1. The paper describes the use of a program on Apache Spark to create decision trees in a random forest applied to data of particularly high dimensionality - in this case variant data from the 1000 Genomes project. The program, called VariantSpark, is proposed as more performant and more accurate than a list of alternatives. The authors hold that this program exhibits novelty in the way parallelization for random forests is implemented, and for the number of features that it can process. The experiment has 4 sections: a. A comparison of run time and accuracy of competing random forest implementations on increasing subsets of a common dataset; b. An exploration of the VariantSpark program's scalability using a synthetic dataset c. An exploration of the effects of choice of mtry and ntrees parameters on the accuracy of VariantSpark d. A direct comparison of performance between VariantSpark and yggdrasil, with the former set to create a single Decision Tree in order to approximate the behavior of the latter.

2. There are a number of typos or grammatical errors that should be fixed to aid readability, as follows: a. Page 1. paragraph 2: "GWAS context" - > "a GWAS context" or "GWAS contexts" b. Page 4, paragraph 5: Various issues in text. To be re-proofed. c. Page 6, paragraph 2: "outpacing" -> "outpace"

3. Page 2. Paragraph 1 states "all of these implementations are optimized ... for HPC". But H20.ai is listed in Table 1 as running on a proprietary cluster.

4. Page 2, paragraph 7 states "execution time of all methods" should specify "all Random Forest methods" so as not to include yggdrasil. Also, please state here which cluster was used.

5. Page 2, paragraph 7 : A 500K dataset is mentioned in the narrative, but not in the associated table 2. If the samples and features columns for table 2 were provided as they are with table 3, it would make fig 1a clearer with respect to which data points were which.

6. General point: Given the title of the paper and the name of the program, and the fact that the y-axis of fig 1b uses it, some treatment of the word "variant" would aid some readers, perhaps with reference to the 1000 Genome VCF files that serve as input.

7. Page 3, paragraph 1: The maximum values of p (50,000,000) and samples (10,000) do not correspond to the values in Table 3 (81M and 1.5K respectively)

8. Page 3, paragraph 2: Fig 2a does not represent the narrative, or support the conclusions, of this paragraph. The narrative describes an experiment where mtry is held fixed, 100 trees are built, while p and n are increased and the time to build the trees is measured. Fig 2a shows the number of cores on the x-axis and the trees-per-hour on the y-axis, as well as showing 4 mtry scenarios. The caption corresponds to the narrative, and not the plot. There is no breakdown of n and p to demonstrate the (respectively) sub-linear and linear scaling described in the narrative. Figure 2b does not appear to support the "linear scalability" of this sections title, with respect to CPU, or even the "scales well" of the narrative. The four scenarios of the "ncore", where the number of cores is doubled from 16, to 32, to 64, to 128, shows rapidly diminishing returns in terms of speedup for 100 trees, at all values of mtry. If an open-ended approach is taken, as with Fig 1a where a 'trees-per-hour" metric is used, then we can say that things look better, but to use the authors' own language, it is sub-linear. There is no explanation of how the number of cores is controlled. Given the values used, it appears that increasing numbers of worker nodes are used (1,2, 4 and 8, but never all 12?). Or was it controlled by setting the number of executors?

In summary, there is a disconnect between the narrative and its claims, when compared to the data presented in the figures. The authors should reconsider what claims they wish to make, and re-build a suitable narrative, with directly corresponding data/figures.

1. Page 4, Paragraph 5: The description of the implementation is poor. It has typo/grammatical errors that confuse the reader, and terms are not clearly defined. What follows is a set of questions that is not exhaustive but indicates points of confusion. a. Is 'worker' intended to mean 'worker node' in the Spark sense? b. Does each 'i' partition build its own tree? c. If not, what decides how many trees are to be built overall, and per worker/node? d. Does the word "node" in the phrase "best split locally for the node" refer to worker node (Spark) or tree node (Decision Tree)? e. "Assume one [executor] per worker": If "worker" means "worker node", this does not correspond to the number of executors in table 4. Given the above, this paragraph could do with re-writing for clarity. Explaining complex code in concise English is challenging, but necessary. The burden on the authors would be lightened if the scripts that ran all 4 sections of the experiment were made available as part of the source code in github.

