Stroke genetics informs drug discovery and risk prediction across ancestries

Previous genome-wide association studies (GWASs) of stroke — the second leading cause of death worldwide — were conducted predominantly in populations of European ancestry1,2. Here, in cross-ancestry GWAS meta-analyses of 110,182 patients who have had a stroke (five ancestries, 33% non-European) and 1,503,898 control individuals, we identify association signals for stroke and its subtypes at 89 (61 new) independent loci: 60 in primary inverse-variance-weighted analyses and 29 in secondary meta-regression and multitrait analyses. On the basis of internal cross-ancestry validation and an independent follow-up in 89,084 additional cases of stroke (30% non-European) and 1,013,843 control individuals, 87% of the primary stroke risk loci and 60% of the secondary stroke risk loci were replicated (P < 0.05). Effect sizes were highly correlated across ancestries. Cross-ancestry fine-mapping, in silico mutagenesis analysis3, and transcriptome-wide and proteome-wide association analyses revealed putative causal genes (such as SH3PXD2A and FURIN) and variants (such as at GRK5 and NOS3). Using a three-pronged approach4, we provide genetic evidence for putative drug effects, highlighting F11, KLKB1, PROC, GP1BA, LAMC2 and VCAM1 as possible targets, with drugs already under investigation for stroke for F11 and PROC. A polygenic score integrating cross-ancestry and ancestry-specific stroke GWASs with vascular-risk factor GWASs (integrative polygenic scores) strongly predicted ischaemic stroke in populations of European, East Asian and African ancestry5. Stroke genetic risk scores were predictive of ischaemic stroke independent of clinical risk factors in 52,600 clinical-trial participants with cardiometabolic disease. Our results provide insights to inform biology, reveal potential drug targets and derive genetic risk prediction tools across ancestries.

(PMID: 34570226). The nuance here is important and should be appropriately addressed by the authors in their results/figures. 4) A substantial amount of the paper focuses on the multi-ancestry applicability of the PRS generated by the authors. While this is interesting, much of this data is more limited in novelty in light of the recent GLGC paper recently published in Nature (PMID: 34887591). 5) While the authors use SUSIE/finemapping techniques to identify putative causal variants, minimal effort is spent in an attempt to identify candidate causal genes at each locus for all identified loci. More recent GWAS manuscripts often integrate multiple lines of evidence (software like PoPs, colocalization with eQTL data in appropriate tissue(s), nearest gene, TWAS, protein altering variants in high LD with sentinel variants, etc) to interpret and identify what they believe is the likely causal gene acting at a given locus. In many cases this may be unclear, and can be stated. 6) This reviewer believes the interpretation of the FGA locus as a direct link to alteplase (a tissue plasminogen activator) is a bit of a reach. While FGA and a large component of thrombus (fibrin mesh) are certainly critical in this pharmacologic mechanism, the nuance of alteplase and its effect in plasminogen in thrombus degradation is completely overlooked. This should be mentioned in greater detail to the reader.
Minor concerns: 1) There appears to be a theme of highlighting novel loci with [] following a numerical statement of overall loci. This reviewer found this presentation distracting. 2) Single cell data from the pFC is briefly mentioned, but it is not clear to this reviewer what this information/data adds to improve the paper. This is particularly evident given how strong other sections of the manuscript are.
3) The PheWAS analysis is somewhat limited and could easily be expanded in light of the plethora of open source/access tools that can be used for this purpose (Phenoscanner, ieu open GWAS project, etc). 4) Figure 2b was somewhat blurry on the PDF version of the article this reviewer read. The resolution on this image could be enhanced 5) In reference to point #1, the discussion mentions that independent validation is available through the use of population studies/clinical trials. This reviewer believes this sentence is misleading and should be amended.

Author Rebuttals to Initial Comments:
Dear Dr. Trenkmann, On behalf of our co-authors we thank you very much for giving us the opportunity to revise our manuscript, 2021-12-19966: "Stroke genetics informs drug discovery and risk prediction across ancestries". We also thank you and the Referees for the helpful and constructive comments. We have addressed these point-by-point below and updated the manuscript accordingly (changes are highlighted in the text).
Major changes are summarized below: 1. The genomics-driven drug discovery section, which was mentioned to be one of the most attractive features of our study, has been further strengthened by the following additions. First, we re-ran the drug target enrichment analysis with GREP using the gene-based association results not only from MAGMA, but also from VEGAS2: all previously described genes were confirmed and two additional genes were prioritized. Second, we explored two additional independent pQTL resources (in plasma and cerebrospinal fluid, CSF) to strengthen the evidence for a causal association of drug target protein levels with stroke risk. All previously described pQTL associations replicated at FDR < 0.05, with consistent directionality for all except one gene (GP1BA), prompting caution for the latter. Third, for genes targeted by inhibitors, we examined associations of rare deleterious variants with stroke and related traits using whole-exome sequencing in >450,000 UK Biobank participants, and observed one significant protective association supporting the corresponding gene (F11) as a potential therapeutic target for inhibitors.
2. In order to address concerns about the confidence in reported loci, we gathered an independent dataset of 89,084 stroke patients (of which 85,546 ischemic strokes; 70.0% EUR, 15.6% AFR, 10.1% EAS, 4.1% HIS, and 0.1% SAS) and 1,013,843 controls, mostly from large biobanks, for follow-up of identified loci. We additionally sought to validate the 42 stroke risk loci reaching genome-wide significance in individual ancestries in at least one other ancestry group among the discovery samples. Moreover, we have now more clearly separated our primary analysis (IVW meta-analysis) from secondary analyses (MR-MEGA and MTAG). Overall, we provide independent validation of the vast majority of identified genome-wide significant associations and graded our loci by level of confidence based on these "internal" and "external" follow-up findings. Despite the notable size of the follow-up study sample, with nearly 90,000 additional stroke patients and the compelling follow-up results, the power to replicate low frequency variants, ancestry-specific assocations, and subtype-specific associations is limited. In regards to the latter, we note that most of the follow-up studies were derived from large biobanks with event ascertainment based on electronic health records, without suitable stroke subtype information. Loci that did not show evidence for replication (13% of our primary loci) were clearly labeled as "low confidence" pending additional follow-up in the future when additional, larger subtype-and ancestry-specific samples become available.
3. We further expanded the polygenic score (PGS) section of our manuscript by adding three types of complementary analyses. First, we harmonized the number of risk factors included in European and East-Asian integrative PGS (iPGS) and confirmed associations of these with ischemic stroke in additional, completely independent evaluation datasets (from the Million Veteran Program, MVP, for Europeans and Taiwan Biobank for East-Asians). Second, we also validated the European iPGS in a clinical trial setting. Third, we evaluated the predictive ability of the iPGS model in additional, unique African-American (MVP) and indigenous African (Nigerian and Ghanaian, SIREN) datasets. While African ancestry samples were too small to generate an African-specific stroke PGS, the European iPGS showed a significant association with ischemic stroke in both African-American and indigenous African (Nigerian and Ghanaian) participants, although expectedly weaker than in European participants. This is the first study assessing the performance of iPGS for ischemic stroke in East-Asian and African-ancestry populations. Moreover, the GIGASTROKE iPGS clearly outperforms an earlier iPGS derived from the smaller MEGASTROKE studies, in all ancestries.
4. We provide different lines of evidence highlighting the added value of the cross-ancestry approach taken in our manuscript (with one third of non-European samples in both discovery and follow-up studies). First, for loci derived from the IVW meta-analyses we show a clear gain in power beyond sample size increase, compared to the incremental addition of European ancestry samples. Second, we show that iPGS models derived from cross-ancestry stroke GWAS had a higher predictive ability than iPGS models derived from ancestry-specific stroke GWAS only, both in Europeans and East Asians.
A number of additional sensitivity analyses have been applied, such as controlling for ancestry specific LD score intercepts, removing overlapping samples from Mendelian randomization (MR) analyses, performing bidirectional MR and Steiger tests to rule our reverse causation, confirming PheWAS results obtained in Estonian Biobank using the Phenoscanner, etc. We have also expanded our description of the bioinformatics follow-up of identified stroke risk loci.
We hope that you will find this revised version suitable for publication in Nature.
Thank you very much for your consideration.
Best regards,

