AI Drug Discovery: Revolution or Expensive Illusion? | Ep. 978

Billions Poured into AI Drug Discovery: Hype vs. Reality Welcome to the Bow Tied Biotech podcast, where science meets business. If breakthrough biotech, bold innovations, and the business behind the science gets you excited, you're in the right place. Make sure to check us out on Substack X and LinkedIn at at Bow Tied Biotech for even more insights and updates. All right, we've got a lot to cover today, so let's dive right in. Speaker 2 Over the last five years, I mean, we've just seen billions of dollars poured into AI drug discovery. Speaker 3 Absolutely. Billions. It's been massive. Speaker 2 Right. And we've seen the huge IP OS, the breathless magazine covers, and just this overarching narrative that artificial intelligence is going to, you know, completely eradicate human disease by next Tuesday. Speaker 3 Yeah, the hype has been completely off the charts. Speaker 2 It really has, but there's this multibillion dollar secret that the headlines are conveniently skipping over, which is that AI doesn't actually cure diseases. Speaker 3 No it doesn't. Speaker 2 In fact, right now the primary thing AI is doing is just getting us to the exact same failure point, just incredibly fast. Speaker 3 Right, because it's the ultimate collision between the clean, predictable logic of computer science and, well, the messy evolutionary chaos of human biology. You had all these tech investors walking into the biotech space and they were expecting to optimize it like a like a logistics network or something. And they are currently experiencing a very severe reality check. Speaker 2 Welcome to today's Deep Dive For you listening. Whether you're prepping for a meeting, tracking the biotech market, or just insanely curious about where technology is actually heading, our mission today is to just cut through the noise. We're diving into a really fascinating stack of analytical notes to figure out if artificial intelligence is truly revolutionizing how we invent medicines, or if, you know, we're just looking at a very expensive illusion. Speaker 3 And this isn't just some academic debate either. The way we value these companies, the way pharmaceutical pipelines are being built, and frankly, the actual timelines for life saving therapeutics, right. All of that hinges on separating the platform hype from the gritty, clinical reality. The Original AI Drug Discovery Pitch Solved Only 10% of the Problem OK, let's unpack this. Yeah, because I think we need to rewind a bit to really understand the original pitch that drove all this capital into the space. Speaker 3 Yeah, let's go back to the basics of the problem. Speaker 2 Historically, discovering a new drug meant essentially running a massive physical trial and error campaign, right? You have these huge robotic library screening hundreds of thousands of chemical compounds. Speaker 3 Just endlessly testing them against a disease. Speaker 2 Target exactly. Just hoping one of them sticks and it sounds like we were promised a supercomputer that could perfectly find a needle in a haystack instantly. Because the old way is incredibly slow, the failure rate is astronomically high, and it costs a fortune. Speaker 3 That's spot on. The foundational problem in biotech has always been inefficiency, so the original thesis for AI was honestly entirely logical. Speaker 2 It makes sense on paper. Speaker 3 Totally. Instead of physically brute forcing our way through millions of compounds, we use machine learning models. We use algorithms to predict protein structures simulate how molecules will bind. Speaker 2 Perfectly designing the chemical. Speaker 3 Right computationally designing the perfect chemical to hit a specific disease pathway, and the promise there was that this would radically proved success rates. Speaker 2 So the narrative was basically that we were moving biology from the game of chance to a game of engineering? Yes, exactly. Speaker 3 But. Speaker 2 Looking at the actual data in our sources, there's a massive structural flaw in that pitch because the AI is doing exactly what it promised to do. I mean, it is designing molecules at unprecedented speeds. Speaker 3 It really is. Speaker 2 But it turns out that specific phase of drug development only represents about 10% of the overall problem. Speaker 3 What's fascinating here is how effectively the tech industry marketed that 10% as the entire battle. Speaker 2 Oh, for sure. Speaker 3 Because getting a molecule to bind to a target in a computer simulation or even, you know, in a Petri dish, that is just the absolute tip of the spear. Yeah. The remaining 90% of the drug development life cycle relies on clinical biology. It relies on pharmacokinetics, which is incredibly messy. It relies on taking that novel chemical and putting it into a living, breathing human being with a wildly complex genetic makeup. Speaker 2 Right, because an algorithm can map a protein fold perfectly, but it can't magically solve the reality of how a human liver is going to metabolize that chemical over a six month period. Speaker 3 Precisely, the AI optimizes the in silico phase, so the computer modeling and it even optimizes the in vitro phase, you know, the test tube, but it hits an absolute brick wall when it gets to the in vivo phase, the living Organism. Speaker 2 So if we have beautifully optimized the 1st 10% of the funnel, the logical next step is to look at what happens when those computationally perfect molecules hit the remaining 90%. AI-Designed Drugs Face Sobering Clinical Reality, Pharma Shifts Deals The real