Genuity’s Thomas Chittenden on Using Genomics and Statistics to Eradicate Disease
This week Thomas Chittenden of Genuity Science tells Harry about the company’s work to use the power of causal statistical learning, Bayesian belief networks, and other advanced mathematical techniques to understand those cascading gene interactions that account for health and disease—and translate them into insights that can provide drug makers with new targets.
Thomas Chittenden, chief data science officer at Genuity Science, says what’s keeping the genomics revolution from turning into an equivalent revolution in drug discovery is that most of our domain knowledge about the molecular biology of the disease has come from a hunt-and-peck approach, focused on one gene at a time. Find some gene relevant to a disease, knock it out, and you see what happens. Such experiments are always revealing, but the reality is that human biology is the product of the interactions of huge networks of thousands of genes—which means most diseases are the product of dysregulation across these networks. This means, in turn, that to figure out where to intervene with a drug, you really need to identify the patterns that cascade through the whole network.
That’s where AI and machine learning come in, and that’s why Genuity has tasked Chittenden to lead R&D at its Advanced Artificial Intelligence Research Laboratory. Chittenden’s team is pioneering new applications of old ideas from the world of probability and statistics, including some that go all the way back to the work of the English statistician Thomas Bayes in the eighteenth century, to look at gene expression data from individual cells and predict which genes are at the beginning of the cascade and are the causal drivers of diseases like atherosclerosis or high blood pressure. The hope is that Genuity can help its clients in the drug discovery business make smarter bets about which drug candidates will be most effective. And that could help shave years of development and billions of dollars in costs off the drug development process.
Chittenden is one of those rare professionals who has more degrees than you can shake a stick at—he has a Ph.D. in Molecular Cell Biology and Biotechnology from Virginia Tech and a DPhil in Computational Statistics from the University of Oxford, and completed postdoctoral training at Dartmouth Medical School, the Dana-Farber Cancer Institute, and the Harvard School of Public Health—but can also explain the actual science in a way that makes sense for a non-expert. On top of that, he’s been thinking hard about how to rein in some of the hype around the power of AI and machine learning in drug development and how to set expectations about what computing can and can’t do for the industry.
Please rate and review MoneyBall Medicine on Apple Podcasts! Here’s how to do that from an iPhone, iPad, or iPod touch:
1.Open the Podcasts app on your iPhone, iPad, or Mac.
2.Navigate to the page of the MoneyBall Medicine podcast. You can find it by searching for it or selecting it from your library. Just note that you’ll have to go to the series page which shows all the episodes, not just the page for a single episode.
3.Scroll down to find the subhead titled “Ratings & Reviews.”
4.Under one of the highlighted reviews, select “Write a Review.”
5.Next, select a star rating at the top — you have the option of choosing between one and five stars.
6.Using the text box at the top, write a title for your review. Then, in the lower text box, write your review. Your review can be up to 300 words long.
7.Once you’ve finished, select “Send” or “Save” in the top-right corner.
8.If you’ve never left a podcast review before, enter a nickname. Your nickname will be displayed next to any reviews you leave from here on out.
9.After selecting a nickname, tap OK. Your review may not be immediately visible
Harry Glorikian: Tom welcome to Moneyball medicine.
Tom Chittenden: Well, thanks for having me, Harry, it’s always a pleasure talking with you.
Harry Glorikian: So Tom and I really want to get into like the personal background and, and, and the work that you’ve been doing. But, I want to start with a really sort of big idea. I want to, I want you to explain to the group what you’re doing with your group on the value of AI and statistical machine learning and drug discovery and clinical trial management. And if you’re right, how does it pay off?
Tom Chittenden: Yeah. Well, it’s a great question, Harry. And I can talk for, I could bore your listeners for days on the topic here, but basically I think what we’re doing the bottom line is that we are using machine learning, artificial intelligence, computational statistics to further our collective understanding of basically human biology. And we now have scientific evidence or experimentally proven scientific evidence that these algorithms or these approaches are actually capable of teaching us the rules, rules that govern the underlying molecular constituents, if you will, that drive cellular behavior and dictate phenotype.
And I think that is the biggest thing in this space. Anyway, that AI and machine learning is going to do for us is that we can’t fully address human disease. Until we have a fuller understanding, a more robust understanding of the rules that govern human biology, and that’s what we’re doing in the lab.
Harry Glorikinn: So if you think about it, right, you’re basically able to understand the characters of the movie and how the movie is moving forward through its frames?
Tom Chittenden: Somewhat. Yes. And, and it’s very analogous to an analogy that I, like to use. Google came out with something called knowledge graphs, probably clear back in 2012 and they have far reaching implications and just about every sector and especially in, the biomedical sciences, but currently, and I like to use the analogy of GPS and Google maps.