Summary: Description of methods could be more systematic. Several readings were required before a clear (albeit still incomplete) picture of the experimental design emerged. This would be offset if there were a way of reproducing the 4 phases of the experiment. While I was able to build Variant Spark and run the built-in examples, there were no instructions for reproducing the experiments themselves. Given the cloud-based nature of the experiment, it would have been possible for the authors to provide Terraform/CloudFormation scripts to reproduce the clusters, and scripts to run the experiment on those clusters. This would have lightened the burden of describing the experiment and presenting its novel aspects. The idea to use Apache Spark to create random forests for high-dimensional data is a good one, and clearly a great deal of work has gone into the implementation. But while the experimental design seems sensible, the paper does not present the experimental data in a way that supports all of the paper's conclusions.

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: (https://drive.google.com/file/d/1DX8zDAjdqJTzgMcDppEFvuB0kM4Mft2Q/view?usp=sharing)

This paper has clearly gone through a major rewrite since the last review. The narrative is clearer and issues of reproducibility on the cloud have been addressed.

There remain a few minor points around readability. I present a non-exhaustive list below, and generally recommend that parts of the paper be proof-read for basic grammatical errors.

Some of the diagrams required multiple passes in order to appreciate the information they were conveying.

Overall I'm happy to recommend this paper for publication. The narrative focuses on the particular use case of high-dimension GWAS studies, and goes to the trouble of explaining current relevant issues around phenotypes influenced by gene interaction and epistasis. This has the benefit of educating readers of a more informatic background, while at the same time giving biologists the raison d'etre for the technology: the key facts that "variant interactions remain invisible to traditional GWAS" while algorithms that address these interactions typically do not scale to whole-genome data. The space devoted to this in the paper is particularly welcome in a cross-disciplinary field where two audiences need to be addressed simultaneously.

I was able to build the software provided on its master branch (with some adjustments like disabling unit tests and switching off the scala formatting plugin). Creating the overall infrastructure was also straightforward, using the AWS marketplace option with the provided AWS Cloudformation Templates. Note that the steps presented in the explanatory video (another very welcome addition) are slightly out-of-date with respect to those that need to be followed today, but not enough to be of serious concern. In principle, using this approach, it would have been possible for me to run large datasets like those used for the experiment (also provided). As it stood, I was able to verify the correct functioning of the running system. All major cloud providers have analogous products, and there are some products such as Terraform which work across multiple cloud providers. Papers such as this one do well in making use of them.

Non-exhaustive list of issues relating to grammar, spelling or readability follows. In general, despite a very well written initial section, some very basic errors were made from page 3 (Datasets) onwards. I don't propose to list everything here, but strongly recommend that it be proof-read.

p 3 para 8: "We generate four subsets [of/from?] this data" Table 2 caption: "... these dataset[s]. Table 4 caption: "All dataset includes" -> "All datasets include" Table 4 data: There appear to be duplicate rows for 256X and 513X p4 para 5: "five genotype dataset[s]" p4 para 6: "In the last two replicate[s]" p4 para 7 : "We build a forest with 10 tree[s] p4 para 8: "applied to all experiment[s]" Figure 1b caption: on last line (beginning "Where, ..." move the comma to after variants, and change 'is' to 'are' for variants. p6 para 5: "highly correlates [with?] the TV and possibly [identifies] p6 para 6 "VariantSaprk" to "VariantSpark" p6 para 6 "of exclusively detect[ed] variants" Figure 3 caption: "...black line [illustrates] if the..." p10 para 2: "polyphonic" -> "polygenic"? List of definitions: "mTry: Number of variable[s].."

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: (https://drive.google.com/file/d/1ThYzklIBObIdWVLw07jmlTHZJwd2fV2Q/view?usp=sharing)