🧬 The Algorithmic Apothecary: AI Frontiers in Pharmaceutical Innovation

📖 Read: 

How AI Agents Are Breaking Eroom's Law and Unlocking a Revolution in Drug Discovery

There is a law in pharmaceutical science that most people have never heard of, and it is quietly devastating. It works exactly like Moore's Law — that familiar rule that computing power doubles every two years while cost halves — except that it runs in reverse. Scientists named it Eroom's Law, spelling Moore's backwards as a kind of dark inside joke. Since 1950, the number of new drugs approved per billion dollars spent on research and development has been cut in half roughly every nine years.

We are not waiting for new cures to be invented. In many cases, we are waiting to find the ones we already have.

The warehouse is full. We finally have the flashlight.

Co-Scientist: A multi-agent AI partner to accelerate research — Google DeepMind 
and 9 other sources

This is Heliox: Where Evidence Meets Empathy

Independent, moderated, timely, deep, gentle, clinical, global, and community conversations about things that matter.  Breathe Easy, we go deep and lightly surface the big ideas.

Transcript

0:25

Imagine for a second, right, that you want to build a simple wooden chair. Nothing fancy at all. Just, you know, four legs, a seat, and a back. Exactly. And back in, say, 1950, maybe it took you a few hours in a garage, right? You have a couple of basic hand tools, maybe a few dollars worth of wood. Tree straight forward. Right. But now... Flash forward to today, you have, I don't know, laser-guided saws. Yeah. You have this complex computer-aided design software that can calculate the exact stress points of the wood. You might even have a fully programmable robotic arm. Right, the absolute cutting edge of manufacturing. Yeah, but despite all that incredible futuristic technology, Somehow, building that exact same wooden chair today takes you 10 times longer. And it costs you 10 times as much. You'd think you were losing your mind. You really would. I mean, it defies all basic economic logic. You have better tools, but the output is worse. But that paradox, that maddening backward trajectory, that is exactly what has happened to, honestly, one of the most important high stakes industries in human history. The discovery of life-saving medicines. Yes. Welcome to this deep dive where we are looking at exactly that and we're unpacking a massive stack of sources today. We've got pharmaceutical economics, some crazy AI technology, and specifically a groundbreaking May 2026 paper from Google DeepMind called Co-Scientist. Yeah, Co-Scientist, a multi-agent AI partner to accelerate research. It's a huge shift. It really is. Because the mission today is to figure out why the traditional way humanity discovers medicines is so fundamentally broken. But more importantly, how a new era of what they call tech bio is radically rewriting the rules. And to understand the fix, you really have to understand just how bad the problem is. Scientists actually have a rather dark sense of humor about that wooden chair paradox you mentioned. They call it Eroom's Law. Eroom's Law. Wait, let me look at how that's spelled. E-R-O-O-M. Right. If you look closely at the spelling, that is literally just Moore's Law written backwards. Oh. Oh, I see it. Wow. Yeah. So we all know Moore's Law, right? The observation that computing power gets exponentially cheaper and faster over time is It shrinks from these massive room-sized mainframes down to, you know, the smartphone in your pocket. Right. Everything gets faster, cheaper, better. Exactly. But Eram's Law observes the exact terrifying opposite in pharmaceutical research and development. Terrifying is the right word. I mean... Think about it. Despite the fact that we have decoded the entire human genome, despite massive leaps in robotic lab automation, and despite a fundamental understanding of cellular biology that would have seemed like pure magic 70 years ago, the number of new drugs approved per billion dollars spent on R&D has been having roughly every nine years since 1950. Which is just staggering when you pull apart the numbers. We are throwing more money, more data, more computing power at the problem of human disease than ever before in our species history. And the machine is just producing less and less. It is the absolute definition of a systemic crisis. So let's talk about the sheer gravity of that bottleneck. Yeah. What does the traditional model actually look like right now? Well, the industry calls it de novo drug discovery. That's the traditional route. De novo, meaning from scratch. You are trying to invent a brand new molecular structure that has never existed before. Right. You're just conjuring a molecule out of thin air to fix a problem. Basically. Yeah. And right now, bringing a single new drug to market that way takes an average of 10 to 15 years. 10 to 15 years for one drug. Yeah. And the cost. An average of $2.6 billion for every approved compound.$2.6 billion. With a B. With a B. And here's the kicker. A huge chunk of that $2.6 billion isn't actually paying for the drug that works. It's subsidizing an invisible graveyard of failures. OK, what do you mean by an invisible graveyard? That is the dark secret of the pharmaceutical economy. There is a staggering 90 percent failure rate for drug candidates once they actually enter human clinical trial. 