We know we currently know how to get from point A to point B. And so we can then monitor traffic flow safe from Cambridge, Massachusetts, up to Medford, Massachusetts, where I live and we can make predictions about how long. But that’s going to take what currently at the molecular level and human biology, we don’t have that underlying roadmap.
And that’s what we’re actually working on. And again, we have Evans as that these algorithms can actually build causal dependency structures. And I’ll explain what that means. I’m sure as we talk. Or we go on here this afternoon, is that they are capable of building causal dependency structures that are actually reflective of the signal transduction, cascades, or how all these proteins interact with one another within the cell and actually drive cellular behavior and dictates phenotype.
So that is what we’re going after is this molecular physiologic map that once it’s built, we’ll be able to track molecular information flow through this map. And at any given point where that information deviates we’ll be able to put a molecular fingerprint on all well known human disease.
Harry Glorikian: So now let’s back up here. Right? You’ve got this, you know, I want to sort of understand how you ended up as, as chief data officer, how you got there. I mean, I think with your background of molecular biology and computational statistics, it sounds like a dream job. So how did you get to where you are?
Tom Chittenden :Yeah, that has been a long arduous road, if you will. And I don’t want to sound like I’m whining here here. Right. But for my undergraduate degree is in exercise science. My first master’s degree is in exercise physiology and biochemistry, and what, as I was going, I started to learn how to ask more involved questions. And along that road, I understood that I needed a greater understanding of molecular biology. I needed a greater understanding of the quantitative principles of statistics to be able to answer or address these questions. And so it’s been a very long road. I hold a doctorate in molecular biology, as well as in computational statistics went on to postdocs in molecular and cellular cardiology.
And then finally, a second post-doc in biostatistics and computational biology. It’s been a long road. I didn’t get out and start making any money. And, my wife will tell you exactly this. We really didn’t start making any money till I was 50.
Harry Glorikian: You know, it’s you’re the perfect poster child for lifelong learning.
Tom Chittenden: There you go. And then I was family and friends accused me of being a professional student for years.
Harry Glorikian: I tell you, I always think to myself that I do want to go back to school. I mean, there’s so much to dig into these days and when you’re working, sometimes you just don’t have time to dig in as deep as you’d like.
So now your organization, you know, it, doesn’t develop drugs right now. It’s, it seems to be focuses more on the services if I’m correct. And, you know, you do. Genome sequencing, manage and, and analytics and in all these different areas. So before we first focus on, so the AI and machine learning and predictive analytics, can you, can you give the listeners sort of a big picture of what, you know, your organization does?
Tom Chittenden: We are, yeah, I think we’re still trying to figure that out as well here. So I’ve been with the company for five years. We used to refer to ourselves as a contract genomics organization. We are now touting that we are a data insights company. So we have offices in Reykjavik, Iceland, Dublin, Ireland.
We were in Cambridge, Massachusetts, and we’ve just relocated across the river here in Boston. And basically we’re enabling drug development, I think where the current state or status of where we are, we have some of the largest patient cohorts in the world. And so we sequenced these patients and then the lab that I oversee we’re responsible for generating usable information from these large high dimensional data sets. Does that help?
Harry Glorikian: Yeah. I’m trying to, you know, get people to understand what you guys do. And then I think it’ll be easier for everybody to understand this concept of predictive analytics and, how it generates value going forward.
Tom Chittenden: Sure.
Harry Glorikian: So, but, Genuity didn’t start as Genuity. It was formerly Wuxi NextCODE. And I think it goes back even further than that too. You said Iceland. So I’m, throwing in deCode genetics, which wow. That’s a blast from the past. Well, you know, when I was at applied Biosystems, decode was a big subject to talk about How does that legacy sort of shape the services or the research that you have today or that you’re moving into today?
Tom: Yeah. So to give everyone an understanding of where we’ve come from is that Kári Stefánsson, and Jeff Gulcher are the co-founders of deCode genetics. They’re in Reykjavik, Iceland. Jeff, and a group of others that were basically the core executive group of deCode genetics ended up branching off with Kari’s blessing and starting something called NextCODE health.
About 2015 Wuxi App Tec, the largest CRO in the world based in Shanghai actually acquired NextCode health. Hence we have Wuxi NextCODE Genomics inc. We were there up until I think probably the spring of this year we just, formally changed the name and restructured the company as Genuity science, but there were some issues that we had run into being owned by the Chinese entity.
We got out from underneath that. We’re still working closely with everyone, so everyone’s playing nice with each other. But that’s basically the legacy. So from deCode genetics, statistical genetics is basically where we come from. So some of the legacy software that was developed at deCode, we still use.
What is different now is this R and D arm that we have at Genuity science and my boss, our chief scientific officer Jeff Gulcher, the co-founder of deCode genetics and the co-founder of, Genuity is my boss. And he has just been absolutely fantastic. He’s a brilliant physician scientist that has basically given me the ball and said, just don’t drop it.