world. Speaker 2 Exactly. And we don't have to guess anymore. We now have several years of AI driven pipelines that have officially entered clinical trials. The rubber has hit the road. Speaker 3 And the data we're seeing is honestly incredibly sobering. Yeah, based on the notes we're reviewing, the results across the board are underwhelming. We are not seeing clearly better success rates. We're not seeing standout clinical efficacy. Speaker 2 So they're failing just as much. Speaker 3 Basically, yes. The drugs designed by AI are failing in phase one and phase two trials at roughly the exact same rate as drugs designed by traditional human chemists. Speaker 2 Wait, I have to push back on that a little bit because I think we might be glossing over a massive efficiency game here. Speaker 3 OK, how so? Speaker 2 Well, the sources explicitly detail that these AI biotech companies are shrinking the discovery timeline significantly. I mean, they they're taking programs from a blank whiteboard to an investigational new drug application in like 18 months instead of four or five years. Speaker 3 Right. They're moving into human trials much faster, yeah. Speaker 2 So even if the failure rate is the same, cutting three years off the front end has to count for something, right? I mean, that is a massive cost saving. Speaker 3 It is a cost saving in the discovery phase, absolutely, but we really have to clarify the fundamental mechanics of where money is actually burned in biotech, OK, there is a massive gulf between speed into the clinic and success in the clinic. The preclinical phase that first 18 months you mentioned is relatively cheap. Speaker 2 Oh, I see. Speaker 3 The real cost of drug development, like the hundreds of millions of dollars is spent running the human clinical trial. Speaker 2 I see where you're going with this. Speaker 3 Yeah. So if you accelerate the speed at which molecules enter human testing, but you haven't improved the underlying probability that they'll actually work, you aren't saving money. Speaker 2 You're just spending it faster, yeah. Speaker 3 Exactly. You are just accelerating the rate at which you burn capital in the clinic. You're arriving at the most expensive failure point much faster. Speaker 2 Wow, that makes perfect sense. To use an analogy here, it's like we built an incredibly efficient AI factory that produces lottery tickets at lightning speed. Speaker 3 Oh, that's a great way to look at it. Speaker 2 Right. Yes, we are printing tickets faster and cheaper than ever before, but the fundamental odds of any single ticket being a winner haven't changed at all. Speaker 3 And the tickets still cost a fortune to actually play. Speaker 2 Exactly. If it cost millions to play them in the clinical trial system, you're just going to go bankrupt faster. Speaker 3 That's the reality. The tech industry loves the metric of shots on goal. You know, generate more assets, take more shots. Speaker 2 Yeah, the Silicon Valley mindset. Speaker 3 Right. But in biotech, every shot on goal requires a $50 million phase two clinical trial. If your accuracy hasn't improved, you literally can't afford to take unlimited shots. Speaker 2 Which brings us to the most honest indicator of reality in any industry. Speaker 3 The money always follow the money, right? Speaker 2 If we track how the capital is moving, we can see exactly how the market is digesting this reality check. The sources highlight a very sharp shift in how Big Pharma is structuring deals with these AI platforms today compared to just three years ago. Speaker 3 The contrast is night and day. I mean that three or four years ago, we were in the middle of a massive FOMO cycle. Fear of missing out. Oh yeah, The major pharmaceutical companies were absolutely terrified that some tech startup in Silicon Valley was going to completely disrupt their entire pipeline. Speaker 2 So they were writing these massive blank check deals, they were paying huge upfront cash premiums just to get broad access to the AI platform itself, completely regardless of what specific drugs it was producing. Speaker 3 Yeah, it was purely an investment in the promise of the technology. Speaker 2 So what does this all mean? Because if you look at the deal structures over the past 12 months, the upfront payments have plummeted. Speaker 3 They've completely dried up. Speaker 2 Right. Big Pharma is no longer buying access to the magic algorithm. Instead, the deals are entirely asset centric and heavily backloaded with milestones. Speaker 3 Yes, they are saying look, we will only pay you if this specific molecule you designed actually passes a phase one safety trial. Speaker 2 Right. Prove it first. Speaker 3 And we'll pay you a bit more if it passes Phase 2. Speaker 2 So from a business perspective, the market's attitude has shifted from valuing the process to valuing the product. The buyer has entirely transferred the clinical risk back onto the AI companies. Speaker 3 The major players learned a hard lesson. They realized that having a neural network capable of generating a million novel chemical structures is absolutely useless if 999,000 of them are toxic to human kidneys. Speaker 2 Wow, yeah. Speaker 3 So the burden of proof is now on the AI companies to prove their molecules can survive the clinical gauntlet. Not just, you know, look pretty on a server rack. Understanding Why AI-Designed Molecules Fail in Complex Human Biology Here's. Speaker 2 Where it gets really interesting though, because we have to ask why the clinical failure rate hasn't, but why are these perfect molecules failing? Speaker 