Expediting the Search for Climate-Resilient Reef Corals in the Coral Triangle with Artificial Intelligence

Round 1

Reviewer 1 Report

I think it is a valuable contribution for conservation matters in relation to tropical coral reefs.

Nevertheless the conclusions can not be further in the behalf of coral reefs conservation "sensu lato" because the authors are focused on one single species of hermatypic coral, and this is always difficult to spread towards a typical Indopacific coral reef.

I'm not an expert in genetic affairs but I know for sure that Genetic methodology is quite nice and consistent, although keep my doubts that this machine would work in the future. 

 

Author Response

Thank you for taking the time to carefully and critically review our article. We do acknowledge the fact in the Discussion that this article is focused on a single model species; it is not meant to be a universal model for all corals. It is our hope, however, that we can slowly but surely begin to incorporate additional species in the future to have a better sense of resilience of the entire reef system. We also hope to more extensively ground-truth the models, as outlined in the Discussion. It it our hope that, if other coral biologists see value in this approach, we can more quickly build models for more coral species (and spanning larger areas) in the near future (hopefully before it is too late for the corals we seek to benefit).

Reviewer 2 Report

This paper considers a machine-learning approach for identifying resilient reef corals in the Solomon Islands. This is a relevant research area and it is within the scope of this journal. 

This is an interesting paper. It is well written and clearly presented.

The novelties of the manuscript could be further highlighted in the section “1. Introduction” of the paper.  Consider also discussing the novelties of the paper in the context of previous work in this area.

The readability of the paper could improve if at the end of the section “1. Introduction”, a brief discussion on how paper is organized is presented.

Figure 5. a) is not clear.  Please present an alternative diagram to depict the ANN model for predicting the coral health index (CHI).

Please add a section “Conclusion” and discuss main findings as well as the principal advantages and disadvantages of the proposed approach.

Author Response

Reviewer #2’s summary: This paper considers a machine-learning approach for identifying resilient reef corals in the Solomon Islands. This is a relevant research area and it is within the scope of this journal. This is an interesting paper. It is well written and clearly presented.

Author response to reviewer #2’s summary: Thank you for taking the time to review our timely article, and we are pleased to see that you support its value. We have done our best to address your comments and concerns, as follows.

Reviewer #2 comment #1: The novelties of the manuscript could be further highlighted in the section “1. Introduction” of the paper. Consider also discussing the novelties of the paper in the context of previous work in this area.

Author response to reviewer #2 comment #1: We sort of hinted at this in one previous sentence, but I think we can more blatantly state it: no one has tried to do this before. Most people just look at temperature and coral cover…..and that’s it. I have now modified one sentence in the Introduction to emphasize the novelty of this approach: "It follows, then, that environmental data (e.g., temperature), ecological (i.e., benthic) data, and physiological data from the corals themselves could be used to devise superior models (vs. temperature alone) that would allow us to predict where resilient corals may be located (Figure 1), yet to date no such comprehensive model has been constructed.”  

I have also mentioned this (with very different wording) in the Conclusions section you recommended that we incorporate (described in more detail below). All in all, this is a good suggestion: to highlight the true novelty of this work!

Reviewer #2 comment #2: The readability of the paper could improve if at the end of the section “1. Introduction”, a brief discussion on how paper is organized is presented.

Author response to reviewer #2 comment #2: This is a good point. At present, the last sentence of the Introduction is very vague because we present the “plan of attack” in the first paragraph of the methods. However, I think it would be good to emphasize it again at the end of the Introduction as you have suggested, so we added a new sentence as follows: “This involved integrating seawater quality, satellite imagery, reef benthic, and physiological data (“stress tests”) from biopsies of the model reef coral Pocillopora acuta sampled from diverse habitats in this Coral Triangle nation. ”

Reviewer #2 comment #3: Figure 5. a) is not clear.  Please present an alternative diagram to depict the ANN model for predicting the coral health index (CHI).

Author response to reviewer #2 comment #3. This is a good point and a valid suggestion: this neural network is too large and complex to show clearly. It is actually not necessary to show it; only the model itself is important. Therefore, I have removed it and instead shown the Python script in the online supplemental data file. I have a crude, low-resolution image there, as well, to give a hint as to what the model actually looks like. Now, only the associated desirability analysis is shown in the main text, which is hopefully easier to read and interpret.

