Choose Language Hide Translation Bar
Level VI

Toward Predicting the Fate of Reef Corals: 2021 Update (2021-US-45MP-855)

Level: Intermediate


Anderson Mayfield, Dr. (assistant scientist), University of Miami


Coral reefs around the globe are threatened by a plethora of anthropogenic stressors, most notably climate change. There has consequently been an urgent push to better understand the fundamental biology of key coral reef inhabitants, understand how these organisms will respond to future changes in their environments, and make data-driven predictions as to the relative resilience of disparate coral populations. In this talk, l showcase a complex, "molecules-to-ocean basins" data set featuring molecular (e.g., proteins), physiological (e.g., growth), environmental (e.g., coral abundance), and oceanographic (e.g., temperature) data to showcase how I have been using JMP Pro 15 and (more recently) JMP Pro 16 to model coral behavior and make predictions as to which colonies/reefs may be at most risk. I also examine which reefs are expected to demonstrate resilience to global climate change and disease. Through this update to my Discovery Summit Americas 2020 presentation, I hope to, more generally, show progress in forecasting coral health and resilience using the ever-growing data sets and analytical power now at our disposal. 


In addition to the recorded presentation itself, I am uploading the following files: the three JMP tables referenced in my talk (each with embedded scripts) and the Powerpoint slide deck. I actually passed over some "hidden" slides in the presentation that provide additional context, as well as additional approaches employed. Of note, I would like to extend a "shout-out" to @DiedrichSchmidt for sharing and customizing his excellent machine learning GUI, which I have used extensively to create thousands of neural net-based models with my coral health data. 




Auto-generated transcript...



Hi everyone, my name is Anderson Mayfield and I'm an assistant scientist at the University of Miami's Cooperative Institute Marine and Atmospheric Studies.
I actually work across the street from University of Miami's marine lab NOAA's Atlantic Oceanographic and Meteorological Laboratory as part of their coral program.
Today I'm going to be talking to you about some of the research I've been working on over pretty much my entire career, which has been trying to develop tools for predicting the fate of reef corals.
So this is actually an update, or a continuation, from what I discussed last year, my Discovery Summit talk, but I know a lot of you will maybe not heard that or even if you did
you may not remember everything, it's totally fine I'm going to give you I'm going to set the stage for the research that I'm doing.
Talking about some problems facing coral reefs, give a little bit of a recap of where I was at last year.
Really excitingly, not probably two months after last year's Discovery Summit talk, I made a kind of a big breakthrough in my coral predictive modeling
progress and this was largely in part due to some of the new features and JMP Pro, so I'm going to share these kind of hot off the press data today.
And that's getting to this idea of can we predict the fate of individual coral colonies?
So predicting the fate of a coral reef, so the entire assemblage of coral, this is actually easier because coral reefs are basically in bad shape across the globe
due to rising sea water temperatures. So corals have dinoflagellate algae living inside of their cells, they photosynthesize like a plant and they translocate
the energy that they produce to their coral host. Problem is when the water gets too hot this messes with their ability to photosynthesize properly.
Corals either digest them or kick them out, the end result is the same corals effectively starve to death.
And this is a problem across the entire globe so predicting the fate of a coral reef, we pretty much already know these these are ecosystems in decline.
But on an individual level you do see heterogeneity, there's coral species that fare better than others, there's genotypes within species that do better than others - that are bleaching resilient when others are bleaching susceptible.
So something I've been interested in my whole career is what drives this resilience and is it something that we can actually predict?
So I liken this to being kind of a coral actuary. Can can you send me a biopsy from your favorite coral and I can do some analyses and say hey,
you know this coral has six months to live, this one is going to be fine for decades, this is this is kind of what I've been trying to do.
Because the way we currently assess coral health is retroactively. We go out there and do our surveys and we basically quantify the amount of dead coral.
So these are important data that need to be collected, but you didn't really do these corals any favors by this sort of approach. This would be analogous
to a doctor telling his patient who had just had a heart attack, hey you had really high blood pressure, you know that patient would have liked to have known that
weeks, months years beforehand, so they could have done something proactive - medicine, lifestyle change, you know better diet and exercise. So
you know I think it's unacceptable to assess the health of humans in that retroactive fashion, so why can't we have this kind of more proactive approach for corals as well?
