EP 155: AI's Edge in Pharma - Lowering Drug Failure Rates

This is the Everyday AI show, the everyday podcast where we simplify AI and bring its power to your fingertips. Listen daily for practical advice to boost your career, business and everyday life.

Why do drugs fail at such a high rate? And what can artificial intelligence do about it? We're going to be talking to an expert who's helping lead the field in helping to answer that question. So welcome. This is Everyday AI. My name is Jordan Wilson. I'm your host. And Everyday AI is for you. It's for us all. It's so we can all learn and leverage what's going on in the world of generative AI, so we can use that to grow our companies and to grow our careers. If you're new here, thank you. Thanks for joining. If you're live, as always, get your questions in. What do you want to know about how AI can help lower drug failure rates in the field and hopefully make safer and better medicines for us all to use? If you're joining from the podcast, thanks for listening. As always, make sure to check your show notes. There's always more information in there, you know, other related episodes, all of that. So before we answer that question and we start to dig into how is AI helping to reduce drug failure rate, let's first go over what's happening in the world of AI news. So Amazon's Reinvent 2023 conference is still ongoing. Some new announcements aside from, you know, earlier in the week with their new Amazon Q. So Amazon did just announce yesterday their Titan AI Image generator. So right now it is available in preview in Amazon's Bedrock Council for AWS customers. And we'll see, we'll see how this works. We'll see if it can hang with Mid Journey Dolly, Stable diffusion and some of those others. So some, some exciting news there from Amazon. Second and kind of relevant to today's discussion. So DeepMind just used AI to discover millions of new chemical materials. So Google's AI division, DeepMind used machine learning to discover 2.2 million new crystals, promising to revolutionize industries such as computer ships, solar panel and just new discovery of materials. Right. So this new tool that they're using called gnome, I believe that's how it's pronounced, bypassed centuries of experimentation and has the ability to significantly improve the efficiency of materials discovery. So researchers there at DeepMind said that this new discovery equated to 800 years worth of knowledge. My gosh, what a great day for DeepMind. I don't know if I accomplished 800 years worth of knowledge yesterday. It was just, just kind of a Tuesday, a Wednesday for me. Last but not least, it is the one year anniversary of Chat GPT. So and some some new news about their board. So yes, it's been a full year finally or already. I don't know what it seems like to you but Chat GPT has been around and we've seen a lot of everything from going from the free model which was not very good to the introduction to the paid chat gbt plus with gpt4 and plugins and the recent introduction of custom GPTs and all the Sam Altman firing drama. But on that and some new news in this. So Sam Altman announced last night on the OpenAI website that Microsoft will now have a seat on the Open AI board as there's new board members but they will have a non voting seat. So let me know in the comments, what do you see happening in the next year for year two of Chat GPT? Drop a comment. Maybe we'll feature it in the newsletter. So all right with that let's, let's talk. Hey and if, if you do tune in for the news because I've been told a lot of people do, there's always more. So just go to your everyday AI.com Sign up for the free daily newsletter that comes out, you know, every, every day around 11:00am Central Standard Time. So but I want to talk about drug failure rates. Why are they so low and what is AI doing about all of this? So it's, it's not just me today. Don't worry guys, we have an actual expert in, in the AI field. So please help me welcome to the show Chris Gibson, the co founder and CEO of Recursion. Chris, thank you so much for joining the show.

Thanks for having me.

Yeah, absolutely. Tell, can, can you just tell our audience a little bit about what recursion is and what work that you do?

Absolutely. So Recursion is bringing together the worlds of biology and technology to try and discover new medicines faster and to bring them to patients at lower cost in the coming decades.

So why do drugs fail? I'm curious, like why does that happen? Right. So you know, all these, you know, great, great minds are creating these, these drugs but it's, it's high failure rates, right?

