Why Drug Discovery Takes 15 Years—and How AI Cuts It to 3 | Andrey Doronichev (4/4)

Intro - 00:00:06: Welcome to the Biotech Startups Podcast by Excedr. Join us as we speak with first-time founders, serial entrepreneurs, and experienced investors about the challenges and triumphs of running a biotech startup from pre-seed to IPO with your host, Jon Chee. In our last episode, Andrey Doronichev shared how he helped scale YouTube to billions of users, why he left Google to pursue a deeper sense of purpose, and how that decision set him on a path toward biotech. If you missed it, check out part three. In part four, Andrey reflects on founding OPTIC, how his team pivoted from a failing startup into a drug discovery company powered by AI agents, and what it took to rebuild both the product and the mission. He shares what makes OPTIC's approach different, how he's thinking about platform design and scalability, and why founders need to be as obsessed with business fundamentals as they are with technical breakthroughs.

Andrey Doronichev - 00:01:14: As I said, the original market fell apart. We found ourselves in the situation where we still had all the money left. We had a team. We'd proven that we're really good at training custom neural nets and building meaningful products around them, but we didn't have a future in the market we started in. We started to look around, and the logical transition was to go into something similar. And because we were already in this business of authenticating information, we saw a huge opportunity in detecting AI-generated content. Around that time, generative AI just started appearing; all the AI-generated images started appearing. And that problem was very important for me personally because of all the YouTube and propaganda and Russia and all those things. In a way, as the head of mobile at YouTube, I had been involved in enabling the biggest platform for spreading information, including lies, including fake news, including all sorts of propaganda stuff. I felt really connected to this idea that we need to figure out how to separate reality from fiction in a future where anyone can generate anything. And so that original product that we had for moderating content on-chain very naturally progressed into a product for media to detect AI-generated content. We rebranded it into aiornot.com. And, again, it was the right place at the right time. It made a splash. We got a lot of press coverage in top-tier news because everyone at that time was asking themselves the question, "Okay, so what happens to the future of news?" And long story short: yes, we built a product that works. No, it's not a good business. It turns out everyone wants, or thinks they want, truth. Nobody is willing to pay for it. So it was really hard to monetize at the time.

But, also, personally, I realized that I'm not man enough to actually be an arbiter of truth because it is a really, really tough job. Like, when Hamas attacked Israel and there were all these hostages and victims and all the horrible pictures coming in, Twitter got flooded with real and fake imagery, and all sorts of different political sides were fighting over what's real and what's a lie. Some of them started using AIorNot to try to say that the other side was faking it. And what really broke my heart was when some far-right politicians took a screenshot from the Twitter of the Israeli prime minister, who published some photographs of victims of the original attack. And that photograph was edited, with faces or name tags being blurred out, and more than 10% of the picture was actually blurred out. And so someone put this particular screenshot into our system, and our system said, "This might have been AI-altered" because of all the blurred-out parts. And then that politician took a screenshot of AIorNot and tweeted it, saying, "Hey, look. Israel's lying. This is AI-generated imagery. There are no victims. There are no hostages." And that tweet went viral. And there was a bunch of press coverage, basically all sorts of real, real hard, messy things happening around us at that time. And that was the moment where I asked, "Am I man enough?" It's an important problem to solve. Am I a person who is willing to stomach that level of uncertainty, scrutiny, and frankly, just a bunch of dirty, unfair coverage that people generate, accusing each other of fake news? And I realized that, honestly, it would be a torture for me.

Jon Chee - 00:05:26: Yep. Sounds like it.

Andrey Doronichev - 00:05:27: So luckily enough, we had a great team and a good product that was growing. And luckily enough, I met a person who was willing to take the role of CEO in this startup, and we spun it off as a separate company. We spun it off and hired a new CEO. Not hired, but he basically took the lead as a founder and CEO. That's the right way to put it because he really committed. He got a new investment, and basically, now it's one of our spokes, essentially. A whole previous team is in this company called AIorNot. And we were left as a founding team at OPTIC with this hypothesis of bio and AI for drug discovery. And that is how we are where we are now. So it was a really interesting and thorny path.

Jon Chee - 00:06:15: I love that because sometimes people think that these entrepreneurial journeys are just up and to the right, and it's that Jensen Huang kind of thing: an overnight success thirty years in the making. Right? And it's all over the place.

Andrey Doronichev - 00:06:31: And many pivots. Right? Slack used to be a gaming company. It had an internal tool that they used. Twitter was something completely different before it became a microblogging service. You can come up with all those different examples of companies that started as something and then ended up pivoting to being something else. One thing that was true for us that didn't change from the very beginning of the company was we started with this good proprietary research where we are really good at massive information retrieval and screening of large quantities of data. So give me a dataset of a billion videos, and we can screen them to find IP violations, like Katy Perry videos in it. Or give me a billion photographs from newspapers, and we'll find those that are AI-generated real fast. And then surprisingly—that was a surprising transition—I bumped into this professor of drug discovery at the TED conference in Vancouver in 2023. And at TED, people get at the table at lunch and they start sharing what they work on from different walks of life, very different areas, but all cool, smart, interesting people. And so this professor, Professor Cherkasov of the University of British Columbia, was telling me about how he has spent his life developing cancer drugs, and I was telling our story. And then at some point, those unusual connections get formed, right? And he was like, "Wait a minute. So you're saying you can take billions of videos and, in hours, screen them to find the bad ones?" I'm like, "Yeah." He's like, "Can you take molecules instead and screen them to find the active ones?" I'm like, "I have no idea what it means, but probably we can." I can imagine as a computer scientist that a molecule is a way smaller amount of information than a video. So, yes, that's what we do. Right? We deal with large quantities of data to find something within those—those search and retrieval problems. And he was like, "Well, right now, there are these ginormous chemical spaces available, like Enamine REAL has 40 billion, and there's like a trillion-molecule synthesizable space out there. And our job as early drug discovery researchers is finding active compounds among those that would be active for certain targets. And right now, I'm running docking on these very expensive machines, and I can only screen a few million of those compounds in a week. It sounds like you can screen billions in hours. We should try to do that."

