The Bioinformatics CRO Podcast
Episode 10 with Mark DePristo
We talk with Mark DePristo, founder and CEO of BigHat Biosciences, about building better antibody therapeutics using machine learning, overcoming the fear of failure, and the pros and cons of working in academia, for large companies like Google, and at a biotech start up. (Recorded on January 21, 2020)
On The Bioinformatics CRO Podcast, we sit down with scientists to discuss interesting topics across biomedical research and to explore what made them who they are today.
You can listen onSpotify, Apple Podcasts, Google Podcasts, Amazon, and Pandora.
Transcript of Episode 10: Mark DePristo
Grant: Welcome to The Bioinformatics CRO Podcast. I’m Grant Belgard, and joining me today is Mark DePristo. Mark, can you introduce yourself please?
Mark: Delighted to be here, Grant. Thank you for the invitation. I’m Mark DePristo. I’m the CEO of Big Hat Biosciences, a San Francisco based AI for drug discovery startup. My background in that space has really been at the intersection between bio and tech for about 20 years. I was an undergrad in computer science and math. And I won a Marshall Fellowship that sent me to England and I got a PhD in biochemistry, and I really got the bio bug after that.
I actually went to Harvard and was an experimentalist for three years. So I learned how to pipette, though pretty poorly. I saw the full stack of that. And from there I went and was at the Broad Institute and most recently at Google applying AI to bio in the broadest way possible.
Grant: Fantastic. So, can you tell us more about what you’re doing at Big Hat?
Mark: Big Hat Biosciences is really focused on radically improving the design of antibody therapeutics. And we’re doing that to enable a next generation of even more sophisticated therapeutic molecules, which I’m sure we’ll get into in more detail. And how we’re doing that is we’re really leveraging recent advances in AI and machine learning as well as synthetic biology techniques to build a new type of wet lab, that’s very high cycle time. So it can do a lot of work quickly. And it’s coupled at every part to AI and machine learning technologies to guide that. So Big Hat is really a close-loop antibody engineering shop. And we drive that technology to basically do data-driven or rational antibody design.
Grant: And can you talk a bit more about what the ideal Big Hat antibodies would be able to do?
Mark: Yeah. So to really answer that question, I think we have to sort of talk a little bit about the history of antibodies. So, the first wave of antibody therapeutics were all based on monoclonal antibody technologies. And what that means is you’re basically repurposing an immune system–it could be mice, it could be human–to give you an antibody that the body can produce. It’s a natural product of the body. And those molecules combine to all sorts of different things. You can make them bind to surface receptors. And this is really the origin of the top drugs today that are biologics that are all antibodies.
The challenge with monoclonal antibodies is not that there’s something fundamentally wrong with them–they’re very good molecules. It’s just that they’re very limited in what they can do. They’re the natural product of the immune system.
So they intrinsically do all the things that the immune system wants them to do. Right. They interact with the immune system, they activate inflammation pathways. They have a very specific way of binding interacting with targets. And they’re really, really big. I mean, these are massive molecules and all of those issues limit what you can do with them.
And so after about 20 years and monoclonal antibodies, there was this huge explosion of so-called next generation formats. And these are antibodies that are engineered to not be natural products. They’re now, more designed around the needs of therapeutics. So there’s all sorts of things like molecular glues that use parts of antibodies to stick things together.
You can stick molecules together, you can stick cells to molecules. You can stick cells to cells with these bridges. You know, we want to make small antibodies. We want to make antibodies that are environment sensitive, so change their behavior and pH. Change their behavior in the presence of other molecules.
Of course, those don’t come out of human immune systems. Don’t come out of mice. We own the engineering of that. And I think Big Hat’s really founded on the problem, that gives rise to Big Hat, the commercial entity, is designing these so-called Frankenstein antibodies is just incredibly difficult. And it’s difficult along a whole bunch of dimensions that we’ve never been particularly good at. It’s really challenging to do rational drug design. No matter whether you’re working on biologics or small molecules, it’s just very difficult.
Two, you have this enormous space of possible antibodies, right? You have this sort of combinatorial search problem on which amino acids to put in the antibodies. That caused the search space to get really big. It can be very hard to find good molecules.
And finally, because most monoclonal antibodies are coming out of organisms–human, mice etc.–they’re pretty good molecules. Like they have to be tolerable in the body. Once you start to engineer those molecules and make Frankenstein versions of them, they don’t really have to function particularly well.