It's extraordinarily high. So the biopharma industry is a multi trillion dollar industry where 90% of drugs that go into clinical trials fail before they ever make it to patients. And it's not because these people don't know what they're doing. There are extraordinary scientists who have been working for decades Some of these scientists will work their entire career without ever having a medicine make it all the way to market. And the reason is because biology and chemistry are so complex. So you've got a trillion cells in your body. Every one of those cells has about 400,000 proteins encoded by about 20,000 genes. And there are trillions of interactions happening every second. The fact that we can find any medicine that works is kind of amazing. But ultimately, in the face of this incredibly complex system, the tools that we've had at our disposal for the last 40 years in the industry have just been too reductionist. We can't study that system in its full complexity. We study it in little tiny pieces. And ultimately, that's why we believe that AI has such a prominent role to play in the future, because AI is so great at taking these extraordinarily large data sets that companies like Recursion are creating and turning them into models that help us understand how different genes and proteins and molecules might be interacting. And ultimately, we believe, and I think we're starting to show that we have the potential to bring down those failure rates. And what I tell the team, if you could go from 90% failure rates to 80% failure rates, where eight of your drugs out of 10 fail in clinical trials, you could reduce drug prices by half.

Wow. And Chris, like, when I hear that 90%, you know, as someone that's, you know, not familiar with the pharmaceutical industry and in the science and the artificial intelligence that goes on behind the scenes, like, I'm shocked by that. Right? Like, 90% seems jaw dropping. But then you just said, hey, even getting it down to 80%, that what that means for the. Or maybe talk about, like, what that means for the world. So lower, lower drug prices. But what else would would just reducing that failure rate by 10% mean for. For the rest of us?

Well, in addition to lowering prices, it means that there would be thousands of diseases that today have no treatment that maybe we could actually make a difference in, in combating. I mean, how many people on your show have lost a relative to cancer? Almost everybody. Or relative to heart disease. And we're seeing extraordinary progress there. I mean, people have probably heard about things like Manjaro and wegovy, from Novo and Lilly, these medicines that are going after these targets that the world really wasn't working on until just a few years ago that seemed to be creating all kinds of interesting shifts, not only in our industry of biopharma, but also, you know, I don't know if you saw Weight Watchers talking about how they think the world of dieting is over thanks to these glip1 agonists. And so we're seeing, we're seeing the impact of one or two really compelling medicines across so many different fields. Imagine if we had really compelling medicines for every disease. And I think that is the future. And the question is, how fast can we get there?

So I'm curious, how have things changed recently? Right? Because I mean, most people know this if they're in a related field, but I'd say the everyday person maybe doesn't know this. Right? Like artificial intelligence has been used in the medical fields for decades. So how specifically have have some of the newer, you know, advancements in artificial intelligence, maybe you know, large language models or, you know, kind of this resurgence or surgeons of generative AI. How is that changing things specifically when it comes to reducing drug failure rates?

So one of the challenges in our industry is getting the right training data. If I'm building a large language model like ChatGPT, I can go train that model across all, all the language on the Internet. And in our world, I liken it a bit to something like autonomous transportation or self driving vehicles. If you want to get a large training set to train a vehicle to drive itself, you need to put people out on roads and have them driving around and recording all of that data. For us in our field, we need to do experiments. And so in the laboratory over my shoulder here, we actually have robots doing millions of experiments every week in real human cells and extracting thousands of different measurements from these cells as we break different genes and add different drugs. And we've generated over 25 petabytes. This is like Netflix scale data sets, 25 petabytes of data about like what happens if you break this gene or this gene or this gene, or add this potential medicine or this potential medicine. And it's so much data that no human could ever even scratch the surface of looking at all of it. And so we use AI tools to look across all of that complex data to find relationships and patterns that we can drive forward. And then we use generative AI tools and other companies do as well, to help design molecules. If we find a new target, design a molecule that can fit in and bind to one of these proteins in a way that we think could be really compelling. What's interesting though is this, as you said, this has been happening for decades in our field. Really the last half decade is where it's made the most impact. But our industry has been pretty slow to adopt a lot of these technologies until just the last year, ChatGPT, as strange as it sounds, seems to be the thing that finally got the biopharma industry excited about AI.