90 percent. So they get all the way to humans and just fail. Exactly. I mean, you might spend five years and a hundred million dollars designing this beautiful molecule that works perfectly in a Petri dish. And it completely cures the disease in a lab mouse. Right. But a human being is not a giant mouse. And a human body is definitely not a petri dish. It is this chaotic, interconnected, wildly unpredictable system of trillions of cells. Right. It's a whole ecosystem. Exactly. So when you put that brand new molecule into a human, maybe it binds to the target receptor perfectly, which is what you wanted. but it also, say, accidentally triggers a cascading failure in the liver. Oh, wow. Or maybe it just gets chewed up by stomach acid before it ever even reaches the bloodstream, or it simply can't cross the blood-brain barrier. So the system is just suffocating under its own weight. And the real-world consequence, if you think about it, isn't just lost profits for pharmaceutical giants. No, not at all. The human cost is that thousands of rare, complex, and neglected diseases are just left completely unchecked. Like, if a disease only affects... say 50,000 people globally. The basic math of a $2.6 billion development cycle means the cure will never be invented. Exactly. The economics simply do not allow for it under the old model. The math doesn't work. Which forces us to look for an alternative route. I mean, if inventing a new drug from scratch is a jammed front door, we have to find another way into the house. Right. So what's the back door? In pharmacology, the backdoor means looking at the drugs we have already spent decades and billions of dollars inventing. We call this drug repurposing or drug repositioning. Okay, drug repurposing. Right. So instead of asking the incredibly difficult question, you know, how do we build a completely new molecule to cure this disease? You ask a much smarter, much simpler question. Does one of the thousands of molecules humanity has already built happen to cure this disease? Oh, that's fascinating. the probability of actually surviving clinical trials and getting market approval jumps from that dismal 10 percent up to roughly 30 percent. Yeah. Right. And the mechanism behind that massive jump in success is purely biological. How so? Well, when you bring a brand new de novo drug into the world, you have to run it through phase I clinical trials. And a lot of people don't realize this, but phase I isn't even about seeing if the drug cures the disease. Wait, it's not? No. Phase I is purely about human safety. You give the drug to healthy volunteers, people who don't even have the disease, to answer some pretty terrifying questions. Like what? Like, does this cause heart failure? Does it completely destroy the kidneys? What is the maximum dose a human liver can process before it becomes completely toxic? Okay, yeah, that is terrifying. Right. But with a repurposed drug, you are entirely bypassing that phase. That molecule has already run the gauntlet We already know its safety profile We know how it is absorbed, how it's metabolized Oh, I get it Because it was already approved for something else Exactly So the primary hurdle you have left is just proving efficacy Does it actually treat the new disease you are targeting? That makes so much sense But historically, finding these secondary uses wasn't exactly some highly systematized science, was it? No, not at all. It was basically blind, dumb luck. It was scientists stumbling around in the dark and bumping into miracles. And honestly, the stories are legendary in the medical field. Oh, they really are. Like, let's talk about sildenafil. Oh, the classic. The classic. Back in the 1980s, researchers developed it as a treatment for angina, right? Which is chest pain caused by reduced blood flow to the heart. So the drug was supposed to open up the blood vessels. Right. And in the initial clinical trials, it was actually quite underwhelming as a heart medication. Like it didn't really do the job. But the researchers noticed a highly unusual trend in the data. The male patients in the clinical trials were overwhelmingly refusing to return their leftover pills at the end of the study. Yeah, they were holding on to them. And when the doctors dug into why, the patients reported a very specific, consistent side effect regarding blood flow. Yeah, they certainly did. And Pfizer quickly realized they didn't have a mediocre heart drug on their hands. They had a blockbuster erectile dysfunction drug. They pivoted and we got Viagra. Incredible. Or look at minoxidil. It was originally formulated as a really potent oral medication for severe high blood pressure. Right. But patients taking it started experiencing hypertrichosis. Meaning they were growing thick hair all over their bodies. Everywhere. Everywhere. So the company reformulates it into a topical liquid, slaps a new label on it, and suddenly you have Rogaine. It's amazing. And those are the classic examples that usually get a chuckle. But serendipity has also led to some really profound, life-saving rescues of drugs that were absolute catastrophes. Like what? The most striking example is thalidomide. Oh, wow. Yeah. In the late 1950s, it was aggressively marketed across the globe as a mild, completely safe sedative. And it was specifically recommended as a cure for morning sickness in pregnant women. And it resulted in one of the darkest, most tragic chapters in the history of modern medicine. It really did. Thousands of children were born with severe birth defects, particularly phocomelia, which involves malformation of the limbs, and it was eventually pulled from the market in total disgrace. Exactly. It became the ultimate symbol of pharmaceutical negligence. The drug was universally shunned, as it should have been. Right. But the story actually didn't end there. Decades later, researchers who were studying completely different diseases made this shocking