They can be deeply unstable. They can aggregate. They can do all sorts of things you don’t want them to do. And removing those things, stabilizing the molecules, removing its aggregation propensity, this is really hard to do rationally. And so we have these amazingly exciting next generation antibodies, things that are transforming cancer therapies, immune therapies, even infectious disease, but we have an unbelievably difficult time creating them because the processes we built for natural product discovery that worked great for monoclonal antibodies don’t really help you on that engineered molecule because they’re not coming out of organisms. And so the problem is fundamentally different and that challenge is sort of what Big Hat is focused on addressing. And the modern technologies we use to address this is sort of why we’re able to do this now. That’s really what’s changed about the world.
It’s not that people didn’t want to do this 20 years ago. They were all very excited. It’s just that 20 years ago, when I was doing this work, making a couple of mutations to an antibody, it would be weeks of work. And today this is something that Big Hat does routinely.
Grant: That’s incredible. So, what kind of balance do you have on your team between ML people, structural biologists? Is it usually people who came out of PhD programs, where they did machine learning for structural biology? Do you get people who trained up through very different fields and then they apply that to another field at Big Hat?
Mark: That’s a great question. So it has to do a lot with sort of what is the structure of the teams and the organization at Big Hat. You know, when we were really forming Big Hat, there was this open question of like, how should we organize the teams? Should we have a computational biology co-organization on one side and an experimental biology work on the other?
And that’s what you typically see if you look at most companies: you’ll see that kind of computational versus experimental division. Often from the very start of the org structure. We really were nervous about going down that route. We really saw the value of Big Hat in part through the integration of the wet lab and dry lab tech.
And we really didn’t want people to be thinking about it being somebody else’s problem to do data analysis or somebody else’s problem to produce high quality experimental work. And so we actually have totally not gone down that route and we’ve really structured Big Hat more around the projects we’re working on.
So in a matrix, we’ve pivoted the matrix to another dimension, and organized the company that way. And what that means is that it allows us to produce teams that are what you could imagine, skill complete. Like all the things that the team needs to be able to do has at least some representative expertise on it. And it’s manifested in, three or four people, that have eight or nine skills that are required. So everyone is multifunctional, but the team isn’t complete unless you have all the people.
It means that we mostly hire people for two attributes: they have some number of skills that they can bring to that team and they work really well with the other people on that team. It’s hard to be the Atlas and push up all the problems at Big Hat on your own. It’s just not going to work because nobody has the expertise across all the tech that we work on. We work on everything from DNA synthesis to active learning technologies on the cloud. That’s just not reasonable to ask for. So we composed Big Hat out of a bunch of puzzle pieces that all fit together to give a picture of what we need to do.
Grant: So I noticed on your website, you talk about being a team oriented, inclusive, remote friendly, and family centric culture. Were you remote friendly from the beginning? Was that accelerated by COVID or is that by design?
Mark: It was both by design and by necessity. Actually, we were remote from our first hire. Our first hire is a guy named Eddie Abrams, who’s the VP of engineering who I worked with at SynapDx before Google. Eddie, he’s a fabulous guy, and Eddie lives in Arizona.
So the choice was simply: do we want Eddie to join us and be remote in Arizona at the start? Or do we not want to do that? And we made the right choice. We took Eddie on. It’s been a fabulous journey with him as employee number one, but that happened a few months before COVID started. So we built Big Hat from the ground up to make sure that our first employee could be productive in a remote-oriented culture.
And so that’s turned out. Obviously we didn’t anticipate how important that decision would be, but it proved great. And now Big Hat is highly distributed all over the country. In fact, I don’t believe that the center of mass of Big Hat is even close to San Francisco at this point. Most of the company is actually stationed on the East coast.
That’s the remote answer. So the remote answer is definitely very remote friendly. I mean, we were just fundamentally remote by definition. I was actually at Big Hat yesterday for the first time in months. So it was great to see everyone again and see what the lab looks like now. But we try to do that because you know you want to eat your own dog food. So I try to stay remote, so I’m sure that the company is remote friendly.
Really inclusive is just this vision, that two things: we want to be flexible. You know, not everyone wants to work the same hours in the same time zone in the same way as everybody else. And in particular, it’s clear that the people who need the flexible hours are often people who have more complicated lives through any number of things. I mean, I have three young kids and this introduces so much logistical complexity in my life. You know, I need to work in an environment that’s flexible and I have a lot of benefits to being able to do that.
So. Big Hat has structured itself around this idea that people should work when they need to work. We’re going to be remote across many time zones. Like we can’t say the business hours are 9 to 5 Pacific time and demand everyone be there. So we really are a more results oriented place. We try to make sure that it’s very clear what everyone’s working on.
We all work on it together. It’s very collaborative. And we care not about face time or how many hours, but really are you contributing to the mission and the goals of Big Hat. That focus on results and flexibility means that we’re able to recruit lots of people in a much more inclusive way than if you’re physically located in one place with very specific timelines and when you should work.
