[00:00:05]
>> Thank you everyone for joining us again at this time I'd like to turn things over to our technical chairs for this final session Dr Christine. Thanks so much right thank you Maggie we have a great set of talks this afternoon I want to remind everybody to pose questions that they have to the speakers that they we will get a chance to after during the q. and a at the end of the 3 talks post I was having a chat and I will be archived and then.

[00:00:37]
Pick from them and that section. So I'm going to start by introducing our 1st speaker. In fact here Erika. Professor Denise cursed or I'm from my own insecurity University of Michigan. Dr Carson received her Ph d. in dynamical systems from Tulane University and the post she conducted her post-doctoral research.

[00:01:03]
With joint employment in mathematics and infectious disease she has been a perfect right University of Michigan for over 20 years in the Department of Microbiology and Immunology where she's now a professor her research focuses on applying mathematical and computational techniques to study questions related to host African interactions and I mean focus of her research is on persistent infections are packed with Africa have evolved strategies to abate or circumvent the host immune system so fat I'm going to have it over kiddies thank you so much for joining us today Christopher that introduction of I let me shame to say that Kristen I at the same campus but we haven't met which is just terrible but will rectify that postcode we don't have coffee yet into Lowe's at some point so thank you to John and Christopher for the invitation today and I.

[00:01:51]
Am going to just want to get the point you're going to my screen All right so what I want to talk to you about today is what we consider a crisp systems biology approach to understanding. Infection interventions and I'm going to used to work at last as my example but certainly things that we're talking about can be applied to many diseases and so I want to just start with acknowledgments I have a large group of fabulous postdocs and graduate students as well as cast members.

[00:02:20]
No Dr piano was on yesterday and even some of the work on mention today she did as well as club leaders at many universities in the country and funding from Ben atan Bill and Melinda Gates Foundation so I want to start by kind of defining what is a systems biology approach and.

[00:02:39]
We sort of think of it as multiple modeller these to address a single question so the way that works is. Well let me if. I lost my mouse is. There again. So so for example we use monkey models we use human they don't we use in cell a whole math and statistics and computation and we use microbiological and immunological approaches to understand the intersection of all of these to address specific problems related to disease treatment and so one of our goals really is to build a virtual human so that we can develop personalized medicine and we do that over multiple scales from molecular to the whole though scale which I'll talk recently about today as well as looking at multiple applications like that things drugs whose directive therapy biomarker discovery as well as something we call virtual clinical trials.

[00:03:39]
So one thing I want to point out today that is maybe a little bit different from the way we've been thinking about things the past 2 days is that there are a contrast between emerging and reemerging diseases so a lot of the diseases that have been mentioned the past couple days have been emergent and those are things like SARS and aids and those are viral driven pathogens and they are sort of newly afflicting humans I'm very interested in re emerging diseases diseases that have been around for a long time and caused lots of morbidity mortality in the world and is defined as their incidence of the the rising are steady and their geographical range could be expanding talking about things like cholera and tuberculosis which are bacterial driven diseases and malaria for example which is a parasitic driven disease and so very different pathogens from what we've been looking at but still as important to think about in terms of prevent.

[00:04:37]
So I'm going to focus on t.v. today as the example that we apply to look at these kinds of questions so before coded TB is the number one cause of death in the world due to infectious disease one and a half to 2000000 people die a year of TB and 3000000 people die of 3 people die every minute so during my talk will have you know 60 people die of TB And if this is a huge problem because we don't understand the trajectories which are either you Clara faction somehow you're exposing you don't get it primary disease where you really are infectious and if not treated you'll die and letting to be where you're able to control disease and you're not infectious However over time there's a risk of reactivation And so and that number of reactivation percent can be increased because of things like immunosuppression and drugs or aging and so this is a this is a real problem and so I'm showing you the bacterium and the formation of these structures which I'll talk about in a minute so basically you breathe in the Mycobacterium tuberculosis and local energy presenting cells and figures that excels take up the bacteria and that and they grow very happily inside but dendritic cells will soon scoot to the closest long draining lymph node where it will elicit an important adaptive immune response where these t. cells are then primed and research relayed back into the long leading to the formation of things known as granulomas and you can see here that not only are these t. cells able to activate activate marketed as we can kill their interest Saylor bacterial load which it was not able to originally but you can see that the formation of this structure where really it's a stagnate a stalemate between bacteria and infection where you've wall that off you're containing it but you're not really clearing it although many cases the granulomas do sterilized.