So there are predictive tools out there, and some of them are actually developed by NOAA, actually most of them are developed by NOAA, my employer, that are trying to make
projections or predictions about coral health.
They're exclusively using temperature because there's just this well this well known well studied link between high temperatures and coral bleaching so it makes sense to hone in on temperature as a predictor.
So this is looking at, actually United Nations a report that came out this year but a lot of the authors are working at NOAA. This is looking at
the year of onset of what they call annual severe bleaching. So this is the year, at which bleaching starts to happen every single summer. So
corals can recover from bleaching, they can acquire new symbionts, they can eat,
they can weather the storm, to a certain extent, but when bleaching starts happening every single year, they lose the ability
to recover, so this is why the annual severe bleaching year is kind of is what I'm calling on at the point of no return. So this is pretty grim because you can see
some of these years in red, 2015, 2020, it's obviously already passed, and this is looking at a map of South Florida, so this is where our field sites are.
But we know this is not actually 100% true. Yes, there are reefs within these pixels that are already bleaching every year but there's plenty of others that are not.
So one of my goals is to improve the spatial and temporal resolution of these temperature-centric models, because these temperature-centric models
don't consider what's there, they don't even consider whether or not there's corals there at all. It's just looking at temperature patterns.
So they don't factor in whether or not there's resilient species there, whether there's resilient genotypes, but then we need to go out, they're not grounded in truth at all, frankly.
We need to go out there and do our ecological surveys we're looking at coral abundance,
biomass, biodiversity, what kinds of coral enemies are out there, certain types of macroalgae and coralavores, but even that I feel is not enough, because you can have a reef that
is absolutely carpeted with corals like this one, you see here in this image.
An ecologist might even say hey this looks really healthy it's high abundance of coral, the water looks pretty clear, there's not a lot of algae, you know to even the trained eye, you might
guess that this is a healthy reef but if I know personally as a coral physiologist that this is a low resilience coral reef because these species are weedy they're the ones that don't keep a lot of energy
in the tanks for when the seawater rises, so this this inherent disconnect between abundance of corals and health.
But if you think about it, this is not that surprising. We don't manage human health this way. I don't go around downtown Miami where I live, and say hey this neighborhood has a million people,
it's two times healthier than this neighborhood over here with half a million people, I mean it might be to an economist healthier.
But if you actually look at the health of the residents, there may not be a correlation. You might say, hey this bigger city has
you know better health care than a small town. So in certain situations there might be some link between population density
and human health, but you can find just as many places, where that's not the case, you know.
Higher population density areas tend to have more crime, for instance, so it kind of blows my mind that this is still the way we manage that
or try to make conjectures about the health of marine organisms is simply by counting what's out there. I think
you know you wouldn't do this with with people, and I think we can do better.
So what I really want to do is not just look at temperature. Temperature is going to be critical for understanding for help, but not in isolation.
Similarly, ecological factors, the organismal abundance, the biodiversity is also going to be critical assess. It it's not going to tell you about persistence and longevity in and of itself.
What I want to do is, I want to factor these things in but also consider the physiological and the actual health of organisms in these ecosystems so
that is a very complex, convoluted model and I haven't actually worked out how it's going to be presented, or how it's going to work
right at the moment, but I have some ideas. Today what I really want to focus on is what I've encircled here so simply trying to use the physiological data to make a predictive model with the capacity to forecast whether a coral is going to resist bleaching or be susceptible.
And for these kind of predictive analyses I gravitate towards molecular biology, because you should see subcellular shifts in behavior
that are reflective of some kind of aberrant physiology before you notice something wrong with the naked eye. So getting back to that heart attack
analogy you're going to document high cholesterol levels or high blood pressure in an individual weeks, months, maybe even years before they're likely to have
cardiac arrest and that's why when you go to the doctor they draw blood. There's we have these well developed biomarkers like blood sugar and cholesterol
for making conjectures about your your future health. They're not going to be able to tell you, you know the day you're going to kick the bucket, but these are
biomarkers and analytes that have high predictive power in terms of your future health prospectus. So
cholesterol is not a particularly good biomarker for coral, but other things may be, and so at NOAA we take this multi-
we call it multi-omics approach, where we look at all the different molecules synthesized by our organisms of interest, corals in my case, to see if we can figure out which ones might
be reflective of a state of stress or a state of resilience. I actually started off in the gene expression world, so the messenger RNAs.