And family centric is just as simple as basically Big Hat forces working hours on people. So, what we mean by that is we don’t want people to be getting emails at midnight. Like if people are sending emails on the weekend, this is strongly discouraged. The expectation is that everyone is working a 40 hour week, and they’re not killing themselves to put in an 80 hour week because we don’t want you to do that.
We don’t want people to burn out. We want to make sure that people who have other obligations in their lives can fulfill those obligations without worrying that Big Hat is somehow unhappy with that. And what it really forces honestly is two things: one is it forces real prioritization, right? Like we’re not going to work.
Thursday night, just because we want to do a little bit extra work that week, right? Like, unless that’s critically important. For example if we have a deadline on that Friday, in which case we do this. But that can’t be your MO because ultimately Big Hat is really a collection of people. And that’s why we’re successful is that we’re able to have really good people who like to work and are very productive in the environment we’ve created.
And in some sense, the worst possible outcome for Big Hat would be to set up a culture where we scour the earth to find all the right people to work with us and then bring them into an environment that’s such a pressure cooker that they burn out almost instantly, and then they’re not productive. And that’s a huge failure mode for tons of startups.
And it’s one of the biggest questions we get. I mean, people will come and say, you guys are such an early startup. Are you expecting me to be here at 9:00 PM every day? And it’s really a pleasure to be able to tell people that no, that’s not at all what the culture is like. I stopped working at 5:30.
Like that’s it. My kids are home. Dinner is happening. I’m going to go do that. So, that’s forces real prioritization, real focus on the company. And actually it turns out to be the right answer for the company as well. You want to be a focused company that only does the critical stuff.
So drawing an arbitrary line in the sand that says 40 hours is okay and beyond 40 hours is not okay. Actually it turns out to be an excellent prioritization mechanism and everyone works on the things that are super, super important because you got a fixed amount of time to work on it.
Grant: So I saw a tweet the other day: “A paradoxical thing about people who consistently choose the most high leverage activity is their efforts have a rough-edged, half-assed quality. Because polishing things to perfection is a low leverage activity.”
Mark: That’s interesting. Who said that?
Grant: Tiago Forte. I’m not sure if I pronounced his name correctly.
Mark: I like it.
Grant: I was wondering if maybe we could go back to the beginning to understand how you got here. Let’s go all the way back to childhood. What got you interested in science? Were you interested in science as a child?
Going way back, I was not a very academic kid at all. I was definitely the problem kid all the way up until about high school. It wasn’t really until eighth grade. In fact, when I really got engaged in academics at all. And that was because I managed to get placed out of my remedial mathematics by taking the pre-algebra aptitude test that Iowa gives out to everyone every year to see if they should be in the algebra class.
In eighth grade, it turned out I should have been in that class, which was really the start of like, Hey, maybe I can do math and maybe I can do this other stuff. But yeah, I grew up in Iowa, in a small town of 50,000 people, roughly half students and half full-time residents. It was a great place. You know, it was super interesting to grow up in the Midwest.
I mean, it would have been easier had the internet been there at the time I grew up, which sort of came in right at the moment I left. But it was a pretty ideal childhood. It was very peaceful. It was very easy. There was no crime, nothing to be concerned about. I could just wander around the town even from a very early age.
But I never really was academic at all. It wasn’t until high school when I really got interested in some things. I went into Northwestern University as a declared English and history double major because those were the best teachers I had by far in high school. I mean, I loved literature. I loved history. I used to love art history of all things. So when I went to Europe at one point, like I just traveled around to all the museums. And honestly it wasn’t until I got to Northwestern that I even saw anything that was sort of interesting on the science side. It was always presented in the most dry imaginable manner.
You know, you’d read these physics books and you’d be memorizing the equations as though the equations and your ability to solve those equations was the thing that was interesting scientifically. It’s only now that I’ve been reading The Feynman Lectures for fun over the last few weeks, and it’s so enjoyable to approach science from this perspective of understanding, as opposed to manipulating equations.
That was really my big journey when I was at Northwestern: the transformation from an English history double major into a computer science and math double major. And I kind of went from English and history, and I got into some cognitive science, which of course I never saw in high school.
Maybe now there’s cognitive science in high school. I actually don’t know what the curriculum looks like these days, but I’d never seen anything like this. This was amazing, this class about how people behave and how to understand the brain. And from there, I was sucked into the computer science classes and the math classes and yeah, I came out really being a tech person, but not a scientist yet. I mean, I was really into math and computer science.