Yeah, that is funny, right? Because it's something, ChatGPT is something in theory, it is so simple that anyone can pick up and use it. But even given the amount of, I'm sure, you know, models and, and deep learning that, that you all are doing, it's something like, or that the industry is doing, it's something as simple as, as CHAT GPT that opened up kind of big pharma's eyes to how AI can help. Why do you, why do you think that is? Like why, why do you think it took something as, you know, quote unquote, simple as Chat GPT to kind of shift the attitude change towards, toward artificial intelligence?

Well, you know, I think our industry is tricky because we do experiments in people and there's an extraordinary ethical and moral obligation if you're doing an experiment in a person. There's a lot of regulation, the fda, the EMA in Europe, et cetera. Like this is a pretty conservative industry for those reasons. And on top of that, we fail 90% of the time in the clinic. And so I think people are a little bit resistant to trying new things because the consequences are so high if those things go off the rails. What ChatGPT did was tap into language. And language is so fundamental to everything all of us do that I think ultimately that's what, you know, it's probably some pharma exec is sitting there and their 10 year old came home and said, hey, check out this poem that I wrote in the form of Dr. Seuss using ChatGPT. And it was that kind of, that kind of just fundamental application of language that I think maybe finally got their imagination going. And to be fair, some of the large pharmas were doing a lot of work in this field before. It's just that it's finally moved like the middle of the distribution into being excited about AI.

Yeah. And maybe, maybe Chris, help, help explain so we can, so we can all better understand the process from start to finish. So you know, you're using artificial intelligence to help create, you know, safer, better drugs that we can all use in the long run. But how does that work from, from after, you know, you kind of said, hey, behind me, this is where all the, the testing and is going on. But what happens after there? Are you creating then your own drugs? Are you partnering with other big pharmaceutical companies to, you know, for them to kind of use, you know, your, your discoveries how does it work, you know, from the testing process forward?

Yeah, we're doing all of that and more. So we actually have five drugs that are in human clinical trials now. So we are literally enrolling patients and testing these medicines with the FDA in the clinic right now. That's our own internal pipeline. We then have partnered with some of the large pharma companies. So we're going after areas of oncology, cancer with Bayer, and then we're going after neuroscience with Roche and Genentech, one of the really preeminent players in the space, and those are diseases like Alzheimer's and als. They're so complicated and large that we really needed to partner with these big companies to have some of the elements that we would need to go into clinical trials. Most of our internal programs are against pretty rare, small diseases. And then we also recently partnered with Nvidia to actually make available some of the tools we're building to the rest of the industry and even to academic researchers. And we'll share more about which tools and how we're going to. How we're going to distribute those tools in the future. But some of the tools we've built for our own internal scientists will be available on the bionemo marketplace with Nvidia in the coming months.

You know, we had a guest on the show a couple of months ago that talked about even using AI in the clinical trial process, because, correct me if I'm wrong, but even once you get it there so right, even if you make, you know, safer and better drugs and get that failure rate down from 90 to 80%, it can still take a very long time, right, to get this, you know, new and improved drug to market. You know, sometimes years. Is that like, how does even that process work? I know maybe you don't have as much hand in that process, but maybe help expand, help explain that for the rest of us.