[00:06:33]
So it's incredibly important to realize that most infectious diseases specifically TB is is a highly multi scale disease I've talked about the granuloma scale but clearly events at the cellular and molecular scale are involved major organs I showed you lymph nodes and things cells traveling through blood and of course there's the whole host and then what's happening at the at the population scale we I'm just going to mention briefly we have an animal model we have a colony of 300 monkeys supported by the Bill and Melinda Gates Foundation that has a beautiful 5050 split between active elate and he'd be that we use to premature eyes and Belding calibrate our models so as I mentioned we think of TB as multi scale in both space and time we do something called a mezzo scale model where we start in the middle and then we build out from there so we put a lot of energy in building a model or scale that reads out of the tissue scale and then we consider the involvement at higher scales as well as lower scales and of course this all fits inside the framework of the whole host.

[00:07:39]
So just briefly what we've done here is we've built an age of base model x. we've built models of all kinds and we've written papers comparing the different approaches but we're really comfortable using this age of base model that has to cast it in discrete elements to it because it really allows us to track mechanistic interactions between the cells of the bacteria and the molecules like cytokines that are that are operating all within the long environment so if you think of this is to millimeter by too long lead or lung tissue you have portals from blood the lungs are highly vascularized where cells can enter as well as molecules and then cells undergo a number of roles and interactions that are all governed by known biology and so what happens here at the sailor scale we see the readout of a grain alone what we do not tell it to draw granuloma a granuloma is the natural manifestation of this agent base model it's called emergent behavior and so here I'm showing you you know has to chemistry slice of a granuloma from a non-human primate that's been stained to look at different cell types and here's our manifestation of a computer generated granuloma now what you can see is that.

[00:08:48]
Of course we match and it's not really necessarily important that we can recapitulate that except for the fact that they really can only get these Once they need crocs see the animal where we can actually track the evolution of these granulomas over time and so that this structure which has been a veritable black box for.

[00:09:07]
4 experimentalists now we can watch evolve over time we can manipulate it through a sort of knockouts and over expression studies that are done virtually sort of predict what's going on and in particular we can study things like drug treatment I'm just showing you this to indicate that while there is a lot of non human primate data it's only given to us sparsely and so of course the modeling can predict a much wider and longer range of outcomes that are still consistent with the with the data that's available.

[00:09:39]
So I want to send you back treatment for TB You know it's a very complex disease TB is different from all of bacteria and that most bacteria double 20 minutes like e. coli I like a bacterium tuberculosis doubles every 24 to 48 hours so the slow growth is really its unique virulence factor if you will and so it takes about 6 to 9 months with multiple antibiotics to treat TB Now most people don't even finish their dente 10 day course of antibiotics you can imagine the compliance issues getting someone particularly in developing countries where maybe access is limited taking in about Explore of them for 6 to 9 months particularly knowing that there are some side effects treatment is long there sport here and there is drug resistance that develops particularly in light of noncompliance these granulomas I showed you are heterogenous jenius We'll talk about that in a minute and and of course patients are Had are genius some people are fast when tabel eyes are as of drugs some people are slower metabolism all these things come together when you want to develop drugs and right now there are many drugs in the drug trials currently So if you think about trying to predict what the best antibiotic regimen could be you have what we define as the regimen designed space so you have these treatment segments Let's say you can treat over the 6 months and 23 month gaps you have 10 drugs although there's even more than 10 you can do you know numbers of drugs per the segment that you decided earlier you can very dose and next project again you can also very frequency very frequency should it be once a week every day every 3 days if you just look at the possible combinations of this you get something like 10 to the 17th options for antibiotics regimens even with models that's too complex so we even have to come up with ways to optimize before we can even begin to solve what optimisation problem really would be here.

[00:11:30]
So we've done that and we were able to take grand Sam our original model and include a plasma pharmacokinetic and pharmacodynamic model that fits and wear the drugs and through transit compartments depending on the drug what we know about the p.k. of those drugs into the into the tissues and track with what happens the killing happens through a something known as emacs and that is the rate at which half of the.

[00:11:59]
Maximal killing is half max there in terms of each antibiotic there's a specific emacs for every into about it in the system there's also drug interactions in our model so we take into account the interactions between the these combination drug cocktails So for example this just shows you that we're able to track the the numbers of doses in a 7 day dose once a day in a week you can see how the drug is changing over time and the different bacterial populations depending on where they are sitting in the granuloma and you can see that there are c. 15 c. or depending on which bacterial class you are and the drug is spending most of its time below those c 50 so no wonder it takes 6 to 9 months to treat.

[00:12:43]
And more importantly what we're able to show in the here is a granuloma that we're going to treat with 2 different drugs one is called for a fan pin and the other as I said I is that and you can see even after a single dose the risk is staying around actually until the next dose but i n h half life is so quickly it's gone notice also that most of the drug in this cranial Obama is on the outside of the granuloma and not within it and that's because we talked about the heterogeneity So you'll see the 2nd dose come on here in a minute on the heterogeneity of this spatial environment similar to what a tumor would be this heterogeneity really impairs the drugs getting in and certainly in this case getting to these bacteria that could be even trapped in the center in these hypoxic areas.