But I had the good fortune to win a Marshall scholarship. You know, this was really a transformative experience for me. One, it was amazingly empowering. I mean, they give it out to 40 people in the United States every year. So really that made you feel like, okay, I have really accomplished something. Then, I am on a trajectory where I can contribute. And it sent me to England with kind of a blank check. They really didn’t care what I did at all. And I tried to be an undergrad again at Cambridge University.
I was originally a natural sciences undergrad, and that was clearly not for me. And I spent a couple of weeks just doing nothing at Cambridge, trying to figure out what I was going to do with myself. And I read an article by a professor at UCSF named Ken Dill, which substantially changed my view on what I should do. He was talking about the problem of protein folding, and the complexity of proteins and computation. This was 2001. And I was like I should do this. This is an area that seems exciting. Like, why don’t I go talk to the biochemistry professors at Cambridge. It seems to be a pretty good place. So I literally walked into my future PhD advisors office. That was a guy named Tom Blundell, who was the chair of biochemistry at Cambridge. I said, I have my own money. Do you want somebody to hang out in your lab? I don’t know anything. And he said, that’s fine. You should talk to this guy, Paul Walker. He’s a great friend of mine now. And it’s funny, we’ve been bouncing around in the same field for many, many years. And he was nice enough to let me in the lab.
And then I suddenly was a biochemistry PhD student, and I had a lot to learn and I spent a lot of time reading a lot of books about science. I mean I really had almost no idea about any of it. So it was incredibly exciting to learn. And after three and a half years of doing that, I popped out with a PhD in biochemistry and mostly focused around methods for solving crystal structures.
So I mean it makes sense. I had this technical skill. I could help with the hard technical problem of interpreting all these spots that you get from spraying crystals of proteins, which was my cup of tea. But after that I had really gotten the bio bug–to be totally frank–and I had this realization that if I wanted to be a serious life sciences researcher, I had to go and do experiments.
So I signed on to a lab to do experiments at Harvard. I joined Daniel Hartl’s lab. I joined originally with a guy named Shamil Sunyaev, but ultimately was more split between Jim Collin’s group and Dan Hartl’s group. And in Dan’s group I became an actual biochemist. Like I actually purified proteins and made Newtons to proteins and we published a great paper with–actually hardly anybody’s on this paper. There’s only four of us on it. One of the guys is still an incredibly good friend of mine. He lives literally down the road from me here in California, Nigel Delaney, and another guy, whose name is Dan Weinreich and Dan Hartl were on the paper where we made 32 mutants of beta lactamase.
So we created the plasma and we went in and made 25 mutations of all combinations of five mutations that include antibiotic resistance in bacteria. And we just asked this very simple question, like how resistant is each possible trajectory from no mutations to all five? And that was a major science paper in 2006, because it was so hard to make mutations.
It was so hard to do that experiment. It took us a year or more to just make 32 mutants and measure their MIC’s. But it was just transformative because you make data. Like you could understand it. It wasn’t just analyzing the data in somebody else’s database. I could make the data and understand what was happening.
And that was totally amazing. I mean, it was great to be in that environment and to work on things like that. But by that point I was really convinced that I didn’t want to be an academic. You know, I think academia is a great place for some people. And I have to say, I continue to be disappointed by the narrative you see in the academic community that the pinnacle of success is to become a professor like the person who’s educating you or mentoring you through the process. And like, do you really don’t want that?
I was pretty happy to leave and joined a consulting firm called LEK Consulting to learn about business. So I spent many, many months really trying to understand–suddenly–the business side of biotech. Like how much should we buy this company for? I don’t know, Hey, how do you value somebody’s pipeline?
Like, I don’t know they got one asset in phase two. Like how much money is a company worth that has this asset, but can’t even sell it. So I was super fascinated by it. It was a great place to work because of all these questions. Ultimately I spent a lot of my time thinking about those questions,
But I didn’t actually remain too long because I got pulled into the Broad Institute, which was at that point a very small place. My friend who I had actually walked into the room with on the first day of biochemistry in Cambridge was at the Broad Institute. He had moved to Boston a year before me and was basically saying, Hey, these sequencers from a company called Solexa slash Illumina have just arrived here.
And we have no idea what to do with them. We’ve got this project called A Thousand Genomes that we’re trying to start up and like barely understand what’s happening here. There’s way too much data. And it’s just chaotic and you do want to come help. And that was really, I think the start of what I would think of as my serious professional career. The Broad Institute was the first place that ever gave me an opportunity to manage a team, to think about a product that isn’t a paper.
I had grown up in the sciences. So the end result of everything was the write up, the paper. Maybe you gave a talk, but that was it. Like suddenly I was at a genomics Institute. Oh my God. I mean, I could publish papers, but there was also like, we just have to sequence organisms and like make software that you could use to do this.