But recently I kind of switched over to proteomic research just because proteins are actually carrying out the work in the cells. So
a lot of times you don't see a strong correlation between the mRNA expression and the protein levels.
mRNAs could still be really good health diagnostic biomarkers but they're not going to be too useful for mechanism. As a physiologist I want to have
my cake and eat it too, so I want to be able to know what's happening in my coral cells, how are the coral cells responding to high temperatures, which proteins are involved in thermo acclimation.
These are the kinds of questions that can be addressed by what we call proteomics, which is the assemblage of all the proteins in a sample.
So you get the mechanistic data, but you also can have these proteins serve as health informative biomarkers, so this is kind of my slide where I try to condense 20 years of research into a single slide just to get everybody up to speed so.
Determining the optimal suite of analytes. So I just talked about the proteins. Do I think proteins are the end all be all and they'll never be a better thing to measure in a coral?
And that's why I put the check here - absolutely not. And you'll see later in the talk that they're actually issues with with using proteins to make health inferences. There's probably something better out there, so this is still a work in progress.
Unfortunately corals are non model organisms, so for all these things, I wanted to measure I had to spend years developing the methodology myself. Molecular benchwork, in particular.
You want to use any sort of analyte, be it a gene, a protein, a lipid, as a biomarker you're going to need to accommodate the natural sources of variation.
And these are things like the light cycle. So because corals have photosynthetic dinoflagellates inside of their cells
they show huge swings in physiology across the light-dark cycle, across the tidal cycle, over seasons, these are kind of the non-sexy, but kind necessary evil
projects you need to do to ultimately get towards modeling coral health. Once you have a grasp of how the corals are getting on their day to day lives, then you do
these environmental challenge studies where you expose corals to stressors
or or in the laboratory or you're looking at stressor regimes across environmental gradients in the field.
And this is actually what the coral field as a whole has done the best of, there's thousands of papers out there on coral environmental physiology and I'm going to talk a little bit about one such example of that later in the talk.
But where we're stuck and were you don't see any work right now is this idea I've been getting at of using coral health data to actually generate predictive models of coral stress susceptibility so that's what I'm trying to make inroads in.
And I think this is because most physiologists and molecular biologists are much more familiar with the kind of descriptive side of things, what happens in the organism as it's dying.
This is not to belittle this sort of research, it's pretty much what made my career to date.
It's easier to publish these kinds of papers. Stick coral reef tank tank temperature up look at the protein levels. I've published 10 papers on that, and you do need to do this type of work to get at
this predictive approach. So but I do, that being said, I do want to kind of make this shift for me personally from being kind of this
undertaker who's effectively writing coral obituaries, essentially what these papers are, they are generating useful data but they're not doing the corals any good.
And we take the data that we got from those inherently descriptive projects and use them in a predictive capacity, try to figure out
which corals are going to be the climate change winners and which ones are going to be the losers. So going from being an undertaker to this, you know vet or the actuary that I mentioned earlier.
And I don't want to give you guys the impression that I'm you know this amazing
modeler. This is something I really just have been teaching myself and picked up on the fly, but I attribute a lot of the success in this and making this kind of seamless transition from doing more descriptive stuff to predictive stuff
this is driven by by JMP because because my familiarity with the Fit Model platform, for instance, really made it easier for me
to kind of make this description to prediction transition so, for instance, what I would have done in my entire career up until about a year or two ago is the example on the left.
You see, these coral proteins in the y box. These are the esoteric accession numbers you don't need to pay attention to what they mean, they're just know they're proteins.
