Artificial Intelligence and Business Strategy
In collaboration withBCG
Frank Nestle, Sanofi’s global head of research and chief scientific officer, was inspired to enter the health sciences field after reading an Albert Camus novel and realizing his calling was to help others. In his current role, Frank oversees the pharmaceutical company’s transition from primary care to specialty care, which includes developing medicines for immunology, oncology, and rare diseases. In this episode of the Me, Myself, and AI podcast, Frank explains how artificial intelligence enables Sanofi to work toward drug discovery in more agile ways.
Read more about our show and follow along with the series at https://sloanreview.mit.edu/aipodcast.
Give your feedback in this two-question survey.
Sam Ransbotham: Artificial intelligence has the potential to scale drug discovery like never before. Find out how one pharma company uses AI on today’s episode.
Frank Nestle: I’m Frank Nestle from Sanofi, and you’re listening to Me, Myself, and AI.
Sam Ransbotham: Welcome to Me, Myself, and AI, a podcast on artificial intelligence in business. Each episode, we introduce you to someone innovating with AI. I’m Sam Ransbotham, professor of analytics at Boston College. I’m also the AI and business strategy guest editor at MIT Sloan Management Review.
Shervin Khodabandeh: And I’m Shervin Khodabandeh, senior partner with BCG, and I colead BCG’s AI practice in North America. Together, MIT SMR and BCG have been researching and publishing on AI for six years, interviewing hundreds of practitioners and surveying thousands of companies on what it takes to build and to deploy and scale AI capabilities and really transform the way organizations operate.
Sam Ransbotham: Today, Shervin and I are talking with Frank Nestle, global head of research and chief scientific officer at Sanofi. Frank, thanks for talking with us. Welcome.
Shervin Khodabandeh: It’s really nice to meet you, Dr. Nestle.
Frank Nestle: Hi, Shervin. Nice meeting you, too.
Sam Ransbotham: Frank, maybe start with your current role at Sanofi. What are you doing?
Frank Nestle: I’m the global head of research and CSO [chief scientific officer] at Sanofi. Sanofi is really a company making medicines available across the world, across 100 countries, with approximately 100,000 employees, providing medicines and vaccines literally to millions of people.
We’re currently going through a very exciting transformation; we’ve been originally in primary care and now are transitioning into specialty care, providing medicines in immunology, oncology, rare diseases — including rare hematology diseases — neurology, as well as vaccines. And my role as global head of research is to really discover and then translate the next generation of breakthrough medicines to patients, who are really waiting, I can tell you that.
Shervin Khodabandeh: Tell us a bit about your background, Frank — how you got started in your career and how you ended up here.
Frank Nestle: I’m a clinician and a scientist. I trained originally in dermatology and clinical immunology, allergology. But I was always driven by the quest to make a difference to patients, and I did that by trying to understand disease mechanisms and then translating those mechanistic insights into therapeutics — hopefully, and ideally, precisely tailored to the needs of an individual patient or a patient population. Now it’s called precision medicine, but that’s what always drove me.
And in terms of the science, I tried to manage to get to those disease mechanisms. One is the science of the immune system. And I’m happy to say, with maybe all your listeners now, that during the pandemic, we all became immunologists.
Shervin Khodabandeh: Thanks to Google.
Frank Nestle: So there’s a lot of know-how out there about what’s the difference between an antibody, a T-cell, and a B-cell.
Sam Ransbotham: I’m not sure there’s a lot of know-how. There’s a lot of experts.
Frank Nestle: So, lots of experts, and that makes actually doing the science I do even more fun because you can totally dominate any dinner conversation. The other topic I could easily talk about for quite a while is dermatology, which is always a nice topic across dinner rounds.
But it’s incredibly fascinating to think about the immune system. It’s essentially a collection of mobile cells circulating through our body, and they move back and forth between tissues where cancer happens — autoimmune disease happens — and the blood. And they’re really accessible via a simple blood draw, and it’s quite exciting to understand how this connected system — my strong belief is all biology is connected — is playing out in cancer, for example, to protect us from cancer, and we have huge successes in checkpoint immunotherapy of cancer, but also if it gets out of control, if it overreacts in terms of autoimmunity. And the fundamental hardwiring for the immune system is really what we apply and exploit in vaccines. Our immune system has evolved to fight pathogens.
Shervin Khodabandeh: Tell us how the new world versus the old world of immunology is being transformed by technology, possibly by better analytics, better AI.