discovery about how thalidomide actually works at a cellular level. What did they find? They realized it has incredibly powerful immunomodulatory effects. But more importantly, it is anti-angiogenic. Anti-angiogenic. Meaning what? Meaning it actively stops the growth of new blood vessels. Now, if you are a developing fetus in the womb, stopping new blood vessel growth is catastrophic. Hmm. Right. That is what caused the terrible birth defects. Exactly. But what if you are a rapidly mutating blood cancer tumor that desperately needs to grow new blood vessels to survive? Oh, wow. You essentially starve the cancer. Yes. You literally starve it. And because of that serendipitous connection, a drug that was a global pariah was resurrected. It received FDA approval in 1998 to treat painful skin lesions associated with leprosy, and later it became a cornerstone, life-extending therapy for multiple myeloma, which is a devastating blood cancer. That is just wild. Through pure observation, A literal poison was repurposed into a lifesaver. It really is incredible. And we don't even have to look that far back to see this in action. Look at what happened during the COVID-19 pandemic. Oh, absolutely. The ultimate forcing function. Exactly. The world was on fire. Hospitals were overflowing. We fundamentally could not afford to wait 15 years and spend $2.6 billion to invent a de novo cure from scratch. We didn't have the time. No time at all. So the pandemic acted as this massive forcing function for repurposing. Researchers realized pretty quickly that it wasn't just the virus killing people, it was the body's own immune system overreacting. Right, it was creating that massive hyper-inflammation called a cytokine storm. Yeah, and the urgency was just unprecedented. So researchers used early computational algorithms to scan existing drug databases, looking for anything that could calm that specific immune response. And in just 48 hours, the algorithms flagged baricitinib. Right. Which was a drug already sitting on pharmacy shelves prescribed for rheumatoid arthritis. The algorithms basically predicted that the same mechanism calming joint inflammation could calm the lung inflammation caused by COVID. And it worked. It did. Within nine months, bericetina went from a computer prediction to an FDA emergency use authorization, saving countless lives. Okay, I have to stop you there, though. Yeah. Because I am hearing these stories, and I'm looking at the math you laid out earlier. An 85% cost reduction Success rates jumping from 10 to 30 percent. Timeline shrinking from decades to months. Right. If it is this cheap, this fast and this effective, why isn't every pharmaceutical company just doing this all day long? Like why ever bother trying to invent new drugs from scratch? Is that 30 percent success rate actually real or is there a catch? OK, yeah, there's a massive catch. And it requires a serious reality check on the statistics. Lay it on me. That 30% figure you see thrown around in biotech headlines, it's a blend. And it is highly misleading. Misleading how? Well, to understand why, we have to look at this pivotal 2018 study from Cambridge University. They tracked over 800 new molecules to see how repurposing actually plays out in the real world. Okay. And what they found is that the success rate depends entirely on the distance of the biological leap you were trying to make. They divided the entire field into two categories. soft repurposing and hard repurposing. Okay, let's define the boundaries there. What makes a repurposing soft? Soft repurposing means you are staying within the same biological neighborhood. Yeah. You are targeting a disease that is a really close cousin to the original disease. Okay, give me an example. So for example, taking a drug that is already FDA approved to treat breast cancer, and testing it in a clinical trial for patients with ovarian cancer. Ah. Right, because those two diseases share very similar underlying genetic mutations and cellular pathways. The drug often works beautifully. The Cambridge study found that soft repurposing has a staggering 67% success rate. 67%? That makes total sense, though. I mean, if you have a skeleton key that opens the front door of a house, there's a very good chance it will also open the back door, right? Because the builder probably used the exact same brand of lock. That is a brilliant way to conceptualize it, yeah. But the real prize and the real challenge is hard repurposing. This is the holy grail of drug discovery. Okay, so what is hard repurposing? This is trying to turn, say, a failed cardiovascular medication into a revolutionary treatment for a neurological disorder like Parkinson's. Oh, wow. Or turning an arthritis drug into a potent antiviral. You're trying to connect completely unrelated biological dots across vast distances of human physiology. Massively. Huge. And the Cambridge study found that when you attempt hard repurposing, the success rate absolutely plummets to just 9%. 