Jon Chee - 00:09:07: Yeah.

Andrey Doronichev - 00:09:08: And at that time, all of my tenth-grade chemistry knowledge came rushing back.

Jon Chee - 00:09:09: Came rushing back.

Andrey Doronichev - 00:09:11: I was like, "Okay, I will need some time to study, actually. I will need some time to understand the problem space." So it was April 2023. We were still working on AIorNot; that was our main product. And I came back from Vancouver, and I told my team about this almost as a joke. I was like, "Yeah, imagine we could cure cancer with AI. That would be fun." And my co-founders were like, "Actually, that sounds super freaking interesting."

Jon Chee - 00:09:38: Hell, yeah.

Andrey Doronichev - 00:09:39: And then my main co-founder, who is our head of AI, just started—as he's a researcher at his heart, right? So he started some research on this side and participated in a Kaggle competition for protein function prediction. Just for the kicks. Right? "Can we do something useful with AI in that life sciences field?" And he won it. He won the gold medal. And so, a couple of months later, he comes back, and he's like, "You know what? I think we can. I think I know what to do. I think we can do stuff."

Jon Chee - 00:10:04: Whoa. That's wild.

Andrey Doronichev - 00:10:05: This is wild, but we are no life scientists. We don't know anything about this. He's like, "Yeah, but this is just AI. It's information. We can be useful for life scientists. We can enable teams to do things way faster, like this Professor Cherkasov guy." And so that was when this conversation with my investors happened, when I came to this review meeting with our partner at Greylock, and I presented our current state of the business. I'm like, "Okay. So we have AIorNot. It's growing at a moderate pace. We have all those issues with being in the middle of scandals, really visible, politically very exposed, and I'm not sure I'm loving it. And we have this crazy one, which is we could use parts of our technological stack and develop it into an AI drug discovery platform." And our partner, David Thacker, he was really great. He was like, "Look, I know that the first one makes more sense if you think about it with your rational brain, but the other one makes your eyes light up, and you're clearly passionate about the problem. It sounds like that's what your team wants to do. That's what you want to do. You should do this." I'm like, "But this is crazy." He's like, "Yeah, it's crazy. You're likely to fail, but this is a thing that will more likely keep you focused and pushing hard for a longer time because you're passionate."

And at that time in my head, as we researched more, I started opening up to this idea because I realized that back in school, when I had those scary feelings about biology and chemistry—that those are so uncertain and so complex, and we're nowhere near writing an equation that would define how a cell works. Right? We can't do that, and we won't be able to do it for a long time. But now I was thinking, "Hey, with AI, we're essentially—what is an AI model?" Right? It's an approximation of an equation. You're essentially creating a model that tries to encompass a whole complicated process, like all of language, in something pretty compact—surprisingly complex, several terabytes, right? All those big language models. You compile all the knowledge of some really complex and fuzzy thing that you would never be able to describe as a set of deterministic rules or code or an equation. But somehow, if you give it enough data, enough compute, it can sort of self-assemble itself into this compact form. And so given the pace of innovation, the pace of progress in the AI field that we're seeing, it's not too crazy to imagine that even though we as humans cannot understand and comprehend all of the complexity of, let's say, even a single cell, maybe if we set up enough data collection loops and we find enough compute and we feed it all in a meaningful way, maybe we are at the brink of what happened to languages, where suddenly, a machine will be able to inherently understand those rules that we don't understand and would be able to come up with something meaningful that generates reasonable stuff out of what feels like a bunch of random data. And so I just had this feeling that maybe, once again, the wave is coming. Maybe this is the right place, right time where it doesn't feel like it yet, but it's about to start happening. That's the bet, if you look at my career. This is the bet a lot of entrepreneurs in tech are making. It's like, "Okay, maybe this Web3 is real. I need to start paddling right now to ride the wave if it happens." And it might not happen. Or "Maybe this VR thing is real right now. It feels like it's about to hit. I gotta start paddling to ride this wave." And so it was a similar thing here. I knew as we researched that companies like Insilico Medicine and Atomwise and Benevolent—all those early-stage AI drug discovery companies from the first generation in the 2010s—started, and a lot of people in the industry are very skeptical and asking, "Where are the AI drugs?" But I also knew that, say, Insilico made it through, and I admire their resilience, and they are now in the clinic, and they are actually delivering a drug, hopefully. And so it felt like the generation of technology has changed, and maybe now is the time. And so altogether, this thinking and the situation with our original product and the encouragement from the investor caused us to, at some point, go, "You know what? We're doing it." And we renamed our email domain from optic to bioptic.

Jon Chee - 00:14:48: Hell, yeah. Let's go.

Andrey Doronichev - 00:14:49: It was not very creative, but a very important milestone. We're like, "Okay, we are actually doing this. We're AI for life sciences. We're AI for bio. We're gonna build an AI-native biopharma company or at least an AI-native biotech, and we're gonna figure out what it means. We're gonna need to hire and partner with some great minds in life sciences. And right now, we're not set up for this at all, so we'll have to change the team and spin off the part that did AIorNot as a separate company and restart this company in a different way." And so in a way, I tell this story as a continuation of one company, but the truth is, that's when the current company was really born from scratch, in July 2023. And then it took some time to set it up. So really, 2024 was when we were truly set up and operational on this new framework.

Jon Chee - 00:15:46: Very cool. And again, it's kind of the twists and turns. You just never know. And exactly what you said about catching waves. You gotta just paddle and do it and see. You gotta learn by doing. Sometimes it just doesn't pan out, and that's okay. But what's more important is that you go out and paddle again.

