Duration 29:47
16+
Play
Video

AI in Drug Discovery By Nataraj Dasgupta, VP Technologies, RxDataScience

Nataraj Dasgupta
VP Technologies at Rxdatascience Inc.
  • Video
  • Table of contents
  • Video
Video
AI in Drug Discovery By Nataraj Dasgupta, VP Technologies, RxDataScience
Available
In cart
Free
Free
Free
Free
Free
Free
Add to favorites
23
I like 0
I dislike 0
Available
In cart
Free
Free
Free
Free
Free
Free
  • Description
  • Transcript
  • Discussion

About speaker

Nataraj Dasgupta
VP Technologies at Rxdatascience Inc.

Over 21 years of industry experience in developing the vision, strategy and execution of analytic capabilities in finance and pharmaceutical domains. Experience at IBM, UBS Investment Bank, UBS Wealth Management, Purdue Pharma, Philip Morris.Core architect of the RWE and Rx Data Analytics solution at Purdue Pharma which was spun out into a new startup, RxDataScience with over $ 3.5M in seed funding. Member of the core founding team and VP of Advanced Analytics.Published author of multiple books on data science, journal articles and research papers in commercial market research, health outcomes, epidemiology and RWE Analytics. Presenter Keynote Speaker and Chairperson at over 20+ machine learning and AI-related healthcare conferences in US, Europe and Asia.

View the profile

About the talk

A key area of application of AI in Healthcare has been the field of drug discovery. Excessive R&D costs, long timeline and constraints such as resources and logistics have created a burgeoning industry that relies on AI for drug development. The talk will start with a background on how AI is being used generally in Healthcare and then specifically how techniques such as knowledge graphs, natural language processing and pattern recognition is used today for drug development

Share

If I'm based in NC North Carolina in the States, and what we do is essentially a lot of data science and machine learning AI related initiatives for pharmaceutical organizations. So what I'd like to do today is start of giving overview FBI indicted in drug Discovery just a bit about myself to start off on the BP of advanced analytics that are SDS. I have over 20 years of experience working in the analytic space. So the way I actually got into the farm is the lights to be at a high frequency trading

on the trading floor at a Swiss bank and Ice to be in that high-frequency trading analytics place. And what ended up happening was that got into Pharma in 2013 and ended up building a phone using the tools that they had on the trading floor. Eventually, what happened was this this platform? It was used for everything from Argentina to commercial market research a Tetra it got relatively popular and the some investors wanted to sort of spin adult and commercialize it so that's how I stay the signs came to be today. If you have for

Global offices for RX8 of science hundred plus staff, I think in four years, it's been it's been a good run. I'm also a published author. So I books on Amazon mainly on data analytics and data science with r multiple research papers that stare on Healthcare and my background is also an analytics academically. So without further Ado in I'm going to get into the main subject. So I started off and I have a ride it some examples to start off. Of AI projects that we have done as a firm now. I mean, I would like to

distinguish between in a machine learning and AI to me a lot of these projects we fall into the realm of machine learning because you know, you're adding another layer of complexity, which is you're not just learning from a training dataset, but you're trying to synthesize mean while just as you wouldn't drug discovery which how to get you in a moment. So do me a lot of these really come down to machine learning and some examples are physician matching for instance can match up in a physician's with patients Market mix modeling differin commercial research then in what order should

I send my sales reps to visit Physicians? Should I send an email servers should I send them is should I send a certain message a tech Residence Inn in Touching of sequence that you can model which is optimized he returns on your product. Alignment physician similarities are there several different ones that are being done in the real world in pharmaceutical organizations against that in Industry. I presume you. Do you know who wants that wasn't really that

active to the extent Finance was has gotten so much involved with the word of a i n So in healthcare, yeah. Yeah, they said has been less prominent mainly due to certain issues for instance regulatory restrictions. There's a huge impediment to research that's caused by the fact that you cannot access certain data sets. If you're in finance, you can get stock market data very easy black box. So when you're presenting, let's say the results of an algorithm and you're using