3 That is the $1,000,000 question. Speaker 2 Because in the software world, when a piece of code fails, you isolate the bug, you rewrite the logic, and you push an update. It's an entirely rational process. Speaker 3 Completely logical. Speaker 2 But reading through the breakdown of biology versus code in our sources, it becomes super clear that human biology actively resists that kind of optimization. I mean, even if you designed the absolute perfect molecule in a computer simulation, it feels like designing the ultimate race car in a video game but then forcing it to drive through a real world swamp. Speaker 3 I love that analogy, and that swamp is a cynical gauntlet where roughly 90% of all drugs fail. And crucially, AI hasn't changed that 90% failure rate meaningfully. And that is the core bottleneck that tech investors continually fail to grasp. Biology is not software. In software, you build the environment, which means you can perfectly simulate it. You know, every single variable. Human biology is a system we did not build, we do not fully understand, and it is inherently chaotic. Speaker 2 To put some concrete mechanics behind that, let's talk about why a drug fails even when the AI does its job perfectly. Let's say the AI designs a molecule. Think of the AI as a master locksmith. It maps the exact 3D structure of a disease receptor, the lock, and it computationally machines the mathematically perfect key to fit that lock and shut the disease down. Speaker 3 Right, so in the simulation it is a 100% success. Speaker 2 Exactly. The binding affinity is flawless. Speaker 3 Flawless. Speaker 2 But the human body isn't just a single door. It is a sprawling metropolis with billions of doors. So you take this mathematically perfect key and you put it into a human being. Speaker 3 And what happens? Speaker 2 Well, it turns out there's an entirely unrelated essential protein in the liver that happens to have a lock with the exact same structural shape. Your perfect key opens the disease door, but it also shuts down the liver. That's off target toxicity. Speaker 3 And that is a brilliant way to explain it because it highlights exactly what the AI cannot see. But honestly it gets even worse than off target toxicity. Speaker 2 Wait, really? Speaker 3 Well, let's say the key only hits the right door. The AI perfectly blocks the disease pathway without hitting anything else in the entire body. Sounds. Speaker 2 Like a win. Speaker 3 You'd think so, but biology has spent billions of years evolving redundancy. Very often when you block a biological pathway, the disease simply routes around it. Speaker 2 Oh wow. Speaker 3 Yeah, the body activates A compensatory mechanism and the tumor just keeps growing anyway. Speaker 2 And no matter how many parameters your generative model has, or how much compute power you throw at it, you just cannot simulate that biological redundancy in silico with our current understanding of the human body. Speaker 3 No, you can't. We just don't have the training data for it. We understand chemical binding, sure, but we do not have a comprehensive mathematical model of human Physiology. Got even close? And until we do, AI cannot predict clinical efficacy. It can only predict chemical properties. AI Drug Discovery: Valuation Bubble or Advanced CRO Service? OK. So if the fundamental output of these AI companies is just a chemical property, like an early stage asset that still faces the standard 90% failure rate in human trials, that creates a massive existential problem for the financial markets. Speaker 3 It creates a huge bubble. Speaker 2 The sources layout a bear case that is gaining serious traction right now and we need to dig into the mechanics of this valuation mismatch. Because if they were just high end research assistants, isn't it a massive problem that they're being valued like highly scalable software platform companies? Speaker 3 If we connect this to the bigger picture, it is arguably the most important dynamic playing out in the biotech sector right now. Yeah, the public and private markets have been valuing these AI drug discovery companies like highly scalable software as a service or SAWS businesses. Speaker 2 Right, giving them a tech. Multiple tech companies get valued at massive premiums because of 0 marginal cost. Speaker 3 Exactly. Speaker 2 Once Microsoft writes the code for a new operating system, it costs them effectively 0, $0.00 to copy that code and sell it to a billion people. The margins are astronomical. Speaker 3 But biotech does not have 0 marginal cost, Not even remotely. If an AI company designs a new drug, they can't just copy and paste it to a million patients. I wish, right? They have to manufacture it, formulate it and push it through years of highly regulated, intensely expensive physical trials. Speaker 2 So if these companies aren't actually tech platforms in terms of their business model, what are they? Speaker 3 Well, the ultimate bear case is that they are simply high end contract research organizations. CR OS. Speaker 2 And for anyone who doesn't track the biotech service sector closely, a CRO is basically the outsourced labor of the pharma industry, right? Big Pharma hires a CRO to run their lab tests, screen their compounds, or manage their trial data. They are very steady, very necessary service businesses. Yeah, but they trade at normal service business multiples. Speaker 3 Yeah, they do not trade at 50 times revenue like a high growth tech platform. Speaker 2 Exactly. Speaker 3 So if your AI is just acting as a highly efficient outsourced research assistant that