And if you do back and let me know about it. And so we started with a very small bioinformatics team back in 2015, I was working with the research computing group, Harvard medical school. And Jeff asked me if I would come across the river and start I don’t think he knew that it was going to blossom into an AI lab.
He, brought me over to expand the bioinformatics team and we just hit the ground running. And so we were in our heyday. We were at about 18, a full blown computational statistics, and bioinformatics grew that then blossom into this, this advanced AI research laboratory here at Genuity.
Harry Glorikian: So you know, not to pick on it, you know, our, you know, the pharma companies and but what’s wrong with the way that they’re doing it. And what are you trying to.
Tom Chittenden: Here. You’re going to get me a trouble with that. Come on. I’ve got some pretty strong opinions of that right now, but it comes down to, and it’s just not pharma, but there are a number of biotechs. And in fact, if you know, for all your listeners that have been following.
You know, the advent or the re insurgence, if you will, of deep learning. And that, that was around 2006 or so, but it really came on the scene from the biotech standpoint, about 2014, 2015, and listening to some of the hype that exists in this space, you would have thought that we had already cured cancer. But we haven’t.
Harry Glorikian : We have, I thought we had no?
Tom Chittenden: Exactly. Yeah. We’re getting closer. But we’re not there yet. So there are still some shortcomings . And when, we say artificial intelligence, we are not building machines, you know, in this space that are capable of human cognition.
These are machine learning algorithms that are actually able to detect patterns within all of this high dimensional Omix clinical, deep phenotypic data. We’ve now integrated real world evidence into what we’re doing. And how do you actually pick out, you know, these algorithms can. Can define the patterns, but can they do it in a reproducible manner?
And to be quite honest with you, we still struggle with that because if we cut our data up, train these algorithms and then test on the data that we’ve left out, we can get really, really good or consistent classification performance. But if we were to do that again and again, and again, we end up getting, consistent classification performance. But the underlying features that are associated with each cut of that data and each run of these individual hour rhythms. Are a bit different each time because we have such a large degree of correlation bias feature dependency within these data sets. That it’s very difficult then to go downstream and define the actual underlying biology at that molecular level to go further with.
And I think that’s what most would agree with me in this space is that’s what pharma struggles with it’s what we still struggle with. So we are working outside of the box if you will. So we’re looking at statistical optimization approaches and, basically inventing the math and statistics on different computing architecture, such as quantum computing, neuromorphic computing, we have been researching or investigating on these architectures for a while now. And it looks promising, but I don’t want the fuel that, the hype. That currently exists in this space because there’s even more hype with quantum computing than there is with artificial intelligence.
Harry Glorikian: I remember listening to one of your talks at MIT about quantum computing guy want to say it was two or three years ago, something like that when you were there. So as you think about improving the process, you know, where do you see the biggest opportunities coming from? And, this is sort of a leading question because I have a sense that we’re in an area where we have so much new data about gene networks at the cellular level. And we’re only starting to translate that into specific insights about drug targets, and we need better ways to do that. And so I’m not trying to say that that’s necessarily the opportunity, but where do you see the biggest opportunities that are coming?
Tom Chittenden: Well, we’ve got two areas that I think that we can improve on. That we can greatly improve on one is the type of data that we’re modeling and the other one are the type of algorithms that we’re actually developing. And so I can talk to you a little bit about both of those, those areas because merging single cell science with very advanced analytical tools, machine learning tools I think is going to do nothing short of redefining our, basically our, current collective understanding of human biology. And we have evidence there. We had a couple of big papers that came out last year. We collaborate very closely with. Not only pharmaceutical partners, but academic partners as well.
And in what we’re doing with the machine learning right now is that we have just now scratched the surface with being able to ask why. Not what correlates with the disease, but actually what is driving the disease? So we are integrating probabilistic, programming and causal inference into everything that we’re doing to get to that point where that these.
I had mentioned earlier, these causal dependency structures are actually reflective of the signal transduction cascades. So two proteins interact with each other on the cell surface and set off a series of events that lead to a, cellular response cellular molecular response is that we are getting closer to being able to model that in a much, much more robust manner.
The second part of that is the data. So we’ve now moved into an era and I think we’re pretty well entrenched in it. We’ve been working on this for about the last three years now and that’s single cell science, so we know relatively speaking and we know very little about biology. We even know significantly less about single cell biology and the last credible estimate that I came across or come across is there are approximately 37.2 trillion cells in the average adult human body, all working in concert with one another. And we’ve been for the most part, ignoring the cells, behave or work in concert with each other. And so by doing that, we had a paper that came out that we published in nature metabolism.