deep learning. It's very complicated to explain that to the center of the individual supposed to eat the Black Box. You really don't know precisely like you would if you were doing statistics or econometrics, how do you model actually works are or how the output actually is produced now? I'm not saying no one will but it's not as simple as doing let your panel regression. It hasn't been a major Topic in healthcare Academia doubts on the true Effectiveness vs. Existing methods. So this is something we come across time, you know on and off is this really

better than what we've been doing for the last 10 years and lost me not enough momentum to sustain long-term interest. Not today. This ecosystem has started changing biotech is the new Big Pharma this retrosynthesis from a purpose and buy Market development small molecules. Excetra is a lot of activity going on which leverage machine learning and there have been numerous success stories rights of benevolent AI called Pharmaceuticals numeral, Texas Township. So these are some of the friends that are really being in the Limelight a few of them at least in

terms of how they have successfully used a high for their business purposes. Know what I would be showing here in the next 20 minutes or so. And then I have questions are sticking to four sections. So I'm going to talk about the a I'm at the largest for drug Discovery the challenges to this methodologies building an organization a high culture. And finally the AI technologies that are used to implement the patients. So first off, you know, this is something I presume mean if your being in the industrial field been working in this field

many of y'all already familiar with it the drug Discovery life cycle. It's non-trivial. It takes 10 12 15 years to create a drug and the first step is Discovery and development. So finding a new drug that will that will treat a certain disease condition. Once that's done. That's a huge step in and of itself, but then you have preclinical research laboratory in animal testing then there's clinical research then there's enough And finally, there's a postmark than the spinosaurus Marcus safety monitoring is you mean that the product has gotten

approved but what this really means? Is that if you look at this chart or around slide new drug development is very expensive and lending and the reason is there's all these steps they they're not like, you know, it's not like running and I'll guard them are doing and I ate your Finance project to know when you're making a drug. There's so many considerations in terms of the effect. The drug will have a cross a diverse set of people that you have to take into account. So

today approximately is 2.6 billion dollars 10 to 15 years of research and after which it still might not get approved. So what has happened is that the efficacy has gone down to your you're going to 10,000 compounds of which one will get approved. Why the R&D costs have gone up. So in 2010 if it costs about a billion dollars to make a drug and the peak sales for 800 million in 2019. It was 2 billion where the Peaks on average is 400 million. So you have the returns that have gone down and the cost of going up obviously, so Unlike many other

use cases such as the ones that I've shown here, right patient-doctor matching patient find her feet when Pathways Medical Imaging Caesar predict prediction. Those are not that expensive in terms of executing those project what is expensive art of the two areas and very expensive is drug Discovery clinical research a preclinical research and those are places where I think machine learning Rai more generally speaking can be very effective. I'll get to that

in a moment. So first, let's see what it takes to create a drug and what if we lie down here is that I'm agreeing with you most of your already familiar, but nevertheless I would like to give a quick background as to the Genesis of a drug and and thereafter. What are the areas where we can use machine-learning Rai to optimize the process of creating the truck? So it always starts out with identifying a disease of interest, you know, whether next day it's headaches. I'm thinking very very trivial example of the researchers must find a biological

Target. It's like it's like a root cause analysis like you're trying to find is there protein or Jean or something on which it be active on that root cause then we will we will hear this this disease condition the finding that targeting enough itself is a challenge but once that Target is validated and sometimes you might not have the correct answer but nevertheless you have to scream through thousands of compounds compounds as Alexa medications to protect into medications. That means called that are cured a disease. It's called the lock and key

analogy, you know, so you have a Target and it should fit perfectly into that Target. So it sits in a drug must fit into the binding pocket of the target to have an effect. This is non-trivial which is why. And it's not even mad because you can't find come and see me find Compounce buddy toxicity the the side effects of many other considerations one has to take into account the target. This one thing the second part is the compounds. The intermolecular space is 10 ^ 60. It is only in fact about to scan through that entire like it's

just not practical. So no one's going to do that. So what pharmaceutical companies do if they're going to have a library of tens of millions of compounds but what happens is only 1 in 10000 eventually tested that are tested eventually make it to the R&D pipeline. So using basic research takes years, but what if we can use if she learning to identify this very good at In a tasks that require hydrations representative work, right? So if I have to scan from millions of compounds,