[00:13:30]
The other thing that we can do is we can apply sensitivity and uncertainty analysis where we can actually vary the properties of the drugs and actually see can we do even better with the drugs that we have so if we take an existing drug like ice and I use that and we vary.

[00:13:46]
Metabolism Rate or the plus the clearance rate of the sailor uptake just by a little bit 20 percent 50 percent or even 90 percent can we improve the that the sterilization time to sterilization or and the treatment failures and so we're able to do that and this allows us to be able to use modeling to really figure out ways to just slightly improve over what we already have so we've got an optimization pipeline that we've created where we have our granuloma we look at the p.k. p.d. of whatever antibiotics that we're looking at and and then we we have an objective function made up of potentially time to sterilization and dosage you always want to minimize dosage and minimize time but you can certainly fit other things into our.

[00:14:33]
Into our formulae here are just a function we use something called surrogate assisted optimization to optimize this which kind of helps you find the maximal on a topo map if you will and then when we make a prediction we tested virtually in virtual clinical trials which I'll talk about in a minute and then our best outcome we test in animal models.

[00:14:55]
So I want to tell you a little story of h r z e is the standard for regimen treatment that's used for t.v. And recently there was a lot of studies that looked at swapping out the age with moxa Fluxus and which is a floor Quinn alone and so we we did the study using these drugs including Moxi and what we found out was that yes definitely when you when you swap out you get improvement over the sterilization time from just the standard treatment but if you swap in it into a different position look at the we can actually potentially get much better sterilization times so we have a non-human primate study now that's got a series of monkeys that are testing and comparing all 3 regimens to see if our prediction is borne out and this would be the perfect thing to do before clinical trial to do it in the computation test in an animal model and then move to the to the human set in clinical phase one so that brings me to virtual clinical trials there's lots of different ways to do work to a clinical trials our goal here is to sort of take things from a granuloma scale where we have multiple grey Lomas and along and move to a population scale to really test our predictions so to do that we actually need to have more of a whole host model so that also allows me to talk about vaccines which requires a more whole host approach so the vaccines for TB are attenuated like Back seen with something called b.c.g. and it is given in most of the world but not in the u.s. or the u.k. it has very low efficacy and it's waning protection over 5 over time it might be good between 0 and 5 years old but over that time it tends to way which is why there's such a variation here however if that invalidates the p.p. diesel contest which we use in these countries and so we'd rather have a skin test which works really well versus a vaccine that's not so educationists there are also new tests for TB test test.

[00:16:55]
Response and the b.c.g. can also give results with that of about 5 minutes left. And so so human studies go into Phase one after mouse studies non-human primates are also being used so the question is Can modeling help decipher differences in that Zene responses and predict better dosing so a few years ago so he rose in a group of us wrote a paper on him you know stimulation immuno Dynamic Modeling So let me just review this ideas what we do is we figure out the best range in a dose what vaccines What's the most immunogenic and then we pass it that to humans by scaling up elementary Glee so there's a lot of problems with that idea and maybe the immune response between the animals and humans are different the scaling may not be the only thing it could be nonlinear scaling and so that may not be the best approach of course we know I've been talking to you about the side of infection but there's lymph nodes and blood involved and so we've built a model now called Ho sim we're not only do we have long as that have been these lungs were taken from c.t. scans and x. y. z. coordinates translated We have our granulomas form of there along with blood vessels as well as multiple lung draining lymph nodes and we now have that model up and running to test these things so for example we we can in effect the Longs and you'll see through the blood the different cell types and how that ends up and you can see over time the grain Ilam is beginning to form so we're in a position now we call this our whole host model because in TB with pulmonary disease most of the story lives in these physiological compartments something important that we see is something called recall response so therefore after time if you introduce infection again you get an immediate and higher recall than you did originally.

[00:18:51]
So what we're proposing is that not only have to get information from the mice so you do some dos prediction relationship in the mouse but then you should actually elementary scale that to humans study it again with models and then go into that humans for phase one and this it should cycle through this process here rather than just having the line go from one to 4 The other thing that's been really useful is we've been using our models to predict differences between what we see in human results and for example in this case non-human primate so there were some beautiful that seems studies done on a new.

[00:19:28]
56 vaccine for TB that was done in one human primates in humans and so we used our model to predict what makes sense to translate between the 2 and what things didn't with believe that helps inform clinicians and experimentalists on what's the best targets to move forward and the setting of the vaccine and finally we've developed a number of tools to allow us to do these translations uncertainty and sensitivity analysis model calibration tools we do something called tunable resolution which allows us to course and fine grain our models that will depending on which group compartment we're focusing on that helps with computation we have primitive identifiability tubes optimization tools and of course we use machine learning and what we've done there is really exciting we'll take limited data sets from animal studies and humans and we will add think that it data sets together and then we will do machine learning on both.