About mid-year year last year that we were working very closely with Mike Simons, he’s the founder of the cardiovascular research Institute at the Yale medical school. What we were able to do is generate, we use the AI to generate working hypothesis. And I think that this is important to talk about as well, is that we’re not curing cancer with the AI.
We’re generating very, very sound robust, if you will working hypothesis that then can be explored and validated behind the bench experimentally and what Mike Simon’s team was able to do, then that information that we had given him, he was able to actually not only inhibit atherosclerosis. Heart disease, if you will.
And, and again, it’s in a mouse model, so I’ve, received a fair amount of criticism with this. It’s a model organism, but it’s a first step being able to do this in humans. So it’s a gradual process. He was not only able to inhibit atherosclerotic. Plaque development in these animals. He was able to, once it did develop in his control animals actually reverse it.
And that is something that was just basically a real breakthrough last year. And it was evidence of coupling the artificial intelligence very closely with these experimental approaches behind the bench and, and it, what it did is it ended up saving a great deal of time and expense for Mike’s group. And he was my first postdoc boss. And he’ll tell you stories about how I spent four years phenotyping the gypsy one or the selectin mouse where we actually. Published a paper earlier that year. And for all your listeners that know of reverse genetics, where you cause a perturbation in DNA and then go hunt for the phenotype.
We were able to do that within six months. What took me four years as my, you know first post-doc we were able to do in six months from an in silico standpoint, we were able to predict, or the network that we develop, these gene network was actually predictive of three vascular phenotypes that when he built the subsequent mouse model or, or actually induced the same genetic perturbation in these animals cause we started with human cells. Those mice presented with those three vascular phenotypes. And that was a big paper that came out last year in the journal of experimental medicine
Harry Glorikian: Yeah, it just seems like every time I turned around everything were. Everything. First of all, we can do things we never could do before and everything is happening a lot faster than I remember it taking before.
Like I can’t read fast enough sometimes to keep up with everything that’s going on. So there are tons of, you know, different types of AI and machine learning methods being used in drug discovery. And I know we can’t talk about all of them in one interview, although. That’d be interesting. You know, if we concentrate in sort of the area that you’re focused on, which it sounds like, and from what I know is you know, causal inference, causal, statistical learning, Basie, and networks and probabilistic programming, and correct me if I’m wrong.
Can you take a shot at explaining it so that, you know, a non-expert can understand.
Tom Chittenden: Yeah, well, I hope so. I’ll take a crack at it because sometimes I struggle with it, what we’re actually doing. And that’s, that’s one of the things that I, you know, try to impress upon the team in the lab is that I don’t want to run the risk.
I think the real risk with artificial intelligence is that we’re going to be exposed to so much knowledge and it’s going to be much easier to acquire that knowledge, that it doesn’t end up in the long run, dumb dumbing us down. In some, some instances, if you will. So these are very, very sophisticated mathematical statistical techniques that are capable of defining what is important.
In a predictive analytic, if that makes sense. So if you’re looking at the difference between two different cancer types and you have 79,000 or 80,000 potential predictors in that model, can you build an algorithm that pulls out a pattern in all of that data that accurately robustly consistently? Is able to predict class one and class two in a binomial or binary experimental design. Now we have done that in multi-class experimental designs or multinomial experimental designs and with the cancer genome Atlas, why I came up with 79,000 molecular features is that’s what we had. Is that we have five different data types across 8,200 tumors representing 22 cancer types.
And we were able to build an algorithm or a set of algorithms that, which I think is, is more important. So we use ensemble, computational intelligence. I’ve never trusted a single run of a single algorithm on a single cut of the data. So we use multiple algorithms on multiple cuts of the data to find what is most informative.
And across those 8,200 tumors, we are at 99.7 percent accurate in being able to classify to actually predict one of those 22 cancer types. And I refer to that. And so in that test set that equates to misfiring or misclassifying only seven out of 8,200 tumors. Right for which are bladder cancers that are consistently misclassified as another tumor type.
And so we believe that the algorithms when applied appropriately, when built appropriately are capable of defining or picking out the misannotation. They can occur the human miss annotation, if you will. And in this case of these types of tumors. And so we’ve got a long ways to go to show that, but it’s, it’s evidence in what I refer to is if you, the analogy is facial recognition, right?
This is disease recognition. And where this is going is we’ve picked out a signature that is capable of differentiating one of 22. Or 21 other cancer types. And what a physician can do with that information is they can sit down with a patient before they’ve even met them, the pre-consult and know exactly what type of tumor that that, actually is now in this case, it’s only one of 22, but I foresee this going that we, we can do this with that molecular physiological road map that we’re working on, that we had talked about earlier. Yeah. Physicians is never going to replace the physician. And I want to make sure that, that I am expressing that here with you this afternoon, it’s, it’s going to be a tool that allows the physician scientist to actually do his or her job much better.