Point-by-point response to the editor and referees
We would specifically like to point out that referees have raised concerns about the confidence in the novel reported loci and we ask that you provide further supporting evidence (and replication as far as that is possible).
 We thank the editor and the referees for raising this point. In order to adress this point we gathered an independent dataset of 89,084 stroke patients (of which 85,546 ischemic strokes; 70.0% EUR, 15.6% AFR, 10.1% EAS, 4.1% HIS, and 0.1% SAS) and 1,013,843 controls, mostly from large biobanks, for external replication (Table A).
Following the referees' advice, we have now followed up genome-side significant stroke risk loci both internally and externally. First, we sought to replicate the 42 stroke risk loci reaching genomewide significance in individual ancestries in at least one other ancestry group among the discovery samples. Second, we aimed for replication in the additional independent dataset (Table A). The biobank setting did not allow suitable ischemic stroke subtype analyses. Based on these follow-up results we characterized the level of confidence of identified loci as follows: high confidence in case of significant internal 'cross-ancestry' and/or external replication after accounting for the number of loci tested, or nominally significant replication in both internal and external replication, or evidence of involvement in monogenic stroke; intermediate confidence in case of nominal significance in either internal 'inter-ancestry' or external replication but not both; and low confidence in the absence of formal replication.
Overall, out of the 60 loci reaching genome-wide significance in the main IVW GWAS meta-analysis, 52 (87%) replicated at p<0.05 with consistent direction, of which 37 (61.7%) with high confidence, and 15 with intermediate confidence (25%). The 8 loci that did not replicate were labeled as "low confidence". Four of these were ethnic specific and 3 were low frequency variants that were monomorphic in some ancestries, limiting our ablity for replication.
Within the secondary analyses (MR-MEGA and MTAG), none of the 3 MR-MEGA loci replicated, although one was borderline significant (Supplementary Table 16). Of the 26 MTAG loci, 18 (69%) replicated with AS or AIS at p<0.05, of which 9 (35%) with high confidence. Of the 8 MTAG loci that did not replicate, 7 showed a consistent directionality (borderline significant for one), and 4 were subtype-specific, limiting our ability for replication with AS or AIS.
While we have clearly labeled "low confidence" variants, we have not removed them from bioinformatics functional follow-up analyses. Indeed, we feel that despite the important worldwide effort that enabled to gather nearly 90,000 additional stroke cases, several issues still affect our ability to replicate some of the identified stroke risk loci:  limits of statistical power, considering a smaller sample size than in the discovery and the winner's curse phenomenon;  we cannot rule out some degree of misclassification in the follow-up samples that were, with two smaller exceptions, nearly exclusively derived from large biobanks with stroke ascertainment based on ICD codes only (Turnbull, Lancet Reg Health West Pac. 2022; Rannikmae, Neurology 2020), while a large proportion of stroke cases in the discovery were recruited and deeply phenotyped in a hospital-based setting;  a substantial proportion of genetic risk for stroke is subtype specific, which is not fully captured in the replication because of the limited availability of stroke subtype data In this revised manuscript we have: -included the replication results and grading in Supplementary Tables 14-16 -revised Figure 1 to provide information on confidence levels for each locus: -amended the text as follows:

Results (p. 16 last paragraph line 455-459 and p. 17 first paragraph, line 460-479): "Independent follow-up of GWAS signals
We followed up genome-side significant stroke risk loci both internally and externally. First, we sought to replicate the 42 stroke risk loci reaching genome-wide significance in individual ancestries in at least one other ancestry group among the discovery samples. We successfully replicated, in a consistent direction, 10 of these loci at p<1.19x10-3 (accounting for the number of loci tested), of which 7 were genome-wide significant in EUR,1 in EAS,and 2 in both EUR and EAS. Additional 15 loci showed nominal association (p<0.05) in at least one other ancestry (Supplementary Table 15). We also note that the referees have mentioned the drug target section as one of the most attractive features of the study and we encourage you to add further supporting evidence in this direction.
 We thank the editor for this comment. To further strengthen the drug target section, we performed the following complementary analyses: 2. We explored two additional independent pQTL resources to strengthen the evidence for a causal association of drug target protein levels with stroke risk. In the initial manuscript, we nominated six proteins with Mendelian randomization (MR) and colocalization evidence, VCAM1, F11, KLKB1, PROC, GP1BA, and MMP12, by leveraging a plasma protein quantitative trait loci (pQTL) dataset. For five of these proteins, we curated drugs targeting those proteins in a direction compatible with a beneficial therapeutic effect against stroke based on MR estimates: inhibitors for VCAM1, F11, KLKB1, and GP1BA and activators for PROC.
To add further support to these results, we conducted two types of analyses, (i) MR with additional two pQTL datasets and (ii) an investigation of protective associations between rare deleterious variants and stroke risk traits.
First, we conducted MR for the six proteins using an independent cerebrospinal fluid (CSF) and plasma pQTL dataset (Yang et al., Nature Neuroscience, 2021). We confirmed significant causal associations (false discovery rate [FDR] < 0.05) for KLKB1, PROC, F11, and MMP12; F11 and MMP12 were further supported by colocalization analyses. Of the four proteins, the causal associations of KLKB1 and PROC were confirmed using the CSF pQTL dataset. Notably, the directions of MR estimates were the same as the initial results for all four proteins, supporting the beneficial therapeutic effects of the drugs curated in this study. As we noted in the manuscript, F11 and KLKB1 are adjacent genes with a long-range linkage disequilibrium pattern and complex coregulation. Whereas the causal associations of KLKB1 on stroke subtypes were initially inferred using a trans-pQTL, this time, we confirmed these associations using a cis-pQTL and observed borderline colocalization for ischemic stroke (AIS). No genome-wide significant pQTL was available for VCAM1 and GPA1B in this smaller dataset.
We also conducted MR using one of the largest plasma pQTL studies (Ferkingstad et al., Nature Genetics, 2021). We observed significant associations (FDR < 0.05) for all six proteins with consistent directionality except GP1BA (for which both concordant and discordant directionality was observed). Colocalization was confirmed for F11, KLKB1, MMP12, and PROC, with directions of effect consistent with the initial results. Together, we confirmed causal associations for six proteins and colocalization for four proteins using independent pQTL datasets.
Next, we investigated protective associations between rare deleterious variants and stroke risk traits. As described in Backman et al. (Nature, 2021), genes for which deleterious variants are associated with lower disease risk are potential therapeutic targets for inhibitors. Since this approach does not use pQTL datasets, we can leverage it to independently validate the beneficial therapeutic effects of the inhibitors nominated in this study. For the four genes targeted by inhibitors, VCAM1, F11, KLKB1, and GP1BA, we examined the associations of rare deleterious variants (minor allele frequency < 0.01) with stroke and stroke-related traits using gene-based burden tests in the whole-exome sequencing data of >450,000 UK Biobank participants (Backman et al., Nature, 2021). We observed one significant association, where rare deleterious variants in F11 had protective associations with venous thromboembolism (OR=0.471 and p=2.46×10 -4 < 0.05 / 4 / 30; 4 and 30 represent the number of tested genes and traits, respectively). The direction of this association was concordant with that of MR estimates, supporting the therapeutic effect of F11 inhibitors on stroke.
The approach used here is different from the phenome-wide association study (PheWAS) in Estonian Biobank because we used rare deleterious variants not captured by genotyping arrays and focused on their associations with stroke-risk traits to validate the potential efficacy of inhibitors. The PheWAS observed no significant association of a cis-pQTL for F11 with non-stroke-related phenotypes. Together, the additional approach used in this revised version and the PheWAS indicate the potential efficacy and safety of inhibiting F11.
 The manuscript has been amended as follows: Results (p. 21, line 584-607): "Third, we used protein quantitative trait loci (pQTL) for 218 drugtarget proteins as instruments for MR and found evidence for causal associations of 9 plasma proteins with stroke risk (4 cis-pQTL, 6 trans-pQTL), of which 6 were supported by colocalization analyses, with no evidence for reverse causation using the Steiger test (PROC,VCAM1,F11,KLKB1,MMP12,and GP1BA,Supplementary   Stroke genetics informs drug discovery and risk prediction across ancestries Mishra et al.
In this manuscript, Mishra and colleagues conducted the largest GWAS of stroke and stroke subtypes to date in five ancestries, which discovered over 60 novel stroke risk loci. The authors performed a range of follow-up analyses to identify putative causal genes and possible drug targets, and improved polygenic prediction of stroke in European and East Asian populations. Overall this work represents a comprehensive set of analysis that advances our understanding of the genetics of stroke. I have a number of concerns/comments that maybe useful.
-Replication: A major limitation of this study in my opinion is the lack of replication of discovered risk loci. This raises concerns about the reliability of the risk loci especially those that were identified by borrowing power from other GWAS via MTAG. While it may be difficult to find additional independent and sufficiently large cohorts for a replication analysis, the authors could more comprehensively investigate the replication rate (at different levels of stringency) across populations, or in left-out cohorts. I think this direction is under-explored.
We would specifically like to point out that referees have raised concerns about the confidence in the novel reported loci and we ask that you provide further supporting evidence (and replication as far as that is possible).
 We thank the editor and the referees for raising this point. In order to adress this point we gathered an independent dataset of 89,084 stroke patients (of which 85,546 ischemic strokes; 70.0% EUR, 15.6% AFR, 10.1% EAS, 4.1% HIS, and 0.1% SAS) and 1,013,843 controls, mostly from large biobanks, for external replication (Table A).
Following the referees' advice, we have now followed up genome-side significant stroke risk loci both internally and externally. First, we sought to replicate the 42 stroke risk loci reaching genomewide significance in individual ancestries in at least one other ancestry group among the discovery samples. Second, we aimed for replication in the additional independent dataset (Table A). The biobank setting did not allow suitable ischemic stroke subtype analyses. Based on these follow-up results we characterized the level of confidence of identified loci as follows: high confidence in case of significant internal 'cross-ancestry' and/or external replication after accounting for the number of loci tested, or nominally significant replication in both internal and external replication, or evidence of involvement in monogenic stroke; intermediate confidence in case of nominal significance in either internal 'inter-ancestry' or external replication but not both; and low confidence in the absence of formal replication.
Overall, out of the 60 loci reaching genome-wide significance in the main IVW GWAS meta-analysis, 52 (87%) replicated at p<0.05 with consistent direction, of which 37 (61.7%) with high confidence, and 15 with intermediate confidence (25%). The 8 loci that did not replicate were labeled as "low confidence". Four of these were ethnic specific and 3 were low frequency variants that were monomorphic in some ancestries, limiting our ablity for replication.
Within the secondary analyses (MR-MEGA and MTAG), none of the 3 MR-MEGA loci replicated, although one was borderline significant (Supplementary Table 16). Of the 26 MTAG loci, 18 (69%) replicated with AS or AIS at p<0.05, of which 9 (35%) with high confidence. Of the 8 MTAG loci that did not replicate, 7 showed a consistent directionality (borderline significant for one), and 4 were subtype-specific, limiting our ability for replication with AS or AIS.
While we have clearly labeled "low confidence" variants, we have not removed them from bioinformatics functional follow-up analyses. Indeed, we feel that despite the important worldwide effort that enabled to gather nearly 90,000 additional stroke cases, several issues still affect our ability to replicate some of the identified stroke risk loci:  limits of statistical power, considering a smaller sample size than in the discovery and the winner's curse phenomenon;  we cannot rule out some degree of misclassification in the follow-up samples that were, with two smaller exceptions, nearly exclusively derived from large biobanks with stroke ascertainment based on ICD codes only (Turnbull, Lancet Reg Health West Pac. 2022; Rannikmae, Neurology 2020), while a large proportion of stroke cases in the discovery were recruited and deeply phenotyped in a hospital-based setting;  a substantial proportion of genetic risk for stroke is subtype specific, which is not fully captured in the replication because of the limited availability of stroke subtype data In this revised manuscript we have: -included the replication results and grading in Supplementary Tables 14-16 -revised Figure 1 to provide information on confidence levels for each locus: -amended the text as follows: -Contribution of population diversity: Although the authors put together a diverse dataset from 5 ancestries, the majority of the analyses were conducted in European and East Asian samples only. I think there is a missed opportunity to investigate the contributions of genomic diversity to the loci discovery and risk prediction. For example, it's unclear whether increased discovery power can be simply attributed to sample size increase, and it's unclear how the polygenic score performs in non-European non-Asian populations and whether increasing the diversity in the training dataset improves prediction across populations (including in European and Asian samples).
 We thank Referee 1 for these comments and questions.
 "it's unclear whether increased discovery power can be simply attributed to sample size increase" We have addressed this point through the following additions: 1-Evaluation of our best integrative polygenic scores (iPGS) in two African ancestry samples that were added with the revision: one large dataset of 107,343 African-American participants from the Million Veteran Program (MVP), of whom 2,227 developed an incident ischemic stroke, and one unique African sample from Nigeria and Ghana (SIREN) with 1,691 ischemic stroke cases and 1,743 controls. Both these cohorts have independent GWAS manuscripts on their unique datasets that are currently being finalized. In order not to interfere with these efforts, considering the challenge to collect samples from these highly underrepresented ancestry groups, we have included their data only in the polygenic score validation (MVP and SIREN) and for replication of our loci (MVP), but have neither meta-analyzed them with the rest of our discovery dataset nor generated African ancestry-specific PGS as we feel this reaches beyond the scope of the present manuscript and could undermine other important ongoing efforts under the aforementioned circumstances.
These important additions of PGS validation in African ancestry populations have been included in the following parts of the manuscript:

Methods (p. 12, line 335-337): "The iPGS model for Europeans was further evaluated in two external cohorts of European ancestry (MVP and pooled data of clinical trials) as well as in two studies of African-ancestry participants (MVP and SIREN)."
2-Incremental addition of non-European samples to the meta-analysis: We have now conducted a secondary analysis whereby we have incrementally added to the previous largest European stroke GWAS (from MEGASTROKE): (1) additional European samples; (2) additional East Asian samples; (3) additional samples of non-European and non-East-Asian ancestry. This analysis shows that the increase in identified risk loci for stroke, proportionately to the increase in sample size, is notably higher when adding East-Asian samples and even higher when adding African, Hispanic, and South-Asian ancestry cohorts.
This has been included in the following sections of the manuscript: Figure 2)."

Extended Data Fig. 2: Increase in power with increase in population diversity
The scatter plot shows the number of loci identified with incremental increase in sample size and population diversity. To address this comment we evaluated the predictive ability of two iPGS models. The first was derived from ancestry-specific GWAS summary data, and the second was constructed using crossancestry stroke GWAS summary data. Five stroke GWAS (AS, AIS, LAS, SVS, and CES) were incorporated in both iPGS models.
For Europeans, the two iPGS models were constructed using the EstBB training case-control dataset, and they were evaluated using the EstBB validation cohort. Compared to the iPGS model derived from European-specific stroke GWAS, the iPGS model derived from cross-ancestry stroke GWAS showed a larger improvement in C-index (∆C-index = 0.006 for cross-ancestry model vs. ∆C-index = 0.002 for European-specific model) and, although confidence intervals overlapped, a stronger association with the incidence of AIS events (HR = 1.14 for cross-ancestry model vs. HR = 1.07 for European-specific model). CI indicates confidence interval; GWAS, genome-wide association study; HR, hazard ratio. ∆Cindex means the improvement in C-index over a base model that includes age, sex, and top 5 genetic principal components.
For East Asians, the two iPGS models were derived using the BBJ training case-control dataset and they were assessed using the BBJ validation case-control dataset. Compared to the iPGS model derived from East Asian-specific stroke GWAS, the iPGS model derived from cross-ancestry stroke GWAS showed a larger improvement in AUC (∆AUC = 0.016 for cross-ancestry model vs. ∆AUC = 0.011 for East Asian-specific model) and a stronger association with the odds of AIS (OR = 1.29 for cross-ancestry model vs. OR = 1.24 for East Asian-specific model). ∆AUC means the improvement in AUC over a base model that includes age, sex, and top 5 genetic principal components.
These results show that iPGS models derived from cross-ancestry stroke GWAS consistently had a higher predictive ability than iPGS models derived from ancestry-specific stroke GWAS both in Europeans and East Asians. The greater ability of the cross-ancestry GWAS-based iPGS model may be simply due to a larger sample size of cross-ancestry GWAS compared to that of ancestry-specific GWAS. However, given that >66% of AIS cases included in cross-ancestry GWAS was European ancestry, a substantial difference in the predictive ability of the two iPGS models would be, at least partially, attributable to the increased diversity of derivation GWAS.