Construct model effects showing the things that I would be interested in temperature, genotype of the coral, when they were sampled, things of that nature, and you can publish a nice paper
looking at these sorts of data sets. Now what I'm doing is I'm basically flipping that on its head I'm taking all those proteins and moving them down into the construct model
box, alongside the experimental and/or environmental data but what's going to be in the y
is going to be what I'm calling here the coral bleaching susceptibility. I give it another name later, but the idea is the same. It's the likelihood that a coral is going to bleach.
So I think JMP has been really instrumental in kind of helping me make this this transition from purely descriptive stuff to kind of getting my feet wet and predictive biology.
So I do want to recap a little bit of the descriptive stuff and feature some of some of my some of the tools within the JMP Pro
package in particular that I'm a big fan of. So I'm going to play around with a little data set we have from this coral here - orbicella faveolata - it's one of the
one of the more resilient corals in the Florida Keys so climate change effects are so bad to where you know we don't have the luxury of necessarily working with all the coral so you basically got to work with the ones that are already inherently a little bit stronger.
So we took corals from different reefs, exposed them to different temperatures, different amounts of time and looked at their protein concentrations, because we need to have kind of an idea of what biomarkers might be, we might might be useful in model building.
How does temperature affect the proteome of this coral, all manner of descriptive things that you might
want to address, so this is very similar to what I would have shown before. This is my model box, I do a lot of univariate stuff where I'm trying to find you know which proteins are
most affected by temperature, by the reef site, things of that nature. I do a lot of multivariate work - principal component analysis, multi dimensional scaling.
Looking for relationships amongst samples, looking at multivariate treatment effects and we're going to be looking at some of this here in a second. So last year, what I showed and I published this paper since then um, this is a multi dimensional scaling plot made
in the JMP Pro 16 multivariate platform - actually what this would have been 15 but all I want to show you here don't get too bogged down in the details. And I apologize if you're colorblind.
But the green, which are the bottom three samples on the right, these are control corals, so they're at the control temperature. The three at the top, are
at 32 degrees, which is pretty much a death sentence for most corals. And you can see after several weeks of exposure, their proteomes, so their protein profiles, start to become more stable and that's important.
You need to know that there are effects of temperature on the proteome before you do any kind of predictive model.
Those data were looking at presence absence data. So whether or not a protein was there or not and it was about seven or 800 proteins.
Since then, the methodology improved, so this is a good thing about working in molecular biology, the methods are always getting better.
So what I did I took the same samples, but rather than measure seven or 800 proteins as presence absence I took a subset of 86.
But the data are now fully quantitative. So I have you know almost tenfold
fewer proteins, but there's much more resolution with respect to the concentration, so the data set I'll be playing around with now is only featuring 86 proteins. So
this is a typical distribution of what these proteins look like they're never normally distributed they're always skewed so this limits me in terms of using exclusively parametric approaches, but that's not a problem
with JMP.
So let's go over into a JMP table.
So this is the first one, I want to show you. You've basically got the two features of this table I really want you to focus on are
its wide nature. So proteomics is expensive. $200 a sample, my budget's limited, in this particular experiment I only had 20 samples.
But I have 86 proteins. What that means is can't do typical multivariate ANOVA to look for treatment effects, but that doesn't mean
that I can't do any sort of useful multivariate analysis. One clever approach I got as a suggestion from JMP developers is to go do my traditional multidimensional scaling. So here's my 86 proteins that then log to transformed I'll put them up here.
And I know from having done this before that four dimensions is going to give me a pretty good solution.
So here's the multidimensional scaling plot, these are the various fates of my sample, so healthy controls, actively bleaching, we'll get more into that later.
But I want to do here is look down at stress, I see okay it's getting a little bit high usually like to be below point one.
I could get the stress even lower by increasing the number of dimensions, but I think four dimensions it's going to be okay, so what I'm going to do
is I'm actually going to save these coordinates. I could save the similarity matrix as well because I'm really more interested in the similarity amongst these samples so I'm going to take these
and they're now going to be my (???), so now, I have a situation where I have fewer y's than samples, so I can do a traditional multivariate ANOVA.
And my model effect, I want to have this done. CHD lab, what does that mean? That's the coral health designation so that's whether or not the coral was resilient or stressed.
