This article reports the release of a new version of the PRSice software for polygenic score calculation. The new version of the software boasts speed enhancements that make it appealing for applications in the growing number of ultra-large genetically-informed datasets including the UK Biobank, 23andMe and others. Also important are features allowing for polygenic score computation from imputed genotype datasets in which genotypes are represented as a probabilities rather than discrete allele counts.

The data on speed are compelling. This alone is a good argument for why PRSice v1 users should upgrade to v2. But I found the article thinner on two other key questions central to addressing whether those not already using PRSice v2 should take up PRSice v2: (1) Does the polygenic scoring method implemented within PRSice2 (additive combination of SNPs with/without LD clumping) deliver comparably predictive scores to other software, e.g. the LDPred and lassosum softwares? (2) What is the value added of being able to accommodate imputed genotype probabilities rather than relying exclusively on discrete allele count data?

I would suggest the following revisions:

Re PRSice2 vs. Alternative Softwares: The authors assert that the method of polygenic score calculation implemented within PRSice2 generates scores that are comparably predictive to two other methodologies, LDPred and LassoSum. It is my understanding that these methods were developed and are in use precisely because they outperform the method implemented in PRSice in terms of the prediction R-squared for the target phenotype. It would improve the article if the authors could provide some empirical evidence for the claim that their software delivers polygenic scores of comparable accuracy to other methods. For example, comparison of PRSice2 scores to scores generated from LDPred and lassosum for a set of traits would be helpful. I like the choices of height and BMI. But it might also be sensible to consider a trait for which existing GWAS are smaller/ polygenic predictions are less accurate, e.g. depression.

Re Imputed Genotype Probabilities vs. Allele Counts: The authors helpfully report that PRSice2 scores computed with imputed data can improve prediction accuracy by about 1 percentage point for height and BMI as compared to scores computed with genotyped-only data. It would be helpful to add an element to this analysis. As I understand it, the authors are comparing a genotyped-SNP-only polygenic score computed from allele counts to an imputed-SNP polygenic score computed from genotype probabilities. But these are not the only two possibilities. In much polygenic score analysis, imputed SNP probabilities are converted to discrete genotypes using a threshold (e.g. probability=0.9) to determine whether a given genotype can be assigned to the SNP. Since this is common practice in the field, it seems to me that it would be helpful to include this approach in the comparison.

Finally, I have one small quibble about language:

In the introduction, the authors assert that polygenic scores have proven clinical utility. This is a bit of an overstatement. I think we can say that "provocative new data suggest the potential for polygenic scores to be useful in clinical settings" or something similar. The recent papers referenced by the authors are compelling. But the term clinical utility has a specific meaning - that application of a tool improves patient outcomes (e.g. see Torkamani et al. 2018 Nat Rev Genet). We are a long way off from that. Instead, the evidence we have supports an argument for the clinical validity of extreme polygenic-scores values for assessing disease risk.

Declaration of competing interests Please complete a declaration of competing interests, considering the following questions: Have you in the past five years received reimbursements, fees, funding, or salary from an organisation that may in any way gain or lose financially from the publication of this manuscript, either now or in the future? Do you hold any stocks or shares in an organisation that may in any way gain or lose financially from the publication of this manuscript, either now or in the future? Do you hold or are you currently applying for any patents relating to the content of the manuscript? Have you received reimbursements, fees, funding, or salary from an organization that holds or has applied for patents relating to the content of the manuscript? Do you have any other financial competing interests? Do you have any non-financial competing interests in relation to this paper? If you can answer no to all of the above, write 'I declare that I have no competing interests' below. If your reply is yes to any, please give details below. I declare that I have no competing interests.

I agree to the open peer review policy of the journal. I understand that my name will be included on my report to the authors and, if the manuscript is accepted for publication, my named report including any attachments I upload will be posted on the website along with the authors' responses. I agree for my report to be made available under an Open Access Creative Commons CC-BY license (http://creativecommons.org/licenses/by/4.0/). I understand that any comments which I do not wish to be included in my named report can be included as confidential comments to the editors, which will not be published. I agree to the open peer review policy of the journal.