9%. Wow. So the moment you try to leap outside the neighborhood, you are basically right back in the mud with a 90% failure rate of tragic death. traditional scratch-built drug discovery. Yes, exactly. Because for decades, finding those hard repurposing connections relied entirely on the limitations of the human brain. Right. I mean, think about it. To find a hard repurposing candidate, a human scientist had to somehow read a paper on neurology and then happen to read a completely disconnected paper on cardiology. And somehow remember both of them perfectly. Yes, and hold both of those highly complex data sets in their working memory simultaneously to spot a hidden microscopic link between the two. And the sum total of human biomedical data is simply too massive for any human brain to process. Right. There are thousands of papers published every day. Exactly. Which means the bottleneck isn't the chemistry. The bottleneck is our human inability to process the sheer volume of biological literature we have already written. If the answers are buried somewhere in the library, we don't need a better microscope, right? We need a better librarian. That is such a good point. And that brings us to the next massive paradigm shift in our deep dive today. We're talking about the transition from blind serendipity to systematized computational seething. Yes. The rise of a field called tech bio. Tech bio. Okay. How is that different from biotech? This is a crucial distinction. So for the last 50 years, we lived in the era of biotech. tech biotech is biology first it is artisanal it relies heavily on manual painstaking laboratory work to validate a specific biological target like the classic scientist in a white coat with a pipette exactly you form a hypothesis based on your limited human reading you go to the wet lab you pipette clear liquids into petri dishes for five years and you just pray you were right something exhausting it is But TechBio completely flips the script. It is data first. It is industrialized. It's powered by AI. And it relies heavily on a concept called literature based discovery or LBD. Literature based discovery. Right. LBD is the idea that the cure for a disease might actually already exist, but it's fragmented into tiny pieces across 50 different scientific papers that have never cited it. each other. Oh, that is wild. And my absolute favorite part of this entire deep dive is the origin story of LBD. Okay, tell me. Because it starts way back in the mid-1980s with a researcher named Dr. Don Swanson. He essentially invented this entire computational field using nothing but paper index cards and library catalogs. Wait, paper index cards? Yeah. In the 80s? Yeah. Don Swanson is an absolute legend in the informatics community. He formalized a brilliant hypothesis. He believed there was a vast ocean of what he called undiscovered public knowledge. Undiscovered public knowledge. I love that phrase. Right. He basically set out to prove that you could make major medical breakthroughs without ever touching a test tube, simply by connecting existing data in novel ways. Okay, let's walk through his famous 1986 experiment, because it perfectly illustrates how AI thinks today. But he did it manually. Manually. So Swanson is digging through the medical literature, spending hours and hours in the physical library. Manually. He starts reading a cluster of papers about Raynaud's disease. Okay, and for those who don't know, Raynaud's is a pretty brutal condition where the capillaries in your extremities spasm, right? It cuts off blood flow so your fingers turn ice white in the cold. Exactly. So Swanson reads that these spasms are strongly linked to high blood viscosity. meaning the patients have unusually thick blood. Okay, so he establishes the first link. Disease A, Raynaud's, is caused by condition B, high blood viscosity. Right. Then Swanson physically walks over to a completely different section of the library. Like down the hall. Literally down the hall. He starts reading nutritional science papers that have absolutely nothing to do with Raynaud's. And he discovers a separate cluster of papers proving that consuming high amounts of dietary fish oil, specifically icosapentaenoic acid, actually increases the deformability of red blood cells. Which would mean? Which significantly lowers blood viscosity. Oh. Wow. So there's the ABC model of literature-based discovery. Yes. A is linked to B. C is linked to B, therefore A might be linked to C. That is so simple but so profound. Even though no scientist had ever written a paper putting Raynaud's and fish oil in the same sentence, Swanson manually connected the intermediate biological mechanism. He did. He published a paper hypothesizing that fish oil could be a potent treatment for Raynaud's disease. and a few years later, clinical researchers actually ran the wet lab trials and proved he was absolutely correct. He cured a circulatory disease using the Dewey Decimal System. He literally did. That is incredible. But think about the sheer amount of time it took one human being, right? Walking around a library, today in the tech bio era, we take Swanson's manual ABC index card process, and we just inject it with modern natural language processing or NLP. Right, on a massive scale. We aren't relying on a human spending their weekends in a library anymore. We're teaching machines to read millions of dense patents, clinical trial reports, and scientific papers simultaneously. And we have to be really clear about how these NLP models actually read because they don't read English the way you or I do. They don't understand the narrative or the prose of a paper. Right. What they do is tokenize the text. They extract specific entities, so a drug name, a specific gene, a protein receptor, a disease, and they map the mathematical relationships between them in a high-dimensional vector space. They build what the industry calls a knowledge graph. A knowledge graph. Okay, how should we visualize that in our heads? Imagine a massive three-dimensional web floating in space. Every single node in that web is a biological concept. Maybe it's the SARS-CoV-2 spike protein. Maybe it's a specific liver enzyme. Don't get the nose. And every string connecting those nodes represents a proven relationship extracted from the literature. Some strings mean inhibits. Some mean