I can do that with AI or machine learning much more ready than I can send with additional basic research. I'm not saying it. It doesn't work at all except that just happens to be more Optima. So the first area in drug development drug discovery, that way we can use our ml is drug efficacy. So once a drug reaches a clinical trial, it is found to have lower efficacy. So due to weak, you know, they could with big biological link between the Target and the diseases to be treated. So basically what this is saying is that in your phone that you have discovered it, you decided on

a compound everything said and done but then once it is gone into the clinical trial phases or even before that you find that there's a week biological link like the target the drug doesn't really solve the problem in a basic. There's a mismatch between the Target and the disease to be treated. The machine learning can be used to find Targets with stronger biological name for instance in I mention one of these artists could be jeans so you can go through the entire. Intersection of jeans and use AI neural networks to predict Target or sets of genes that can

meet. So one of the examples that's here is in a prediction of a novel therapeutic targets using Jim Disease Association Dayton drug targets using data on genes known to be associated with diseases this paper actually go through it. It's predicted 1000 jeans as potential new targets. So the second space we're so that's one of the spaces in drug efficacy the second space where it's useful is virtual screening by which I mean a screening of the compound for instance there two types. There's a Target

basic screening and then there's a phenotype extreme in Target based is the effect of the compound on the target exam example, the protein phenotypic screaming screening is the effect of the compound on the whole system as an organism know as I mentioned are tens of millions of compounds. So in these cases machine learning can be used to prioritize which ones are going to test instead of doing a full-scale expensive lab testing a screen. Of course, you're not doing all the time but nevertheless with machine learning intern gets out of her almost like a probability

musescore in like a calibration. They're many real world examples on this artist in recited a couple of them here just to get give it give some examples. I research a screen 12 million commercially available compounds that treat heart disease and purchased 200 using machine learning of them three compounds Advanced preclinical and animal studies and the completed free design cycle in 1 year that with other device take six years. So very end of the concept to this very simple and very natural space to use something like the I

turn it on it for examples for you. So the third one is high troop with microscopy. So he was microscope to understand the effect of a drug candidate On Target, but the human eye can always detect very in a my new changes and you might need more position than that can be afforded even with high throughput microscopy. So there are against several examples where researchers have used computer vision to precisely measure sickle-cell changes that are not readily observable by the human eye using something called convolution neural networks of the

various different types of neural networks in if you think of deep learning, right, so there is CNN's there GNN stairs are in Encinitas of them have their own space CNN are basically used for image recognition or computer vision. So if there's a car like Tesla is going to use CNN send another one, but he should be is CNN's have their own Spence if a Time series analysis long short term memory and answer what So anyway in this particular case CNN's have proved to be very useful and the last one is the dino about drug design which really speaks to

can we generate new compounds with desired biological properties. So essentially what this is saying is instead of screaming all the compounds, right? You do have this entire library of compound cancer screening all of them. Can we just generate a new compounds with the desired property know all of this. I'm really just came in overtime. But these are vast stop in and of themselves, whatever y'all claims here is the fact that you will be the weirdest places where you can use machine learning AI for drug development in case

of generating new compounds the two of the most common methods that are a list of generative adversarial networks Gan some reinforcement learning. So you can think of a generator as it creates a story of a training CET training set is the set of molecules with a toxicity profile and the generator the function of the generator is essentially to create synthetic molecules. So it's just creating molecules right with a given toxicity profile what the discriminators going to do is it when the estimate How likely meet the new molecule King from his training set. So it's fantasy. I created a

new compound. Does it really is it similar to the two instead of known molecules that staring my friend said and it started for the risk-reward kind of mechanism that happens then like in various various other generators rewarded when its fools the discriminator discriminators rewarded when its card key labels molecule, so you'll find in a ml a lot of these lot of algorithms used to this kind of risk reward at McKenna's if you know you get declassification, correct? The weight increase you get the classification wrong the rates changed. So based on that. It's almost into going back when