Reviewer #2 comment #4: Please add a section “Conclusion” and discuss main findings as well as the principal advantages and disadvantages of the proposed approach.

Author response to reviewer #2 comment #4: We had actually meant for our final two Discussion paragraphs to highlight many of these things, but since this paper is so dense with statistics, I think it’s important to have a “big-picture” conclusion as you suggested. I have therefore crafted a Conclusions section as follows:

“Herein we devised a series of machine-learning models known as neural networks to predict an expensive- and difficult-to-derive resilience metric, the “coral health index” (CHI), from cheaper, easier-to-measure survey parameters. Although models featuring only very rudimentary environmental parameters that do not necessitate in-water surveys or SCUBA training (e.g., type of reef) did not yield accuracies that were deemed high enough to have utility in environmental management and bioprospecting for resilient corals, neural networks featuring both environmental data and survey-derived benthic parameters were of sufficient predictive power to be useful in proactively identifying areas with high densities of resilient P. acuta genotypes. This first approach at integrating coral reef data across multiple scales of biological organization within an artificial intelligence for the purpose of predicting shifts in coral health consequently shows promise, though it is important to emphasize that this article describes a single species in an isolated, understudied region of the Coral Triangle; a more concerted, wide-scale effort by coral scientists elsewhere (& with additional coral species) will be needed to grow this technological and computational capacity to where it can robustly inform reef-scale decision making.”

 This also allowed us to re-emphasize the novelty of the approach (per your earlier suggestion).

Reviewer 3 Report

This is interesting study by Mayfield and coworkers. I can understand the objective of this work, and the picture quality is also good. I have no problems with basic science discussed. The ML approach used seems reasonable to model the systems considered. The conclusion part of the work is also consistent with the results analyzed. It is better add a conclusion rather than "Caveats & future work" since it is not known at this moment whether the future work outlined in this ms can really done??

Author Response

Reviewer #3’s summary: This is interesting study by Mayfield and coworkers. I can understand the objective of this work, and the picture quality is also good. I have no problems with basic science discussed. The ML approach used seems reasonable to model the systems considered. The conclusion part of the work is also consistent with the results analyzed. It is better add a conclusion rather than "Caveats & future work" since it is not known at this moment whether the future work outlined in this ms can really done??

Author response to reviewer #3’s summary: This is obviously a good point because you are now the second reviewer to mention the utility and need for a Conclusions. Please now see the revised article for the newly written Conclusions section, which also reiterates some of the caveats. The issue is not that the future research cannot be done; it has more to do with “buy-in” by the coral field, who I suspect will prefer to do things the old, easier way (temperature and coral cover only). But even though the model of Figure 5 is super complex, those data are actually EASY to obtain and require NO lab work. They DO require SCUBA training and a boat (both expensive, unfortunately). I had WANTED it to not even involve SCUBA diving (which is why we tested the super simple environmental data FIRST). In fact, costs and ease-of-analysis was a large goal of this work, so you will hopefully see that the Conclusions are written from this perspective: how to make the best model that requires the LEAST effort to give you useful data for coral health predictions. I had only briefly touched on this costs-benefits notion in the Abstract and Introduction, so hopefully now that it is more carefully spelled out in the new Conclusions section, the reader will be left with a better idea of what needs to happen next and not left feeling defeated or that this is not an approach that should see widespread adoption.

Reviewer 4 Report

In this work, the authors have proposed a machine-learning approach for identifying resilient reef corals in the Solomon Islands. Neural networks were capable of accurately predicting the Coral Health Index from environmental and ecological features alone.  

I appreciate that the authors have done very extensive work with a solid methodology including an in-depth framework and comprehensive theoretical study. The work is quite complete to be a successful publication but misses some points in the engineering problem formulation. The manuscript can be considered a candidate for a journal publication. The main problem is related to the method of information presentation, and storyline development. The following suggestions are recommended for the authors to prepare the revised version before its final acceptance.

* The title is very poor and does not reflect very less about the knowledge addition in this area. It is strongly suggested to brainstorm on the title to make it attractive.