promotes. Some mean binds to. Ah, so it's a map of verbs and nouns, basically. Yes, on a scale of millions. Yeah. Platforms like Molliervis and Premagee have built these graphs containing millions of nords and billions of connections. Yeah. So suddenly researchers aren't relying on remembering a paper they read five years ago. They are querying the sum total of human biomedical knowledge in seconds. The AI can traverse the web and find a path connecting a failed heart drug to a brain receptor through a dozen intermediate biological steps that a human would just never spot. But tech bio isn't just about reading old papers to find new uses for old drugs, right? The technology has evolved to the point where AI is revolutionizing how we invent new drugs from scratch. It absolutely is. To understand this leap, we have to look at a company called InSilico Medicine and a drug they developed called RentoCertib. Yes, RentoCertib. This drug targets idiopathic pulmonary fibrosis, or IPF. And IPF is a horrific progressive disease where the lung tissue becomes increasingly scarred and stiff, basically making it impossible for the patient to breathe. Awful. But Insilico's approach with Rentesertib is an absolute masterclass in the shift from biotech to tech bio. Instead of spending years manually synthesizing thousands of slightly different physical compounds in a lab to see which one might halt the lung scarring, they utilized an AI architecture called generative adversarial networks, or JANS. Okay, I really want to break down JANS carefully because the mechanism sounds like pure science It does. The architecture of a Jan is just brilliant. It consists of two separate neural networks locked in a continuous high speed debate. Like an argument. Basically. You have a generator network, which acts kind of like a master counterfeiter. Its job is to imagine and draw completely new hypothetical molecular structures. Then you have the discriminator network, which acts like a forensic detective. And the discriminator is trained on the strict laws of physics and chemistry. So the generator draws a molecule and hands it to the discriminator. And the discriminator analyzes it and says, no, this carbon bond is physically impossible. Or no, this molecule is too unstable and would instantly degrade in stomach acid. Or no, this shape won't fit into the IPF receptor. So it's super critical. Very. The discriminator rejects the fake, and the generator uses that exact feedback to draw a slightly better molecule. And they fight back and forth like this millions of times an hour, constantly refining the design. So it isn't just searching a library of existing molecules. the AI is literally imagining a perfectly shaped entirely new molecular structure from scratch tailored to the exact atomic shape of the disease target that's exactly it and the performance metrics on this AI pipeline are almost unbelievable for anyone trained in traditional pharmacology like how fast are we talking well remember that in traditional discovery just getting to a preclinical candidate which is the final molecule you lock in to test in clinical trials takes an average of four and a half years and requires chemists to physically synthesize and test thousands of molecules in a real lab. Incilico used their Jans to nominate their candidate, Runtusertib, in just 18 months. Wow. That compresses years of human labor into months. And because the AI had done so much of the trial and error digital simulation, Incilico only had to physically synthesize and test 78 molecules in the real world to find the winner. Just 78. Out of thousands. Yeah. The total cost of that discovery phase was $2.6 million. Yeah. Compared to the traditional hundreds of millions spent at that early stage, it is basically a rounding error. It is a complete paradigm shift. I mean, we are compressing time and money at an unprecedented scale. Okay. I do have a major pushback here. Okay, let's hear it. Gs and knowledge graphs are incredible mathematical tools, obviously. They're amazing at pattern matching, at connecting node A to node B, and at imagining complex 3D shapes that obey chemical laws. Okay. Sure. But actual scientific breakthroughs, like the kind that cure the hardest, most complex human diseases, require more than just matching data points, don't they? They require a leap of intuition. They require formulating entirely new hypotheses about why a biological system is acting the way it is. Absolutely. So how do you get an artificial intelligence to actually think like a human scientist? rather than just acting as a superpowered three-dimensional search engine. You've just hit on the absolute frontier of artificial intelligence right now. You're pointing out the limitation of pure pattern recognition. And that exact limitation brings us to the core text of our deep dive today. The DeepMind paper. Yes. In May 2026, Google DeepMind published a paper in the journal Nature titled Co-Scientist. a multi-agent AI partner to accelerate research. What DeepMind has built here is a monumental leap beyond a search engine, and it is a massive departure from the standard AI chatbots people are used to. Right, because with a standard LLM, you type in a prompt like "cure cancer," And it just spits out a statistical prediction of the most likely next word based on its training data. It just gives you a superficial summary. Exactly what CoScientist does not do. CoScientist is built on the Gemini model, but it is structured as a collaborative coalition of specialized AI agents. The coalition. Right. It is a multi-agent framework. These agents don't just answer questions. They are managed by an adaptive supervisor agent