Authors' response to reviews: Reviewer #1: Choi et al. have proposed a new extension of their PRSice method. The new method, PRSice-2, main advantage is speed as most of the code is written in C++ and PRSice-2 avoid creating intermediate files.

Major Comments:

1. The authors claim that their method is faster and more memory efficient than LDpred and lassosum. However, the authors need to compare these methods in case of prediction accuracy as well.

Thank you for your suggestion, which we think has now made our Technical Note more comprehensive. We have now included a full simulation analysis investigating the predictive accuracy of PRSice-2 compared to LDpred and lassosum (see Figure 3 and Supplementary Figure 2).

1. I like to see experiments where the authors compare PRSice-2 with PRSice performance.

We have now performed a comparison between PRSice-2 and PRSice-v1.25, both in terms of speed and memory (predictive accuracy is the same given the same underlying approach). Results can be found in Supplementary Figure 1, Supplementary Table 1 and Supplementary Table 2

Minor Comments:

The authors need to comment regarding the case where we have multiple populations in a study. For example Luna et al. Genetic epidemiology 2017 work discuss how to solve this problem.

The authors need to mention some of their method limitations in the discussion section.

Thank you for your comment. We agree that differences in allele frequencies, linkage disequilibrium and factors such as genetic drift and natural selection between populations can reduce the generalisability of PRS analyses across populations and produce misleading results, as suggested by Martin et. al. (2017) and as described in our ‘Guide to performing polygenic risk score analyses’ (Choi, Mak, O’Reilly. 2018. bioRxiv). We have now described this issue in our discussion, citing Duncan et al, Luna et al, Martin et al and Choi et al, and we caution users to take extra care when performing cross-population and family-wise PRS analyses.

Reviewer #2: This article reports the release of a new version of the PRSice software for polygenic score calculation. The new version of the software boasts speed enhancements that make it appealing for applications in the growing number of ultra-large genetically-informed datasets including the UK Biobank, 23andMe and others. Also important are features allowing for polygenic score computation from imputed genotype datasets in which genotypes are represented as a probabilities rather than discrete allele counts.

The data on speed are compelling. This alone is a good argument for why PRSice v1 users should upgrade to v2. But I found the article thinner on two other key questions central to addressing whether those not already using PRSice v2 should take up PRSice v2: (1) Does the polygenic scoring method implemented within PRSice2 (additive combination of SNPs with/without LD clumping) deliver comparably predictive scores to other software, e.g. the LDPred and lassosum softwares?

Thank you for your comment and we agree that this is an important question. To address this, we have now performed a comprehensive simulation analysis to demonstrate the predictive power of PRSice-2 Vs LDpred and lassosum (see Figure 3 and Supplementary Figure 2).

(2) What is the value added of being able to accommodate imputed genotype probabilities rather than relying exclusively on discrete allele count data?

We thank the reviewer for this comment. We have now also performed an analysis to compare the predictive power of PRS constructed from genotyped data, or from imputed data either in terms of best-guess genotypes or dosage values. Briefly, the R2 for the Height PRS increased from 0.145 when using genotyped data to 0.152 when using best-guess imputed genotypes, and to 0.153 when using dosage data; likewise the R2 for BMI increased from 0.0475 when using genotype data to 0.0529 when using best-guess genotypes, and to 0.0535 when using dosage data.

I would suggest the following revisions:

Re PRSice2 vs. Alternative Softwares: The authors assert that the method of polygenic score calculation implemented within PRSice2 generates scores that are comparably predictive to two other methodologies, LDPred and LassoSum. It is my understanding that these methods were developed and are in use precisely because they outperform the method implemented in PRSice in terms of the prediction R-squared for the target phenotype. It would improve the article if the authors could provide some empirical evidence for the claim that their software delivers polygenic scores of comparable accuracy to other methods. For example, comparison of PRSice2 scores to scores generated from LDPred and lassosum for a set of traits would be helpful. I like the choices of height and BMI. But it might also be sensible to consider a trait for which existing GWAS are smaller/ polygenic predictions are less accurate, e.g. depression.