These updates have been reflected in the revised manuscript:
 Supplementary Table 48: "Predictive ability achieved by PGS models derived from crossancestry vs. ancestry-specific stroke GWAS"  Results (p. 24, line 675-677): "Of note, iPGS models derived from cross-ancestry stroke GWAS had a higher predictive ability than iPGS models derived from ancestry-specific stroke GWAS both in Europeans and East Asians (Supplementary Table 48)."  "it's unclear how the polygenic score performs in non-European non-Asian populations" We evaluated the predictive ability of the GIGASTROKE-based iPGS model for Europeans in two external non-European non-Asian datasets.
First, we used the Million Veteran Program (MVP) longitudinal cohort to estimate the performance for subjects of African-American ancestry. Survival models were estimated from data spanning the beginning of 2011 (ie start of MVP) to end of 2018 (date of the latest National Death Index [NDI] data). Participants were censored at event (AIS), administrative censoring (Jan 1st, 2019), or death. Participants who had an AIS prior to enrollment in MVP were excluded. Out of 107,343 African-American ancestry MVP participants at baseline, 2,227 developed an incident AIS during follow-up.
In parallel we also evaluated the predictive ability of the GIGASTROKE-based iPGS model in European MVP participants. Out of 403,489 European ancestry MVP participants at baseline, 8,392 developed an incident AIS during follow-up. For Europeans, our iPGS showed an evident association with AIS incidence (HR per 1 SD increase, 1.19; P = 6.9×10 -52 ) and a certain improvement in C-index (∆C-index = 0.010). For African-Americans, the GIGASTROKE-based iPGS model showed a significant association with AIS incidence, but the strength of the association was weaker than for Europeans (HR per 1 SD increase, 1.11; P = 1.8×10 -5 ). Also, the improvement in C-index was smaller in African-Americans (∆Cindex = 0.003) than in Europeans (∆C-index = 0.010). ∆C-index means the improvement in C-index over a base model that includes age, sex, and top 5 genetic principal components.
Second, we evaluated the predictive ability of the iPGS model in the indigenous African (Nigerian and Ghanaian) SIREN study, comprising 1,691 ischemic stroke cases and 1,743 controls. Our iPGS showed a significant association with the odds of AIS (OR per 1 SD increase, 1.09; P = 0.010) and an improvement in AUC (∆AUC = 0.007). ∆AUC means the improvement in AUC over a base model that includes age, sex, and top 5 genetic principal components.
These results revealed that our iPGS model, which was derived using the cross-ancestry GIGASTROKE summary data and data of EstBB participants of European ancestry, was significantly associated with AIS in populations of African-American and indigenous African (Nigerian and Ghanaian) ancestry. However, our iPGS showed stronger association and larger improvement in predictive ability for Europeans than for Africans. Future directions for filling the gap in the prediction power between ancestries are the inclusion of larger numbers of non-European subjects in the cross-ancestry stroke GWAS and construction of ancestry-specific PGS models for non-Europeans. Therefore, continuous efforts of international collaboration will be of great importance.
These updates have been reflected in the revised manuscript: -Presentation of significant loci: As there are population-specific GWAS and cross-ancestry metaanalysis for multiple stroke subtypes, it's sometimes difficult to figure out the number of overlapping loci between subtypes and how the total number of (novel) loci was calculated. For example, the authors reported 56 genome-wide significant loci in cross-ancestry meta-analysis (IVW and MR-MEGA). In the next paragraph, per-allele effect sizes were compared across 60 IVW meta-analysis loci. It's unclear where this number (60) came from. When exploring the overlap between stroke risk loci and vascular risk factors, the authors reported 57 of 88 loci had pleiotropic associations, but why not 89 as reported in the abstract? I'd suggest (i) double check all the reported numbers are consistent and explain how they were calculated; and (ii) find a better way to present the discovered loci and their relationships between populations and subtypes to improve of flow of the manuscript (additional figures/tables may be needed in supplementary).
 We thank the referee for these suggestions and apologies for the confusion. To navigate the loci better we have now complemented Supplementary Tables 4 (now Supplementary Tables 4A) with  Supplementary Tables 4B describing GWAS  ; and NOVEL refers to the loci that were not previously reported to be associated with stroke and any of its subtypes at genome-wide significance." To improve the flow of manuscript and tone down the MR-MEGA and MTAG results, we have separated the sections describing the 60 loci identified using the primary IVW meta-analysis from the section describing loci identified using secondary analyses (MR-MEGA and MTAG). Moreover, we verified all numbers and clarified the text wherever required.
We have rephrased the following sections of the manuscript for more clarity:  ) (medrxiv 2021). Using BBJ data alone (corresponding to a subset of the GIGASTROKE East-Asian sample), for both ischemic stroke itself and phenotypes with similar sample sizes like myocardial infarction, the authors identified i) a limited set of high PIP variants and ii) a low number of 95% credible sets overall.
In addition, the limited overlap of EUR and EAS credible sets was seen over multiple phenotypes, leading to the conclusion that high PIP variants are not overlapping across populations. In fact, across the three biobanks, of the 4,518 identified variant-trait pairs with high posterior probability (> 0.9) of causality, only 285 replicated across multiple populations, thus displaying, as in our study, a surprising lack of overlap among fine-mapping results from different ethnicities. Many but not all of the variants with high PIP in one population but not in the other two populations were explained by the fact they are rare or monomorphic in the other two populations.
Beyond situations of population-specific associations, as in the study by Kanai et al, we believe that the main reason for the limited overlap of credible sets between EUR and EAS populations is limited power due to low sample size, especially in stroke subtypes and in EAS populations, enabling use to fine-map only variants with large or at most moderate effect sizes. Nevertheless, we believe that our fine-mapping approach has enabled the identification of some high-confidence causal variants shared across populations, at an important stroke risk locus, some population-enriched fine-mapped variants in EUR, and allelic series of high-impact variants across populations at several loci. We have now added the following sentence in the second to last paragraph of the discussion: "Despite major efforts to enhance non-European contributions to GIGASTROKE we still had limited power for identifying shared causal variants through cross-ancestry fine-mapping."  "It's also unclear how to functionally interpret the findings from the MENTR analysis which seemed to pinpoint a range of tissues and cell types with no clear convergence of biology": In the submitted manuscript, we showed that the alternative G allele of rs12476527 (5'UTR of KCNK3, coding for K+ channel protein; increasing stroke risk) was predicted to increase KCNK3 expression in the kidney cortex tubule cells, despite no eQTL of this variant being reported. This is a lead SNP associated with pulse pressure (p=2.553×10 -21 ) and systolic blood pressure (p=5.892×10 -50 ) in the previous cross-ancestry study in >750,000 individuals (ref1), suggesting that the kidney involvement is plausible.
We also showed that three variants (rs12705390 at PIK3CG, rs2483262 at PRDM16, and rs2282978 at CDK6) were predicted to regulate the expression of a long non-coding RNA and enhancer RNAs. Unfortunately, currently no functions are available for these transcripts; however, many predictions were for many types of endothelial cells, as shown in Supplementary -Data release: it seems that only summary statistics for meta-analysis will be publicly available. I would encourage the authors to release summary statistics for all population-specific GWAS and GWAS for all stroke subtypes. Population-specific data will be hugely useful as the field continues to develop methods (e.g. cross-ancestry fine-mapping and polygenic prediction) that can integrate data from multiple populations for improved genomic discovery and prediction.
 We thank the referee for this suggestion. We will make summary statistics available for crossancestry and ancestry-specific meta-analyses and for stroke subtypes, as suggested by the referee.
The data availability statement has been clarified accordingly: "Summary statistics for the GWAS meta-analyses of stroke (cross-ancestry and ancestry-specific, for any stroke and stroke subtypes) will be deposited in a public repository and made available by the time of publication. All other data supporting the findings of this study are available either within the article, the supplementary information and supplementary data files, or from the corresponding authors upon reasonable request." -Two-sample Mendelian randomization: Is there sample overlap between the GIGASTROKE study and some vascular risk factor GWAS, thus violating the assumption of two-sample MR? For example, both GIGASTROKE and the blood pressure GWAS included the UK Biobank sample?
 We thank the referee for this question and we apologize for not having addressed this issue in the first version of the manuscript. Indeed, sample overlap in Mendelian randomization analyses can lead to winner's curse bias thus inflating the effect sizes of the selected instruments and introducing weak instrument bias. The magnitude of relevant bias in the obtained estimates has been found to be dependent on the magnitude of sample overlap and the strength of the instruments, but for binary traits, the obtained estimates are robust against weak instrument bias even in a one-sample setting, as long as the overlap is restricted to the control sample (Burgess et al, Genet Epidemiol. 2016 Nov; 40 (7): 597-608). In our analyses we used the largest available datasets for the exposures of interest, which indeed led to overlap with the GIGASTROKE European sample. However, the overlap is expected to lie predominantly at the level of controls, as the vast majority of studies contributing stroke patients in GIGASTROKE were not included in risk factor GWAS, except for population-based cohorts, in particular UK Biobank that contributes only a small subset of GIGASTROKE stroke patients (<3%). Due to the complexity of the meta-analyzed datasets, it was not possible to estimate the exact magnitude of overlap on the case sample, in order to provide a reliable estimate of the magnitude of the expected relative bias based on the simulations performed by Burgess et al. Notably, there is no overlap in the datasets used for selecting and weighing the instruments in the East Asian cohorts and the GIGASTROKE East Asian sample.
To fully address the Referee's comment, we have run GWAS analyses for all the exposures of interest within the UK Biobank cohort, after excluding the individuals that were included as stroke cases in GIGASTROKE. As previously recommended, we applied a three-sample Mendelian randomization approach, which is more robust to sample overlap, winner's curse bias, and weak instrument bias (https://www.medrxiv.org/content/10.1101/2021.06.28.21259622v1.full.pdf). Specifically, we (1) selected the instruments from the original GWAS meta-analyses for the exposure traits, (2) weighed them on the basis of the UKB GWASs excluding stroke cases from GIGASTROKE, and (3) tested them on the GIGASTROKE cohort.
Importantly, the Mendelian randomization estimates derived from this approach were remarkably consistent with the estimates derived from the original analyses, thus confirming the robustness of our results to any potential sample overlap between GIGASTROKE and the cohorts used in the GWASs for the exposure traits. This is demonstrated in the correlation plot presented below comparing the results from inverse-variance weighted Mendelian randomization analyses derived from the two approaches. Figure. Correlation plot comparing the effect estimates (betas) derived from inverse-variance weighted Mendelian randomization analyses using weights from the potentially overlapping GWASs for the exposure traits and using weighted from a non-overlapping UK Biobank sample (Pearson's r=0.96). -SuSiE: Please elaborate on how LD patterns were checked to identify false positives or use a more objective way to filter out false positive signals.