Andrey Doronichev - 00:16:04: Yeah. I love those surfing metaphors when it comes to startups, and I think I overdo them. But for those of you who are listening who are into surfing, in some cases, it's tempting to catch the shoulder of the wave where the wave has already broken. You already see that a couple of people are riding it, and you're dropping in on them. Don't do that. You take a shoulder of the wave because it's already obvious. You know where to catch it. It's easier to do it this way. You won't get the peak. You won't get the most thrilling ride, and in the case of surfing, you're not allowed to do that because you would drop in on that person. But extending this metaphor to startups, in some cases, it's easier to jump on the market that is already obviously there. Right? Something's happening. You know it's happening. E-commerce. Okay, great. There are examples of successful companies out there. "I will start another e-store. I will run the playbook, and I will be reasonably successful here." And you're gonna make a lot of money doing this. But that idea of riding a wave that is about to form, that is not seen yet—to me personally, it's the most thrilling one, when you're jumping on something that is about to happen, that until now is science fiction. And those few courageous minds who started embracing it before even myself, those people are, in many cases, being laughed at and not taken seriously for years. And then it turns out they were right all along. And suddenly they're heroes, but for a while, they are just not taken seriously. And I feel like AI for life sciences is having this moment where suddenly, even when we started, the first conferences I went to, like the first JPM and first BIO, you would talk to a lot of industry veterans, and they were like, "Yeah, those AI companies, you guys are running around with this thing that doesn't work." And at this year's BIO, pretty much everyone I talked to, they're like, "Yeah, okay. This is happening. We need to figure out how to enable an AI strategy, but this is happening."

Jon Chee - 00:18:02: Yep.

Andrey Doronichev - 00:18:02: And so this happened within the last two, three years, when it suddenly became common knowledge that, yeah, some way, AI will be transformative to this industry. And it's just a matter of how exactly. Maybe not de novo single-shot molecule generation—if you talk to hardcore scientists, a lot of them are still very skeptical that you can do it. There are fundamental reasons for it. I still believe we need orders of magnitude more data to actually be able to do something like this reliably. We have a trillion tokens to train GPT-3. All of the publicly available academia data on small molecule-to-protein interaction is maybe 20 million data points. So it's six orders of magnitude of a difference. Someone has to generate this data first. So there are objective reasons why the magic did not happen. But while magical AI, like AlphaFold, which suddenly solves something that wasn't solvable before, is thrilling, there's also a huge field of boring AI that does things that were done before. You could already write a clinical trial protocol by hand. But suddenly, if you get an AI that does it just as well as a human and saves you tons of time and money, or helps you do due diligence around the drug, or does search and evaluation work for you to find a molecule you want to in-license—those kinds of things feel like marginal automation improvements. But if you can do those things and be way more efficient and move faster and save money, suddenly your chances of getting your breakthrough medicine into the hands of patients at the end of this long journey are increasing. And so this is how we see ourselves here. We're a very low-ego team in this sense. We're not here to transform the industry in the most profound ways. Maybe we will, and that's great. But really, the mission as it's formulated for us is more down-to-earth, and that is: we want to harness modern AI research to bring real drugs to real patients sooner and cheaper. That's it. And there are so many ways to do it. And as a team, that's what we are here to do.

Jon Chee - 00:20:18: That's amazing. And you talked about the boring AI. It all adds up. Screening is one thing. There are so many other things that need to become more efficient, upstream, downstream, all across the board. And that is critically important, and I love that the mission is kind of broad in that sense because there's a lot to it. And so can you just tell me a little bit about BiOptic's technology and platform and how you're effectuating this mission in its current state?

Andrey Doronichev - 00:20:51: Yeah, of course. As I said, the way we started was with screening technology. We had the screening for videos and for photos and for this and for that. And we were like, "Okay, can we do a similar thing for molecules and screen for active molecules?" And our original academic collaborator, Dr. Cherkasov, who came to us and enabled this whole transformation of OPTIC into BiOptic, he had specific ideas. He had this target he's been wrestling with for a while, and like many in academia, he's a bright mind and a big believer in this particular idea. And he wanted to drug—and he still wants to drug—this DNA-binding domain of one of the transcription factors, which was a hot thing, as far as I know, at some point, and then people became discouraged because you cannot really achieve selectivity there. But he believes in this, and I respect that. And so he wanted us to solve this with AI for this particular receptor that he was working on. And so we tried to adjust and fine-tune our technology to do this particular screening for this particular target. And it started showing some promising results, and I didn't know much about the industry or the way biotech startups are funded. That was all new for me. So I started pitching this idea and shopping it around. At first, it sounded like a single-asset company. I was like, "Oh, so we have this AI technology. We'll use it to drug this particular hard-to-drug or impossible-to-drug target. And the potential cancer drug that comes out of it is a multi-billion dollar asset. Great." I thought it was a great pitch. I pitched it around, and biotech investors were like, "We don't fund single-asset companies. This is just not interesting."

Jon Chee - 00:22:30: Yeah.

Andrey Doronichev - 00:22:31: I was like, "How so? We could solve this cancer that kills so many people, and that could be a multi-billion dollar drug." And, of course, that was learning by doing, but pretty fast, I was like, "Oh, so that's the economics. That's how much we have to dilute. That's how much capital we need and how many years we need, even with all this AI technology, to get through clinical trials and actually get to a functioning drug." I was like, "Okay, I guess this is not a very good pitch." So we were like, "Okay, but we still have a screening technology, and that is target-agnostic. We potentially can use it for all sorts of targets." So I was like, "Okay, well, then we are a platform company." So the first idea was a single-asset company. Not a good business. Second, "Okay, let's try to turn it into a pure software platform, like TechBio. Researchers come to us with their targets, and we do virtual screening for them." Sounds like a great thing to do. Just build a SaaS, the software, and then you estimate the total addressable market, and then you calculate how much you have to charge for the software. And then you look at the existing market leaders like Schrödinger, and you're like, "Wait a minute. Also not a good business." Schrödinger was a $200 million company when they were already the absolute leader of this market of computational software for R&D. That's a tiny valuation for a company that achieved that much success in the market. And it turns out that, yeah, your total addressable market is a few hundred people out there, and they will not pay you that much. Okay, you gotta come up with a better strategy here. So the SaaS path—we also tried to build the whole nice UI around our neural nets. We tried to grow it this way and that way, the way software grows. It turns out, okay, not a really good strategy. And that's when it hits you. You have to own part of the IP. That's how you get rich. Alright. Research collaboration deals with pharma. Let's go sign some deals and help some programs quickly screen their things and optimize them. At that time, our first model was called BiOptic B1. And that was pure screening. It's a ligand-based screening that enables you to find hits for existing targets, for known targets for which you already have an existing ligand. So, essentially, it's a patent-busting kind of use case or finding a best-in-class drug kind of use case. And, of course, to us, as noobs and wide-eyed wanderers who entered the industry, we were like, "We actually generated this molecule, and it's a completely novel molecule." We had a cutoff point of 0.4 Tanimoto similarity. So whatever we generated was completely novel, and we then tested them in wet labs, and they worked. It's like, our thing actually works. It predicts molecules that are completely novel, and they're active, and you can confirm this in the wet lab. To me, it was like gold. It's amazing. And then you talk to a bunch of people in the industry, and they're like, "Yeah, whatever. You found a hit. So what? From there to mouse, from mouse to IND, and from IND to first dose in a human, it's an endless journey. You're claiming you're gonna make the journey from here to London faster by making an elevator ride down to your lobby faster."