Frank Nestle: There’s actually a great story [about] how the great convergence of life sciences with data sciences and engineering has played out to increase our knowledge about the immune system, and it all converges on the topic of single-cell immunology. We can now assess and analyze the immune system of a patient with just a few thousand cells, and we do this by applying single-cell immunology technologies, where we can study 2,000 genes per cell. So think about the magnitude of gene modifications we can study if we take 100,000 cells, and across these 100,000 cells, we study 2,000 genes per cell. And this has been possible because of progress in engineering, microfluidics. It has been possible because of AI. For example, at Sanofi, at our institute, we are writing code, we’re writing an AI algorithm, to then essentially analyze those data and to rediscover cell fates, and we annotate, then, if a cell is a B-cell or a T-cell or some new immune cell we never ever heard about.
The exciting fact is also that we can not only get cells from the blood, from circulation; we can go into, for example, the joints of a patient with arthritis, or even the cerebrospinal fluid of a patient with multiple sclerosis. So all of a sudden, we can go into the compartments where disease plays out and then literally put this disease under … you could call it a molecular microscope.
And if you think about the fact that the cell is the common denominator of a physiological system, of the immune system, then you could call it essentially the atoms of the immune system. We can study the immune system literally at atomic-level resolution.
Shervin Khodabandeh: And the use of AI here is both to understand the mechanism as well as to maybe come up with new drugs?
Frank Nestle: Exactly. So the first part is, I’m always saying, first you have to explore and create a map, a landscape, of physiology and of pathology, and this is what we are currently doing. And then, once you have all that insight, you create the theory behind it — how the mechanism plays out — and then you make drugs.
First, we want to understand how a disease mechanism plays out, and that’s exactly where we use these AI algorithms. Imagine you have 100,000 cells, and they have 2,000 genes up or down, and we don’t know if it’s a T-cell, a B-cell, or some new immune cell. So if you run a learning algorithm on these data sets, that algorithm gets better and better to actually tell us, so you essentially reorganize, or rediscover, the cellular annotation of the immune system.
Shervin Khodabandeh: It seems like in some applications of AI that we — Sam and I — have been talking about, it’s an existing business process that gets improved a bit, or optimized a bit. And, Frank, you’re talking about applications that wouldn’t even exist without AI, I would assume, because the scale that you’re talking about, with thousands of cells and thousands of genes, it appears to me these would have been unsolvable problems without the use of AI.
Frank Nestle: I tend to agree. Always, the question is, “What do you mean by AI?” There’s a whole spectrum, from AGI, artificial general intelligence — and we are not there yet — to various applications of machine learning, where you … When I started off — many years ago — to use machine learning approaches, these were just simple clustering algorithms or these were, for example, random forest-type machine learning approaches. But now we can obviously do much more in terms of applying AI.
The discovery of medicines is then a whole different order of challenge. And this is where, essentially, you get into this question of molecular design. Medicines come in different sizes and shapes. If they’re below 500 daltons, then they’re called small molecules, or above, they’re large molecules. Small molecules are your typical pills you take, and large molecules could be biologics: antibodies you inject. And as we know from COVID, you can inject antibodies to protect you from COVID, but you can also take small-molecule pills to basically block, for example, the viral replication of the COVID virus.
So how do you get to those molecules? You first have to understand your target. So, for example, if you think about a target in multiple sclerosis or in systemic lupus, you have a certain target; it’s a protein. You have to understand the structure of that protein, and this is called structural enablement. We can now use fantastic technologies, such as cryo-electron microscopy or X-ray resolution of those targets, and we can discover their structure.
But what we can also do is, we can run virtually literally billions of small molecules against those targets and then discover small molecules hitting those targets in a way that it has a functional impact. And these are typically called then allosteric inhibitors, so an inhibitor which essentially binds to a target, to a protein, and does so in terms of an output that is functional. So it can then, for example, block a certain target, block a certain protein.
And this is then the first step to a medicine because we have, all of a sudden, a target. We have a disease where it plays a role. We have a tool compound — for example, a small molecule. And then we can take this starting chemistry to ultimately then get to a real medicine.
Sam Ransbotham: You’re talking about trying all these things. Are you trying them in real time, in the physical world, or are you trying them virtually?
Frank Nestle: It’s a combination of the real and the [virtual] world. I’m always talking about AI as the chemist around the table. For example, if you want to come up with a new structure of a small molecule, a medicine, we used to do crowdsourcing. We have more than 300 chemists in our organization, so you could send that crowdsourcing request out there, and you would have 300 chemists’ brains engaged in finding the perfect molecule.