[00:20:22]
So all those papers are available on our website and I'll just end with a summary of what I think our strengths and barriers of the current approaches I believe messed mechanistic models are key to identifying processes driving outcomes prevention efforts for diseases are important for those that are and Demick as well as those or are an emergent within his modeling can significantly and identifying that seem targets drug regimen optimisations drug development and drug targets virtual clinical trials can speed time and save costs right now there are $7000.00 critical clinical trials for TB That scene's each $1.00 cost a $1000000000.00 We've examined them and we know that at least 5 of them will not work modeling can link with Coast infection dynamics to blood readout so people to biomarkers from blood modeling can help figure out what's going on inside to that compartment.

[00:21:16]
And help with biomarker discovery modeling can distinguish between animal and human studies as I mentioned but these 3 are really important goals that also can serve as barriers if they're not considered 1st there's got to be collaboration between experimentalist clinicians and computational model Earth and grant support to foster that considering Kobler Mitty's most of the places where these diseases are endemic co-morbidities are extreme even if you think about 1000 heart disease in the us was a major barrier in people that were getting disease here and HIV and TB in Africa were major barriers and then the other thing is that you know as you build these more more complex models we do need access to high throughput computing options because we all have super computers in our universities or in our labs.

[00:22:07]
So that's all I have to say thank you so much. If you very much and that's really interesting I mean I have adopted John is going to introduce our next speaker you know have questions after him. Yes I have the pleasure of introducing. Pin Kumar who is a regents professor and also William Norris chair in the Department of Computer Science and Engineering at the University of Minnesota his current research focuses on bringing the power of big data and machine learning to understand the impact of human induced changes on the earth and its environment Professor Kumar's work has been honored by the a.c.m. sig Kay the innovation of war which is the highest award for technical our excellence in the field the knowledge discovery and data mining that's cute each and the he's also recognized by the 2016 I triple each computer science Sydney Fehrenbach award one of the highest award sent high performance computing Dr Kumar We look forward to your talk now of thank thank you for the introduction and for the invitation.

[00:23:24]
And you know my background is in computer science in my interest isn't basically bringing the power off machine learning in Big didn't technologies to a moment of problems and you know some of the. Problems on the one interested in is is this is find a sort of see if you can understand how the climate especially the extreme weather may be changing the spawns to human actions or natural variations and more generally how ecosystems like our forests lakes and there were bodies and in the other and Ramadan system that we depend on are again changing in response to climate change and in human actions so we are also trying to make predictions about the future and I guess that's the commonality and I would ask to sort of share my experience in bringing.

[00:24:18]
The machine learning and scientific modeling together in these context and sort of what lessons could be learned from experience in these domains the kind of big planning and I was listening to the park since morning since yesterday and and then and this morning and I can sort of see that there are huge number of commonalities and.

[00:24:41]
Many number of many many issues rather come up in various talks including the one just. Just just prior to my thought here it's clear that in pandemic modeling you're looking at phenomena that's happening at multiple scales you know all the way from little scale to. The daily and weekly and seasonal and of course population scales there are.

[00:25:05]
Multiple disciplines and also you have obviously quite Him science that is yesterday you. Mentioned that you need to sort of evil to create data from small amount of data and since again. You're treating the Atlantic when there are forecasting of course you see that they have been created there's a need to adapt to data that's becoming available like good assimilation sort of the example and above all you need to handle we have to do 90 which is from in came up in a question earlier but in terms of the concept dresser all of these are issues actually that sort of come up very much in any scientific discipline and many of these challenges have to be handled and you're sort of trying to see whether bringing machine learning and scientific morning together can help us address some of these challenges better than using minimal.

[00:25:56]
Technologies by themselves so that I'm just going to for it will be a very sort of a brief high level overview and I'm here of course to learn as much as to share what I what we sort of know here so in the sense of if you're trying to model it a logical system one moment of system whether reply McCue looking at basically some sort of a physics based model or by a physical model of inanimate and system which can be represented as a system that has some input which don't really varies in space and time we have an output but the system has a hidden state and commenters and there are some physics basic questions that sort of would at any given time given the inputs in the state will create a new.

[00:26:41]
Output and also some of those equations would also modify the hidden states and of course these maybe could have been equations or could be because if it models some on some combination of these and you can sort of see the spare time unfolding in many many domains I'm giving you some examples here from ideology of aquatic sciences and we've got a mix but if you look at the.

[00:27:04]
End of the turning you will sort of see something very similar your mind to. You in the band is not in the output for example might the outbreak. Well and so of infections and such but the input would be again the parameters that our input would be sort of the situation there are sort of unfold in the society where I'm at or perhaps could be policy based solutions so whenever you have mortars like these you're trying to make predictions and you sort of lead him to go out in the sense that the models all models are wrong somewhat useful basically the models are never perfect they almost have incomplete representation of phenomena or certain aspects are completely missing and again these are because of be scale speed accuracy there are because you cannot model everything even if you want to model everything and even if competition was a non-issue you would be able to find a better driver so that means you're always making some approximations or better or worse and as a result you have inherent bias is.