So whoops.
So these this is BLR, that means bleaching resistant, I'll use this term throughout the talk. BLS, bleaching susceptible, and this is looking at a plot of these four dimensions for these two groupings of samples so I'm going to run a multivariate ANOVA.
And I see a marginal effect of health state, and this is again something you would want to know.
If these bleaching resilient corals have a significantly different proteome than the bleaching sensitive corals. So it's not strongly significant is kind of on the border, but
with an environmental data set like this, I would happily report this in a publication. Another thing you can do that's probably the better approach is to take those same four coordinates
The MANOVA is going to get tricky with small data set, a small wide data set like this, where you've got a bunch of different environmental parameters, so instead I'm actually going to use Partial Least Squares platform that I'm a big fan of.
And I'm going to take my experimental data, these are things like where the coral was from, the temperature it was exposed to, the genotype and I'm going to do a response surface.
And then what's running.
I'm actually going to bump the KFold down to four just because the data set is so small.
And this is not exactly the output that I wanted, I think I must have included something a little bit different but that's Okay, so I've saved the script that actually want to show you.
Is here and really I'm not even too concerned with the number of factors.
You could see here, I chose three.
Really, all I want to show you is this, it looks very chaotic and messy, but this is known as the correlation loading plot, and I think this is really an unsung underutilized tool in the entire JMP Pro package.
Just because of how much you can gain from this figure, so this is showing where my samples fall out with respect to one another, then I can layer on my environmental factors, genotype, site temperature time.
In this case, I'm using the coordinates so the y's are going to be more difficult to interpret, but I could have done this with the raw data as well, and then you could really start to look at how
your analytes, your environmental variables and your samples all kind of mesh and I think this is a really powerful exploratory tool that that may be less familiar to some of you out there.
So that's one of my favorite multivariate descriptive tools, I like to use in the JMP Pro package. So while, that was all experimental data, so while we had those experiments going on
these corals they're my patients. They're sessile. I know exactly where to find them go out there on the reef.
I can sample them in different seasons. I'm taking tiny little biopsies so it's not going to be perturbing their their health to any great extent.
So I have these coral samples through space and time and I know when they bleach,
where they bleach, which ones bleached and which ones didn't. So now, I can go back backwards in time through the archive and hindcast looking at corals
that haven't bleached yet and trying to see if I can make predictions about which ones did and which ones didn't.
So just to give you a little bit more info on the sites they're down here in the upper Florida keys, we have two inshore sites, The Rocks and Cheeca Rocks, two offshore sites Little Conch and Crocker Reef.
And it's important to know that the inshore reefs tend to be much more robust you still see big healthy corals not everywhere, but some places.
Offshore you're dealing with these little press, so we have a nice gradient of resilience in situ.
This is going to be important for model building you wouldn't want to build or test your models using exclusively strong corals or exclusively the weak ones, you're going to want to have a mix of both.
This is just a quick Graph Builder plot I put together, and this is looking at color scores, so you want to be a five. Five basically means you're totally healthy.
Zero means you're stark white and you're about to die. In reality three to four is getting pretty stressful. So you can see here in August
this is in the middle of a bleaching event, we have some pretty market bleaching at these four sites, even the more resilient inshore sites were bleaching to some extent.
But then, by October they've recovered and then December they've recovered, but you see this kind of anomalous behavior and I think this is driven by a disease.
So this is getting back to kind of my my jargony coral health designation that I mentioned earlier, so.
This is basically, you know I want to have this ultimately or eventually be maybe a continuous variable, maybe a one through 10.
But right now we're trying to predict one of essentially three categories. You're either bleaching resilient or resistant, you're bleaching susceptible or you're intermediate. So the bleaching resistance is going to be this green line at the top, your color score doesn't really change
over the course of the bleaching event. This is another Graph Builder plot, so this is looking at the temperature on the bottom half.
So an intermediate coral is going to bleach you know a little bit, but then it recovers.
The bleaching susceptible one is either going to bleach markedly and recover or, in some cases, and may not recover it all. So these are the three essentially the phenotypes that are going to be the y's and my model.