And so I really built out a team there that created the GATK, which is called the genome analysis toolkit, which is now a pretty widely used piece of software in genomics. And I grew that team. It was originally very small, maybe one or two people, and they grew up. At the end when I left there were about 20, but it’s huge now. I think there’s a hundred people on that team, owning all of the analytics at The Broad Institute.
And I would say that was just a fabulous experience. You know, I got to build software that was high scale. And in general, genome sequencing covers a lot of data. So that was fascinating. You know, it was super hard, statistically: I learned everything about stats and machine learning, really not from all the theoretical stuff that I’d done before, but there’s nothing like banging your head over and over into the error modes that are on the next gen sequencer to make you really appreciate all the different ways that you can build statistical models and all sorts of machine learning things. That’s really hard. And it’s the most nightmarish environment for all of this stuff. Systematic errors and all sorts of complex structures to do them.
And that’s hard, but at the same time, you can sequence the genome of an organism, at low costs. So it’s totally worth every hour you could put into solving that problem. I spent five years thinking about that problem at the Broad. And it was great. It was so satisfying to create GATK.
But after five years there it was clear that I didn’t want to become a professor. That was sort of the only out I had. I was super senior by that point. I just didn’t want to do that. You know, I didn’t want to write grants and papers for the rest of my life. I found releasing GATK software updates, like a thousand times more satisfying than writing a Nature paper. I had to figure out what to do next. And I knew I wanted to go back to business. I loved being on the business side at LEK, but I knew that I didn’t want to be a comp consultant again.
I mean, I liked building stuff. And so I joined a guy named Stan Lapidus at his startup called SynapDx, which was a little bit outside of Boston in a place called Lexington. And he was running an amazing startup. I mean, they were looking for biomarkers of autism in the blood. And so I got to join this company and for two years ran a multi-omics trial, trying to look for any possible way of diagnosing earlier risk for autism.
And it was a great experience. I mean, I did every kind of imaginable blood base sample you could find. Unfortunately there is a signal in the blood–I mean, we know that it’s in the DNA, it’s in the small molecules–but there’s not enough to be clinically useful. So at the end of SynapDx, which was really, I have to say, fabulous. Despite the negative outcome, it was among the most important experiences I’ve had.
One of the guys who was on the board there was Andy Conrad, who was starting this Google life sciences thing inside of Google X and asked if I wanted to go there. So we packed up everything and moved out to San Francisco and were suddenly at Google X. And then that became Verily. And I ultimately moved to Google Brain.
So I was in Google brain from 2015 to 2019. So really at the heart of the AI revolution inside of Google. It was a shocking place to find myself, but it was totally an amazing experience. That’s my story. In a nutshell, like I can dive into any one of the pieces in any more detail that you’d like to hear about.
Grant: Yeah. I’m just wondering for the last hop from Google to Big Hat what prompted you to start your own company?
Mark: I think there were two major drivers there. One was, I loved being at SynapDx. It was amazing. All my colleagues were amazing, great people. And we had a small company. It was only 20 people and we did amazing things.
We had a whole giant clinical trial. Created these amazing machine learning systems in the cloud to analyze tons of data. It was just fabulous. And it really convinced me that there’s a lot of things that you could do with 20 people and you can move the needle on really important problems.
And I’ve always liked the commercial side of things. When I was young, I did all sorts of commercial stuff. Like really I loved mowing lawns and such. Actually the worst investment I ever made in my life is that I was really into Magic the Gathering cards. Way before they were popular. So I had everything, I had hundreds of these cards, they’re all worth thousands and thousands of dollars.
Given that I was like a baseball card collector. It’s kind of embarrassing that I didn’t see that, but I loved it. I mean, when I was a high school student, I would go to the local university and trade Magic the Gathering cards all weekend. I actually never really liked playing.
All I really liked to do was trade. So I would find out who had what cards and liked what cards. And I would just arbitrage card values all day long to collect more and more of the stuff. So that was really, I think, the start of my real interest in commerce, knowing what other people wanted and who had what you could really collect up some amazing stuff.
So I always had that bug and I saw how much you could do in a startup with Stan. And it was great being in Google, but Google is big. I mean, that’s just the truth of it. It was a startup, but it’s a hundred thousand person company now. And so no matter how happy I was in Google, I mean, my colleagues were amazing. We had unbelievable support. It’s still too big for me, it’s hard. I don’t like spending all day sort of talking to all the different people, trying to convince them about what we’re doing. You’ve got to like that if you are in a big company. Because you got all your peers and all the people up in the org above you, who are going to want to talk to you about what you’re doing. And that I just found too, too much overhead.
