots for each meta.

It should also be noted that the TOPMed imputed data has more coverage. In particular, when comparing the final SNP count of the cross-ancestry meta-analysis (AFR+AMR+EUR) we see that the TOPMed results have over a million more observations.

• 1000g_p3: N SNPs = 16,841,575
• TOPMed: N SNPs = 18,033,320
  
  

AFR+AMR+EUR

	without WIHS3	with WIHS3
manhattan
qq

AFR+EUR

	without WIHS3	with WIHS3
manhattan
qq

AFR-specific

	without WIHS3	with WIHS3
manhattan
qq

ProbABEL GWAS results were deflated

WIHS2 and WIHS3 1000g_p3 were both ran using the ProbABEL GWAS software. This is apparent from looking at the summary stats headers and also the file names (e.g. palogist in the name which is a ProbABEL term for logistic regression). The lambda values for WIHS2 and WIHS3 were 1.017 and 1.01, respectively. Though these values do not raise any red flags, we do observe in both the Manhattan plots and the QQ plots below (click to expand) there are some striking similarities. For example, it is evident that the points in the QQ plots for both WIHS2 and WIHS3 are systematically shifted below the line y=x. This characteristic of being deflated could explain why the 1000g_p3 metas were not as markedly inflated as we observed within the TOPMed meta-analysis results.

1000g_p3 ProbABEL GWAS plots

plot locations:

• WIHS3 (N=729)
  s3://rti-hiv/gwas/wihs3/results/acquisition/0002/afr/archive/1df/aa/final/wihs3.aa.1df.1000G.hiv_acq.maf_gt_0.01.rsq_gt_0.3.assoc.plot.all_chr.snps+indels.*.png
  
• WIHS2 (N=844) s3://rti-hiv/gwas/wihs2/results/archive/wihs2.aa.palogist.status~snp+ageatbl+EV.plots.snps+indels.*.png.gz
  
  

	QQ	Manhattan
WIHS2
WIHS3