And this is convoluted by design don't worry, this is basically showing the different ways, you could ultimately go through the model building process. So on the left
you'll see which data are you going to use to train and validate the model.
The model itself, what are you going to use test the model and then what you're ultimately trying to predict. So if you're just using lab coral data
to make a predictive model in which you're testing it with more lab world data,
that's going to give you the power to predict coral health in the lab, but that's not really what we're after. I mean that's okay for publication, but we want to know, we want to be able to predict what's going on in the field.
But then you run into this issue of how, how do I feel about using lab data to make a predictive model
that's attempting to forecast what's going on in field, you know that's always going to be a precarious issue of using this lab data to make predictions about this less
constrained field involved. But we're going to try it anyway. So again, this is using this colony health designation as the y.
The predictors, as I mentioned earlier, are ultimately going to include more than just the proteins, but for the sake of simplicity we're going to use the proteins today. And this is
using JMP Pro 16.1 I have X G boost add-in that I really like and I really recommend. This is more of a note for myself, but I do want to mention one important thing, and it gets to
one big drawback of proteomics and it's stochastic in its nature. So remember I sequenced 86 proteins in my laboratory samples So those are the ones I'd like to use and my predictive model.
When I went out there and measured proteomes and my field samples I didn't get those 86 proteins I got a completely different set of proteins.
So, then, I had to go and try to figure out which proteins were actually basically found and all my samples, so this dramatically whittled down.
The number of proteins that I could use, it actually went from 86 to 31 all the way down to five proteins and that worried me because I really wasn't sure if five proteins would be enough to give me any kind of predictive power but let's see.
So that data set is here so
basically I took a subset of 31 proteins that were found in all my field samples, but what I want to do today is, I really only want to focus on those proteins from the corals that I sampled in July
2019. I only want to know these coral samples because this is before the bleaching event. I want to know is there something I can detect in these corals before they bleach
So, then, that dropped my 31 down to five so I'm going to try to make a predictive model with these five proteins. And I'm going to use my favorite new feature of JMP Pro which is known as model screening, because this is going to allow me to test
multiple different models in parallel. So I'm going to put my coral health designation for the lab here. I'm going to put my five proteins here.
I'm going to give all these different options a chance you know what I actually don't have the XG boost installed on this computer but that's Okay, I think we can,
will still be able to find a decent model. I'm going to use a Kfold cross validation of five. I do want some of these options here.
And let's see what we get.
This is going to take a moment to run beca8se it's going to check 20 different models in parallel, using all these different factorial combinations and we have some diagnostic data here, but I want to get right down to the validation sets.
So they're essentially ranked here, so we have a neural boosted model that seems to be performing well. Because I'm concerned with accuracy I tend to gravitate towards the misclassification rate.
We have a generalized regression model, let's check that out so I'll select it. It's looking pretty good.
I'll run it.
You see all the diagnostics here, I do want to take a little bit of time
to look at the Profiler
because this is going to tell me
which proteins are contributing the most.
So what I had said earlier, I apologize, I need to reset it now.
I usually have the desirability to where one of my treatments is maximized, be it the bleaching susceptible or the bleaching resilient. Then what I can do is, I can say hey show me what it takes to be a bleaching resilient coral, so I'm actually going to minimize the bleaching sensitive
Well, what it's only showing me show me one option. I think, you know what, it's already it's already been been set to that so that's that's okay
it's showing me in this case
how the proportion of samples fall out with respect to bleaching susceptibility as I changed the levels of these proteins, and this could be useful for things
like genetic engineering, if you want to try to find the proteins that are going to be
most involved in the resilience of of coral. You're going to want to play around with the Profiler but I'm more interested in today is actually the predictions itself, so what I'm going to do is I'm going to publish the prediction formula here.
You'll notice
one of the proteins dropped off, so one of them was not deemed to be useful in the actual model building.
And this is something I've only ever done in the last two or three days, believe it or not. So I've got my July corals in this data table, and I want to compare.