[00:28:04]
That in result in your predictions and because you have phenomena that you're trying to model at multiple scales and you can only model it at certain scale or limit or type of scales you model the other aspects using some parameters now and again you can sort of look at any pandemic.

[00:28:21]
Model and you see many many parameters well but that happens for any physical model as well and these are known planters have to be calibrated if you have some limited operation there and they use them to figure out what value of these parameters really fit the operations and of course the the this is sort of the this is a caricature but the but it's going to get a little ice very widely to many many scientific situations firstly and moment of situations or for example one of the parameters if you're trying to model the basin of a river or for.

[00:28:55]
Hydrology for being able to flow in a flood and in the spawn still rainfall for example some of the parameters might involve knowing how the quality of the soil is a different levels. And of course a lot of those things are not even visible so you have to estimate what they are from the data and then of where you're trying to estimate its monitors is often time it will be easy to overfeed them to the observations and because the reality is much more diverse much more happen genius and again you can see many similar situations in the pandemic.

[00:29:31]
Of the one alternative that people have tried to. Pursue in response to the limitations of the physics models and because of the excitement of this black guy that modern science models have created in his best interest in the earth is to see by the big could be used as an alternative to modeling with an inconsistent because you know there are you could use them perhaps if you have plenty of there are you could possibly use them directly in the old world and as I've been a sort of a.

[00:30:04]
I been fallings at least some of the literature of the pandemic planning and I've seem to have been sort of papers coming up which are sort of trying to make the projections purely from data without having to have any underlying theory behind it so people. Have been troubling that we're going to be planning as are these projections well so there are there's a huge tradition of these there is ice models dating back to you know many many decades but most recently for example beat learning has become extremely powerful tool and which is already found hugely successful in many commercial applications and are increasingly being pursued for applications and domain but in many of them they have intra disappointing results because these.

[00:30:47]
Sort of the are complex the learning of wasn't a very powerful can be very good if you're lots of data without that they could have disappointing results they can have some very very poor results they can create in addition to that they can create physically inconsistent results they can fail to generalize wanted scenarios they can you know the un able to provide valuable insights into the sponsor the community the scientific community and the people specially those who are interested in building machine learning into the community together have been pursuing this dichotomy between scientifically based models and they did as I said model and if you look at each one of them surely based models are limited by our current size in the standing order we're able to model the phenomena of that level of detail.

[00:31:33]
But it isn't models can do really well if you have lots of data but if the data isn't the representative it happens the case many many times on the limited in those situations so the new paradigm that people are sort of trying to pursue is can we bring a hybrid can we build a hybrid of these 2 models can sort of bring it as ice models which is guided by scientific theory and direction of course has been pursued by many people we have been talking about it.

[00:32:01]
As far back as 2014 but should not in more recent times of become very very popular and huge numbers of people in different disciplines are following it every discipline you can to find some examples of people pursuing it more recently just a few months ago he wrote a survey paper which had about 300 papers on the in just the last few years that are such programs and conferences and workshops devoted exactly to the subject so this is a very very vibrant.

[00:32:31]
No misunderstanding and this is my last right hour and I know I should be short of time here so the scientific community which is sort of pursuing this ecological progress thing and momentum for casting is sort of trying to figure out how do you bring the physics or the scientific theory and machine learning together and they're trying to pursue some of these questions because we just wrote Mr them here and as I listen to the talks over the last few days I can see each of these questions sort of being relevant for the discussion we've seen of the pandemic planning hour and there's a lot to learn.

[00:33:07]
From what we have done and the number of Sciences and what perhaps could be done in the in the pandemic planning for the week with an order to a collaboration there separately possible and just to take a minute to sort of go to these questions some of the biggest question that we are trying to pursue here in the us who are in good pure United or if you just got a machine learning Marling is this question is can be and for outperform your physics model you're going to signal a pure mission minimal That means you can provide and you provide better accuracy with less of the vision that I can produce result that are physically consistent and be generalized to novel testing scenarios to sort of.

[00:33:47]
Handle the consecrate and things like that can you handle it complex that is a collection of process is where the process and fully the different scales and happens are either looking at the head logical system or the rainfall happens on a daily scale the soil moisture changes much more slowly the snow back again you know take weeks and months to form and to go right now we have no one on it and in the entire country so I went to the so called snout and which was apparently caused by the North Pole using a lot of ice so the whole process that are unfolding at very different.