Especially in Google. You consider that at Google you’re talking to people who don’t even know about life science. You’re saying, Hey, look, life science is important. You could end up in that situation. So if you really want to stay focused on moving the needle on life sciences stuff to sort of bottom out into conversations about why life sciences even work could be a little demoralizing.
But it was a great place despite that. And ultimately I think what really convinced me that I had to leave was when I was at Google, I had the opportunity to sort of apply machine learning at arbitrary scale to any public data set I wanted. So we pulled in everything. I mean, all I did all day long when I was at Google was think about how I could hit some life sciences problem with the deep learning hammer.
And it was really like can I transform this kind of data problem? And can I bring ImageNet or Transformers or whatever kind of system we have in place to solve it? And so we did that all day, every day for years. And what was really clear toward the end was that we made the most progress when we could compare to some kind of reality or experiment that was much more sophisticated than the raw data for your models. And that’s why DeepVariant, which was the flagship effort of the team, which is interpreting genome sequencing data and in some sense did so well, because you could reference these gold standard datasets to generate tons of training data points.
But as soon as you left that, you didn’t know what truth was in almost all of biology. So it was very hard to train models. So you would make some progress, but you could just forever be frustrated by the inability to have the right data to validate your models.
And I don’t mean trivial, like let’s split the data. Like that’s not what I meant. I mean, everybody does that. That’s table stakes of being able to do splits. It’s more like, how do I know my model generalizes, if I’m going to use it in the future? And the truth of the matter is: there’s no amount of cross validation that’s going to make you convinced that the data you’re going to collect today is going to be similar to the data you already collected.
And so Big Hat was simply a product that I wanted to do something next, where the lab was integrated from the start. It wasn’t bolted on to the AI systems after the fact. And that was fundamentally never going to happen at Google. Google was in silico in the first place. From a life sciences perspective, doing experiments was incredibly difficult for many reasons, but at the end of the day, like I wanted to have experiments integrated into all this stuff. And so that observation that there was so much that could be done if you just had an integrated lab with the AI was why I left.
And I spent months and months once I left saying, okay, out of all these areas, where would you get the biggest bang for your buck if you could integrate a lab? And we want a lab that doesn’t look like everybody else’s lab. I want a high throughput lab. I want a high cycle time lab. I want my lab to be measured in like Hertz, right? Not number of data points. Like I want a totally different scale. And so that really led toward Big Hat.
And of course I met my co-founders at both SynapDx and then really at Google. And so we all sort of had a similar issue, which was: we need better experiments. Experiments designed for the needs of the AIs.
Grant: That’s fantastic. So other than antibody engineering what else do you think will be the most transformative applications of machine learning and biology in the coming few years?
I think there’s four big areas that I would expect to see major advances. One is the space that Big Hat is in which we broadly think of as molecular engineering. Most of the technologies we have for creating molecules, small molecules, big molecules, material science, all this stuff, they’re like they’re screening technologies, right?
Like I can generate diversity from some natural thing. I can go grab the fungi from the forest of the Amazon and screen it for anti-microbial properties. And that story is everywhere. Right? Like I can do that for material science. I can do that for biologic molecules. And nobody really likes that. You know, it’s just totally disempowering because we want to engineer, like we do cars, not screen for random molecules that we don’t fully understand.
And so AI tech is key. To break through that transition from screening to rational design, you need technologies that look like the things that Big Hat is heavily in, predicting the properties of molecules. So given a molecule, what do you think it’s going to do? What are all the properties? And then the other side of it is, given I made a whole bunch of molecules already, what’s the right next set to make that we think would be even better than what we’ve made before? And those are really the two AI techs that Big Hat cares a lot about.
And I think that’s going to play out everywhere. Material science, small molecules, big molecules. I think you’ll see huge advances in target selection with AI in the next few years. The argument there is simple: we’re very good at target selection today from genetics, right? So at the Broad, it was just like unbelievably sophisticated GWAS stuff happening everywhere.
And it’s usually like that’s table stakes. Like you already have unbelievably sophisticated statistical machinery that has nothing to do with machine learning to define your genetic variants associated with disease. What’s been difficult, really is integration of data, right? It’s hard to take the genetic variance and expression levels and cell state stuff and all these sort of heterogeneous data types that actually could tell you really more specifically about how to think about drug design, and really whether this is a good target, what exactly you want to do with the target.
And I think AI is big there. It’s really not clear how to do data integration without going into some sort of AI like tech. The statistical model is unclear. So I think you’ll see a huge boom because you’ll be able to integrate data. You’ll be able to leverage subtler signals to make more informed decisions about the targets.