I want to see the predictions it makes from that data table, so let's go over and see what it guessed based on that generalized regression model. And I'm just going to quickly run through here and tell you, because it's only 12 samples, whether or not the guess was right or wrong, so this is
right. Sorry the r doesn't work on this computer, believe it or not, this is right.
Yes, there's definitely an easier way to do this, but because this is wrong.
Right, so now it works again right.
Right and right.
So in this subset of 12 samples using this for protein generalized regression model.
It got the bleaching likelihood right in 11 out of the 12 samples, which is actually really cool, I mean I've run the simulation various ways over and over again that's actually
better than what I've typically been batting so that's whatever 92 93% so that is really exciting, especially because that's only for proteins. So
let's, already talked about that, so yeah so my average from doing this sort of simulation is about 80 to 90% accuracy.
Which for me is incredibly exciting. To a physician that would actually be terrible. If you're only right
you know, in terms of the prospect of a cardiac arrest event and 80% of your patients, that would be a huge failure, but I think
given that this is kind of new for corals, I'm Okay, with the accuracy of 80 to 90%. And this what I didn't mention
that subset of five proteins, which then got whittled down to four, that was only looking at the host coral. I've removed the symbiont proteins, because I I literally just got these data, a few weeks ago.
When I add those symbiont proteins in, I think the accuracy is only going to go up and also remember this.
I was using lab data to make predictions about field coral behavior. So you wouldn't expect it to be 100% accurate. What I should have done and what I'm going to do, probably within a few hours of finishing this talk,
I've got all the field data. I've got all the lab data. I can go ahead and use both of those data sets in the training and validation.
Make a model in which I predict field coral behavior. And I'm confident that accuracy with that approach is only going to go up. You could argue,
why would you even want to use the lab coral data at all when you've got this field data set? And you know, I think that could be a very valid point.
From a biological side you know the actual nature of these proteins is incredibly interesting to me and I just uncovered what they were last night.
Running out of time, so I'm not going to go into them, but suffice to say that's going to be incredibly interesting as well that we're able to mine out. One of them is evolved and prey capture, which I think can be super cool for for
trying to develop a mechanism.
This is, these are really small data sets these are dozens of samples. So do I think I made the world's best coral stress test?
No, this is, this is very much a work in progress and there's other methodological issues that need to be worked out, but really what I ultimately want to do is develop this kind of a stress
test so we can go out there and do what I'm calling coral reef triage. I could take my little samples
do my proteomic or other molecular analyses, input the data into JMP,
spits out these resilience guesses and I tell manager hey this reef over here looks to be in really good shape and let it be for now, you know, keep an eye on it.
This one over here all the corals were deemed bleaching susceptible by my model. If there's anything you could do to mitigate or to give those corals a chance - clean up the water quality, curb
overfishing - those are the ones where I think you should triage your effort. So really that's what I want to be able to do with these kinds of data sets. So of course the data sets need to grow.
I do want to compare them to this temperature based models that are out there, although the scale is inherently different.
And I really just want to be able to either have this kind of a visual or just a map where I just say you know here here are the bad spots here the good spots.
I'm going to do all the dirty work, I want to be able to distill this into something that a manager or lay person could easily digest and I think once the kinks have been worked out and once we have a little bit you know bigger data sets,
incorporate more coral species, extend the geographic range, we really could be off and running into using taking this coral health predictive modeling approach into
you know, as a decision-making tool. So, not just for monitoring, but for looking for resilient genotypes that might be useful for coral farming for restoration.
You could also use this approach for tracking the success of mitigation projects and things of that nature, so I think there's a lot of potential for this kind of
coral health predictive modeling approach for kind of this more proactive marine management, and I hope
this is something that we can we can continue to grow at NOAA, University of Miami in the coming years, especially as, as these ecosystems become ever more imperiled.
So with that I'd be happy to take any questions. I'm again I want to mention on I'm not a modeler by training, so I definitely welcome
any suggestions. Don't take any of this as being dogma, you know if there's a egregious misinterpretations of data or anything like that, you won't hurt my feelings and I definitely will appreciate any questions or feedback.
So with that I'll end my talk and be happy to hear from you. Thanks.

Thanks @abmayfield , looking forward to seeing the video during the Discovery Summit!