that operates non-linearly. Non-linearly. What does that mean in practice? So when the human scientist gives co-scientist a broad research goal like, say, find a novel pathway to reverse cellular aging, the supervisor doesn't just write a long essay. It breaks the massive problem down into distinct, manageable tasks. Right. It spins up different AI agents with distinct personalities and instructions, runs them all in parallel, and coordinates their interactions. You know, I was reading the architecture diagram in the paper, and it struck me that it operates almost exactly like an AI version of a television writer's room. Oh, I love that analogy. Yeah. You don't have one person sitting in a dark room writing the whole script from start to finish. You have a room full of specialists. You have the ideas... Person pitching wild, out-of-the-box concepts. You have the harsh, cynical critic shooting down the bad ideas. You have the meticulous copy editor fixing the structure. And you have the showrunner, which would be the supervisor agent, managing the chaos and keeping everyone on track. That analogy maps perfectly onto the architecture DeepMind created. They structure this virtual writer's room into three distinct, highly rigorous phases. Let's walk through the phases. Phase one. Phase one is the generate phase. The supervisor activates the generation agent. This agent uses its massive context window to read across vast swaths of biomedical literature and proposes novel hypotheses. to solve the human's problem. Okay, standard generation. But to ensure the AI doesn't just fixate on the most obvious, widely published theory, the supervisor also spins up a proximity agent. A proximity agent? What does that do? Its sole job is to map all the hypotheses the generation agent is creating in a latent space and mathematically force them apart. Oh, weird. Why force them apart? It clusters the ideas and basically says, these five hypotheses are too similar to each other, generate something completely different, It forces a wide, highly diverse exploration of the scientific landscape, preventing the system from getting stuck in a narrow rabbit hole. That's smart. So phase one is about casting a massive, diverse net. But an AI generating a thousand wild biological ideas isn't science yet. It's just brainstorming. Right. It's just throwing spaghetti at the wall. Which brings us to phase two. debate. And reading the DeepMind paper, this is where the system gets incredibly intense. Phase two is the crucible. This is where hypotheses are stress tested to the breaking point. The supervisor brings in the reflection agent. You can think of the reflection agent as a ruthless, highly skeptical, virtual peer reviewer. The cynic in the writer's room. Exactly. Its entire prompt instruction is to find flaws. It tears the generated hypotheses apart, checking for factual inaccuracies, logical leaps, or contradictions with known biological mechanisms. But the true genius of phase two is the ranking agent, which orchestrates what DeepMind calls a "tournament of ideas." The Tournament of Ideas. Explain the mechanics of this, because it borrows heavily from DeepMind's previous very famous AI models, doesn't it? It does. A few years ago, DeepMind shocked the world with AlphaGo, an AI that mastered the ancient board game Go, and AlphaStar, which mastered the highly complex video game StarCraft. Right, the gaming AIs. Both of those AIs achieved superhuman performance, not by being fed human rules, but through self-play. They played millions of games against themselves, constantly learning from their own mistakes. DeepMind took that exact principle of adversarial self-play and applied it to scientific debate. So instead of playing a board game, the AI agents are playing a game of scientific argumentation. Precisely. The ranking agent takes two competing hypotheses. Let's call them hypothesis A and hypothesis B. It pits them against each other in a pairwise comparison. The AI agents debate the merits, the evidence, and the physical viability of each idea. How do they keep score in an argument, though? This system uses an ELO rating system. This is the exact same mathematical formula used to rank human chess grandmasters. If hypothesis A defeats hypothesis B, A's rating goes up and B's rating goes down. The system runs thousands of these micro-debates, bubbling the most robust, scientifically sound ideas to the top of the leaderboard. Okay, this sounds fascinating, but I have a massive problem with this. I am going to push back hard. Bring it on. We have all interacted with large language models. We all know they suffer from hallucinations. Yes, they do. An AI can be incredibly eloquent, highly persuasive, and confidently state things that are completely factually wrong. If an AI hallucinates a fake chemical reaction or invents a fake protein structure and confidently tells a human scientist to go spend $100,000 in six months testing it in a physical lab, That is a catastrophic failure. Absolutely catastrophic. So how do you prevent a highly persuasive AI from winning the tournament of ideas with a brilliant sounding lie? You have hit on the single biggest danger of applying generative AI to the physical sciences. DeepMind was acutely aware of the hallucination problem. A language model, by itself, is just predicting text, right? It doesn't actually know physics. That is why the vast majority of co-scientists' computational power is not spent on generating ideas. It is spent on verifying them against external reality. External reality, like how? During that tournament of ideas, the reflection agent isn't just debating using its internal text weights. It is