* The abstract is missing the storyline and the novelty of the work which should be addressed in the beginning. The abstract is also missing interesting conclusions. The results reported in the abstract are also very limited. Authors should not overuse well-known information in the abstract, even as background information. The abstract should be briefly written to describe the purpose of the research, the principal results, and the major findings. The authors should revise it.

* The selection of keywords can be improved.

* In the Introduction, the literature review was not logically organized, and all literature cited seem separate descriptions without connections. The readers can't know what the state-of-art methodologies or gaps the current study plans to resolve or fill, are and how significant or what contribution the current study is. The novelty of this paper is not clear. The difference between the present work and previous works should be highlighted. Research gaps and objectives of the proposed work should be justified.

* The authors can improve the discussion of the results. A reasonable physical explanation should be provided for the observed trends, not only report what is graphically seen in the figures. More physical insight into the Discussion section is needed.

* Proposed methods should be compared with the state-of-the-art existing techniques

* Write the conclusion more precious.

* What is the accuracy and scientific background of the Coral Health Index?

* Principal Component Analysis, ANOVA, or MANOVA is not machine learning method, but they are actually from multivariate statistics.

* How the prediction capability of the Coral Health Index will vary if an artificial neural network or support vector machine can be applied to the dataset?

Author Response

Reviewer#4

Reviewer #4’s summary: In this work, the authors have proposed a machine-learning approach for identifying resilient reef corals in the Solomon Islands. Neural networks were capable of accurately predicting the Coral Health Index from environmental and ecological features alone. I appreciate that the authors have done very extensive work with a solid methodology including an in-depth framework and comprehensive theoretical study. The work is quite complete to be a successful publication but misses some points in the engineering problem formulation. The manuscript can be considered a candidate for a journal publication. The main problem is related to the method of information presentation, and storyline development. The following suggestions are recommended for the authors to prepare the revised version before its final acceptance.

Author response to Reviewer #4’s summary: Thank you for taking the time to review our article and provide input and insight that could improve its content, and specifically readability. We have done our best to address your specific concerns, as follows:

Reviewer #4 comment#1:* The title is very poor and does not reflect very less about the knowledge addition in this area. It is strongly suggested to brainstorm on the title to make it attractive.

Author response to reviewer #4 comment #1: We had previously taken a very conservative, even “humble” approach to the writing style because this work describes only a single species in a relatively small region of the world. However, we have taken your suggestion to give it a more appealing title that will increase its readership; after all, why put in such a huge effort, when in the end only a few people read it because it had a boring title?! I have therefore changed the title to: “Expediting the search for climate-resilient reef corals in the Coral Triangle with artificial intelligence.” This better reflects the goal, presents a larger geographic area in which we are working, AND uses a more general term (artificial intelligence) that more people will be familiar with. However, if you can think of yet a better title, we are certainly open to suggestion.

Reviewer #4 comment #2: * The abstract is missing the storyline and the novelty of the work which should be addressed in the beginning. The abstract is also missing interesting conclusions. The results reported in the abstract are also very limited. Authors should not overuse well-known information in the abstract, even as background information. The abstract should be briefly written to describe the purpose of the research, the principal results, and the major findings. The authors should revise it.

Author response to reviewer #4 comment #2: You are correct in pointing out that our abstract was, frankly, quite boring to read, despite this (in my opinion) being an exciting topic about which we would like others to gain an interest. We have now re-written it to not only include key conclusions, but also the novelty of the approach. I have broken it up sentence-by-sentence to show how we accommodated all of your suggestions.

Sentence 1 (background): “Numerous physical, chemical, and biological factors influence coral resilience in situ, yet current models aimed at forecasting coral health in response to climate change and other stressors tend to focus on temperature and coral abundance alone.

Sentence 2 (approach): To develop more robust predictions of reef coral resilience to environmental change, we trained an artificial intelligence with seawater quality, benthic survey, and molecular biomarker data from the model coral Pocillopora acuta obtained during a research expedition to the Solomon Islands.