Harry Glorikian: No, I mean, you know, if you can accurately diagnose someone, treating them at least starts to go down the right road faster. So I’ve heard you talk about opening up the black box and you know, so that you’re not just providing predictions about associations but coming up with hypotheses about the underlying molecular connections. Right. So without giving away sort of your secret sauce you know, can you explain, you know, how you might do that? What the unique and proprietary is here with again, without giving away the sauce?
Tom Chittenden: Yeah. So, just to take it out of, that black box. No, what everyone refers to as black box algorithms. we just completely remove that. We are working with math that was first developed probably 270 years ago. So the base, there are conditional probability. You know, this is what we’re doing. Can we actually look at. The probability of event based on the knowledge of conditions that might be related to that event.
And so the problem comes in with this and why we’re struggling with this. And I like to use the analogy, please forgive all my analogies and trying to make it as entertaining as possible.
Harry Glorikian: I do it all the time. Don’t worry.
Tom Chittenden: Okay. Yeah, we we’ve had 130 years of collective engineering. That’s gone into the current state of the automobile. If your car doesn’t start in the morning, you can rule out rotating the tires because it has nothing to do with the ignition system. And we know that because we have been, we’ve had a hand in that collective engineering now with human biology, we’ve had three and a half billion years of natural engineering that has gone into the current state of the cell.
So actually modeling in large part observational data is extraordinarily tricky is that we can find what’s associated with a disease, but oftentimes that’s problematic. So if I go back to the analogy of the car, I park in the left-hand side of the garage, Right. And before the pandemic hit, if I was collecting observational data over a year, there would be a perfect correlation to the position of that car in the garage or my car in the garage with the ignition system I’d start that car. And there would be that, perfect correlation now, as, as ridiculous as that analogy sounds, that is a perfect correlation. And we quite often find these perfect correlations that actually ended up being spurious correlations because of all the, everything in the human genome is there for a reason.
And to some degree or another. It all correlates all these different entities within the genome correlate to a certain degree with everything else in the human genome. And so I think it’s very important that we start building the math and the statistics to be able to address those questions. So can we predict the state of a gene based on the state of another gene. And that’s how we start developing these networks are building these networks. And what we’re finding is that when we go in experimentally and change the state of that gene, it actually changes the state of the network. And again, this is math that has been, you know, first developed 270 years ago by Thomas base.
Harry Glorikian: Right?
Tom Chittenden: So there’s, there’s not a whole lot of secret sauce on this right now know the way that we apply it and the understanding these, these are all very, I would say biologically relevant algorithms. So we have a Genuity, we apply a great deal of biology, domain knowledge into what we are doing.
That basically I give, they see these algorithms, if you will, to give them a headstart, to actually be able to predict the differences between different States.
Harry Glorikian: So what do you think? I mean, we’ve been talking about this, I’m watching, you know, tons of companies and tons of universities sort of going in this direction. I still think we haven’t graduated enough people to do it. As widespread as we like, but what are the biggest, like, is it scientific, technological, social to adopting these methods, to drive drug discovery in the right directions? I’m leading in a certain way, because I believe that this is the future also, but you know, I just, I hear a lot of people talk about it.
I know that. It’s not happening the way that I, I think it should or could happen. And maybe I’m just impatient.
Tom Chittenden: Right? So I’m going to step out on a limb here, Harry, a little bit. Now I’m not a clinical psychologist by any means, but I think that this has to do in large part with the unknown. No, the human ego fear of change.
Harry Glorikian : Right?
Tom Chittenden: So if man were meant to fly, though, that you’re getting back to that old adage, is that now we’re facing something here and there’s a fair amount of the unknown and there is so much. Misinformation or uninformed information that exists out there that is standing in the way we’re running into ethical issues.
I don’t think we’re going to run into that biggest ethical issue with, with in this space in the biomedical sciences is what happens when we finally are able to build these algorithms, can, they can tell us. Exactly what’s going on with the underlying dysregulated biology that then, very accurately informs us on basically how to right the ship.
Now we’ve got some ethical issues about, okay, we know how to right. The ship, but can we do that? Right? And so we’re going to have to be working. We’re going to need to work with politicians and others in other fields. To dispel a lot of the hype that exists out there and it sells books here. The AI stuff it’s sells books.
Singularity. I usually get answered that question. When is singularity going to happen? And I don’t want to disappoint, but I strongly believe that it’s not going to happen that point in time where AI leads to the demise of, of humanity. I don’t believe, I think it’s going to be a very gradual synergistic marriage. Between human and artificial intelligence that actually shapes the trajectory of human evolution. That’s where I believe that this is going, but it’s going to be a very gradual process.