[00:34:30]
And you assimilate new information and data because any system has multiple parameters not just the one that you are trying to predict and some of those values have become available to you from time to time and being able to assimilate that information is extremely critical for example to real good weather forecasts same thing is good you can assimilate new information as it becomes available partially information that you can belittle in this physics knighted machine learning model yesterday Martin Marty talked about being able to create data out there for public consumption which is necessarily at a much higher delusion than what's available from this model in the machine learning community people talk about sort of a solution going to data to much finer data and you make the solution method methods which is guided so they can be able to handle some of these complex and the mental application so there is an obvious list this question that are being pursued in our group but also in many many other groups as well and I see huge relevance of what's going on here what's of the pandemic planning and I'm looking forward to the discussion in the breakout sessions and of course in the question answer session us I'm going to stop here and again if any questions later Great thanks Dr Kumar for a very stimulating talk of turning the mike over to Christopher for us to introduce our final speaker.

[00:35:56]
All right next item to announce our final speaker and Dr Reed a call well and a distinguished university professor at the University of Maryland at College Park and John Hopkins University Bloomberg School of Public Health. Let's see her interests are focused on global infectious diseases water and Hell Dr Colwell developed an international network to a draft emerging infectious diseases and water issues including safe drinking water both developed and developing world a collaboration with the safe water network headquartered in New York City and she also served as many of you know as the 11th director of the National Science Foundation from 198-2004 but then our energy future I'll pass it to you Dr Colwell Sorry about the computer me some of.

[00:36:43]
The previous talks in super. Didn't east from the molecular expanding to the most recent So talk to me on the modern system but I'd like to do because I I think this all fits together exterior Well you know just to show how we're taking a disease color and this is a global disease a water related disease burn we are in the southern stand on that in fact we were in 2 pandemics we have been in the 7th pandemic of colors since the 19th sixty's and then overlaid with Kogan but I'm going to show you I hope that you will be able to see that what we've learned from college we're able to apply to prediction color occurs in about 50 countries 6 millions of people we I'm seriously labeled the Bengal Delta is the homeland but 100 years ago cholera was in the u.s. in Europe it's actually the bacteria we made the discovery some years ago that the bacteria are naturally occurring in your environment and they carry out carbon and nitrogen.

[00:37:56]
Cycling's lake they really can't be eradicated we started our work very early on in the Chesapeake Bay and made a series of just covering one of which was quite critical and that is that the bacterium sociate it with a host which is the Coca-Cola 120 and in the world oceans and is them distributed throughout the world in fact we have isolated a real call reef from Iceland where there has never been columns to the bacterium exist in the car to exist on the coated lot of it is its host and I would prefer to call it a vector it's a vector the does move why did it swims now less the bacteria are associated in the God on the Gill and certainly on the a case that you can see in this gravity email playing a symbiotic role probably simply only cannot suppress your young producing a very powerful enzyme that causes the exact break and release the extensible article we developed very early on about one point years and 30 years ago a crude model of the transmission of the real collar read from the environment to humans mainly just unsafe drinking water now the criticism was that we were doing all this work in the Chesapeake Bay and had no relevance to the poor bunga dash which is where color is endemic So you know it's not easy we started 25 years ago and showed quite clearly that when you take water it has the villagers in remote areas buying the dish do.

[00:39:30]
You take the water of this contaminated and direct translation and indeed to the model that we enhanced by the work in Bangladesh clearly show that yes there is person to person transmission is highly dominant because of the very poor sanitation but the environment is really the source of the bacteria now yes occurred to us that.

[00:39:56]
At that time we were doing this work very early on when the fact that the launch and in the eighty's and ninety's. Being it was being used to measure various parameters of the environment chlorophyll sea surface temperature and sea surface height were free parameters measured by the centuries long on the lance that's not a white man's ought we made this discovery and published this 21 years ago showing that there was a quite strong linkage between our cholera case and good way of Bengal and.

[00:40:36]
The sea surface temperature so so this is my understanding that 1st you have this chlorophyll measurement and then you have 6 to 8 weeks when the. From the fight of my can do these ultimately apply to plankton So with this correction you get this very very interesting that there are a lot of other parameters that have been environmental branded associated with the outbreaks of cholera flooding rainfall heavy rains the surface temperature of I'll just mention salinity is a very strong factor and yet just being we know that charge was that the contamination was directly related but we found in our work that it's really in inverse relationship that don't cover that Drea present naturally and then has to cause by the increase neutrino so that the number has become much greater chance margin numbers of cases we don't tell us with improved satellite sensing with the data we've been gathering we've improved the model significantly and so with some of the more sophisticated set of satellites we have moved from 2025 years ago using ransack to the.

[00:41:56]
Present time when we use a variety of satellite data gathered and improved with the model that we've been able to develop shown here is an analysis that we did of hurricane Matthew which traversed. Haiti and and caused another massive recruitment crescents of cholera from the 2010. Introduction. To shows an introduction that I'll talk about that minute you know any case what we did was register active week that 2015 cholera.