Sentence 3 (results-1): This machine-learning (ML) approach resulted in neural network models with the capacity to robustly predict (R²=~0.85) a benchmark for coral stress susceptibility, the “coral health index,” from significantly cheaper, easier-to-measure environmental and ecological features alone.

Sentence 4 (results-2): A GUI derived from a ML desirability analysis was established to expedite the search for other climate-resilient pocilloporids within this Coral Triangle nation, and the neural networks specifically suggest that resilient pocilloporids are likely to be find on deeper fringing fore reefs in the eastern, more sparsely populated region of this under-studied nation.

Sentence 5 (conclusions, novelty, and future work): Although small in geographic expanse, we nevertheless hope to promote this first attempt at building AI-driven predictive models of coral health that accommodate not only temperature and coral abundance, but also physiological data from the corals themselves.

Reviewer #4 comment #3:* The selection of keywords can be improved.

Author comment to reviewer #4 comment #4: This is a good suggestion, as we barely talked about dinoflagellates, and “ocean health” is very vague. We have now added terms that more accurately address what we actually did.

Reviewer #4 comment #4: In the Introduction, the literature review was not logically organized, and all literature cited seem separate descriptions without connections. The readers can't know what the state-of-art methodologies or gaps the current study plans to resolve or fill, are and how significant or what contribution the current study is. The novelty of this paper is not clear. The difference between the present work and previous works should be highlighted. Research gaps and objectives of the proposed work should be justified.

Author response to reviewer #4 comment #4: This is a good point, and you are the second (or maybe third) reviewer to mention it. The fact is: no one tried this before. The idea is: if temperature can give us some predictions and benthic data can give us some predictions, what if we COMBINED the temperature data, the benthic data, AND data from the corals themselves? We now emphasized this in the abstract, as well as the conclusions. We also modified one sentence in the Introduction to better promote the novelty: “It follows, then, that environmental data (e.g., temperature), ecological (i.e., benthic) data, and physiological data from the corals themselves could be used to devise superior models (vs. temperature alone) that would allow us to predict where resilient corals may be located (Figure 1), yet to date no such comprehensive model has been constructed.” In other words, to summarize our Introduction, it is as follows:

“We have some predictive models for coral reef health, but they actually aren’t very good in some places” (beginning of paragraph 1)

“These models do NOT consider the corals themselves” (end of paragraph 1)

“We should combine the OLD way (temperature+benthic survey) with the NEW way (physiological data from corals (beginning of paragraph 2). Then I give an example of why the old way is bad.

“Here is an example of one way you could try this new, more comprehensive approach” (end of paragraph 2).

Hopefully, with the revised abstract, the revised Introduction (including a new sentence at the end that describes more specifically what was done), and the new Conclusions section, you will be left with a better sense of what we actually accomplished, as well as why it could be useful for coral reef management.

Reviewer #4 comment #5:* The authors can improve the discussion of the results. A reasonable physical explanation should be provided for the observed trends, not only report what is graphically seen in the figures. More physical insight into the Discussion section is needed.

Author response to reviewer #4 comment #5: You are correct in that, for many findings, we did not actually attempt to explain them. For instance, we made no mention of why coral cover was higher (and algal cover lower) between regions of the country, so I have now added a sentence to mention this: “Although it is tempting to link the higher coral cover (& lower algal cover) of the remote, eastern half of the country with its markedly low number of human inhabitants (see. OSDF for population & other demographic data), we regrettably did not measure key parameters, such as nutrient levels, that could directly attest to human impacts. As mention in more detail below in the context of the CHI predictive models, nutrient data should certainly be incorporated, if at all possible, in future such AI modeling analyses of coral health.”

            I also think the depth story is interesting, i.e., more resilient corals on fringing fore reefs, and have provided some discussion around that point, as well. In fact, this is worth mentioning because this was NOT expected; this coral is more commonly found in stagnant lagoons: “In much of the Coral Triangle, P. acuta is more commonly found in more stagnant lagoonal environments [18] given its relatively more delicate branches than closely related congenerics (e.g., P. damicornis & P. verrucosa). The fact that more resilient genotypes of this species are predicted to be found on fringing fore reefs, then, was an unexpected finding and could be related to differences in water circulation, food supply, or other, undocumented variables.”