The results from this sensitivity analysis are presented in Supplementary
 We thank the reviewer for this question and apologize for not being clearer in describing our methodology. We illustrate our approach to identify false positives on one specific example: In EAS there was one locus for AS (SH2B3, chr12:111910219) that produced 8 credible sets, which could potentially indicate a false positive finding. To identify potential LD mismatches, we used the diagnostic vignette accompanied with SuSiE (https://cran.rproject.org/web/packages/susieR/vignettes/susierss_diagnostic.html) and compared the expected zscores vs. observed z-scores. Figure 1A below shows the SuSiE result before curation, variants are outlined by color according to the credible set. Figure 1B shows an outlier 12:112645401 which was in high LD with the lead variant (r 2 =0.782, z=7.29) but had observed z=0. 88, while the expected z-score was 6.69, flagging an inconsistency.
In the EAS summary statistics, this variant was missing from 67.25% of the samples (N_EAS:112645401=86,331, N_EAS_total=263,592), thus the power for this variant was greatly reduced, leading to statistical fluctuation. Hence, we removed this specific variant and reran the fine-mapping. The results and diagnostic plots are shown below in Figure 2. Only one CS was identified with consistent observed z scores vs expected z scores. This strategy was repeated for all loci. We now include a sentence in the Methods section: Methods (p. 6, line 149-152): "Fine-mapping results were checked for potential false-positive findings using a diagnostic procedure implemented in SuSiE. In short, we compared observed and expected z-score for each variant at a given locus and removed the variant if the difference between observed and expected z-score was too high after manual inspection." -PheWAS: Consider using PheCodes to reduce noise in ICD codes and combine information from diseases that are individually rare to increase statistical power.
 We thank the referee for this suggestion and we have now rerun the PheWAS analysis using phecodes that were generated with the PheWAS R-package (see Methods section). We see that all of the previously significant associations remained and are now even stronger (Supplementary Table  36). For the rs2289252 SNP we still detect strong association with pulmonary embolism and see significant association with phlebitis and thrombophlebitis. We also detect borderline significant association with acute pulmonary heart disease and pulmonary embolism and infarction (Extended Data Fig. 8).
We have edited the manuscript accordingly -iPGS analysis: The overall predictive performance of PGS looked fairly weak, requiring an extreme cutoff of 0.1% to reach an OR of 3. Is it possible to assess the prediction performance of iPGS using completely independent training and evaluation datasets? Currently training and evaluation sets were created from the same biobank, which increases the risk of overfitting to specific sample characteristics of the biobank and does not fully assess the generalizability of the iPGS. Why the European iPGS was constructed from 14 secondary traits but the East Asian iPGS was built from many more (37) traits? Would that be better to make a fair comparison between the two iPGS (i.e., using the same set of GWAS for training and evaluation)? It would be helpful to separate the contribution from stroke GWAS and secondary GWAS to the predictions.
 We thank the referee for these questions:  "Is it possible to assess the prediction performance of iPGS using completely independent training and evaluation datasets": In the initial submission our evaluation datasets were entirely distinct from the training datasets (and from the GIGASTROKE GWAS dataset), but the referee is referring to the fact that training and evaluation datasets stem from the same study (Estonian Biobank for Europeans and Biobank Japan for East-Asians). Following the referee's recommendation, we evaluated the predictive ability of the GIGASTROKE-based iPGS model for Europeans and East-Asians in completely independent evaluation datasets, stemming from entirely distinct studies. For comparison, a MEGASTROKE-based iPGS model constructed in a previous study (Abraham et al., 2019) was also assessed in the same independent evaluation datasets.
First, we evaluated the European iPGS models using European-ancestry participants of the Million Veteran Program (MVP) longitudinal cohort. Survival models were estimated from data spanning the beginning of 2011 (ie start of MVP) to end of 2018 (date of the latest National Death Index [NDI] data). Participants were censored at event (AIS), administrative censoring (Jan 1st, 2019), or death. Participants who had an AIS prior to enrollment in MVP were excluded. Out of 403,489 European ancestry participants at baseline, 8,392 had an incident AIS during follow-up. The GIGASTROKE-based iPGS model showed a significant association with the incidence of AIS (HR per 1 SD increase, 1.19; P = 6.9×10 -52 ) and a certain improvement in C-index (∆C-index = 0.010).
Compared to the MEGASTROKE-based iPGS model, the association was stronger and the improvement in C-index was larger. ∆C-index means the improvement in C-index over a base model that includes age, sex, and top 5 genetic principal components.
Second, we evaluated the European iPGS models using clinical trial data (across the spectrum of cardiometabolic disease, described in our initial submission) in European ancestry participants (51,288 European participants of whom 960 developed an incident AIS over 3 years follow-up).
The GIGASTROKE-based iPGS model was significantly associated with the incidence of AIS (HR per 1 SD increase, 1.19 [1.11-1.27]; P = 3.2×10 -7 ) and a certain improvement in C-index (∆C-index = 0.008). Compared to the MEGASTROKE-based iPGS model, the GIGASTROKE-based iPGS model showed a stronger association and a larger improvement in C-index. ∆C-index means the improvement in C-index over a base model that includes age, sex, and top 5 genetic principal components.
Thus very similar confirmatory results were obtained in the two completely independent evaluation datasets. These suggest an improved predictive performance of the GIGASTROKEbased iPGS model compared to the MEGASTROKE-based iPGS model.
Third, we evaluated the East-Asian iPGS models using East-Asian ancestry participants of the Taiwan BioBank (TBB) study (87,940 participants, of whom 1,391 developed an incident stroke). The GIGASTROKE-based iPGS model was significantly associated with the risk of AIS (OR per 1 SD increase, 1.18; P = 1.8×10 -9 ) and a certain improvement in AUC (∆AUC = 0.003). Again, compared to the MEGASTROKE-based iPGS model, the GIGASTROKE-based iPGS model showed a stronger association and a larger improvement in C-index.

 Methods (p.13, end of first paragraph line 347-348): "The iPGS model for East Asians was further evaluated in an external study of East Asian-ancestry (TBB)."
 "Why the European iPGS was constructed from 14 secondary traits but the East Asian iPGS was built from many more (37) traits": In the initial submission, we had constructed the European iPGS model using the same set of risk factor GWAS as in a previous study using an integrative PGS approach in Europeans (Abraham et al., 2019), in order to more specifically assess the added value of the new (GIGASTROKE) over the older (MEGASTROKE) GWAS dataset. At the same time, we had constructed the East Asian iPGS model by selecting stroke-related phenotypes from East Asian or cross-ancestry GWAS summary data collected in jenger (http://jenger.riken.jp/en/result) or pheweb.jp (https://pheweb.jp). As a result, the number of risk factor GWAS incorporated in iPGS models were different between Europeans and East Asians.
Following the referee's comment, in order to minimize the difference in the derivation methods of iPGS models between ancestries, we have now re-constructed European and East Asian iPGS models using the same set of 12 risk factor GWASs ( For Europeans, compared to the set of risk factor GWAS used in the previous study (Abraham et al., 2019), we could obtain GWAS summary data with larger sample sizes for several traits (i.e., AF, T2D, SBP, DBP, BMI, and height). We re-constructed the iPGS model using the EstBB training dataset, and subsequently, evaluated the predictive ability in the EstBB validation cohort (and, subsequently, as described above, in MVP and the clinical trial data). Compared to the previous version of iPGS model, the revised version of iPGS model showed slightly stronger association with the incidence of AIS in EstBB (HR per 1 SD increase, 1.26; P = 2.0×10 -15 ), with a slightly larger improvement in C-index (∆C-index = 0.027). This slight improvement was possibly due to the larger sample size of risk factor GWAS. ∆C-index means the improvement in C-index over a base model that includes age, sex, and top 5 genetic principal components.
For East Asians, we used jenger (http://jenger.riken.jp/en/result) as a resource of GWAS summary data for the selected 12 risk factor traits. The 12 GWAS were included in the previous set of 37 risk factor GWAS. We re-constructed East Asian iPGS model using the BBJ training casecontrol dataset, and subsequently, evaluated the predictive performance of the iPGS model using the BBJ validation case-control dataset (and, subsequently, as described above, in the TBB data). Compared to the previous version of iPGS model, the revised iPGS model showed similar strength of association with the odds of AIS (OR per 1 SD increase, 1.33; P = 9.9×10 -26 ) and similar improvement in AUC (∆AUC = 0.019).  Fig. 9)." Fig. 9)."