Jon Chee - 00:25:46: It's nothing.

Andrey Doronichev - 00:25:47: To the nine-hour flight. I'm like, "Oh, okay." So then we started doing those research collaboration deals. And, of course, it's a great thing to do because at first, it gives you exposure to different research programs out there and challenges your software across different targets. And it turns out you have better targets and not-so-good targets. In some cases, they don't work so well. But whenever—we learned from those research collaborations that the platform works really well. At that time, we started developing BiOptic B2, which was a model that is pure sequence-based and enabled us to find binders for completely novel targets. And besides, it allowed us to expand beyond small molecules into peptides across a number of academic collaborations. So we have a dozen programs running. Six of them are public on our website, and then we have a few more where we're just a fee-for-service or research collaborator for a bigger company or academic institution, which we are not publicly talking about. But basically, in all cases, we use our AI platform to speed up an existing program for someone else. And then we have a few that we funded ourselves as a next step. But with those research collaborations, we found that our stuff works. We find cool binders. Hopefully, this year, there's gonna be a number of papers that will come out about that. To be fair, we only started signing those deals exactly a year and a couple of months ago. So within a year, we discovered a number of active hits, optimized them, and some of them are first-in-class. Some of them are best-in-class—enough to get a few good papers, enough to be accepted to ASCO as one of the poster presenters this year. And so I feel like we've proven that at least the R&D automation part of our platform works. Oh, and then the BiOptic B3 model is the ADME-Tox optimization tool that we built on top of those first two. So we started layering the stack and basically automating stages of drug discovery as we go through them.

But also—so this was the core, the foundation of BiOptic: B1, B2, and B3. And those really are the models for this one tiny bit of the long, long process. So they didn't really go into biology. We didn't really do anything around target discovery, and I feel like this is the most complicated part. We did automate part of chemistry, and if you come to us with a target name, we can really, really fast go from here to "here are some really good wet-lab-validated hits." Then let's optimize them for potency, let's optimize them for selectivity, all in an AI-automated loop. And then let's optimize them for ADME-Tox and get close to getting into an animal model with this molecule in a couple of months rather than a couple of years, which I feel is cool. But as you know, this is a very tiny first step in a long journey. And so that was the most important learning for me and the team in the first year: how complicated, how hard drug development is, and how many steps there are, and how many things there are to automate.

But another big thing that I learned was that when we started on this journey into life sciences, we were thinking about the science part. Right? It's R&D. Whatever parts of the development of a molecule that there are, we're thinking what is specifically a scientific problem. And with me working in this market, trying to sell and sign deals and bring funding and spin off some of our internal programs and get them funded and talking to a lot of founders in the field and understanding all the stories of hundreds of biotechs that failed, I came down to one specific truth, which is pretty grim: this market is tough. And technical success, scientific success, getting a good molecule, getting published, getting your program to the clinic is nowhere near the same as getting a commercially successful drug out there. And that's when it hit us that on top of all the scientific complexity of chemistry and biology, which is already ridiculously complex, a biotech founder has to deal with even more complexity of funding and regulation and commercialization of a product which nobody even thinks about. You're just busy trying to go after this target you're excited about and optimize the molecule. And then you find out that that wasn't a good market to go after and that maybe you should have focused on a different indication but the same target or whatever. And then once you have your clinical trials and you've made it through, you got through IND and you got to clinical, great. That's huge. Right? But then there's a whole new set of problems, which is, "Okay, well, how do you design your endpoints, and how will they impact the way you launch your drug later on?" Because if, at clinical stage three, you design your endpoints in the wrong way, then it might be hard for physicians to prescribe your drug because they will not really see the value in the way the endpoint was defined. Like, all those complexities started hitting me, and I was like, "Oh my god. Actually, if our real goal, if our real mission is to truly help companies to get there faster and cheaper or be ourselves an AI-native biotech that gets a truly AI-native drug out there in the market, the amount of complexity we have to solve goes way beyond science." And actually, it's an intersection of super complicated fields of science and business and regulation—at least those three. And so just training really good scientific models, as we started, is not enough. And we have to go beyond that. We have to build a truly intelligent reasoning engine that can, in one, quote-unquote, head, reason through all those complexities and use all sorts of tools, including our models and third-party models, like AlphaFold and whatnot. Use our data, but also third-party data, like publicly available academic papers, but also the GlobalDatas of this world. Just like a human organization would. Right? A real biopharma or biotech startup would have a bunch of people using all those things and basically having a separate smart person for separate parts of these complexities, and then hoping that this team of different smart minds can collaborate and share information efficiently enough to make good decisions. If we really, truly want to build an AI-native version of that, we have to cover all those complexities and not just the scientific complexity of it. And that is what brought us to this latest version of our platform, which really is an AI agent. That's what we call them now. When we started, we called it an "orchestrator," basically an LLM-based system that tries to reason through whatever stage you're at right now. It could be designing the experiment that you need to run next. It could be sending emails to CROs and negotiating wet lab quotes for the next assay you need to run, or it could be making requests to GlobalData and Google Search and clinicaltrials.gov to analyze the universe of competitive drugs and trying to understand if there's a program you can in-license instead of developing from scratch if you wanna go after this indication. So this is the ultimate vision of what we're building right now. This truly intelligent system that can reason and use all sorts of scientific and business tooling underneath that basically creates an automated, self-driving biopharma.

