Dr. Noubar Afeyan, Founder and CEO, Flagship Pioneering; Chairman, Moderna
Interviewer: Alan Murray, Chief Executive Officer, FORTUNE

Fortune Brainstorm A.I. 2021: How A.I. in health care promises to save lives

Conferences

it's in healthcare that AI may have the greatest impact on human condition, helping diagnose and treat disease saving and improving the quality of life. Just some examples AI is helping bio pharmaceutical companies quickly sift through existing compounds to match them with emerging diseases and it's aiding in the development of new drugs. It has revolutionized medical imaging and this is only getting better. Our next guest. And I'm so excited to introduce this next session brings macro and micro level insights into life science technologies, both as a serial entrepreneur and a biochemical engineer. His company was at the forefront of one of the most powerful and impactful collaborations in recent times developing one of the covid 19 M R N A M. RNA vaccines in record time. Dr New Bar Afyon has more than 100 patents and has co founded and helped build more than 60 science and technology startups. He's the co founder and ceo of biotech incubator. Flagship, pioneering and co founder and chairman of a company you may have heard of Moderna Dr Afyon is joined in conversation by Fortune's Chief executive Alan Murray. Welcome. Mhm. Yeah. Yeah. Yeah. New Bar, thank you so much for joining us. Uh and by the way, thank you all for joining us. We've been trying to do this for two years. It's really great to see all of you in in one room and great to have a robust virtual audience as well. So so thanks we have a lot to talk about. You've been talking about the vaccine for a while now, I'm not going to start there, but I would like you to explain to the group what flagship pioneering is because it's a pretty unique organization. Sure. So thanks for having me. Uh Flagship pioneering is based in Cambridge, the 21 year old organization and it's basically an experiment uh in two things. One whether one can make breakthrough innovation systematically over and over again and the second whether one could create First in category companies in a systematic way. So we're an institution were about 230 people. Today we have our own internal R. And D. Capability by which we invent new platforms. We then convert them into companies that we own and operate for a few years and then we spend them out. So think of it as a company creation company. But we have our own capital. We have our own teams that lead these companies. So it's a it's effectively what traditional startups would have to organize one at a time in terms of a combination of science, management and capital. We have all three in one entity have had for 20 years and over all these cycles of continuously doing this, we've learned well over 100 companies have come out of this now how to do this in a more institutional way. And I think you said to me at one point you're in the business of solving impossible problems. Well finding them and then trying to solve them. So it's one of the frustrating things when we describe what we do is that the world wants to believe that it's either worse finding solutions to big problems are finding problems for solutions that already appear. And we do neither. We basically try to reiterate solutions and problems until we find matches. And then if the match points to a major value proposition then we'll build that company. So we're quite agnostic as to what technology and what problem we're going after. Once we discover them, we try to validate whether those are big global issues but we don't start by obsessing over either. Well the vaccine was certainly the solution to a big global problems. So thank you for everything you did in that area. I think we all appreciate it. Um How has data and machine learning and ai transform the way you think about like the life sciences areas you work in well. Um So back up and say that kind of the machine learning AI component is which is kind of intensified in the last few years, has its roots 25 30 years ago as well. And life science has always been about data in that we're studying systems that precede us right most other human endeavors study things that humans made. We don't and so and we never did. So what the only way you can do that is to get better and better measuring and then interpreting and then coming up with models often inadequate models with which to make slight perturbations which we call drugs or or other such things. So so it's always been about data and the more we've been able to measure both more precisely and in more diverse ways. The more data we generated, the more we've needed better and better models. The machine learning advances in the recent past simply have given us better models to do that with. And that has increased if you will are leap size. Now there's another side of this week leap leap size. Yeah, the size of the steps we can take. So you can, you know, basically if you can take a whole system and keep it intact and look at it from many different perspectives and then say this is what's happening, then the kind of questions you can ask is very different than if you have to disintegrate the system and study time little portions of it. Which is what molecular biology has been doing for a long time. I think about drug discovery. What we do is we take a complex human and we disintegrated into ideally a single protein that we then make a drug against. And then we're worried that