 Method (p.13, first paragraph line 344-347): "To derive the East-Asian iPGS model, we incorporated 5 ancestry-specific and 5 cross-ancestry stroke GWAS (AS, AIS, LAS, SVS, and CES) from the GIGASTROKE project, and 12 GWAS of vascular risk traits (Extended Data
 "It would be helpful to separate the contribution from stroke GWAS and secondary GWAS to the prediction": To separate the contribution of stroke GWAS and risk factor (secondary) GWAS, we compared the performance of iPGS models with and without risk factors GWAS.
The Table below shows the predictive ability of iPGS models for Europeans, which was evaluated in the EstBB validation cohort. The iPGS model derived from stroke GWAS only showed a significant association with the incidence of AIS (HR per 1 SD increase, 1.14; P = 1.6×10 -5 ) and a modest improvement in C-index (∆C-index = 0.006). By incorporating risk factor GWAS in addition to stroke GWAS, the iPGS model achieved stronger association with AIS incidence (HR per 1 SD increase, 1.26; P = 2.0×10 -15 ) and higher improvement in C-index (∆C-index = 0.027). CI indicates confidence interval; GWAS, genome-wide association study; HR, hazard ratio. ∆Cindex means the improvement in C-index over a base model that includes age, sex, and top 5 genetic principal components.
The following Table below shows the predictive ability of iPGS models for East Asians, which was evaluated in the BBJ validation case-control dataset. The iPGS model derived from stroke GWAS was associated with the odds of AIS (OR per 1 SD increase, 1.30; P = 3.1×10 -22 ) and achieved a substantial improvement in predictive ability (∆AUC = 0.016). By incorporating risk factor GWAS in addition to stroke GWAS, the iPGS model showed slightly stronger association with AIS (OR per 1 SD increase, 1.33; P = 9.9×10 -26 ) and slightly higher improvement in AUC (∆AUC = 0.019). ∆AUC means the improvement in AUC over a base model that includes age, sex, and top 5 genetic principal components.
These results suggested that the contribution of risk factor GWAS was important to improve the predictive ability of iPGS models both in Europeans and East Asians, but the degree of importance seemed to be different between the two ancestries.
These results can be found in Supplementary Tables 41 and 45.
 PGS in clinical setting: Is there any reason not to use the genome-wide PGS or iPGS constructed by P+T, LDpred or PRScs in this analysis?
In the clinical trial setting we had used a GRS approach based on independent genome-wide significant risk loci for stroke, as in a previous study using a similar approach based on an earlier stroke GWAS (Abraham et al., 2019), in order to more specifically assess the added value of the new (GIGASTROKE) over the older (MEGASTROKE) GWAS dataset. Moreover, we found it interesting and complementary to present both an elaborate iPGS and a pragmatic GRS-based approach. However, in this revised version we have now also used the clinical trial data for iPGS evaluation.
In this trial setting, the GIGASTROKE-based iPGS was significantly associated with the incidence of AIS (HR per 1 SD increase, 1.19; P = 3.17×10 -7 ) and a certain improvement in C-index (∆C-index = 0.008). Compared to the MEGASTROKE-based iPGS model, the GIGASTROKE-based iPGS model showed a stronger association and a larger improvement in C-index. Mishra et al. present a genome-wide association study (GWAS) of 110,000 cases of prevalent or incident stroke with extensive computational follow-up analyses. This sample is 40% larger than the previous largest study (MEGASTROKE), and importantly, includes more individuals from diverse ancestries. While the current study reveals incrementally more genomic loci, credible sets, genes, and pathways-which I would argue is mainly of interest to an expert statistical genetics audience-I consider the following findings of particular importance and expect them to appeal to readers across disciplines: -Dozens of potential drug candidates for stroke are highlighted in the study by combining stroke genetics with public transcriptomic and proteomic data. The large stroke GWAS sample size and innovative use of multiple drug discovery methods lead to compelling evidence for F11 and KLKB1 as drug targets, with the former likely having minimal off-target effects.
-Genetic risk scores from the cross-ancestry GWAS provide meaningful improvement in stroke incidence prediction in addition to standard clinical risk factors, for individuals with EUR ancestry and individuals with EAS ancestry. (Granted I am not qualified to assess whether the selected clinical risk factors used were appropriate) I believe the study is methodologically sound and the conclusions are largely supported by the evidence presented. There are a small number of statistical checks I expect the authors to perform to verify their results, as outlined in the "minor comments" section below.
My only issue with the current work is its presentation of newly discovered genomic loci. The authors acknowledge their study lacks a formal replication dataset-which is understandable for a GWAS of this size-but despite this limitation, they are not conservative with their definition of stroke risk loci and genes.
We would specifically like to point out that referees have raised concerns about the confidence in the novel reported loci and we ask that you provide further supporting evidence (and replication as far as that is possible).
 We thank the editor and the referees for raising this point. In order to adress this point we gathered an independent dataset of 89,084 stroke patients (of which 85,546 ischemic strokes; 70.0% EUR, 15.6% AFR, 10.1% EAS, 4.1% HIS, and 0.1% SAS) and 1,013,843 controls, mostly from large biobanks, for external replication (Table A).
Following the referees' advice, we have now followed up genome-side significant stroke risk loci both internally and externally. First, we sought to replicate the 42 stroke risk loci reaching genomewide significance in individual ancestries in at least one other ancestry group among the discovery samples. Second, we aimed for replication in the additional independent dataset ( Table A). The biobank setting did not allow suitable ischemic stroke subtype analyses. Based on these follow-up results we characterized the level of confidence of identified loci as follows: high confidence in case of significant internal 'cross-ancestry' and/or external replication after accounting for the number of loci tested, or nominally significant replication in both internal and external replication, or evidence of involvement in monogenic stroke; intermediate confidence in case of nominal significance in either internal 'inter-ancestry' or external replication but not both; and low confidence in the absence of formal replication.
Overall, out of the 60 loci reaching genome-wide significance in the main IVW GWAS meta-analysis, 52 (87%) replicated at p<0.05 with consistent direction, of which 37 (61.7%) with high confidence, and 15 with intermediate confidence (25%). The 8 loci that did not replicate were labeled as "low confidence". Four of these were ethnic specific and 3 were low frequency variants that were monomorphic in some ancestries, limiting our ablity for replication.
Within the secondary analyses (MR-MEGA and MTAG), none of the 3 MR-MEGA loci replicated, although one was borderline significant (Supplementary Table 16). Of the 26 MTAG loci, 18 (69%) replicated with AS or AIS at p<0.05, of which 9 (35%) with high confidence. Of the 8 MTAG loci that did not replicate, 7 showed a consistent directionality (borderline significant for one), and 4 were subtype-specific, limiting our ability for replication with AS or AIS.
While we have clearly labeled "low confidence" variants, we have not removed them from bioinformatics functional follow-up analyses. Indeed, we feel that despite the important worldwide effort that enabled to gather nearly 90,000 additional stroke cases, several issues still affect our ability to replicate some of the identified stroke risk loci:  limits of statistical power, considering a smaller sample size than in the discovery and the winner's curse phenomenon;  we cannot rule out some degree of misclassification in the follow-up samples that were, with two smaller exceptions, nearly exclusively derived from large biobanks with stroke ascertainment based on ICD codes only (Turnbull, Lancet Reg Health West Pac. 2022; Rannikmae, Neurology 2020), while a large proportion of stroke cases in the discovery were recruited and deeply phenotyped in a hospital-based setting;  a substantial proportion of genetic risk for stroke is subtype specific, which is not fully captured in the replication because of the limited availability of stroke subtype data  Main comments: The authors claim to identify 89 risk loci for stroke and stroke subtypes (61 novel) based on a genome-wide association threshold of 5x10-8.
However, the authors did not adjust any of the input GWAS for their LD-score regression intercept. While they are correct there is no systematic inflation across ancestries, the European GWAS has an inflated intercept of 1.10 and contributes the largest sample to the cross-ancestry meta-analysis. I would expect the standard errors of each ancestry-specific set of GWAS summary statistics to be adjusted for any inflation before meta-analysis to minimize false positive discoveries.
 We thank referee 2 for their suggestion. We have now additionally performed cross-ancestry GWAS meta-analyses accounting for the LD-score regression intercept observed in each ancestry specific GWAS meta-analysis. With the exception of two loci (THADA, p=6.40E-8, AS, and LPA, p=7.53E-8, LAS), all other loci remained genome-wide significant in our analyses. Of note, some loci that did not reach genome-wide significance in cross-ancestry meta-analyses controlling for LD score intercepts for the main phenotype were genome-wide significant in LD score intercept corrected cross-ancestry meta-analyses of other phenotypes and/or were also identified in ancestry-specific analyses (footnote of supplementary table 4). LPA is biologically relevant to stroke pathology, with variants in this gene previously associated with stroke risk.(Langsted, J Am Coll Cardiol 2019) Moreover, in our follow-up of genome-wide significant loci we also observed a strong association of the LPA locus and a nominally significant association of the THADA locus with AS and AIS in independent datasets.
The manuscript has been edited accordingly: Supplementary Table 4.