Well, we're starting to deploy AI in those kinds of ways as well. It's amazing. If you take the amount the industry spends and divide it by the number of new drugs approved, every year, the industry spends about $2.5 billion of R& D for each drug that gets approved. And it takes, on average, about 12 to 15 years from starting a project to getting a drug to market. And remember, when most of that work, most of that cost is in the failures of the drugs that never make it to market. And so we think AI is likely to be deployed or other technology tools. Maybe it doesn't have to always be AI, but technology tools are likely to play a really prominent role in many of the hundreds of steps it takes, including how you run a clinical trial. How do you identify patients? We actually partnered with a company called Tempus, recently started by Eric Lefkowski, who was the founder of Groupon. He started this company to gather cancer patient data after his wife went through a really traumatic oncology experience. And he's built now one of the largest data sets in the world of cancer patient data so that we can find cancer patients to enroll our trials faster. I don't think there's a silver bullet, one software program that's going to all of a sudden unleash the next magic medicine. It's going to be hundreds of tools built in a full stack from discovering new fundamental truths about biology all the way to even how we market and distribute medicines. And the companies that can put together that full stack and create compounding efficiencies. One tool, two tools, three tools that make a little bit of a difference at every step. Ultimately, I think that's how we're going to see AI tools and other technology tools fundamentally shift this industry so that in 10 or 20 years, anyone who needs a medicine has it. It's inexpensive, it doesn't have a lot of side effects, and maybe even one day our industry can shift away from treating disease to preventing it, which ultimately would be way better. Right, if we could actually shift one day to just preventing diseases in the first place.

Yeah, that would be great. And you know, I, I do have to mention, you know, you brought up Tempest. Hey, I just, I just always have to shout out like awesome people doing work in Chicago, because I think Tempest is like three blocks, you know, from where I am now. I've had, you know, family and friends work there. But actually a great question that I wanted to bring up here from Douglas. So, Douglas, thanks for the question. So asking out of curiosity, are you using public large language models like Chat, GPT or Bard, because you kind of. Chris talked a little bit about the different ways that you're using AI. So, you know, presumably there's a lot going on, kind of more proprietary, kind of the things that are happening in the lab. But then what about from there? Yeah. Are you using, you know, some of the same systems that the rest of us are using as well?

Absolutely. So great question, Douglas. So we built a lot of our own AI tools for a lot of these very niche problems, but we've generated so much data that we face this problem of not knowing which disease or which to drug program to start next because we found 5 trillion relationships from the data sets we've generated. So we actually use GPT4 on Azure with a large number of tokens, like, like an extra sort of enterprise level of tokens to take those trillions of relationships, find the very best ones and ask does somebody, has somebody already figured this out? So we know that there are hallucinations with LLMs and we don't want to use them as a central portion of our research, but to give our scientists a prioritized list, we like to go after completely new stuff and, and we use GPT4 to essentially say if this drug and this gene, have they ever been talked about before, do we know about this? And if the answer is no, and we did a lot of prompt engineering to kind of whittle this down, a lot of benchmarking, we could get to the point where we could essentially 8020, a GPT4 can give us an 8020 answer, 80% accurate about whether or not something we have found is novel. And if it's novel and the relationship is strong and there's a lot of unmet need in that disease, we use the LLMs to answer all those questions. Then we hand it to our scientists and that helped us take those trillions of relationships and whittle it down to a couple hundred thousand that are going to be the programs that we try to focus on at recursion. So yes, we're using it there. And I should also say, like all the other businesses in the field, our EA teams are using it, right? Our finance teams are using these kinds of tools. We're using it across all the other kinds of normal business things that all of us have to do in any kind of company we're building.

Are you still running in circles trying to figure out how to actually grow your business with AI? Maybe your company has been tinkering with large language models for a year or more, but can't really get traction to find ROI on Gen AI. Hey, this is Jordan Wilson, host of this very podcast. Companies like Adobe, Microsoft and Nvidia have partnered with us because they trust our expertise in educating the masses around generative AI to get ahead of and some of the most innovative companies in the country hire us to help with their AI strategy and to train hundreds of their employees on how to use Gen AI. So whether you're looking for ChatGPT training for thousands or just need help building your front end AI strategy, you can partner with us too. Just like some of the biggest companies in the world do. Go to your everydayai.com partner to get in contact with our team or you can just click on the partner section of our website will help you stop running in those AI circles and help get your team ahead and build a straight path to ROI. On Gene, what are the, what are some of the next largest hurdles, right, to get from that, that 90% to that 80%. Because I'm sure when you said that and you know, you said, hey, the, the implications of that are potentially paying half as much for, you know, for the drugs that we all use, which I'm sure a lot of people would, would be very interested in that. So what are those next hurdles? Because, yes, advancements in AI technology. It seems like, Chris, you know, from someone a much more non technical perspective that they're coming so fast it's hard to keep up. So what are the next hurdles and just how much do some of these more recent innovations and advancements in AI help in clearing those hurdles?