Jon Chee - 00:33:52: That's really fascinating, and I think the part that really stood out to me is that it was built upon your just firsthand experience of, "Here's the problem set, it is way bigger than we thought," and then you just keep peeling back another layer of the onion. You're like, "Holy crap, it's even bigger, and it's even bigger." But I love that your ambition continued to match. Because sometimes you can keep peeling it back, and you're like, "This is impossible."

Andrey Doronichev - 00:34:21: Yeah, which might be true, by the way. A lot of people tell me this.

Jon Chee - 00:34:24: Yeah. I mean, everything seems impossible, and then, like you said, the switch flips, right? Exactly what you said, that for a long time, AI and science was like, "Yeah, whatever, computational people, whatever." I was on the other side. I was the wet lab person and was highly suspect of everything comp bio.

Andrey Doronichev - 00:34:44: Of course.

Jon Chee - 00:34:44: But then, eventually, I came to it. And it feels like it happens overnight, and everyone's in. But I like that you were unpacking this and unpacking this. So I think sometimes as scientists, we can think that's all that matters. As long as the science is sound. But you just forget, at the end of the day, there's a physician who needs to prescribe it, and there's gotta be a reimbursement mechanism.

Andrey Doronichev - 00:35:08: A reimbursement code, right?

Jon Chee - 00:35:09: Right. And you're just like, "Yes, the science was sound, but that thing in the future could totally just throw a wrench into this fantastic science." And I love that you're trying to tackle it holistically because I always thought it was a holistic problem that's multivariate and takes tons of different stakeholders. If we stay in these tribes of, "I'm the science person, the business people handle the business stuff," and not cross the lines and try to figure out how we address this in totality, together, holistically, I think we're gonna impede the pace at which we can gain these efficiencies. But I love that. And so, with this now agentic newest version of BiOptic, where are things at? Are these already unleashed on the world, or is this something where you're still cooking it up in the kitchen? What's the latest and greatest?

Andrey Doronichev - 00:35:58: Just to comment on what you just said about the complexities of it and how you have to think about the problem holistically. In the beginning of this podcast, we talked about the past where scientists were generalists and how Newton and da Vinci could draw and solve physics and mathematics and all the other things and probably write poetry too and paint. And then as the frontier of knowledge pushed forward, the only chance to actually get to the frontier of knowledge and then beyond it was to specialize. You have PhDs spending their lives just focusing on this one receptor, this one mutation, whatever—some really narrow space within this extremely complex field. And as a result, it's hard to make a business in life sciences because you need to find exactly the right set of very specialized minds who are super good at their particular thing, and then also enable effective collaboration and collective thinking across those fields. And sometimes it's like, okay, you can do it with scientists so they speak the same language, but you put a business person in the room who thinks in ROI and dollars, and they just don't gel. And then you put a software engineer in the room, and it's completely broken. They just can't talk. And this is where I think this age of AI gives us a huge chance to change things. Because in a way, by building reasoning machines that can take the totality of complexities across different fields, you essentially have one, quote-unquote, head in which it's trying to make a decision across all those complexities. It's very new. It's something that didn't exist until now. And the promise of this agentic AI is exactly this: that suddenly, your problem will be thought through—and I say "thought" in quotes. I don't want to equalize human intelligence, especially scientific intelligence, and machine intelligence just yet. But one thing we learned is that there are multiple ways to think, just like there are multiple ways to fly. Birds fly in a certain way, and we thought that this is the only way to fly. But then we found out that, no, actually a machine can fly less efficiently energy-wise, but it can also do it, and it's useful. Similar with thinking. Yes, humans think a certain way, and machines don't think this way, and yes, they're in many ways dumb compared to humans right now. However, I think at this point, the evidence is there that machines learn to think in a different way from organic brains. But over time, they're getting better at this. And maybe they will never be able to match us, but maybe in some other ways, they will be better than we are. And we're already seeing signs of that. So this is what I mean. I don't want to be criticized as one of those AI la-la-land people.

Jon Chee - 00:38:42: Yep. Totally.

Andrey Doronichev - 00:38:44: But I do believe that this holistic thinking of a problem that you mentioned is something that was extremely hard for objective reasons, that everyone specialized, and you had to enable this cross-person collaboration. And now, suddenly, there's an opening which enables that, that suddenly makes it possible through those automated reasoning agents. And that's where we are right now, to answer your question. We're entering this era where we went from generalist to specialist. As a caveman, you were a real generalist. You could hunt, you could sow, you could do all sorts of things to survive. Probably none of us would survive that. We are not generalists enough. Then there was a time when specialists really mattered. And I think now AI, in a way, is leveling the field and bringing back a new spin on generalists. Creative generalists that can empower their knowledge with an AI platform that suddenly turns them into a one-person army. And so this is our current thinking. It's not that it's a completely self-driving pharma that doesn't require humans anymore. No. But I think we are trying to envision an organization that instead of a team of 20 scientists highly specialized in each field, we need one really, really strong scientific mind and a very strong AI platform underneath that, in collaboration, can do what in the past a 20-person team used to do. And that's our current vision.

So that was a long beginning of the answer to your question. What does it look like right now? So the way it looks right now is that we are a very small and scrappy team, given that we are running a dozen programs in parallel that does a lot of things through automation. And then we have a team of, right now, four life scientists who essentially are each the orchestrator of a team of agents. And those agents right now, I will say, are not extremely smart. They still make stupid mistakes. And I think Andrej Karpathy, the famous researcher in the AI field, said it's currently "artificial jagged intelligence." It's good at certain things. It's horrible at other things. It's embarrassing. And that's what it is right now. But if you look at our internal process right now, you really see a scientist collaborating with a team of agents that run searches or read literature or email CROs and do a lot of things that historically were done by humans at a glacial pace. In our organization, they're happening super, super fast right now for that reason. We can find new binders for a novel target overnight, and our AI agent will negotiate assay prices with a number of CROs and will get to the point where we can send money and actually run experiments on the molecules we've just generated without owning any lab equipment. And, you know, yes, that's been possible in the past too. It's just that it wasn't efficient. You weren't fast enough with this. But through those, as I call them, boring AI automations, we're suddenly making things almost as fast and as efficient as if we had everything in-house with one lab facility within our company.