Results (p 15, line 423-429): To our knowledge, our results include the most comprehensive and largest description of stroke genetic risk variants to date in each of the five represented ancestries. In cross-ancestry meta-analyses 53 loci (51 loci after controlling for ancestry specific LD score intercepts) reached genome-wide significance (Supplementary Table 4),
Methods (p 2, line 30-34): "We applied the covariate adjusted LD score regression (cov-LDSC) method to ancestry-specific GWAS meta-analyses without GC correction to test for genomic inflation and to compute robust SNP-heritability estimates in admixed populations. 64 We conducted crossancestry GWAS meta-analyses without genomic correction and with correction of the LD score intercept for genomic inflation observed in individual ancestry-specific GWAS." Secondly, throughout the manuscript, the authors investigate three stroke subtypes (LAS, CES, SVS) and two stroke summary traits (AS, AIS), but do not adjust for multiple testing of these phenotypes. It is clear these stroke traits are correlated, but the authors do not quantify this correlation, and (from a statistical point of view) treat their five phenotypes as one. The authors should either 1) quantify the (in)dependence of their traits and adjust any significance thresholds for the number independent components (e.g. using genetic correlation and PCA) or 2) qualify statements regarding the number of discovered loci and genes. Conceptually, however, we are looking at subtypes of the same phenotypic entity. Moreover, as we have now introduced follow-up analyses (both internally and externally, in nearly 90,000 additional stroke cases), we propose to describe the level of confidence of identified risk loci primarily based on the level of replication.  AS and AIS (N=32,903 and 16,863) in longitudinal population-based cohort studies. For the meta-analysis combining both incident and prevalent stroke studies, a few incident stroke studies were removed, because they were already part of a meta-analysis of stroke GWAS used as an input of the overall meta-analysis (WHI, Hisayama, REGARDS, JHS)." Mendelian randomisation can be quite susceptible to horizontal pleiotropy and reverse causality. However, the authors do not test for reverse causality in their MR analysis of vascular risk factors on stroke traits. As they have GWAS summary statistics for all phenotypes, they should perform a bidirectional MR. This step is particularly important when assessing the causal role of BMI and WHR, where alternative pathways could explain the observed association with stroke phenotypes.
 We thank the Referee for this comment. We share the Referee's opinion that Mendelian randomization estimates may be influenced by reverse causality, especially when such large datasets are used for performing the analyses. As recommended, to minimize risk for reverse causality, we now performed: (i) the Steiger test for directionality in the original analyses (ii) bidirectional Mendelian randomization using the stroke phenotypes as exposure.
The Steiger test compares the variance of the exposure (r2) and the outcome explained by the variants included as instruments to the model. The Steiger test relies on the assumption that if an association derived from Mendelian randomization is due to reverse causality, then the r2 of the exposure explained from the instruments will be lower than that of the outcome (Hemani et al, PLoS Genet. 2017 Nov;13(11): e1007081). In our main analyses, for all of our exposure-outcome traits, the Steiger test confirmed that the directionality of the tested associations was correct and robust to potential reverse causality, both in Europeans and East-Asians (Supplementary Table 24).
As suggested by the reviewer we additionally performed reverse Mendelian randomization analyses (Supplementary Table 26). Using the same criteria for instrument selection as for our exposure traits, we ended up with 21, 23, 1, and 7 genetic instruments for Any stroke, Any ischemic stroke, Large artery stroke, and Cardioembolic stroke, respectively. No variants fulfilled our instrument criteria for Small vessel stroke. Furthermore, because of the substantially lower power in the East Asian GIGASTROKE sample, we performed this analysis only for the European sample. The reverse Mendelian randomization analyses indeed showed some associations between the stroke traits and the risk factors we used as exposures in our original analyses. However, out of the 17 significant associations, all but 1 failed the Steiger test (association between cardioembolic stroke and venous thromboembolism). This analysis showed however high heterogeneity (p=3.8x10-114), was not consistent across the different methods, and the Egger intercept was significant (p=0.009) implying directional pleiotropy towards a positive association.
All in all, our sensitivity analyses confirm that our Mendelian randomization results are robust to potential reverse causality.
These results from the Steiger test have been added to Supplementary We also re-ran the drug target enrichment analysis with GREP using the gene-based association results not only from MAGMA, but also from VEGAS2 (in response to a comment from Referee 1). All previously significant enrichments are maintained, and two additional genes were prioritized: F2 and TFPI, targets of Lepirudin and Dalteparin respectively, both involved in the coagulation process and used for treatment of recurrent thromboembolism.

in EUR, EAS, AFR, HIS, SAS specific GWAS to combine the respective ancestry-specific pathway association p-values."
Line 170 -the phrase "checking the LD pattern" does not make it entirely clear which criteria were used to justify removing specific credible sets.
 We thank the reviewer for this question and apologize for not being clearer in describing our methodology. We illustrate our approach to identify false positives on one specific example: In EAS there was one locus for AS (SH2B3, chr12:111910219) that produced 8 credible sets, which could potentially indicate a false positive finding. To identify potential LD mismatches, we used the diagnostic vignette accompanied with SuSiE (https://cran.rproject.org/web/packages/susieR/vignettes/susierss_diagnostic.html) and compared the expected zscores vs. observed z-scores. Figure 1A below shows the SuSiE result before curation, variants are outlined by color according to the credible set. Figure 1B shows an outlier 12:112645401 which was in high LD with the lead variant (r 2 =0.782, z=7.29) but had observed z=0.88 , while the expected z-score was 6.69, flagging an inconsistency. In the EAS summary statistics, this variant was missing from 67.25% of the samples (N_EAS:112645401=86,331, N_EAS_total=263,592), thus the power for this variant was greatly reduced, leading to statistical fluctuation. Hence, we removed this specific variant and reran the fine-mapping. The results and diagnostic plots are shown below in Figure 2. Only one CS was identified with consistent observed z scores vs expected z scores. This strategy was repeated for all loci. We now include a sentence in the Methods section: "Fine-mapping results were checked for potential false-positive findings using a diagnostic procedure implemented in SuSiE. In short, we compared observed and expected z-score for each variant at a given locus and removed the variant if the difference between observed and expected zscore was too high after manual inspection." Line 269 -Which population/source was used for R2 calculations of proxy instruments in the protein MR & colocalization analysis?
 We apologize for not having specified this before. We have used 1000G EUR. The methods section has been amended accordingly (p. 10, line 259-261): "For the lead variants of pQTL that were missing in the stroke GWAS summary statistics, the proxy variants with the largest R2 were used if the R2 was greater than 0.8 (1000G EUR)." Figure 2 -error bars are not defined (95% confidence interval or standard error?)  We apologize for the omission. Error bars correspond to 95% CI. This has been added to the legend of Figure 2.
The x-axis of Figure 4D is not defined.  We apologize for the omission. The x-axis corresponds to time from inclusion in the trial in days. This has been added to the figure and figure legend (now Figure 4e).
Supplementary  This has been corrected.
The LDSC label and some of the vascular trait labels in Extended Data Fig 3. are not properly visible in the PDF file.  This has been corrected.
Cathie LM Sudlow appears twice in the second tier author list.  This has been corrected.
The total participants in the Dutch PSI cohort is reported as 1375 in the Supplementary Appendix, but the number listed in Supplementary In this manuscript, Mishra et al perform a GWAS of stroke and stroke subtypes in over 110,000 cases and 1.5 million controls of diverse ancestries nearly doubling the sample size of the largest prior stroke GWAS (MEGASTROKE). The authors identify many novel stroke loci, and perform a series of downstream analyses including conditional analysis (GCTA COJO), multi trait analysis (MTAG), fine mapping, enrichment analysis, Mendelian randomization, and a PRS analysis. Most compelling to this referee was the drug discovery analysis, identifying a series of putative stroke drug targets combining multiple lines of evidence. The manuscript represents a landmark paper in the understanding of stroke genetics with substantial novelty in many of its features, and is well written. Most of the conclusions are supported by the data, with appropriate methodology, and the authors cite relevant prior work. However, this referee has several concerns (major/minor) outlined below.
Major concerns: 1) While this manuscript is a multi ancestry meta-analysis, no independent replication for novel genomic loci are provided. While the authors' efforts in aggregating as many cases as possible should be commended, at a minimum some form of internal replication would improve the manuscript. For example, as utilized in a BP GWAS from 2018 (PMID: 30224653), genome-wide significance plus a p value of 0.01 in 2 (roughly) evenly divided strata of participants would reduce concerns for false positive findings. These concerns are additionally important in light of the fact that multiple GWAS in distinct stroke subtypes (AIS, LAS, SVS) are performed without further correction for multiple testing.
We would specifically like to point out that referees have raised concerns about the confidence in the novel reported loci and we ask that you provide further supporting evidence (and replication as far as that is possible).
 We thank the editor and the referees for raising this point. In order to adress this point we gathered an independent dataset of 89,084 stroke patients (of which 85,546 ischemic strokes; 70.0% EUR, 15.6% AFR, 10.1% EAS, 4.1% HIS, and 0.1% SAS) and 1,013,843 controls, mostly from large biobanks, for external replication (Table A).
Following the referees' advice, we have now followed up genome-side significant stroke risk loci both internally and externally. First, we sought to replicate the 42 stroke risk loci reaching genomewide significance in individual ancestries in at least one other ancestry group among the discovery samples. Second, we aimed for replication in the additional independent dataset (Table A). The biobank setting did not allow suitable ischemic stroke subtype analyses. Based on these follow-up results we characterized the level of confidence of identified loci as follows: high confidence in case of significant internal 'cross-ancestry' and/or external replication after accounting for the number of loci tested, or nominally significant replication in both internal and external replication, or evidence of involvement in monogenic stroke; intermediate confidence in case of nominal significance in either internal 'inter-ancestry' or external replication but not both; and low confidence in the absence of formal replication.
Overall, out of the 60 loci reaching genome-wide significance in the main IVW GWAS meta-analysis, 52 (87%) replicated at p<0.05 with consistent direction, of which 37 (61.7%) with high confidence, and 15 with intermediate confidence (25%). The 8 loci that did not replicate were labeled as "low confidence". Four of these were ethnic specific and 3 were low frequency variants that were monomorphic in some ancestries, limiting our ablity for replication.
Within the secondary analyses (MR-MEGA and MTAG), none of the 3 MR-MEGA loci replicated, although one was borderline significant (Supplementary Table 16). Of the 26 MTAG loci, 18 (69%) replicated with AS or AIS at p<0.05, of which 9 (35%) with high confidence. Of the 8 MTAG loci that did not replicate, 7 showed a consistent directionality (borderline significant for one), and 4 were subtype-specific, limiting our ability for replication with AS or AIS.
While we have clearly labeled "low confidence" variants, we have not removed them from bioinformatics functional follow-up analyses. Indeed, we feel that despite the important worldwide effort that enabled to gather nearly 90,000 additional stroke cases, several issues still affect our ability to replicate some of the identified stroke risk loci:  limits of statistical power, considering a smaller sample size than in the discovery and the winner's curse phenomenon;  we cannot rule out some degree of misclassification in the follow-up samples that were, with two smaller exceptions, nearly exclusively derived from large biobanks with stroke ascertainment based on ICD codes only (Turnbull, Lancet Reg Health West Pac. 2022; Rannikmae, Neurology 2020), while a large proportion of stroke cases in the discovery were recruited and deeply phenotyped in a hospital-based setting;  a substantial proportion of genetic risk for stroke is subtype specific, which is not fully captured in the replication because of the limited availability of stroke subtype data In this revised manuscript we have: -included the replication results and grading in Supplementary Tables 14-16 -revised Figure 1 to provide information on confidence levels for each locus: -amended the text as follows: 2) From this referee's understanding, the authors identify a series of novel loci in their primary IVW analysis, and then perform MTAG to further identify loci with correlated and/or causal stroke traits. This MTAG analysis identifies a series of additional loci, which the authors then claim as novel stroke loci in the abstract/paper. In the opinion of this referee, this representation of the data is misleading. MTAG results should be interpreted as largely exploratory, and not be claimed as distinctly novel stroke loci. This is particularly true in light of the fact that there is likely substantial overlap among samples in these analyses. While MTAG accounts for sample overlap, these results should not be considered "novel stroke loci" with the same level of supportive evidence.