actively reaching out to massive, established, human-curated databases like CHEMBL, which contains millions of bioactive molecules, and UNIPROT, which catalogs protein sequences. So it's fact-checking itself in real time against actual biological data. Constantly. It doesn't just ask, does this hypothesis sound plausible? It queries the databases to ask, does this specific molecule actually exist? Does it violate the laws of thermodynamics? Does this contradict known protein folding structure? It grounds the textual hypotheses in rigid mathematical physical reality before those ideas are ever presented to a human. That is reassuring. Which finally brings us to phase three, evolve. Yes, phase three. Once the tournament of ideas has concluded, the evolution agent takes the surviving hypotheses from the top of the ELO leaderboard. It looks at the feedback from the reflection agent, and it refines the ideas, sometimes merging the strongest parts of two different hypotheses into a single, highly optimized theory. It takes the best of everything. Right. Finally, the meta-review agent synthesizes this entire chaotic debate into a clear, comprehensive, highly cited research proposal for the human scientist to review. you it is a breathtaking piece of software architecture a virtual writers room playing evolutionary chess with biological concepts fact-checking itself against the laws of physics it's beautiful on paper but biology is inherently messy it is wet it is unpredictable involves living breathing cells that frequently do things that look impossible on a computer screen oh absolutely a brilliant hypothesis that wins an ai chess tournament is completely useless if it fails the moment you put it in a petri dish Does co-scientists actually work in a real, physical wet lab? That is the ultimate litmus test. If the AI only works in silicon, it's just a toy. We have to move from the server farm to the cells. And the DeepMind paper doesn't just present a theoretical software framework, it backs it up with real-world validation. Oh, excellent. They gave co-scientists to leading human researchers and told them to use it on their hardest problems. What we are seeing in the paper is the dawn of the lab in the loop concept, sometimes called self-driving labs. Lab in the loop. I like that. This is where computational AI models and robotic automated wet labs form a continuous feedback loop. The AI designs the experiment, robotic pipettes execute it physically on living cells, the data is fed immediately back into the AI, and the AI iterates its hypothesis in real time. That sounds incredible. The paper highlights specific scientists who ran this loop, and the results are stunning. Let's talk about Gary Peltz at Stanford. Yeah, Gary Peltz. He studies liver fibrosis. And for context, liver fibrosis is severe, chronic scarring of the liver tissue. When the liver is constantly damaged like by alcohol or a virus specialized cells called hepatic stellate cells go into overdrive, they produce massive amounts of collagen that turn healthy tissue into rigid scar tissue. It is notoriously difficult to stop. It is. So Peltz brought this exact problem to co-scientists. He asked the AI to find a way to halt that specific scarring mechanism. The multi-agent system sifted through mountains of genomic data and literature, ran its tournament, and highlighted an overlooked existing drug as a prime repurposing candidate. Repurposing strikes again. Exactly. It hypothesized that this drug would interrupt the signaling pathway of those stellate cells. When Peltz took that AI-recommended drug into his physical lab and tested it on actual liver cells, the drug blocked 91% of the scarring-linked response. 91% that's almost a total block. It is an astronomical success rate for a first pass experiment. That 91% blockage points toward an entirely new gene regulating approach to treating chronic liver disease. And the connection was found by a machine. Unbelievable. Then there is the story of Ritu Raman at MIT. She is tackling ALS Lou Gehrig's disease, which is a brutal neurodegenerative disease that destroys motor neurons. Right. What I found so fascinating about her experience with co-scientists is the way she described the interaction. She didn't use the AI like a glorified calculator. She said, "Science is a team sport." The AI acted as a true intellectual collaborator. Raman noted that the AI helped her structure her own chaotic thoughts and digest literature far outside her immediate subfield. But the most impressive moment was when the AI actually spotted a structural weakness in her hypothesis. Oh, he critiqued her. Yes. It realized the experiment involved complex RNA dynamics that neither Raman nor the AI could fully resolve alone. The AI explicitly prompted her to pull in an outside human expert, an RNA specialist named Ryan Flynn, to complete the puzzle. Wow. So the AI didn't just solve the problem. It facilitated human to human collaboration by identifying a knowledge gap. Exactly. It acts as an intellectual bridge. And then we have Omar Abudaya, who is working on something that sounds like pure magic. Reversing cellular aging. Reversing aging. Okay. He is researching epigenetic reprogramming, looking for ways to reset the cellular clock. He used co-scientists to synthesize decades of fragmented literature to find novel genetic targets. And he explicitly stated that using co-scientists felt like having a team of 50 people at his disposal. A team of 50 people. It slashed their data analysis time from months to a matter of days. And the genetic leads the AI identified actually showed signs of rejuvenating cells when tested in the physical lab. These