somehow when we put it back together, it won't reproduce in clinical trials, well, it shouldn't reproduce in clinical trials except rarely. And that's what we observe. But if you could not disintegrate the human and actually do uh drug discovery or drug development using data from an entire human. Better yet millions of humans. Uh and those measurements are informative enough as to the model of what's happening. It's a different ballgame. And that's what we're beginning to switch into. And how quickly is this happening? What what what is the next decade going to look like in terms of advancements you can make using these technologies. I mean of course, you know, we're in the business of optimism and so it feels like we're in an exponential period which means that a year ago we couldn't imagine what we're doing today. And two years from now. We'll laugh at what we're doing today and that's the definition of kind of a period where things are changing very rapidly. For example, one of our most recent projects, the company that's become known as Valley Health. But it was just a platform in our labs a couple years ago has literally collected millions of humans equivalent of disease and normal data and is using that to motivate the kind of decisions it makes to what drugs to discover, not just choose. And that type of vertical integration and kind of digital native version of conceiving of a pharmaceutical company was not possible two years ago, three years ago now we're experimenting with it. Now the version one point oh that will create a lot of value. But then there'll be a two point and three point Oh, so the learning cycles are happening. Another example I can give you is that several years ago, one of my colleagues was in the audience of Actovegin form the company that is called now generate buy medicines. It's what it's doing is essentially trying to de novo design brand new proteins. So we're given 100,000 plus or minus proteins in our human body and that defines us. Uh and yet the biotech industry wants to bind to those proteins, change those proteins use them as vaccines, like the spike protein etcetera. The question is, what can computers do in that field so far? They've been able to observe and and analyze these proteins. What we think we can now do again, based on machine learning based models is to de novo generate an arbitrary protein for an arbitrary function. Now the difference between DNA sequence and a protein function is vast when it comes to information processing and yet ourselves do that every single day. It turns out our DNA in our genome does not know what a protein is, let alone that its freedom actually folded, let alone that it binds to something. And yet we have tons of interesting proteins. Now we're building models that can similarly correlate DNA sequence to a protein function, which means that you can specify protein function. I want to buy into this thing at this place and not everywhere else and it'll spit out a DNA sequence, you can try in the lab and learn in a few cycles how to make a completely novel protein that has never been possible in the 35 years I've been in biotech and that's becoming possible as we speak, huge and that sort of starts to take you in the whole area of personal uh personalized medicine. I mean if you can where you can create medicines particularly designed for my DNA to create the proteins, I need to cure myself. Yeah, I mean that you know the personalized part of this has always been a promise that's been used to motivate at say investment and societal interest. It's overblown. I think it's overblown because I think that many, many things we don't yet know how to do can be solved with non personalized but more I'll call it deterministic medicines than probabilistic medicines and by that, I mean if you actually make something for a purpose where you know what's causing the effect, not because you took it out of context and said this is the protein but actually it derives from the intact system. Then I think these new techniques, particularly some of these new classes of molecules that are information molecules like M. R. N. A. You can begin to start thinking about drug design, not as kind of a fancy word but actually something that should work. So I mean it's one of the things people who are not in this industry don't realize is that every day an R. And D. Person goes to work in a pharmaceutical company in general. We have many of them Basically. It's a probability game. So you have like a 5% chance of succeeding. And for the next seven years you probably have less than 50% of succeeding. Well that's a different business model than when people are designing. You know, iphones and these things are supposed to work now. What's it gonna take for us to go from? You know, having a low probably success to a pretty high property success information models and the ability to do things through information molecules. All of that is beginning to be available. So we think that's what's going to change it. So the personalized version is yet another extreme individualized to you. But we share a lot in common when it comes to many many diseases with the exception of cancer where our genomes are basically going separate, diverging from each other. They are the personalized may matter most but we've got a lot of other diseases. Obesity. Alzheimer's where the direct personalization may not need to be an extreme to get much higher problem is generalized answers. I mean commercially it's much easier