Please see above response

Re Imputed Genotype Probabilities vs. Allele Counts: The authors helpfully report that PRSice2 scores computed with imputed data can improve prediction accuracy by about 1 percentage point for height and BMI as compared to scores computed with genotyped-only data. It would be helpful to add an element to this analysis. As I understand it, the authors are comparing a genotyped-SNP-only polygenic score computed from allele counts to an imputed-SNP polygenic score computed from genotype probabilities. But these are not the only two possibilities. In much polygenic score analysis, imputed SNP probabilities are converted to discrete genotypes using a threshold (e.g. probability=0.9) to determine whether a given genotype can be assigned to the SNP. Since this is common practice in the field, it seems to me that it would be helpful to include this approach in the comparison.

Please see above response

Finally, I have one small quibble about language:

In the introduction, the authors assert that polygenic scores have proven clinical utility. This is a bit of an overstatement. I think we can say that "provocative new data suggest the potential for polygenic scores to be useful in clinical settings" or something similar. The recent papers referenced by the authors are compelling. But the term clinical utility has a specific meaning - that application of a tool improves patient outcomes (e.g. see Torkamani et al. 2018 Nat Rev Genet). We are a long way off from that. Instead, the evidence we have supports an argument for the clinical validity of extreme polygenic-scores values for assessing disease risk.

We thank the reviewer for highlighting this and we entirely agree, that as worded, this could have led readers to a conclusion that we do not agree with ourselves (ie. we also believe that PRS are a long way off clinical utility at the individual-level). We have now changed the introduction as follows (note mention of ‘stratified medicine’ in the revised version, as opposed to personalized medicine):

“Polygenic Risk Score (PRS) analyses are beginning to play a critical role in biomedical research, being already sufficiently powered to provide scientific insights and with the potential to contribute to stratified medicine in the future [1-9].”

The authors have addressed my comments.

I have to say I am surprised by the results of the simulation analysis that finds polygenic scores computed using the PRSice2 and LassoSum methods are substantially more predictive of their target phenotypes as compared to the LDPred method. I think I understand why LassoSum should perform better than the other methods. It's based on multivariate regressions rather than a series of univariate regressions. But I am unclear how PRSice2 is deriving its advantage over LDPred. Some explanation of this finding through discussion of the differences between these tools would help clarify the result.

Declaration of competing interests Please complete a declaration of competing interests, considering the following questions: Have you in the past five years received reimbursements, fees, funding, or salary from an organisation that may in any way gain or lose financially from the publication of this manuscript, either now or in the future? Do you hold any stocks or shares in an organisation that may in any way gain or lose financially from the publication of this manuscript, either now or in the future? Do you hold or are you currently applying for any patents relating to the content of the manuscript? Have you received reimbursements, fees, funding, or salary from an organization that holds or has applied for patents relating to the content of the manuscript? Do you have any other financial competing interests? Do you have any non-financial competing interests in relation to this paper? If you can answer no to all of the above, write 'I declare that I have no competing interests' below. If your reply is yes to any, please give details below. I declare that I have no competing interests.

I agree to the open peer review policy of the journal. I understand that my name will be included on my report to the authors and, if the manuscript is accepted for publication, my named report including any attachments I upload will be posted on the website along with the authors' responses. I agree for my report to be made available under an Open Access Creative Commons CC-BY license (http://creativecommons.org/licenses/by/4.0/). I understand that any comments which I do not wish to be included in my named report can be included as confidential comments to the editors, which will not be published. I agree to the open peer review policy of the journal.

Authors' response to reviews: We contacted the first author of LDpred, Dr Bjarni Vilhjalmsson. He informed us that LDpred can become sensitive to small deviations in LD estimates when there are large sample sizes in application to a trait with high heritability. We also noted that there is a new version of LDpred (v1.0.6) now available. Repeating our analyses using a smaller base sample size of 50000, and using the latest version of LDpred, we noted that the performance of LDpred substantially improved. As a result of this, we repeated our entire analyses using the latest versions of PRSice-2 and LDpred and have updated our results accordingly. The overall results remain qualitatively unchanged: PRSice-2 is still markedly the fastest PRS program (more so than previously) and it has comparable power to lassosum and LDpred, with predictive power higher than LDpred and lower than lassosum.