I think the flip side of that, the other really complicated thing on the drug design side, is not the targets, but the people. So, which drugs are likely to work on which people? It’s a huge problem. I think you’ll see a lot of more sophisticated patient selection approaches arising from machine learning, where you’re saying, yeah, we can give this drug to everybody, but we think that this subset is going to be most enriched. Let’s do that trial on the subset because it’s faster, cheaper, better. We’ll have better signal and all this stuff.
Really if you could imagine this world where you could perfectly choose which patients will respond to your drug, but this just means that you don’t have to have giant trials. You can have small trials that have big effect sizes. So the pressure there to figure that out is so huge. I think you’ll see tons of AI applications. And again, I’m not that this is not a novel insight. Lots of people are out there doing that.
And finally, there’s the biomarker problem, right? I want to predict who’s going to get the most benefit from this drug. And I have two drugs, which of the two are most likely to work for you? You didn’t respond to this. What’s the next best choice? Like all of that kind of stuff. It’s going to be transformed by AI.
You’ll do imaging based and biomarker based assessments systematically to help you make decisions. And my prediction is that it’ll reach the patients last in some sense, which is disappointing, but not unreasonable. Because I think what you’ll see is that’ll come out in patients that are in trials or patients that are really already in a medical situation.
As opposed to personalized medicine, right? What you’re going to see is things like smarter cancer drug selection, and that’s already going on now. And then maybe one day 20 years from now, somebody will tell you to take Motrin instead of Aspirin because of your genetics.
But I think we’re still a ways away from that world. But if you’re in a clinical trial, I’m sure that you’ll be routed soon. Based on the AI deck.
What lessons do you wish you’d learned earlier in your career?
There are many, many things that I wish I had known. You know, I was reflecting on this when you sent me this question earlier, it’s not obvious to me that you can learn these things just by listening. Like, I can tell you what I wish I had known. And in some sense, intellectually, I probably did know it, but there’s a world of difference between reading in the book and knowing it in your bones. I think there’s a couple of takeaways. One is: technical skill isn’t everything. Most things actually look more like a threshold than a continuous scale of return.
It’s like, there are very few areas–I think maybe pure math is like this–where raw skill translates to success all the way up into whatever quintile you happen to be in. If you’re in the top 50 in the world, you could tell the difference between the top 50 and the top five. You could do that in sports, but I don’t think you could do that in software engineering.
There’s a table stakes. Like you’re just productive enough. And granted, some people are super productive and others are a little bit less, but my experience is that you can do it or you can’t. It’s very binary. Like, do you have the skills to do that? So it’s really important not to think that the world is only optimizing along these one trajectories.
It’s like most things are like, you want enough competence that you can function in that environment and now you’re good. So I think that was a surprising thing in retrospect. And many people actually fall into the trap of thinking that technical skill is everything in some sense, like academics it leads in this direction, right?
Like you’re the world expert on like one protein and all its mutants. And your technical skill is maximized in this tiny little area, but no one wants to work with you because you didn’t spend any time learning how to manage your people. Right? So that’s a good way to think about it.
So that was one observation that I’ve had. What was surprising to me too. Similar to that is: being right isn’t always so important. There’s a lot of time where you should lose the battle to win the war. And that’s not so easy to do because you have to really understand what the wars are so that you can lose the battles on purpose.
It’s like, if you have really interesting ideas that are undermining previous people’s work in your field, how do you handle that? Like, it’s very easy when you’re new to these areas to be like, everyone else is wrong. I’m right. Look here. I will rub everyone’s nose in it. That’s not ideal. The outcomes that you get, if you have consensus, as opposed to battling it out so that you seem like “I’m right and they’re wrong.” It’s better to not be so focused on making sure everyone knows you are the right one. And that’s again an easy trap.
Another thing that I thought was interesting is being smart is like a table stake. You’ve got to have at least some amount of smarts to play at a certain level in academics or business, but success in those areas are determined by other factors once you have enough.
You know, and you could see this go off some people, right? Like the smartest person, the person with the best idea, isn’t necessarily in charge, isn’t necessarily the leader in the field, isn’t necessarily the most influential. And it’s not always obvious when you’re young, at least it wasn’t obvious to me. Like why is that habit? Like, what does that mean? Like, is it not a meritocracy? It’s just a multi objective. There’s things that matter a lot for success in real endeavors that aren’t just about being smart. Just being the smartest guy in the room doesn’t mean that you should be in charge. You might actually have major deficits in other things that actually matter more for success.
And that’s a hard lesson to learn, and it’s not always obvious where you learn that. Those were sort of early things. I think there’s things that I’ve learned more recently that I think are useful. One is: I would be very careful about becoming a trophy collector. In some sense, getting a Marshall Scholarship was one of the best things that ever happened to me because I remember thinking to myself at one point, I’m not sure I’m ever going to win anything as prestigious as this again. This is it. I was like, Okay, I don’t want to collect anything else.