Well, it's not a very sexy answer, but I think it's a data problem, right? And there are, you know, if you look at so many of these tech companies in the world making all these incredible tools, data availability is, is not the primary limiter. The scaling laws apply. But it's mostly compute that we see being deployed against existing data sets to help take us to the next level. We have done a lot of work to demonstrate that these scaling laws apply. By scaling laws, I mean more data and more compute give you better answers, right? And that's why we see more data, more compute together they give you better answers. In our field, data is the limiter. The right data is the limiter. And that's why we have. If I were to take you on a tour right now, you'd see this factory full of robots doing millions of experiments every week. And so for us and so many other companies, it's about building the data set. Because there doesn't exist some large public data set that's nicely relatable, that's free and open source that we can use to train these models. And so it's about generating that data and accessing that data. That's the hurdle that we're all facing in the field. And it's why there's a few companies like Recursion who have really for the last decade been working to build those data sets and find ways to accelerate our ability to scale that, that piece of the puzzle.

You know, Chris, I like, I like the comparison that you made earlier with kind of what, you know, recursion and other in your fields are, are are trying to do when it comes to better using AI to create drugs that fail less kind of to the self driving or autonomous vehicles, right, because at that point you had to create the data, right? You couldn't tap into a data set before. How challenging is that process, you know, for, for those of us that aren't, you know, in the field and trying to understand that, how challenging is that to, you know, create that data? And then what, what happens, right, is is there kind of that inflection point that, that once you hit, right, it's like, oh, once the self driving car is, you know, better than the average driver, then things, you know, completely open up. Like what, what does that process look like? And is there that point that, that you will get to or the industry will get to that then all of a sudden the floodgates will open?

Well, I think we're nearing that point and that'll be the first molecule that was discovered with the help of AI, being approved by the FDA and being applicable for real, real patients on the market. And we've got five drugs in clinical trials right now that are on their way there. There's a number of other companies who have medicines in clinical trials. And when the first medicine where AI helped make it possible is being used every day to help patients with some disease, I think that will really be the unlock point for our industry to more broadly accept this. Because today as with any industry, there's tons of skeptics, right? Even in the self driving vehicle space. Some of those companies have already outperformed humans. But we still, they're better than average human drivers. But we still actually see a lot of resistance in our, in, in the world to accepting self driving cars, right? When one of those cars gets into an accident, it still makes the front page news. Even if human drivers get in what, 30,000 accidents a day or whatever it is. I think it just is going to take time and real proof points. And as I said before, it takes 12 to 15 years in our industry to go from start to finish recursion is only 10 years old. So the fact that we're on the precipice of having medicines in the clinic and maybe one day in the near future approved, I think that'll be the point where it starts to be real for people. That'll be our, our version of we're better than the, you know, the average human driver.

Sure. I, I don't like doing this. I don't like asking guests to, to look into their crystal ball, but I can't help it. In, in this case, Chris, because, you know, even going from that 90 to 80%, you know, hopefully, you know, cutting down, you know, the, the price of drugs, which, you know, and treatments as well, because that can be extremely expensive. But what does that future look like? Maybe once we do hit that point or once there is that first successful drug that, you know, you know, has an artificially or has a molecule kind of created by artificial intelligence and you know, kind of. Oh, then the floodgates open. But what could that look like in the long term, right, Is that, you know, going from that, you know, 10 to 15 year old drug approval process to months. I mean, what could the future of better medicine look like with the help of artificial intelligence?