But if you have AI at play and use generative algorithms, then you can literally go through not millions but hundreds of millions of potential structures and then optimize them. It’s just an order of magnitude larger of what we used to do. For example, we did high-throughput screening just with a few hundred thousand molecules. Now we can do it with hundreds of millions of molecules, and we can do this virtually.
But then that virtual screen will get you just to a hit — we call it the first iteration of a potential medicine — and then this has to be optimized. And it’s actually this cross talk between the machine and the human, which is happening all the time.
Shervin Khodabandeh: And the human in this case is a chemist?
Frank Nestle: The human would be a chemist, yeah, exactly.
Shervin Khodabandeh: So chemists still have a job here?
Frank Nestle: Absolutely. Absolutely.
Sam Ransbotham: Yeah, I was going to ask what that crowd of chemists thinks, that suddenly 300 chemists are not asked for their input. What do they think about this?
Frank Nestle: Exactly. So they actually enjoy the challenge. Just simply what we call the ligandable space, or the chemical space, we’re exploring is increasing constantly. And through the increase of that chemical space, we can come up with completely new molecules. There’s nothing more interesting for a chemist to be faced with than a new molecule they haven’t even thought about and then put that into motion.
But it’s a long road from this original hit to ultimately a clinical candidate. It takes typically four to five years. And then you have to clinically translate this; this takes another eight years. So it’s a long journey from this original hit, but this is where, essentially, a molecule is born. A medicine is then born later in the clinic, when we do proof-of-principle studies in patients.
Sam Ransbotham: It seems a little bit difficult though. If you’ve got suddenly many, many more hits, doesn’t that create a big impact on your process and your workflow downstream from that?
Frank Nestle: You might think so, but actually the reverse plays out.
Sam Ransbotham: Oh, really?
Frank Nestle: So our experience tells us that these models, they predict structures, and they help us to reduce that enormous space to a much smaller space. So I’ll give you an idea: We typically had to synthesize about close to 5,000 molecules with the typical drug discovery paradigm we were applying a few years ago. If you use the support of a predictive model, like an AI model — and there are multiple steps in the value chains to get to a molecule to ultimately apply models — you can reduce this to 500.
And this is the big promise now — and it is really important to understand that — is that predictive power of an AI algorithm being fed and trained with a lot of data sets, refining the number of molecules we need to conceptually come up with and then test in our assay system. So actually, what it does is it reduces the investment we need for synthesis — and it’s expensive to synthesize compounds — but then also to test these compounds. And ultimately what this can lead to is that we reduce the timelines. And if you understand the mathematics or the economics of drug discovery and development, it’s all about timelines. If you can shave off one or two years from the 10 to 13 years it takes to get a medicine to a CVS near you, then we can dramatically alter the economics of drug discovery and development.
Shervin Khodabandeh: So going from the 5,000, which in the old world I assume would’ve had to be tested by trial and error, to the 500, you are limiting the universe that then is going to be developed and tested. Is it possible that in that process of excluding the other 4,500, you might throw away some good candidates?
I’m curious how the learning happens here, because in most other AI systems I’m close to, there is a truth data where the algorithm learns from.
Frank Nestle: Yeah, what’s the truth?
Shervin Khodabandeh: Right. If you’re throwing away structures that the engine plus the chemist might think is not going to even work, are you possibly throwing away some potentially unborn good drug candidates that never saw the light of day?
Frank Nestle: Yeah, that’s a good point, and it also depends on the different steps of the value chain. When we make a medicine, we are testing different attributes. For example, just one attribute might be very potent binding. The next attribute might be highly specific binding, not hitting other targets. The next attribute might be, it’s safe; it doesn’t, for example, lead to cardiac malfunctions or other side effects in the liver. The next attribute might be, a molecule is well absorbed by the gut. So the next attribute is distribution in the body, into the organ system you need.
So all of these different attributes are optimized by our dedicated models, and these models are trained. I’ll give you a specific example. We use what’s called Caco cell lines to mimic what gut absorption looks like, and we mimic that by understanding how well the cell line is absorbing, or taking up, a molecule. So we’re running thousands of these Caco cell lines all the time, and we are then studying — the truth is how well the molecule is absorbed in that Caco cell line, and that gives us then a high ranking of a molecule. And the system then learns from this iterative process of what a good Caco cell line uptake molecule is. And that ultimately then gives us a better and better model. So this model would be then one of the many different models we would use to predict those 500 we would ultimately test. Does that make sense?
