Streamline Your Processes with Digital Bill Format for Life Sciences

Digital bill format for Life Sciences

hi folks and welcome to this session of the big data and ai conference my name is bill wong and i'll be talking about ai healthcare life sciences i'm from del canada and i lead dell's artificial intelligence and data analytics practice coast to coast which includes healthcare and life sciences the bulk of the presentation will focus on our experience and working with some of those organizations and we've got time we'll talk about some of the things we're seeing coming down the pipe so uh agenda a quick update on digital transformation in these industries then we'll talk about our experience with an organization with drug discovery and their use of ai hospital in ontario and their experiences in accelerating ai and then again some time if we have some time we'll talk about some of the future challenges we're seeing so every two years dell does this survey on digital transformation and uh it's across 18 industries which includes healthcare life sciences and they interview or give surveys to 4 300 decision makers that's director c level across 18 countries and if we focus on health care here you'll see uh no surprise that you know with the pandemic you'll see less digital laggards now and more people in the middle there evaluating and adopting digital transformation initiatives now some of the challenges and in healthcare and not surprisingly i've seen this firsthand it's tough to get at some of the data in healthcare there's so many different formats there and that's probably one of their top three barriers across the industry 94 percent of businesses we're seeing face various to uh digital transformation but in healthcare the top three here are the ability to extract valuable insights here on data making sure privacy security is maintained and a lack of budget the good thing about this industry is they have great confidence in ai and three-quarters are are confident some are confident or very confident on the benefits of ai in their industry with life sciences as you'd expect they're they're a little more savvy with dealing with data but certainly issues with privacy and security are big concerns so we're going to talk about our experience with working with ubc and the vancouver prostate center and in the area of drug discovery it is a very challenging problem and if you're in the industry you'll know firsthand just how long and complex this would take i've had the opportunity to work with a number of researchers and outside of the complexity here and you can see you know all the tasks here and the characteristics of what makes a good drug uh it's it's such a you know challenging environment the volume of data difficult because it requires lots of resources in terms of compute it can be risky you know the the drugs that they examine very uh intensive and requiring capital and labor to test these drugs now traditional drug development done in the lab the estimate that most folks are citing is uh 2.6 billion dollars and this is over you know 10 to 15 years and the success rate is uh you know it is less than one percent so you can imagine um the challenges people have their risk preference but for trying to create these new drugs now over the past few years people have tried to use computers and they are in accelerating this but they really haven't developed that many new drugs here and here if you had a 200 petaflop supercomputer the estimate that it would take 10 to the 35th number of years which is a really big number older than the universe itself just to examine a molecule with just 30 atoms and whether or not this would be a good drug candidate we examine all the possible permutations here so so that's the complexity that we're dealing with so even with computers we don't have enough resources now in this um what they call in silico where we try to do drug discovery via computers rather than actually testing uh in a wet lab here molecular docking is a very popular technique here and it's the concept of a key and lock where you have the right key you will lock the normal function so a protein or a virus has a target or binding site a small molecule which they call ligand can come in and create a complex there where this new combined molecule will not function and that's the hope of finding these drugs and in order to do that they develop scores to evaluate the strength of the binding now here's one of the challenges here on the left hand side there's a zinc database so these these chemical databases or some people call them biobanks that have been growing and since 2015 the number of molecules that you can get from these databases have increased uh exponentially whether these databases are in the billions now and one here on the bottom left hand here enamine rs database contains 38 billion molecules and people when they test molecules as a possible drug they don't test them all they test just a handful and and most drug endeavors campaigns will not test 99 of the available molecules in a database now on the right hand side what the ubc team was doing was they had access to hundreds of compute nodes here but with all their compute resources which spanned research center ubc provincial resources they were estimating for them to do molecular docking of a billion molecules to take two and a half years and so again not something they could do feasibly the big challenge that not only ubc but all these research facilities have is how can we screen and do molecular docking when there's billions uh to test in a cost effective and timely manner so enter deep learning and so they've introduced and i believe they were first to introduce this concept of deep docking by augmenting molecular docking with deep learning algorithms they did test other algorithms but they found that deep neural networks were the best ones and predicting uh the scores or the binding scores and what we want to do here is with this technology is to go back to these chemical databases and to actually test all these and that's what they did so they test against a publicly available billion plus a database called zinc 15 1.36 billion compounds there and on the table here that you see here uh other research institutions were also doing this and their molecular docking exercise one institution had access to 45 000 cores another one 27 000 plus gpus and here we have at the bottom ubc with only four gpus they were able to do a campaign of over a billion molecular docking of the compounds there and so they were the first that we're aware of that published the results of a billion plus docking campaign and as soon as that paper was released was cited 200 times in in just 12 months and it actually got released just as coped was starting in early 2020 and then we had the opportunity to work together and what they wanted to do is okay we're the first in a billion and people are kind of using our technique now let's kind of up the ante here and they wanted a docking campaign of testing 40 billion molecules now with their current infrastructure they had it would take them more than 10 years they leveraged our hpc and ai lab where we have a supercluster there of of cpus and of gpus and they used our gpu platform and we were able to uh advance their research and um into uh while accelerating here their their compute requirements