Harry Glorikian: Well, and I, but I do believe that people need to have, first of all, you have to, you know, maybe it’ll take a generation or so, but you have to be comfortable with that, the two coming together. And you have to be curious. Because the system is going to throw something out and you’re going to be like, what the, I never thought about that. Right. And then at some point it’s going to get better than you are in doing certain tasks.
Tom Chittenden: Okay. Yeah. And Harry, that’s a very good point because I believe that it is better than us at certain things right now. And that is the ability to generate working hypotheses. So when I, 20 years ago, this was actual heresy. Wait, I cut my teeth on the old BNA micro-ray days, right? Where we started being able to capture it with a single assay capture the entire, excuse me, transcriptome in a single assay. What that ends up doing is it’s not hypothesis testing.
It’s hypothesis generating, but then what the old schoolers at the time, and this is strange for me to say that at 57 now, but when I was a youngster cutting my teeth on this is that the old schoolers didn’t see that it allows us to ask questions that we probably would not have been able to ask. Before the age of all of this high throughput, high dimensional Omix type approaches that we’re using with the AI is that it is hypothesis testing, but it’s generating a hypothesis in allowing us to ask the right questions.
Harry Glorikian: Right? Well, I, I believe that is, is, and I say this to people and they look at me strange, but it’s, I think we’re seeing the modification of the scientific method. I’m not generating the hypothesis to start. It’s giving me my top choices.
Tom Chittenden: Right.
Harry Glorikian: And then I’m sort of going down those roads because there is such a variety of data that’s coming in. That it’s impossible. If you could, if you took 10 people in a room to try and, you know, synthesize that data and the interactivity of that data. To be able to say, you know, this is sort of, here’s some areas you should look into. I love this stuff.
Tom Chittenden: Yeah. I absolutely agree. And I don’t know, I need to give credit to somebody out there and it may have been a 60 song. The actual answers lie within the questions. And can the AI get us to the point where we are asking the right questions? Right. And I believe that that’s what AI in this space is going to be able to do for us. And it’s already doing that. We would have never gotten to those, those publications last year.
And earlier this year, the cell STEM cell paper, where we’re actually looking at classifying various human disease States in cardiovascular disease by cellular differentiation trajectory. So now we’ve applied at the longitudinal experimental designs and we can see. We’re not being able to, we’re not to the point where we can track a single cell over time, but we can take collections of cells and that signal is so strong that we can link those collections or clusters of cells in a longitudinal experimental design that actually forms a disease trajectory, if you will, and classification of human disease. And again, it gets back to this underlying molecular physiologic blueprint of human biology.
Harry Glorikian: So I’m gonna now ask you to be the futurist, right? Zoom out and paint a picture for people listening and near term. And then long-term futures look like how much do you think AI machine learning could help speed up drug discovery in the next five years? And then, I mean, I don’t want to go out as far as 50 years, but you know, say five years in the next, maybe 20 years because it’s moving so fast, it is moving quickly.
But I think that we need to get a good to get a handle on this and those that are working in the space. I think it’s very, very important that we help educate the lay. The lay population of exactly the capabilities of what we are doing. So where I see this in the near term is using some of these existing computing architectures that we have these high performance computing architectures with the known established statistical optimization approaches that we have. We marry that with single cell science and we start garnering a much greater in-depth understanding of human biology. And this is greatly going to advance our ability to identify the right target drug target that then to build more efficacious targets on top of, as you probably well know, 75% of all of pharma, R and D that’s associated with pharma can be directly attributed to failed clinical trials. And I hope that this does not sound arrogant on my behalf, but I believe is that we’re going after the wrong targets.
We’re still within that, we still haven’t moved in the causal inference fully. We haven’t jumped into it feet first. We’re taking our time making sure that it’s right, but we’re still going after things that associate we’re addressing the symptoms, not the causes. So in the near term, within the next five years, I see that happening is that we’re going to.
End up now, it’s going to afford the means to develop much, much more efficacious drugs. We’re going to be able to repurpose existing drugs. So there may be a drug that’s associated. I always use the analogy of of asthma. We might be able to apply that in a cancer state where inflammation is driving tumor agenesis, you know, it’s those types of things that we’re going to see that are going to start surfacing.
Now, looking out further with this, I think we are going to get, and this is what we are coming. Well, what we’re realizing is we’re very limited right now with statistical optimization. And so in, in the, in the longterm 10, 20 years down the road, we’re gonna, we’re gonna actually, and if it’s, if it’s not me, Harry, there’s somebody here that we’re working closely with, that’s going to end up doing this and, you know, actually get a handle on statistical optimization approaches in integrating other computing architecture, such as quantum computing, which will afford quantum machine learning. In fact, we have a paper under review that is the first successful demonstration of classification of human cancer patients from a multi-orgasmic standpoint with a quantum machine learning algorithm and we’re having we’re, having a really tough time getting this thing published.