Reviewer #4 comment #6:* Proposed methods should be compared with the state-of-the-art existing techniques

Author response to reviewer #4 comment #6: Please see the Introduction for a treatise on the older methods. No other attempt has been made to make predictions of coral health in the Coral Triangle in a similar way, though I have mentioned how this could be done in the “caveats” section of the Discussion. I also added a new sentence that addresses something mentioned in the Introduction: the “gold standard” for coral health predictions=NOAA’s Coral Reef Watch. “Extent and magnitude (severity) of bleaching of corals of differing CHI values during marine heat waves could also be compared to levels projected by temperature-exclusive models, such as NOAA’s Coral Reef Watch, though it is important to note that the approach herein aims to make predictions at the colony-reef scale, while Coral Reef Watch predicts at a multi-km scale.” Indeed, this is the next step: to see if these models are actually significantly better than the temperature-exclusive ones.

Reviewer #4 comment #7:* Write the conclusion more precious.

Author response to reviewer #4 comment #7: We had actually refrained from having written a conclusion at all, but have now done so at the request of other reviewers. Hopefully now, there is a better sense of what we accomplished.

Reviewer #4 comment #8: * What is the accuracy and scientific background of the Coral Health Index?

Author response to Reviewer #4 comment #8:  The coral health index is based on many years of field and experimental data from work with Pocillopora acuta, which is the model coral species for research. There are at least 30-40 papers we have published that were essentially laying the groundwork for this response variable. The idea is: the more things you measure, the more accurate the prediction will be, which is why we measured so many different parameters. We reiterated a lot of calculations from prior works to help the reader, but we urge you to seek out articles in the references for details. In short, the predictive capacity will vary by location. Predictive models of the CHI in Taiwan, where it was developed, are >90% accurate, whereas in the Solomon Islands, they only reached 85-90% max, and that was for very complicated models. Ideally, one would do a ton of experiments on SOLOMON ISLANDS corals, then calculate a Solomon Islands-specific CHI, but the country does not have the resources to do this, so we basically took a CHI developed for Taiwanese corals and applied it in the Solomon Islands. That’s why the accuracy was lower than in Taiwan.

Reviewer #4 comment #9: * Principal Component Analysis, ANOVA, or MANOVA is not machine learning method, but they are actually from multivariate statistics.

Author response to reviewer#4 comment#9:  You are correct. This paper is not exclusively machine-learning; we actually used a mix of machine-learning, univariate, and multivariate statistics. We wanted to emphasize the machine-learning in the title, abstract, and discussion, since that is the component that is actually making the predictions.

Reviewer #4 comment #10: * How the prediction capability of the Coral Health Index will vary if an artificial neural network or support vector machine can be applied to the dataset?

Author response to reviewer #4 comment #9: Great question! Actually, we have a new dataset using proteins (NOT the CHI), and sometimes the best model is the support vector machine (NOT the neural network). I actually PREFER SVM. Do you know why? Because if you repeatedly rerun the SAME model, you get the SAME result. For complex neural networks, you need to test different “tours” in which the model starts and stops in different places. That means the R2 varies from run-to-run, sometimes by a lot. That’s why you’ll see instances in the methods where we mentioned that we did 10, 15, 20, or more simulations. This was to ensure that the neural network R2 was stable. In other words, if the SVM and NN give you a similar result, you definitely should use the SVM (in my opinion). In this dataset, I don’t think the SVM performed well, and I actually don’t know why, but in my experience the more variability and complexity in the dataset, the more likely it is that only a neural network will suffice.

Round 2

Reviewer 4 Report

Dear authors, thank you so much for providing such an insightful and friendly response to my comments. In fact, some of the comments that you have made are quite interesting. I also agree that SVM can be preferred if R-square of ANN and SVM is equal, as the performance of ANN can be tricky even for the same tuning parameters. Nevetheless, all approaches are correct as long as it has a scientific justification. I would still suggest you to improve the title. Now, the part "With Artificial Intelligence" seems to be disconnected from the title. One suggestion, just a 'suggestion', would be:  Expediting the search for climate-resilient reef corals in the Coral Triangle: Analysis assisted using artificial intelligence. Congragulations on this article.