if it's generalized. Yeah. Yeah. So so cast I want to open it up for questions by the way. If you have a question uh raise your hand. I see one right back there. Uh Why don't we go ahead and take that now? Thank you. Uh I'm Abby Mukherjee with Sarah Adventures. We get bombarded with this uh new startups that are trying to disrupt how the medicine development or biomarkers for cancer have to be detected. While there have been some advancements in computational biology. My predicament is, for example, clinical trials, the people of color are extremely underrepresented but take that and try to find the biomarkers or some models that they are basing off this data, which is completely off. We have been seeing the trends when you take your comprehensive blood panel. The numbers are way off for asian versus an african american versus a caucasian, male or female. Um, so how do you, how do you think we can solve this? Systemic and systematically when there is so much of challenges in terms of data collection exists at the grassroots level. So, I mean, it's a very particular question, but I'll try to give you at least how I would think about this. First of all. Um, we have a regulatory system where it's not even clear that if you solve this problem, you can make any money doing it because we don't develop drugs against biomarkers today by gen generally that's not allowed and it's frowned upon. And in the diagnostic space you go, we have to go through a long journey to be able to use biomarkers and diagnostics. So I would say that if, but for that you just need more data and more inclusive sampling, which I think you can do these days, especially since there's so much uh interest in people and learning things about themselves. You can easily create I believe and people are doing this ways to get people involved in this process where they feel collaboratively included in a research project that's going to generate markers for their own disease. Just look at Alzheimer's where there's been so much controversy around biomarkers and Alzheimer's and yet every single family who has a member with with Alzheimer's is absolutely happy to help find bite marks. So I think that the measurement and data problem can be addressed. The problem you end up in is that with these biomarkers is the chicken and the egg. It's not clear you can actually create value from it even though scientifically it's incredibly valuable. And so that's where I think the system breaks down. Uh you you talked about data, we have a question from the virtual audience about the reliability of data. Um It says I recognize the challenge for reliable data sets in health, good clean data on primary endpoints. Should we introduce ai when it is not yet perfect. I mean yes because our alternative is even less All right meaning human intelligence more imperfect. I mean it's it's look I mean I think at the end of the day Um again I've I've worked most of my career except the last few years in actually making instruments that measure things in biology. That's the first company I've built for 10 years. And what you realize is that you make these measurements and then humans Use pretty basic statistical techniques to basically simplify the data classifiers, et cetera and make sense out of it. So, for example, the top 10 genes that turn on if you have cardiovascular disease is an irrelevant concept. Biologically there's no such thing as top 10 in biology. That's a human construct Because we also have the top 10 reasons people buy diapers. And so the techniques that are being used, I'm not kidding actually are much the same techniques. These are statistical regression techniques that we've used in biology. But if you could actually train a system to without a model of every single interaction, be able to say these are the variables that are driving behavior without a human superimposing top tens and all that. That's a different ball game. Even if the data is imperfect and you can say about it might mislead you. It's certainly going to do better than simplification the way humans where we are. So, I think it's, you know, two years from now we're gonna wish we had done things with with low quality data because by that time algorithms will figure that part out and the more better data we get, the better, the further we'll get. We don't have much time left. But I'd love for you to just take us 10 years into the future. What what is going to surprise us about how health care is practiced boy I that's that's even though I have a weird name, it's not Nostradamus. So I'm not I'm not gonna I'm not gonna answer that question but I but I will tell you this because we have such little time I leave the audience and thinking the application of AI in the field that I'm in. Like science field is a particularly intriguing one because unlike in most fields where it is interpretive and pattern recognition driven in our field that will be generative and so just like life science deals with with things we don't understand. AI will generate things we also don't understand and we'll be able to use life science approaches to decipher those things. And that's a very interesting other use of life science and AI than people are thinking about now. So the idea of how you generate a new natural reality, I hate calling it artificial. And so the life science space will expand based on the generative nature of Ai fascinating. I wish we had more time to get into that. But that was a great introduction to the topic. Thank you so much