I'm going for it and finally optimize our narang down onto an optimist a different parameter send this particular case we using GI ensenarles to do what's called. Do you know about drug design? Another several different saying success stories one that caught my attention that are spoken with them a different is a phone call Todd Pharmaceuticals. So they have of course, there's 10 ^ 60th and mention what do something very interesting call Quantum molecular machine learning. No one gets into the details. They have fairly good literature on their site and so on

but it's an example where you're using mode only, you know, not only deep learning but you're going to Step Beyond to try and find the Pinterest know all this is well and good but there are several challenges which are sort of impediments to the to the process of doing drug Discovery using Ai and some of that sound the top ones are for instance the gaps in sub listen to a few but I've just highlighted someone decide on the on the I do just you know, I like the main ones so gaps in understanding

of biology or chemistry. There's a lot of literature that's a lot of talent and then they still got especially when it comes to for instance how to use a iPod cases. So when I say there's not enough skilled mlra eye experts. I'm not saying you won't find a eye experts right? So there is no Dana University and with drug discovery that really is going to make the difference. So you need thought of someone who's in first in the area, but also at the Domain understand, you know, that

does not always have if you're if you're if you're a scientist in the bio space chances are you're not really in a merged into a I so they need to be sort of a meeting of the minds. So that's one The third and fourth a data related issues for instance data restrictions data quality issues data Access Data sharing. So what these really relate to are essentially the you can say the the data governance part of the of the entire ecosystem right there so many restrictions or regulations around what kind of day that you can use

and who can use that mean even if your if your organization has the date that you might not be able to access them. There's departmental budgets, of course, but there's also another factor that something related to the other in a DDD expertise that I mentioned. So there's a divide between MLA and domain Express. So domain Express. Let's see your email your your in the healthcare field chances are you've been using us as our state and amazing very very trivial example and your you believe in statistical methods and then comes from someone who's done a lot of AI

and and the person comes over and says you should be using artificial neural networks. So there is even though there's a need for meeting of the Minds The Mines maybe in different places at all times because one stinking and ends are the best way to do it and the other person's thinking statistical methods are the best way to do the biggest of them in their own rights in their academic experience other professional carriers have seen the door spacers work when for that particular use and finally there are regulatory so that you know that something I every organization would have to overcome

and find me the regulator eat. I'm just using a map for discovery. If you're making a new drug, and you're spending it's a 2 plus a billion dollars and in at the time comes for the FDA to approve it. You don't want to get caught off-guard where the FDA says this is the Deep learning algorithm and your own network. You don't know how many in what you've done here is getting your own tube activation functions. It's all kinds of things. So and you erased the drug not getting approved

of the fact that your mother is so complex. Never mind how well it works. Your mom is so complex that the that the research Community has questions around the around the Accuracy Century of of that model. So so so what scientist thing to do is stay with the safe side effects are going to use your traditional method of drug Discovery if there's going to be some animal parts as well. But if it is likely not to form the bulk of the underlined fundamental basis. So I don't have some challenges now in the

next few slides in so that basically what I've done is in I give an overview of where the high-level fart in front Discovery way. I can be used and thereafter the challenges. No, but in order to in order to implement A to Z Nation or your team's infested, I think it is essential to build an AI culture within your organization. You might already have the resources you might already have the Italians and so on but I know it starts with very simple steps one is in a learning the subject and most

often what I find is not work with I worked at various pharmaceutical organization inside work with their anal groups are groups and I work with a scientist. So one of the biggest challenges that I come across is that It's not that I know it's not that they're they're opposed to a i r m l. It's just that there is not enough familiarity or awareness of what a i r m l can do today the lack of clarity. So to speak in reality is not that hard, you know, and there are many many different today. There's just so many courses where you can at least get a favor of what animal is

a I've done this myself in a many years back. It's not hard to spend it takes time and you put your mind to it it like 2 months 3 months for months by the end of each of the score with each of them. You get new a better understanding of what animal is about what are the limitations and so on inside the top participants on this form calc angle where you have Emma competitions are not all animal research as they're just people who have worked on these problems hands on. Reviewing the current trends and successes to seeing if Lexie wants to use AI in your space

understand what other firms are doing and how they are using. It's not you don't always get real world examples, but nevertheless there still a lot of news that comes out where you can see what other friends are doing what other departments that might be similar to yours and that shit that gives you some idea as to where it's going to be pissed and I can tell you and I work with Ben this as well and it's nothing about 4:11 this but it's essential to its essential to develop your own personal understanding of the limitations and applications of machine learning because