Right? Cause there’s so many things here that, that we’ve stepped out into the unknown. Right? I think that adjust here probably before the end of the year, we’re going to know whether we’re going to get it published or not, but we’re working on a COVID 19 project with the university of Strausberg that we have integrated quantum support vector machines on a very small cohort.
And we believe. Or I’m saying, I guess what I should do is tamp this down a little bit. I get kind of excited. One of them is that we have a promising signature for why patients actually move from being admitted to the hospital on oxygen to then being admitted into these ICU. ICU units and needing a mechanical ventilation to address the acute respiratory distress syndrome or arts.
And so we have used in combination with established approaches, we have used quantum machine learning to help define that informative signature. So I think we’re going to see more and more and more of that. And then looking out a little bit further down the road. What’s every bit as promising as quantum machine learning is neuromorphic computing because this affords longitudinal experimental designs.
And so what I had. Talk to you a little bit about or shared with you when you’re your listeners earlier about that cell STEM cell paper in the longitudinal experimental designs is to be able to do that in a much, much more robust way with, with neuromorphic. Processors and spiking neural networks.
So I usually get answered that question, you know, whereas deep learning is deep learning, going to be the future of AI that, you know, the easy question is yes. Or the direct question is yes, but I think it’s going to be a form. It’s going to be a spiking neural network or typology that’s. Just associated with that, that we can actually integrate longitudinal experimental designs and move away from the, the cross sectional study designs that are useful.
But then, you know, it can be problem as well.
Harry Glorikian: So we’re going to need a whole new breed of people to help take this forward. Cause I I’ve, I talked to people all the time and there is, there is not enough people that are, let’s say, trained in this. In these two areas that you need right. To, to drive this forward we’re just not graduating enough people. I mean, Google keeps stealing them. We need to bring them over to our side.
Tom : Right. And, and I think that that’s going to fall largely on academia, the major university programs, developing that domain expertise. And when I say domain expertise, you know, this is going to touch every, every sector of modern society.
But when we’re looking at healthcare, Yeah. Is, can you take, you know, somebody that has been formally trained in molecular biology, but also give them the quantitative skills to actually move forward. So for the days of the, the molecular biologists that was trained in the era of. One gene hunting and pecking, as I refer to it is that you take a gene out, you scramble it, you put it back into a system, you evaluate it.
That has all been very useful. That basically the, the knowledge domain that we have currently is in a large part based on, on that approach. But can you take that molecular biologist? Couple it, you know, or, or give them the quantitative skills to start building these algorithms to start in. And again, at the end of the day, so that we can ask the right questions on the flip side of the coin.
And in bioinformatics, you usually have those that come from the biology domain or the computer science domain. So those that are coming from computer science are going to have to acquire that ball that, that deep. Biology domain expertise to be able to marry all of this.
Harry Glorikian: Yeah. At ABI, we used to have to take two guys, put them in a room, right. Man, it took them a long time to figure out how to talk to each other. But you know, getting to some, you know, harder numbers, I mean, My belief is that you could that I’m, you know, you could cut out three years or four years on the front end of the discovery process with the technologies we’re talking about.
Tom Chittenden: Are you talking drug, drug discovery? Yeah, cause that’s kind of all over the map right now. And I, think most would agree that it’s 10 years and over $2 billion, they actually bring a drug to market. Now I’ve been talking with folks out there that are coming up with all sorts of ingenious in, not ingenuous. Approaches, if you will, to be able to do this. So David Berry at Volo health, you know, he’s working on things that, you know, can we do that? You know, it can, can we significantly cut the amount of expense and time down to two years to bring a drug to market? That’s left to be seen. We need to wait to see if that can, if that will happen.
But I think that that’s the direction that we’re actually going and what I’m working on or what we’re working on on the team is to make sure that what we are finally or eventually bringing to bear is actually going to be more efficacious. Right. It’s going to be effective because we are hitting those mechanisms of action.
Those, the dysregulated biology that’s leading to disease. And from where I sit, that’s the most important, not how long it takes to bring a drug to bear, but can you bring an effective drug to market in a much more efficient manner? And I think that’s what AI is going to do for us,
Harry Glorikian: But if we’re both Sprite, I see a shift in the business model.
I see a deflationary effect that’s coming from the technology itself because you’re always getting more for the same amount or less.
Tom Chittenden: Right. Which would you agree though, that that pendulum has shifted in this direction, if you will, the, all of that AI hype. Right now is that it needs to shift more towards where we were, where we were coming from.