Shervin Khodabandeh: Yes. Yes, it does, because what you’re saying is, the molecule you would select has to have different attributes. And these attributes are real-world-tested on other molecules that might look similar to these molecules.
Frank Nestle: Exactly. Yes.
Shervin Khodabandeh: So some kind of a clustering, or molecules that look like this, or have these kinds of chains, or whatever, or these kinds of ligands, work better. And so that’s how you’re infusing that learning into the algorithm, if I understand.
Frank Nestle: Exactly.
Sam Ransbotham: I think what’s important there is, I think we were guilty of slipping down into this binary classification thing, where it’s either a good drug or a bad drug, whereas you’ve got hundreds of different attributes you’re looking at, and you can play with each one of those, and you can test on each one of those so that it’s not just that crude up/down, yes or no, that we were kind of crudely thinking before.
Frank Nestle: We call it an optimization process. It’s a high-dimensional optimization process, where often, if you then remove a certain molecular piece of the molecule, you might improve the uptake, but you might decrease the potency. So it’s actually a give and take. And this is what chemists are very good about, and this is where they enjoy — because it’s such a high-dimensional space, and we have so many data available — to have that partner, which is AI, to tell them what the AI thinks.
Shervin Khodabandeh: This has been fascinating, Frank. So, we’re going to transition to a segment, which is a series of rapid-fire questions. So I’m just going to ask you five questions. You probably don’t know them yet. And so just give us the first answer that comes to your mind. Are you ready for this?
Frank Nestle: Sure, sure. Absolutely.
Shervin Khodabandeh: Tell us your proudest AI moment.
Frank Nestle: The proudest AI moment was when we, for the first time, could annotate single-cell immunology cell fades, with our in-house-developed AI algorithm at Sanofi.
Shervin Khodabandeh: That’s fantastic. When was that?
Frank Nestle: That was 2018.
Shervin Khodabandeh: What worries you about AI?
Frank Nestle: What worries me about AI is that people don’t understand what AI is. They always think about artificial general intelligence as like machines replacing humans. That’s not at all what I’m seeing. If you see where self-driving cars are at the moment, there’s a lot to be solved until you get even close to it, but what AI really is, is for very specific questions, with a good data set, high computational power, and a good algorithm, to solve problems a human brain couldn’t do. And these little, small contributions of AI are making all the difference, certainly in terms of what I am seeing in the R&D value chain.
Shervin Khodabandeh: Your favorite activity that involves no technology.
Frank Nestle: Cycling.
Shervin Khodabandeh: The first career you wanted in childhood.
Frank Nestle: I always wanted to go down the route of being a writer and a director — actually, a theater director. And then I was reading Albert Camus, La Peste [The Plague], and I was studying philosophy and literature. And there’s a person, Dr. Rieux, who’s fighting the pest. It’s very timely, with the pandemic. And he’s fighting this l’absurdité, the absurd existence, by doing good, by helping.
And I’m not as philosophical, but when I entered medicine, it was just a way of doing good and doing something useful. And now we do it at scale by hopefully finding the next medicine, transforming patients’ lives.
Shervin Khodabandeh: It’s a great antithesis to Albert Camus.
Frank Nestle: Yeah, exactly.
Shervin Khodabandeh: And your greatest wish for AI in the future.
Frank Nestle: Make it more explainable, both at the level of getting it out of the black-box situation, make it explainable in terms of understanding what it does, but also explain it to people so that there’s not this misunderstanding of AI. A lot of people just project their fears or their misunderstanding on that unfortunate small acronym. And once you use it in certain contexts, it’s actually transformative.
Shervin Khodabandeh: Thank you, Frank. This has been incredibly insightful, and I’m sure, very valuable for all of our listeners. Thank you so much for this.
Frank Nestle: Thank you for having me.
Sam Ransbotham: Yeah, great to talk with you. Next time, Shervin and I talk with Stéphane Lannuzel, Beauty Tech program director at L’Oréal. I’m always up for a good episode about cosmetics. Please join us.
Allison Ryder: Thanks for listening to Me, Myself, and AI. We believe, like you, that the conversation about AI implementation doesn’t start and stop with this podcast. That’s why we’ve created a group on LinkedIn specifically for leaders like you. It’s called AI for Leaders, and if you join us, you can chat with show creators and hosts, ask your own questions, share your insights, and gain access to valuable resources about AI implementation from MIT SMR and BCG. You can access it by visiting mitsmr.com/AIforLeaders. We’ll put that link in the show notes, and we hope to see you there.