But another thing that we learned and that we are now piloting is that, as I said, we are optimizing for the final result, and the final result is getting drugs there faster. We are not coming from a particular research field. I have no scientific ego that I really want to drug this particular target or that I really care about oncology specifically. We're target-agnostic, we're indication-agnostic, we're modality-agnostic in a way. We're better at small molecules, but we're expanding into more modalities. But what I said about understanding the complexity of the market is that one thing I learned is that while we're trying to generate novel molecules and many people around us try to do the same thing, there actually are tons of great molecules out there. You know? And they're just shelved, or they're just starving for funding, or they're about to be cut because a new CSO came in and just cut a bunch of programs. And I realized how much waste there is because each of those molecules is years and sometimes decades of work of incredible bright minds, millions and tens of millions of dollars spent on experiments. And then in many cases, they are worth billions if you could actually get them into the market, and they just get shelved. And so one thing that we're really excited about right now, and I got inspired by the story of Karuna from PureTech. And I spent some time with the co-founder of PureTech who shared a lot of wisdom on the business side of the industry and business strategy with me. And thank you very much for that, Eric. And so the Karuna story from PureTech was incredible to me. It's like, okay, shelved research, failed drug, really smart and first-principle thinking by PureTech's team who were like, "Okay, we can in-license this thing even though it failed, and we can solve the toxicity effects through this other research that we understood." So combining the scientific knowledge and the business reality that this drug is very cheap right now, we can combine those two. We can fund clinical trials and boom, we can sell it for, what, $14 billion or something. So this story of "one man's waste is another man's treasure" in this market right now fascinates me. So a big thing that our agents are doing right now, where we're spending a lot of compute at the moment, is actually running this search and evaluation agent that, on one hand, reads a lot of scientific literature trying to find gaps. New research suggests that you can solve this problem with that other molecule that we find by sifting through shelved molecules or existing research, and then combining those into new scientific ideas that could be fundable. And the idea is that it's essentially like you can think of it as an AI-native Roivant. It's this idea of not starting from scratch, from de novo, but finding existing clinical-stage assets out there and finding hidden gems among them. And because this is our bread and butter—screening massive amounts of data—as you see from my career, it goes as a single theme from back in the day to the information access work, and then Google and then YouTube. I kept doing the same thing. I'm fascinated by vast amounts of information out there and just trying to find a useful way to organize it and turn it into value. So in a way, what we're doing right now with our search and evaluation agent is exactly that. And I feel like there are tons of value to be unlocked. And more importantly, it's a way to actually, truly get something into the market faster that otherwise would have taken a decade of research. Why wouldn't we save something that's out there? Or why wouldn't we find something in international markets, which are currently harder to navigate? Suddenly, for our search and evaluation agent, it's doable. It reads Chinese just as well. So now that it's working and it's constantly searching for things, we're getting a lot of interest from a lot of companies who will reach out to us and go, "Hey, can we use your agentic platform for our search and evaluation team, or can we use it for our investors?" And we're talking about those deals right now as we speak. We've signed a couple already.

Jon Chee - 00:46:24: Whoa. That's really cool. I was gonna say, have you read the book For Blood and Money?

Andrey Doronichev - 00:46:30: No, I actually didn't.

Jon Chee - 00:46:33: No worries. But the story of Keytruda.

Andrey Doronichev - 00:46:35: Oh, yeah. I know the story. I didn't read the book, though.

Jon Chee - 00:46:38: No worries. That wasn't to put you on the spot, but there was a moment of it being shelved and not seeing the light of day, and then we all know Keytruda.

Andrey Doronichev - 00:46:46: So I agree, of course.

Jon Chee - 00:46:48: Right. There's so much work that has been done, and it gets shelved for whatever reason. And I think we have a very inefficient machine. A lot of it is brute force, and things don't get done for whatever reason. And I love that this is asking, "How can we just make it all more efficient and make it global in nature?" So that's really cool. And from a business model perspective, you mentioned search and eval. People are trying to use your tool for their own search and eval, and it sounds like you're doing internal work for your own pipeline. So you're doing a little bit of both. You're like, "Here's our tool. It's available if you wanna use it," but you guys are also using it for yourselves. As they say in software, you're eating your own dog food.