hands a molecule over to big pharma to take all the clinical risk, you are a CRO. Speaker 2 A very advanced one, but still. Speaker 3 A technologically advanced CRO, but a CRO nonetheless. And if the market completely re rates these multibillion dollar AI companies as service providers, the valuations are going to absolutely implode. Speaker 2 So the obvious question is what actually breaks this bear case? How does an AI drug discovery company justify its existence as a revolutionary platform and defend those massive valuations? Speaker 3 It comes down to one word, evidence. Speaker 2 Show me the data. Speaker 3 Exactly, and not anecdotal evidence. We don't need one lucky drug that happened to make it through phase three trials. The only thing that proves the AI is fundamentally changing the paradigm is a statistically significant, repeatable pattern of clinical success that is directly attributable to the algorithmic design. Speaker 2 So we need proof that the lottery tickets printed by the AI actually have a higher win rate than the tickets printed by human chemists. Speaker 3 Exactly right. Speaker 2 If a traditional pipeline sees one out of 10 drugs take it through the clinic, we need data showing the AI pipeline gets three or four or five out of 10 through. Speaker 3 Precisely. And looking objectively at the clinical readouts we have today, we aren't even close to that yet. We are still waiting for the definitive proof that an algorithm can untangle the chaos of human biology. AI's Efficiency Creates a Catastrophic Clinical Trial Bottleneck So bringing this all together for you listening right now, what is the actionable take away here? Based on the deep dive into these sources, it is crucial to reframe how you look at AI in this sector. Speaker 3 It really is. Speaker 2 AI in drug discovery is incredibly powerful. It's successfully mapping proteins, it's finding new binding pockets, and it's generating novel chemistry, but it is an incremental tool for efficiency. Speaker 3 It is not a magical revolution for clinical outcomes. Understanding that distinction between preclinical efficiency and clinical efficacy is your shortcut to being well informed. Speaker 2 The next time you see a headline claiming an AI generated A breakthrough cure in just 45 days, you now know the right question to ask. Speaker 3 What happens on Day 46? Speaker 2 Exactly. You know that 45 days only covers the easy part. The real test is the 10 years of human trials that follow the expectations and the capital flowing into the space. Just need to acknowledge the grueling reality of biology. Speaker 3 It's a necessary maturation for the industry. I mean, AI will absolutely be a foundational piece of how we discover drugs moving forward, but it is a tool wielded by scientists, not a replacement for the scientific method. Speaker 2 But before we wrap up, I want to leave you with one final, incredibly provocative thought to Mull over something that builds on this dynamic but points to an entirely different crisis on the horizon. Speaker 3 Oh, this is the big one. Speaker 2 Our sources show that AI is undeniably making the top of the funnel, the discovery of early stage molecules, incredibly fast and cheap. It is generating an unprecedented volume of new drug candidates. Speaker 3 The pipeline of assets heading toward human testing is swelling to a size we have literally never seen before in the history of pharmacology. Speaker 2 Right. But think about the physical infrastructure required to test all of those candidates. Yeah, if we flood the pharmaceutical pipeline with thousands and thousands of cheap AI generated virtual molecules and every single one of them still has to navigate the slow physical highly regulated human trial process, well. Speaker 3 What happens to the system? Speaker 2 Exactly what happens? Speaker 3 You're looking at a fundamental capacity breakdown. The FDA does not have infinite reviewers to process thousands of new investigational new drug applications. Speaker 2 And it is not just the regulators. What happens when we literally run out of human patients? Yeah, to run a phase two or phase three trial, you need hundreds, sometimes thousands of human beings with a very specific disease profile who meet strict eligibility criteria and are willing to take an experimental compound. There are only so many patients available. Speaker 3 Especially in rare diseases or highly targeted oncology, where multiple companies are already fighting to recruit from the exact same tiny pools of patients. Speaker 2 It's a massive looming irony. By using pristine computer code to completely eliminate the bottleneck at the very beginning of the discovery process, we might not be revolutionizing the speed of medicine at all. By solving the 1st 10% so efficiently, we might just be creating a catastrophic, unprecedented traffic jam at the edge of the biological clinic. Speaker 3 It is a classic systems engineering problem. You didn't eliminate the bottleneck, you just moved it further downstream. And a lack of human trial patients is a bottleneck that no algorithm on Earth can code its way out of. Speaker 2 Something to keep in mind the next time you read about the AI health revolution. The algorithms are ready. The physical world might not be able to keep up. Thanks for joining us for this deep dive and we'll catch you on the next one. Speaker 1 You have been listening to the Bow Tied Biotech podcast. Make sure to check us out on Substack X and LinkedIn at at Bow Tied Biotech for even more insights and updates. We appreciate your support and see you next week.