Well, you know, we tend to overestimate what we can achieve in one year and underestimate what we can achieve in a decade. I'll go out to a decade, I think a decade from now. If you look at the top biopharma companies in our industry, there's going to be some new names, maybe ours, there's going to be some older companies. Some of these Biopharma companies are 100 plus years old. I think some of them are going to fall by the wayside just like we've seen in, you know, like look at the market caps of like a Tesla versus maybe a more traditional auto brand. So we're going to see a shakeup in the industry. But for patients, what this is going to mean is better medicines available more quickly. I don't think months. I think there's still going to be a necessary FDA approval process that will take years. But if you can go from 12 or 15 years to 2 to 5 years and you can go from 90% failure rate to say 50% failure rate in the next decade, it means medicines are no longer going to be really, really expensive for folks 10 to 20 years from now. It means that there's maybe even going to be new models of care. Maybe you don't have to go get a prescription. Maybe it's part of Amazon Prime. We're already seeing this actually Amazon has this service now where you can get your medicines delivered via Amazon Prime. Maybe we get rid of PBMs, these middlemen in our industry that basically increase the price of drugs really substantially. And your, your doctor just puts your order in and the Amazon drone drops it off. And, and, and I hope in 10 to 20 years, Jordan, that we're in a position where this means we can move from treating disease to preventing disease and that'll be the next real frontier for Our industry, How do we make sure you never get the disease in the first place?

This, that would be great. You know, hey, I'm, I'm looking forward to this, this, this future that you just kind of described, Chris. But I mean, we've talked, we've talked about a lot. We've talked about the problem with, with collecting data and, you know, kind of drawing that comparison to autonomous vehicles, and we've talked about why drugs fail and, and how artificial intelligence can, can help improve that process. But maybe what's, what's the one takeaway from that you hope people, you know, maybe not even people in the field, but what's the one takeaway that, that you hope people can have from hearing you speak today specifically on, on how it impacts the future of, of their medical care? What is that takeaway?

Well, look, there's so much doom and gloom even just the last like 10 days looking at this existential world of AGI. I think the work that we and so many others in our industry are doing is such a great example of the potential for AI for good. And so, and I would just say to everybody out there who's working in an industry where maybe it's come to an industry like this where you have the potential to deploy these tools in the service of a mission like alleviating suffering at scale. This is such an exciting place to work. I'm just, I'm a nerd. I love being in the science. I love working in these complex problems because I think it's going to do so much good impact for humanity in the coming decades.

Oh, I love that. I love that. Well, hey, thank you so much, Chris, for joining the everyday show and helping us all better understand what's going on and how companies like Recursion are using AI to bring down that failure rate in drugs. Thank you so much for your time and coming on the show.

Thank you.

Hey, and as a reminder, maybe you joined late. Maybe you're, you're driving your, your car and someone was beeping and you missed the important parts. Don't worry. Every single day we go over this. So make sure if you haven't already, go to your everydayai.com sign up for that free daily newsletter. Yes, we always recap the news and trends and, and new tools. But we also every single day break down the conversation that we had with our expert guests and tell you even more about what it means for you growing your career, growing your company, or hey, maybe in this case, growing your longevity in life. So thank you for joining us and hey, you all voted hey in in the newsletter yesterday. I said, what do you want to hear on Friday? So tomorrow, because you you wanted it, we're going to be talking about the Amazon Q new model. So hope to see you back tomorrow and every day for more Everyday AI. Thanks, y'.

All.

And that's a wrap for today's edition of Everyday AI. Thanks for joining us. If you enjoyed this episode, please subscribe and leave us a rating. It helps keep us going for a little more AI magic. Visit your everydayai.com and sign up to our daily newsletter so you don't get left behind. Go break some barriers and we'll see you next time.