Andrey Doronichev - 00:47:30: Yeah. So, look, this is the nature of bio platforms, I feel like, because the economic reality is that, let's say our search and evaluation agent is the best investor in the world that can find the best assets, find all the hidden gems in the world, and just put them on a platter in front of us. What do we do with all of them? Can we actually fund all these clinical trials? Can we actually get them all into—of course not. Right? It's so extremely capital-intensive. It's so still risky. No matter how good your AI is, there are so many reasons why it might fail and never make it to be a drug. So, given all that, can we actually just do an internal pipeline and be happy with that? No. We'll only use a tiny sliver of it. So if we actually want to be efficient—and AI lives and dies by data and feedback loops—we can only provide feedback loops and data on this small number of programs we can actually stomach ourselves, and we're a tiny company right now. We're an early-stage startup. So it's only reasonable for us to actually have partners who can increase the usage of this platform that we have. And there are so many unmet medical needs. There are so many targets out there. There are so many ideas out there that there's plenty for everyone. That is where, for now, our first step is really to focus on expanding the platform through partnerships and enabling this feedback loop. Think of it as a Genentech AI; it's almost like a digital employee. Instead of hiring another search and evaluation person for your team, you're getting an army of 100 interns that work 24/7 that might not be as intelligent as your super senior people, but they have all the basic understanding of science. They can read in every language. They never sleep, and they do this work for you night and day. And then your actual BD team or your search and evaluation specialist could review and accept or deny their work. But the best part is that when they reject their work and provide some reasoning, your AI interns, quote-unquote, become smarter, and they understand your process better and better and better. And over time, they become real smart. And so for us, that creates a network effect where the whole platform becomes smarter. It doesn't necessarily become smart in a particular methodology because everyone has their own thing, their own strategy for search and evaluation. That is proprietary. That never gets shared with the platform. But the overall thinking on how you spot meaningful research from research that doesn't have enough good data—that's common knowledge, and that's what actually feeds back into the system. And that's how we already run all sorts of benchmarks where we compare ourselves to some general horizontal systems, like OpenAI's Deep Research or Perplexity, which are designed for a broad set of tasks from finding a restaurant table to finding a better dress for your prom to finding a molecule. It's a very broad set of tasks, of course, and you cannot succeed in all of them. And we run evals where we compare our very specialized, verticalized search and discovery agent or search and evaluation agent with those horizontal ones, and we are way ahead—in some cases, double-digit percentages on some benchmarks. But that is because we keep feeding it with feedback. And, yes, we have a team of four PhDs in-house whose job is to essentially train those models to become smarter. But also, by running pilots and collaborating with teams out there, we're also making it smarter for everyone. So that is stage one of our strategy. At some point, if we are super passionate about a certain therapeutic area and we actually have enough confidence to commit and start funding our own clinical-stage program by in-licensing a drug and so forth, cool. We will also eat our own dog food in that sense. For now, we do have one program that we use as a testbed for our agentic platform. We actually discovered some really cool research. It's preclinical because we need something for which we have enough capital to in-license and actually push forward.

Jon Chee - 00:51:42: Yeah.

Andrey Doronichev - 00:51:43: If we do clinical right now, we just don't have enough capital for it, and I don't think we can raise that much money that quickly. But for this program, our agentic system discovered really interesting, super-niche research, and we now have a team of scientists who are super excited about this. We're in-licensing it, and we're currently raising for it. It's in gut-selective pain management for IBS and IBD, a macrocyclic peptide, a really interesting program. And I feel like it's a cool niche where not only does it verify that our agentic stuff works, but also, we might get a decent drug program out of it. So, literally, we're doing it right now.

Jon Chee - 00:52:23: That's so cool. And I love how this all came back from—again, we talked about how you asked me, "Are my experiences in a non-life-science role interesting or valuable?" This is where it all comes together and is interesting and valuable. Right? It's kind of this thing where you're like, "All these learnings that we've had on the YouTube video side, and then, boom, let's apply this tool here." And then you're starting to see it work. It's awesome. Obviously, it's gonna take time. It'll take—you need to keep feeding the machine, and you have to keep going. But absolutely, this is awesome. I'm fired up.

Andrey Doronichev - 00:53:04: I wonder how many people made it through all the beginnings of it to get to this part, which is the actual life sciences part.

Jon Chee - 00:53:11: I know. Believe me, people do.

Andrey Doronichev - 00:53:14: If you're one of those people, thank you very much for hanging in there.

Jon Chee - 00:53:18: Yeah. That's super cool. Because there's an aspect of this where it's just "one man's trash is another man's treasure." I love that aspect. It's more efficient. It's more sustainable. Let's not just let all this resource—time, money, blood, sweat, and tears—languish. Let's see if we can make something of this versus trying to—you know, there's always value to doing something new and de novo, but there's a lot of work that's already been done. Let's see what we can do with this.

Andrey Doronichev - 00:53:43: And I think this is where it's important to stay low-ego. Right? Because being mission-driven in this sense helps a lot. At the end, me and my co-founders, we always go back to the mission that didn't change since we started BiOptic. And we're like, "Okay, harnessing AI to bring real drugs to real patients faster and cheaper." Are we actually doing it by continuing to invest in our chemistry models? Yes, to a certain extent, but what we learned is that this is not the most impactful part. We save some time and money, but it's peanuts compared to what it costs to run the program. Okay, are there other things that we can do to speed up, let's say, clinical trials? Okay, short of simulating a cell and then a tissue and then the whole human patient body—which for now sounds like complete science fiction, very, very far out—short of that, maybe we can speed up clinical trial protocols or the consent process. There are some things we could do. But the truth is, it will still take a lot of time. So, okay, what else can we do to actually get the drug faster to the patient? Oh, hey, we could find something that otherwise would not have launched, which is halfway there, and we can actually recover this, and in two years, it might become a drug. Well, that sounds like a reasonable strategy. So it all boils down to the first principles in our mission. And every time we assess those ideas, it's like, how should we address this? And being low-ego and being able to say, "Okay, our previous hypothesis was cool, but we have to move on and use a different approach." I feel like this is a very important aspect of it.

Jon Chee - 00:55:19: Very cool. I mean, I respect and just commend you on this. Pivots are hard. And you continuing to pivot and then finding—it seems like at least there are early signs that this thing is catching. You're paddling now, and you're like, "Oh, this might be something here." So that's really awesome to see that you made the pivot, and you're now getting back and paddling again. And, honestly, I'm fired up. I've learned a lot, and thank you for sharing. The early tectonic shifts underneath you could not have been easy and were very stressful. So thank you for sharing that with all of us. And I love what you guys are doing now and weathering the storm and continuing to push through. And, I guess, as we're starting to round this out, in one to two years, what's in store for you and OPTIC?