Results (p. 19, line 546-548):"Of note, Mendelian randomization analyses performed with binary exposures should be interpreted with caution due to the potential violations of the exclusion restriction assumption. 17 "
Throughout the manuscript and the figure/tables, we now interpret our findings derived from Mendelian randomization for binary exposure traits in terms of genetic liability to these traits. Accordingly, we have rephrased our results, as well as the table and figure legends to address this issue.
4) A substantial amount of the paper focuses on the multi-ancestry applicability of the PRS generated by the authors. While this is interesting, much of this data is more limited in novelty in light of the recent GLGC paper recently published in Nature (PMID: 34887591).
 We thank the referee for raising this point.
We would like to respectfully point out that our approach to the derivation of polygenic score (PGS) models was substantially different from recent genetics studies including the GLGC publication (PMID: 34887591) because we found that widely used state-of-the-art algorithms, such as P+T, LDpred and PRS-CS, generated PGS models with a limited ability to predict AIS events. Therefore, we sought to incorporate not only AIS GWAS, but also other stroke GWAS (AS, LAS, SVS, and CES) and GWAS of 12 stroke-related phenotypes ( In brief, we created tens of candidate PGS models using the state-of-the-art algorithms (P+T, LDpred, and PRS-CS) with varying parameters for each input GWAS summary data. Then, for each input GWAS, we selected the best PGS model that showed the highest predictive ability in terms of area under the curve (AUC) or C-index using a training dataset of AIS cases and controls. Subsequently, we combined the best PGS models derived from GWAS of various traits with a machine learning method, elastic net logistic regression. The elastic net regression model was referred to as integrative PGS (iPGS) model in the present study, and was trained using the training dataset. Finally, we evaluated the predictive ability of the iPGS model using independent validation datasets.
For Europeans, we constructed the iPGS model using the EstBB case-control dataset of European ancestry, and assessed the predictive ability of the iPGS model using the EstBB validation cohort of European ancestry. For comparison, we evaluated the predictive ability of a PGS model derived from AIS GWAS alone and an iPGS model derived from GIGASTROKE-based stroke GWAS (10 GWAS of AS, AIS, LAS, SVS, and CES for each of cross-ancestry and ancestry-specific). The full iPGS model was derived from 22 GWAS, including 10 GIGASTROKE GWAS and 12 GWAS of stroke-related phenotypes.
The PGS model derived from AIS GWAS alone showed a significant association with the incidence of AIS (HR per 1 SD increase, 1.08; P = 0.008) and a modest improvement in C-index (∆C-index = 0.003).
The iPGS model derived from 10 GIGASTROKE GWAS showed a stronger association (HR per 1 SD increase, 1.14; P = 1.6×10 -5 ) and larger improvement in C-index (∆C-index = 0.006). Moreover, the ful iPGS model derived from GIGASTROKE and stroke-related GWAS showed an even stronger association (HR per 1 SD increase, 1.26; P = 2.0×10 -15 ) and larger improvement in C-index (∆C-index = 0.027). For East Asians, we constructed the full iPGS model using the BBJ case-control training dataset of East Asian ancestry, and evaluated the predictive ability of the iPGS model using the BBJ case-control validation dataset of East Asian ancestry. We compared the predictive performance of the full iPGS model with that of a PGS model derived from AIS GWAS alone, and that of an iPGS model derived from GIGASTROKE GWAS.
The full iPGS model showed the strongest association with the odds of AIS (OR per 1 SD increase, 1.33; P = 9.9×10 -26 ) and the largest improvement in AUC (∆AUC = 0.019), followed by the iPGS model derived from GIGASTROKE GWAS and a PGS model derived from AIS GWAS alone, in a descending order. These European and East-Asian analyses consistently demonstrated that the iPGS models had a superior predictive ability compared to the standard PGS models derived from a single GWAS.
To our understanding, the GLGC publication (PMID: 34887591) reported the predictive performance of the PGS models derived from single GWAS, and the approaches to the derivation of PGS models were substantially different from the present study. Our approach enabled us to overcome the limited predictive ability of standard PGS by constructing iPGS models and applying this for the first time to a cross-ancestry setting. As such, we believe that the present study would have a particular novelty for stroke risk prediction. 5) While the authors use SUSIE/finemapping techniques to identify putative causal variants, minimal effort is spent in an attempt to identify candidate causal genes at each locus for all identified loci. More recent GWAS manuscripts often integrate multiple lines of evidence (software like PoPs, colocalization with eQTL data in appropriate tissue(s), nearest gene, TWAS, protein altering variants in high LD with sentinel variants, etc) to interpret and identify what they believe is the likely causal gene acting at a given locus. In many cases this may be unclear, and can be stated.
 To further elicit the most likely causal gene at each locus, and to expand on the extensive bioinformatics follow-up of stroke risk loci summarized in Supplementary Table 30, we performed PoPs for all 60 genome-wide significant loci in the IVW meta-analysis. We used the lead SNP and SNPs within ± 500kb for the analysis and selected the gene with the highest PoPs score based on MAGMA results for EUR and EAS separately. As an additional filtering step, we only considered genes for which the PoPS score was in the top 10% of PoPs scores over all genes in the respective analysis. Consistent results across EUR and EAS are reported.
The manuscript has been amended as follows, to summarize more clearly the numerous bioinformatics follow-up analyses we conducted: 6) This referee believes the interpretation of the FGA locus as a direct link to alteplase (a tissue plasminogen activator) is a bit of a reach. While FGA and a large component of thrombus (fibrin mesh) are certainly critical in this pharmacologic mechanism, the nuance of alteplase and its effect in plasminogen in thrombus degradation is completely overlooked. This should be mentioned in greater detail to the reader.

Results
 We thank the reviewer for raising this point and agree that the link between the FGA locus and alteplase (a tissue plasminogen activator) is via the conversion of plasminogen into plasmin, which in turn degrades fibrin, rather than a direct action of alteplase on FGA and that this link should be made more clear. Accordingly, we now expanded on the Results as follows Results (page 20, lines 572-574): "This encompasses the previously described PDE3A and FGA genes, 24 encoding targets for cilostazol (antiplatelet agent) and alteplase (thrombolytic drug acting via plasminogen 25 )," We would like to avoid going into more mechanistic detail as the discovery of FGA as a drug target for alteplase and other thrombolytic agents has previously been reported (Malik, Nat Genet 2018) and because of limitations in space. However, we have added the respective reference from DRUGBANK, which provides more background on the underlying mechanism (https://go.drugbank.com/drugs/DB00009). We would be prepared to expand on this further if the reviewer and editors feel this is required.
Minor concerns: 1) There appears to be a theme of highlighting novel loci with [] following a numerical statement of overall loci. This referee found this presentation distracting.
 This has been edited for more clarity The Phenoscanner results helped to further confirm that we detect no significant association with nonstroke-related phenotypes for the F11 cis-pQTL rs2289252 and confirmed the seen significant associations with venous thromboembolic disorders.
We have edited the manuscript accordingly: Results (p. 21, line 607-611 and p. 22, line 612-621): "To further validate the candidate drugs and estimate their potential side effects, we investigated whether the drug-target genes were associated with stroke-related phenotypes using a phenome-wide association study (PheWAS) approach. 34