It’s it’s, it’s all AI is, is that we’ve got a magic wand and that is not anywhere close to reality. No,
Harry Glorikian: But I think it depends on where you’re applying that in this giant, you know, process that we’ve got. Right. And I am seeing leaps in different areas at different points. Like almost when I least expect it, something happens. So I’m, I’m constantly encouraged that like, you know, people say, well, it’s going to slow down or it’s going to stop, or we’re going to go into another winter. I’m like, I’m not so sure about that. Right. Because I see the tech industry driving things forward and we borrow all that stuff or we use all that stuff and they’re not slowing down.
And so I see it helping us now, biology, doesn’t always, you know, function or do what you want it to do. And there’s a lot of discovery to go on there, but I haven’t, I’m bullish that in parts of medicine, we’re going to see it have a dramatic difference when it comes to imaging capabilities, when it comes to pathology.
And that’s going to feed the real time data that you’re trying to use. Right. So I just see. Things happening, happening on a multi-factorial level. I’m not trying to hype it. I’m just saying, I don’t think it’s slowing down. I think it’s, it’s, it’s found a cadence that it’s going forward on. And even when we think it’s slowing down or coming to a stop, somebody is trying to find a different way to come at it.
Tom Chittenden: So, Harry, I, I’m not disagreeing with you. I think the cadence needs to slow a bit. There are so much AI hype right now. I think that that’s going to start falling away, right? Yeah. We are still waiting for that first AI driven therapeutic to come to market. And I’ve heard that there’s some companies and I’m not mentioning them that have actually done that or achieve that or working on that.
I think that that’s where that’s going, but a lot of what has come up within, you know, the last five years or so to a great, to a great degree. If I may say this to your listeners is just absolute nonsense, is that it’s, it’s going to take a lot of hard work and the AI, I think can expedite the process.
As long as it’s applied in the right manner. And there’s an old adage out there and I, you know, statistics don’t lie, right?
Tom Chittenden; But they can be naïve, really applied. And horribly misinterpreted. And so we’re in this era, this era right now with artificial intelligence that, okay, is this real? Is it not? And we haven’t taken a time from my own personal opinion here.
We haven’t taken enough time to experimentally validated, especially in this space right now. Now I’m not talking to image analysis. Cause we we’re we’re we’ve now moved into digital. Pathology and imaging building these algorithms off of various biomedical imaging modalities. And they’re highly accurate, but they’re much, much less involved than what we’re doing at the molecular level.
So these are. Narrowly more complex or complicated when we’re addressing human bias. So, you know, the Googles out there that have advanced that fantastic, that has real application in digital pathology that can help with clinical trials, inclusion, exclusion criteria, and do it in a much, more cost-effective manner.
I absolutely agree, but I think that cadence of where I think it’s going to slow and I think it needs to slow. Is that we really need to understand and have a strong understanding of what it is currently, the state of the art, what is the current state of the art, and that has not been defined. And so, again, I’m not disagreeing with you.
No, no, no. Yeah,
Harry Glorikian: I agree. But I, you know, after I feel like I’ve gone through a couple of cycles now in different areas and what I always see is the hype happens. A lot of money goes into this space. Nobody knew what they were doing when they started. Some of those things go bust. A lot of them may go bust, you graduated a new, it’s almost like a new university because nobody knew what they were doing.
And now the second time around, they’re like, ah, we tried it that way. It didn’t work. Let’s try this way. And you get a new crop of people because we’re not teaching any of this really in school. You’re not learning it until you get into a company and actually cut your teeth on the real work.
Tom Chittenden: Yup.
Harry Glorikian: So I think sometimes we need the hype cycle. Otherwise we won’t build enough of a foundation of people to, to take it to the next one.
Tom Chittenden: Yeah. And then in that, that Gardner hype cycle curve, you know, I most would, place deep learning kind of at the pinnacle of those inflated expectations. I don’t know. I think that we’re going and especially what we’re doing and others are doing by taking the time to experimentally validate what we’re we’re doing is that where we’re not going to run into that trough of disillusionment.
Is it we’re gonna, it’s gonna move straight across, but it’s going to take time and it’s going to take a lot of effort and there is no magic wand with any of this.
Harry Glorikian: No, I mean, as an investor, I’m glad there’s no magic one because you know, you’re then you’re always worried about losing, right? So you’re you’re you gotta find multiple bets.
That you believe in the team that can then get it to the other side, Tom, you know, as always, it’s a great pleasure and, you know great to talk to you and great to have you on the show. And I can only wish you incredible luck in everything you’re doing because people like me will benefit.
Tom Chittenden: Yeah, well, fantastic, Carrie, and again, thanks for the opportunity.
I hope I haven’t bored all of your listeners to tears here, but it’s always a real pleasure talking to you and, and, you know, thanks for the forum here because I do it to, it’s an incredible time to be working in this space.
Harry Glorikian: Yes. Excellent. All right. Well, stay safe.
Tom Chittenden: You too. Thank you, Harry.