Andrey Doronichev - 00:56:10: So in one year from now, we hope to have a number of good logos on our website using our platform for search and evaluation. We'll continue investing in some of our programs, like early-stage programs. And I think it's important. We're trying to really, really be the holistic company that does everything from low-level, early-stage, building a model from scratch, running experiments from scratch, all the way to reasoning through a commercialization strategy using AI. So I really, truly believe that someone has to build this holistic, AI-native pharma that does all those parts. You probably can build a business solving one of those parts and turning it into a fee-for-service. But if you really want to build AI-native pharma, you have to build all of those things together. So we continue investing across the board, but the main goal for the next year is to grow this agentic part because I feel like it is the thing that connects all the dots together—business and science and regulation—and truly pushes things forward faster. So that's our main focus right now, agenticpharma.ai. And we're seeing a lot of really, really strong positive response right now, so we have a number of pilots running in parallel. So I'm excited about that. And then we're gonna continue publishing. There are a number of papers that are gonna come out, some with our research collaborations, some with real big labs of great universities, and then some from our own internal research team, both on the life science side where we discovered some molecule with our AI platform, and there's a method and whatnot. But also just pure AI research. For example, for this agentic AI, which is at the forefront of the industry right now, we developed some really nice benchmarks and some approaches that show how a verticalized agentic search and discovery platform, like the one we built for molecules, is better than a way bigger model that's less focused, and how they compete and in which cases one is better than another. So those kinds of papers are gonna come out this year, and I'm also excited to just see a response from the community. Because after all, as we talked earlier, we're all poking a hole at the frontier of knowledge. And while we're all competing in a way, we're also all collaborating in a way. And we're trying to be a good citizen in the community, both of AI research and also life science research. And so we're gonna continue publishing. I'm excited about the future.

Jon Chee - 00:58:33: I'm excited too. That's freaking awesome. I commend you. Again, I'm rooting for you guys, and I think it's a noble cause. So I'm team BiOptic for sure.

Andrey Doronichev - 00:58:43: Thank you.

Jon Chee - 00:58:44: In traditional closing fashion, I have two questions for you. So first question, would you like to give any shout-outs to anyone who supported you along the way?

Andrey Doronichev - 00:58:52: Yes, absolutely. We are so blessed to have incredible people in our corner, specifically Dr. Artur Cherkasov, who enabled this whole transformation and enabled us to start this crazy journey into drug discovery. Thank you very much, Art. Dr. Anna Kostikova, who's been an incredible mentor and helped us so much along the way. And then I'll just give a blanket shout-out to a little closed-off biotech community called Biotech Barbecue here in San Francisco. It's a little chat group with a bunch of people from business and science that all have been super nice to us and have been helping us along the way. So thank you guys so much. And, of course, our investors. I could not have hoped for more supportive investors who supported us through all those pivots and different crazy changes and still continue reading my investor updates that include so many words that many of them don't understand now. And they're still reading them and responding to them. So thank you so much. Finally, my team. Guys, you're awesome. Thank you for being with me through all of this. I truly believe that we will succeed. So thank you all.

Jon Chee - 01:00:00: That's amazing. I can just imagine the transformation of your investor letters, but I love that they're still engaged. That's so good. You're like, "It was a completely different business, but we're still in there." It's like a little crash course in life science.

Andrey Doronichev - 01:00:16: Yeah.

Jon Chee - 01:00:16: Last question is, if you could give any advice to your 21-year-old self, what would it be?

Andrey Doronichev - 01:00:21: Focus on health earlier. And that's broad advice, but my journey to life sciences also came from a personal internal transformation of values. When I moved to Silicon Valley, the motto out there was "move fast, break things." And so I did: content platforms, big bucks, crazy hours, all that. And by now, by the age of 40, I was like, "Oh, actually, there's a different mantra, which sounds like, 'First, do no harm,'" and that's what the health industry lives by. And, actually, "first, do no harm" starting with yourself and not working yourself to death and actually taking care of your personal health is a great mantra to live by early on. And so this advice would go two ways, personally in terms of business, but also in terms of personal health. Health is important in all sorts of ways. And the earlier you start focusing on it, the better it is.

Jon Chee - 01:01:23: Man, honestly, I need to hear that because I think I was burning the candle at both ends for a lot of my early career. And like you said, I'm feeling the repercussions of it. But you know, when you're 21, you're like, "Yeah, I'll be fine." I'm like, sleep three hours and just figure it out. But it catches up to you is the one thing I will say. So absolutely take care of yourself. But Andrey, thank you so much. Thank you for sharing your story. Thank you for telling the whole journey. I've learned a ton. I'm feeling inspired and fired up, and I'm excited to take this back to Excedr and fire up the whole team. So thank you again. And, you know, we're both in San Francisco, so let's grab some coffee sometime.

Andrey Doronichev - 01:02:07: Yeah. Let's hang out. Thank you, Jon. It's been a real pleasure.

Jon Chee - 01:02:10: Yeah, absolutely. I'll talk to you again soon.

Andrey Doronichev - 01:02:12: Cheers.

Intro & Outro - 01:02:14: That's all for part four of our conversation with Andrey Doronichev. From rebuilding his company to launching OPTIC, Andrey's story shows what can happen when bold pivots meet long-term conviction. If you're enjoying the show, be sure to subscribe, leave a review, or share it with a friend. Join us for our next series featuring Sandra Shpilberg, Co-founder and COO at Adnexi, an AI-powered platform that profiles and evaluates patient advocacy groups by therapeutic area to enable biopharma to be patient-centric from clinical trials to treatment launch. Sandra is an accomplished life science executive and serial entrepreneur. Prior to Adnexi, she founded and led multiple ventures, including Sanibel Health, a special purpose acquisition company focused on accelerating healthcare innovation, and Censa Health, a digital patient-finding platform that improved clinical trial recruitment. Earlier in her career, Sandra held executive roles at leading biopharmaceutical companies, including Nora Therapeutics and BioMarin Pharmaceutical, where she led the commercial launches of multiple drug therapies. With deep experience across biotech, digital health, and patient engagement, Sandra brings a rare founder-operator perspective on building innovation that actually reaches patients, making this a series you won't want to miss. The Biotech Startups Podcast is produced by Excedr. Don't want to miss an episode? Search for the Biotech Startups Podcast wherever you get your podcasts and click subscribe. Excedr provides research labs with equipment leases on founder-friendly terms to support paths to exceptional outcomes. To learn more, visit our website, www.excedr.com. On behalf of the team here at Excedr, thanks for listening. The Biotech Startups Podcast provides general insights into the life science sector through the experiences of its guests. The use of information on this podcast or materials linked from the podcast is at the user's own risk. The views expressed by the participants are their own and are not the views of Excedr or its sponsors. No reference to any product, service, or company in the podcast is an endorsement by Excedr or its guests.