um to just uh 10 days so it is is a really a great exercise a great opportunity a privilege to work with them and really what they've started is this new era of ai accelerated uh molecular docking virtual screening of billions of compounds their vision is to build a platform where other researchers can come and use their platform to test their their possible vaccine so that's the drug discovery use case and i was with ubc this is a hospital in ontario they have a research facility and the use case that they're looking at was to try to reduce wait time and if you've worked in healthcare this is a universal use case very common requirement in all emergency wards and what they want to do here you look at the diagram on the bottom is as a patient arrives here is to triage data about that patient and the data that they want to get access to were health records existing health records and if you've worked with health records they can be in very unfriendly formats to try to get any kind of useful information there and then to automatically have an ai algorithm assess and deliver a diagnosis push out an order for test and then come back and see how the patient is after that now in order to do that what the institution did was to build a data platform to house all the data the the toughest data were the medical records and its proprietary data formats and healthcare probably has the most data formats that are proprietary very difficult to integrate that with uh let's say friendlier formats like relational so that was where the bulk of the time was spent was trying to get the data out of the emr databases uh once they get that to put it in a data platform where that data can be shared and then put an optimized platform where you can get access timely by the clinicians here and and so what was proposed here and what's on the left-hand side the architectural principles and this is what the hospital wanted they have currently um dozens of ai projects and each project has its own infrastructure its own tools everybody scurries and begs browse and steals compute and storage platforms and so what they want to do is okay we can't really control nor do we want to limit you on the type of tools you want to use but when it comes to accessing resources such as data such as compute such as storage let's try to standardize there so the initial was to create an ai sandbox where the architects developers would come in and the environment would be provisioned for them using platform as a service and there are some representative tools there and they would use a data catalog which is becoming much more popular in in building these data platforms and so happens that boomi uh has one of these ones all right moving on to a new use case and this one isn't in canada but it's a common use case that we're starting to see in healthcare that neither just ai nor classical simulations is enough to really help accelerate the drug discovery pipeline and it's really a combination of analytics that's required to provide insight in the covid case ai was used to kind of predict its molecular structure and while that's being done they would take their their hypothesis and run simulations to see if a potential drug or vaccine would block the viruses replication um this is an example of where we're seeing uh a lot of research pipewise is a combination of ai plus singulation and because of that uh what this slide attempts to show is on the left-hand side that the classical hpc type of workloads more and more they're going to these high-end workloads to do not just simulations by ai as well on the right hand side is what dell calls the ready solution so what this does it allows you to coexist um basically what tools that we generally see in research and academic environments driven by bright cluster manager with the ai tools that such as docker kubernetes etc and try to integrate that into one platform and so this one platform allows you to have both environments right cluster manager or you could have it through ansible based packages there and it would deploy these environments as required so regarding personalized medicine and if you're involved in any of these uh huge amounts of data but as i've seen firsthand the inflexible uh medical records they're not designed for precision medicine this is where where people are going to spend 80 plus percent of their time is to try to get that data out in a usable form and in a timely fashion again let's talk talking about a little bit about some of the futures here and again working with ubc uh what the pandemic has taught them is that um and even though they were ahead of the curve for the most part most of the world was was unprepared and what they're proposing here is kind of a defense system for other possible pandemics pathogens out there so again they would use ai to accelerate their workflow and and these are existing viruses today and and thankfully they're not um a pandemic around the world but there are unfortunately diseases that are much more deadly than uh than coated the first one here up to 75 percent of those infected uh perished so a very high fatality and the one below is their fatality rate of 30 percent this is a study i'm citing because it came out of toronto it's at a conference a healthcare conference at a stanford university and they actually studied this conference and called it a clever way of kind of assessing how people react with ai and in this experiment what what the researchers did was give folks x-rays and these are people in hospitals they're either specialists radiologists or generalists what they call non-experts regular physicians and they labeled the diagnosis of the x-rays either provided by ai or a human and the assessments were on the left-hand side that the radiologists were very harsh much more negative when the diagnosis was incorrect and on the right-hand side what also what it tries to show is that um given a a diagnosis generally they rated uh humans the ones that were labeled done by humans as of higher precision than if it came from an ai program so what what the paper's trying to do is uh say that you're going to have to consider these kind of reactions that the experts are kind of the harshest critics that document this new technology and there are many reasons why why that would be but it seems like the generalists the non-radiologists were much more apt less critical on an eye diagnosis so um that was a bit of a rush just these slides we just want to give you a broad brush of some of the organizations that we're working with and what's coming down through the pipe and as for kind of our best practices if you're in this this field what we always recommend is to have this data-driven culture and while we do have that in many of the hospitals in healthcare we do find that there are a lot of silos there still so the big challenges is to try to consolidate and try to harness everybody's efforts to try to have an architectural big picture approach we do see again because of the silo approach and trying to get something done quickly it's not necessarily optimized case in point that hospital that has has dozens of ai projects but everybody is sub-optimizing their resources right now and final point here is that regardless of the use case each one you find is quite new in its uh challenges so we'll stop there now and uh let's take a question