I can focus on something else. I could focus on my ability to attract the right people who want to work with me and like, they don’t give prizes for that. That was really an important lesson that I had the benefit of learning early, not through sophistication on my end. I just got lucky to see that up close.
But you can see a lot of people over time chasing those trophies. They start to become obsessed with: I want to be on the 30, under 30 list. I want to be on the 40, under 40. I want to make sure I’m at the hottest, coolest startup with the coolest logos.
And you can really end up in a situation where you’ve obviously chased the trophies instead of caring about the race and that’s not good. Like you can really see that people get very unhappy because they’ve pursued the awards at the expense of the fields that they’re in. In the field, where they happen to get the awards, but they don’t really care about advancing.
And I think that the flip side of that is you’ve got to know yourself, know what you want to do. It’s easy to be scared of what you want to do. It doesn’t really matter. If it’s what you want to do, if it’s what’s going to make you happy, maybe you should just do it anyway. Even if it’s not gonna lead to a trophy at the end, I’ve had many people I respect say, wow, they’ve done amazing things. Rick Moran is a good example of this. I like thinking about these examples, he had kids and he decided that he wanted them to have a life where their father was around. So he paused his acting career at the height of his fame and spent 20 years doing this.
I think a lot of people would look at that and go, man, that’s crazy. I’m like, that takes real commitment. That is a person who knows what they want and I have nothing but respect for that. And that’s what you should really aim for is the confidence to make decisions that are like that. And that’s really hard and you’ve got to not be scared.
I mean, I’m sure he was terrified to make such a decision. And that happens all the time. I mean, I’ve sat back and I’m like, I’m going to walk away from Google? As the head of this AI division? That’s unbelievable. I get paid tons of money to hang out and do whatever I want. But at the end of the day, it’s like, I’m never going to be happy here. Like I can’t be here for another five years because this isn’t what I want to do.
So even though it was painful in many, many ways and super scary, you’ve just got to try it. There’s tons of famous quotes along this, but it’s ultimately like what failure modes are you trying to avoid? The failure mode, I think you’ve really got to ultimately avoid is that you don’t do anything with your life, right? Like you had the opportunity and you didn’t do it. And it’s amazing how fast it goes.
Grant: I think that’s one of the most common regrets people have on their deathbed.
Mark: Totally. And you can see how it’s so easy. It goes super fast and you get more and more comfortable and the choices to go do things that you think are interesting and important to do get harder and harder, but you’ve got to do it. Otherwise you could end up in a terrible place. So my biggest advice for everyone is you’ve just got to do it. It’s scary, but you’ve got to do it.
Grant: The Nike slogan.
Mark: Yeah. There’s a reason for this. Like being an athlete, it’s scary. You’re going to stand up and compete in a stadium of 30,000 people, half of whom don’t even like you. They just want you to do terrible. But you’ve got to walk out on the field and participate.
That’s really hard. And I’ve often now been jealous of the athletes, who have grown up in environments that challenged them to do that. Cause I have to do that now on the business side and pitch Big Hat. I’ve got to raise money. Like if I don’t raise the money, the company, it’s going to go under. This is very stressful.
You have to know how to approach that, to know that it’s going to be okay. Sometimes you lose it. Sometimes it doesn’t work. Sometimes you’re humiliated. Like it just happens. There’s a really beautiful quote from Michael Jordan. It’s just a litany of his failures and he’s like: 70 times I’ve been passed the ball to take the game winning shot and missed.
And you’re like seventy times? That’s a lot, but it’s what’s going to happen. If you want to play in positions where you might make the game-winning shot. You’ve got to be prepared to lose. And I did mine. It’s funny. The most valuable thing that happened to me was to join SynapDx and have it totally blow up. Because it’s not so bad. It was terrible because I had to lay people off. I’m not minimizing how in the moment, this was terrible, but still it’s okay. A couple of years later I’m like that was bad. That was a learning experience.
Grant: Life goes on.
Mark: Exactly. But like the bad’s aren’t as bad as you think they’re going to be. And in many ways, that was a key thing I could tell people. Look, I’ve been in a failed startup. At Big Hat, we’re trying our best, but don’t be so afraid of the failure mode to not go after the success because it’s not that bad. If you fail, you fail. It’ll be okay.
Grant: I totally agree. This has been really fun.
Mark: Yeah it’s been a total pleasure. Thank you so much for the invitation. I’ve been increasingly impressed with what I’ve learned from these podcasts and being able to really contribute is so satisfying. So thank you for the invitation and it’s really great to be here.
Grant: Thanks Mark. Really appreciate it.