[00:42:37]
Research resurgence in Haiti and we did a calculation and then on the left you know from our more sophisticated model the prediction of the risk of cholera and then you see on the right the actual cases that occurred in 2015 we we didn't now see this you know in Yangon the outbreak in 2017 was the worst call Rupa Dimmick in history really are millions died we then again did the analysis refuse my Dibley the risk map is only half blue and red and then the actual cases on the Right now this was picked up by the British aid agency.

[00:43:20]
In January of 2018 in fact it was by a telephone call where we were asked by the good chap of the British rating agency if well we would collaborate and we would provide predictions for Yemen and they would locate positions medical supplies safe water reception and so we have done the corroboration we're able to provide cool weeks in advance and this is on a rolling basis every week or every 2 weeks depending on on whether the risk is high or low and this.

[00:43:56]
Allowed us. To provide more for young women now working with the British aid agency just the British we're a logical agency NASA providing funding for our work and Unicef and so this continues to the present and we provide these wrist maps for the British aid agency and for Unicef on a.

[00:44:23]
Monthly basis and then we a bucket up to every 2 weeks and they're really depending on when the projection is so severe I'm going to miss last article well beyond just remove Very quickly to you know skip this you know I'm going to cope. With Colvin. Now going into details we're all familiar with it we're able using a d.n.a. sequencing technique to isolate identifying you know guys like my junk my and see which samples the the color starts to Barra's and it's very based on all the work has been showering that would sample here if you read from the top to the water and 41 patients read across the red line is the throat swab was a little the yellow orange line shows how long after going on the wires is are released into the stool and we're currently measuring then using this technique.

[00:45:28]
Will be analyzing sewage we're able to pick up as many others are doing right now this is turning out to be a global phenomenon but you know as an area of predicting health in a community sense for them all what we've been able to do is modify the model that we've used for the common for cogen substituting or introducing other parameters such as humidity and dew point of that years and using satellite data.

[00:45:56]
Populations were dropping and that has given us the ability of Iran to project. And I'll close here sure and not what you have seen in the talks. In integration really driven from the molecular to. The. Space. How the environmental factors are very tightly integrated with disease and production is clearly possible with the various business to be able to.

[00:46:34]
Thank you. Not sure but you know you know right this time we're going to move to our panel discussion so the article invited her to stay more vital of the prior speakers as well. Continue to make themselves available as we address some audience. Questions and so on one time back over to the sessions with.

[00:47:07]
Great thank you. Crystal would you like to lead off yeah I believe and I'm I'm maybe not can't turn from Atlanta here all right so let's see the 1st question I have I think a comment from a 1st trucker Dr Kirshner Thanks for the talk would you elaborate a bit on how you merge synthetic and real data for are doing and how do you treat them as equivalent so that's a really good question.

[00:47:39]
And I could point to the paper where we did it but what we do is is try to simulate the exact same study as the one human primate study and then and then you know generate the same type of data and then merge it so that it really increases basically in size of our study and that gave us that the actual statistical that we needed to do the machine learning on.

[00:48:05]
So I can follow up now this is a question for Dr Kumar do you think key g d s p g m l takes away and I'll ask you to define those takes away from the surprise findings side of data science in cases such as endemic protection and moderates do you think that this t g d e s n l can help especially since the models could get very complex and heterogeneous with multiple domains that ball you know has you also to help for those who are not in the know what that t.g. ts t.g. I know that yes yeah I think the T.V.'s t.d.m. no less wealthy.

[00:48:50]
Acronym there was sort of being used for the purely guided data scientists like the machine learning being guided by scientific theory which are. As opposed to building if pure black box machine learning algorithm because the tradition in machine learning with that you give me that era and I will build a model language translation for example uses very little linguistics Fury's and yet you can pick your smartphone and point to any language in them anywhere in the world and it does a pretty decent job of translation so.

[00:49:22]
I guess the question is is what about here you got it at a science is it taking away the surprises or is it would you still get surprising results and I I think I guess this question can be sort of looked at from 2 different perspectives and one of them is that if you are constraining your machine learning paradigm by traditional theories is it a chance that you will not find something surprising because you're always going to buy the theories and my answer to that is that what it what it helps you find is the missing theory in the missing physics because you can never have perfect model you always are making approximations and on and if you're here you got a machine learning model is doing better than your mechanistic model and there are a number of reasons for it and one of them could be that it's not representing.