the moment it comes to like anyone can say anything so but as a decision-maker, you have to have some sense of where it can be really apply and ask those difficult questions. When in a certain position comes for and lastly and I often get this question? Where can we apply a I like you know, is there a space space to apply machine and tell me this is such a nice almost like a trivial thing because what you're doing and Excel in a given the data mining part can be done using machine

learning. I mean you can just call Progressive essentially. There are ways to introduce machine learning Rai in all types of activities and it's not as difficult as it might seem to be in and out now the results we don't know right the results might be almost the same as what you got by during an average across a certain aggregate but nevertheless if you if anyone anybody wants to introduce a I apply machine learning in the day-to-day work. It's feasible. They're just some in a sort of pointers one is simple can be powerful

which is if I want smoke with someone who was at Novartis and he said that the big is and who is the head of the group and he said the biggest success they had was when to change the timing of an email instead of sending it out on Monday morning. They sent it out on like a meal of the week afternoon and he was quite disappointed because he said that made such a huge difference on Monday morning if they're not really looking into it. I mean that it doesn't. They're requiring

general characteristics. It should be easy to accomplish have high visibility and abroad well-recognized used is the most calm music for very complex tasks our machine learning for very complex. But if we cannot explain what we're doing, let's say 2 in in in layman terms. Speak to someone who has no idea about that field. Then I take it has become already too complex to gain a natural attraction and then are many examples not once have time to get into them. But basically it should be easy to accomplish have high visibility

easy to explain and finally, you know, you shouldn't ignore store analytics, which of the thing that's done on every John reputative basis or every week. If you can make the same for using machine learning or a higher than you know, that can also have an equally important or significant impact. So developing analytics pet friendly environment is basically talks about interview should also in encourage team members and her steam users to to engage. We should build a culture of Wellness within the firm

and encourage them to learn in order to make it truly, you know, truly Democratic I said, so to speak there many different Technologies, you know, I have personally work and see you in the eye in my background has mainly on the practitioner side and that I think has been very useful. Cuz once you try all of these hands on you get such a such an understanding of what order are they in your breast fall short, right? So anyhow, you know, I won't go too far and endured many different things here,

which I presume of that can share the lights off towards the top of the day if an AI technology Can be used in so long and I would just like to end up by saying, you know, there's a very very promising statement for a promising saying by Carl Sagan, which space somewhere something incredible is waiting to be known and I find that very encouraging, you know, if you're in the middle of a crisis that just seems to be re starting all over again, and hopefully in 20 years from now 30 years from now he is going to be so far develop that finding a vaccine

would be a matter of a couple of months perhaps. So anyways, so with that I'd end up. So thanks for your time. And thanks for in a listening to the entire presentation. Thank you.

Cackle comments for the website

Buy this talk

Access to the talk “AI in Drug Discovery By Nataraj Dasgupta, VP Technologies, RxDataScience”
Available
In cart
Free
Free
Free
Free
Free
Free

Ticket

Get access to all videos “Global Artificial Intelligence Virtual Conference”
Available
In cart
Free
Free
Free
Free
Free
Free
Ticket

Similar talks

Sanji Fernando
VP Technologies at Optum
+ 4 speakers
Matthew DiDonato
Senior Staff Product Manager - Artificial Intelligence at GE Healthcare
+ 4 speakers
Nataraj Dasgupta
VP Technologies at Rxdatascience Inc.
+ 4 speakers
Slava Akmaev
Chief Technology Officer at Scipher Medicine
+ 4 speakers
Shahidul Mannan
Head Of Data Engineering & Innovation at Mass General Brigham (Partners Healthcare)
+ 4 speakers
Available
In cart
Free
Free
Free
Free
Free
Free
Hudson Mahboubi
Sr. Manager Of Data Science at Workplace Safety & Insurance Board
Available
In cart
Free
Free
Free
Free
Free
Free

Buy this video

Video

Access to the talk “AI in Drug Discovery By Nataraj Dasgupta, VP Technologies, RxDataScience”
Available
In cart
Free
Free
Free
Free
Free
Free

Conference Cast

With ConferenceCast.tv, you get access to our library of the world's best conference talks.

Conference Cast
561 conferences
22100 speakers
8257 hours of content