[00:50:12]
The phenomena that I level of detail and as and as difficult as a hypothetical answer but a very concrete lustration of this is in the context of what we did with the u.s.g.s. which is a sponsible for monitoring all of the water bodies in the us there are so $100000.00 of them and they try to figure out the the behavior of temperature how the temperature is changing the next in the sponsor to climate change and the human actions because that determines the water quality the kind of fish that can live in them and it turned out of the history of the argument the operational model that they were using was much much worse than this the joint machine learning and a model machine learning model that did most of the learning from the physics model and yet the new the bigger to refine it and you can sort of figure out that there were certain aspects of the modeling actually were very big and they could parties really improve them so indeed yes you could still get surprises and you could have these as these methods produce operational results on a national scale one for example with us yes.

[00:51:19]
We are talking about replacing the entire set of actions for all of the other bodies that they do in the entire country using this paradigm now Greg it's very interesting thank you. Chris the. Other question for Dr Colwell you men and I think one thing my spark Marianne has wondering if you could.

[00:51:42]
And I share with I think it might be for him and how I don't know him and comparing my clinical data on my Marion yes movie I apologize for not having organized the slides that we have very serious medical emergency in the family the last couple of days but out there the the answer to the question is that we've been able to show particularly you know working closely with the team in Frederick Maryland wonderful team of public health as well as the wastewater folks found over the last the struggle months we've been able to terminal peaks you know the numbers of viruses in the race water it.

[00:52:26]
Is about $7.00 to $10.00 days before you begin to actually receive cases appearing in the community so it is so early warning system for the community to be prepared in addition. We also now have been storing samples since last year when we started our work and we are able to identify the variance so we're now being able to detect the presence of the variance in the wastewater in these 50 different locations in the in the state of Maryland and all these include not just the our waste treatment plants the dormant to raise.

[00:53:05]
And outlive the the way smart Let's get a system assisted living facilities so that it's possible as we've been able to show but one of the conduit used that we were able to detect the presence of the Mars coming from one dormitory and then everybody in the dormitory could be tested rather than having to close down the entire campus so it's become a very useful tool.

[00:53:30]
My last. I have to go to like to emphasize is that I wanted to see this kind of community analysis done for a long time because I didn't have time to show that we could have done it with Orange County child when you were out there using race water to mean and from the sewage pond because of the drought and we were able to determine the presence of wheat even coming wastewater.

[00:54:00]
Agents swipe your outdoor viruses at Novara. A trip to spray about such medication we'd never dream of as a really important early warning system to the community and I think now everybody is doing this kind of analysis that this will become a tool that assured poor public you know.

[00:54:21]
That's fabulous. Follow up a question for Dr Kirshner you mentioned the challenges of our metrics dealing in applying Mauser non-human primate data to predict what will go on into humans and not on what I think this is an issue not only for to vote for the most distant from biomedical science right because a lot of work is done on animal lots and I would wonder what kinds of alternative approaches might you suggest as well alternatives to automatic stealing to allow us to use mouse or other animal data to help predict what might happen if humans.

[00:55:01]
So and that question goes hand in hand with another one I saw that said you know how could the studies I talked about today translate to virus infections but I thought TB was the example I use today but I think in general you know the approaches that we did are definitely amenable to any of any infectious disease.

[00:55:19]
I think it we've done similar studies with HIV where we had a blood in a lymph node model so it you know similar studies like that could take place for viral infections. So about Illmatic scaling the idea is that I think it's a non-linear phenomenon so even though you know we have little we're big and we try to scale it the idea is that there is this big black box in the middle and so that's where we believe the modeling can come in and I think if you did dosed response studies in the animal and then fed that to dose response studies in a computer simulation for a primate level model whether it's human or not even primate then I think that you're better suited to predict the dose that's going to work in the human clinical trial and I think.

[00:56:10]
The idea is that I just don't think it's a one step I think it's a multiple step process and yeah so obviously they're great thank you. All right we have another minute and I'll kick your start break out the last question is for Dr Kumar I can much lesions in animal biodiversity gave her happy clappy integrate with climate change.

[00:56:36]
Of course yes and I again you know if you have reserves that you're producing on a global scale that control of the very helpful in many ways and I just give an example in some of our own work we have build did a base of about 700000 water bodies around the world below 50 degree that is you just the most complete Did a step up and it's dynamics as to how these water boy have been changing growing in training over the last 20 years and if somebody has a biodiversity data on a global scale it can be put right next to it you can figure out exactly how the changes and these water bodies you know in the links are shrinking and it was going to be somebody else who printed water it could be done so basically as some of this methodology is many of them that you know what he would have by this but just kind of machine learning paradigm but able to produce global scale results it creates opportunities for us to Mars the students have to buy diversity.

[00:57:33]
Things as well now I'll be happy to talk to you for the behalf to talk to the questioner for the one on one. With the streets example where I think we can pass off back to a man you know thank you very much to our state 1st yes thank you so much was that there was an had a burden for in virtual round of applause just a properly bank everyone for their contributions thank you so much to the speakers for taking the time to be with us in this year from their experience and that we've all learned a great deal from this conversation.