human stories really prove the multi-agent framework works. But we cannot fully grasp DeepMind's wet lab success without bringing in their other major breakthrough, which works in tandem with these systems. Yes. Alpha Proteo. Alpha Proteo. This is the sister technology to their famous AlphaFold model, correct? Yes, it is. AlphaFold was a revolution because it could accurately predict the 3D structure of almost any existing protein based on its amino acid sequence. But Alpha Proteo takes a massive leap forward. Alpha proteo doesn't just predict what already exists. It designs novel protein binders from scratch. A protein binder? Yeah. Meaning a molecule designed to physically stick to a specific target, like putting a cap on a pen. Exactly. If you have a virus, like the SARS-CoV-2 virus, it uses a spike protein to break into human cells. If you can design a protein binder that perfectly fits over that spike, like a glove fitting over a hand, you neutralize the virus. And Alpha Proteo designs these gloves from scratch. From scratch. And the physical lab results were off the charts. They tasked Alpha Proteo with designing binders for several viral proteins. When they took the AI's digital designs and physically synthesized them in the lab, they achieved experimental success rates between 9% and 88%. Wait, to really understand how wild an 88% success rate is, we have to look at the baseline, right? We do. Traditional human-led methods of protein design require massive, agonizing trial and error. You mutate a protein slightly, test it, fail, mutate it again. If a human team achieves a 1% or 2% success rate in the lab, popping champagne is completely justified. 1% or 2%. So hitting 88% on the first try. It means the AI isn't just guessing. It means the AI understands the physical reality of molecular binding, the Van der Waals forces, the hydrogen bonds, at a fundamental mathematical level. It has successfully bridged the gap from digital simulation to physical biology. Which is just staggering. It really is. If we extrapolate the convergence of co-scientists' reasoning, Alpha Proteo's physical design, and automated robotic labs, we are moving toward a world governed by digital tornadoes. Digital twins? Define that for us. Imagine having a perfect, mathematically accurate digital replica of a human liver cell, or even an entire organ running on a massive server farm. Before you ever spend a single dollar buying physical chemicals, or subject a single animal or human to a physical trial, you can unleash co-scientists to run a million simulated experiments on that digital twin. Wow. You can compress decades of evolutionary trial and error into an afternoon of computing. It completely removes the physical constraints of time and resources from the early stages of biological discovery. Okay, I have to pause the celebration here. Uh-oh. Because I am listening to this and I'm putting the pieces together. We have AI agents debating each other at light speed. We have wet lab results blocking 91% of liver scarring. We have AI designing custom proteins with an 88% success rate. We have huge cost savings in compressed timelines. It sounds like a utopia. Right. Why isn't my local pharmacy completely stocked with AI-designed miracle drugs right now? So... Why hasn't E-Room's law completely collapsed already? What are the roadblocks? Because in science, there is always a catch. There is a profound catch. It is a double-barreled roadblock, and it is holding the entire tech biorevolution back. the roadblocks are both clinical and legal let's start with the clinical reality check because we really have to talk about failure okay earlier we talked about in silicone medicine but let's look at another pioneer in AI drug design Xantia a few years ago Xan she had designed a drug called DSP Lum 181 and intended to treat obsessive-compulsive disorder, or OCD. It targeted a specific receptor in the brain. And when they announced it, it was heralded as a massive historic breakthrough. It was the first AI-designed drug to ever enter human clinical trials. And the AI did its job fast. The exploratory design phase took less than 12 months. It was a massive win for the speed of AI generation, for sure. But what happened next is rarely put in the press releases. The drug entered phase-I clinical trials in humans, and it quietly failed. It was discontinued. Wait, why did it fail? If the AI designed the molecule perfectly to fit the brain receptor, what went wrong? It failed because of the vast canyon between the chemistry problem and the biology problem. AI right now has largely solved the chemistry problem. An AI can design a beautiful, stable molecule that binds incredibly tightly to a specific receptor in a plastic Petri dish. Right. But the biology problem is infinitely more complex. We touched on this earlier. A human body is a chaotic environment. In pharmacology, we talk about ADME absorption, distribution, metabolism, and excretion. Yeah. Right. A drug might bind perfectly to the OCD receptor in the brain, but maybe the molecule is too large to cross the blood-brain barrier in a living human. Or maybe it binds to the brain receptor, but its shape is just similar enough to a heart receptor that it accidentally triggers cardiac arrhythmia. Oh, I see. So to use an analogy, the AI is a master locksmith. It can mathematically design a perfect key for any lock you show it, but making a perfect key doesn't matter if you don't fully understand what happens inside the house when you unlock that specific door. Exactly the problem. Speeding up the discovery of a molecule does not guarantee that the molecule will actually heal a complex dynamic human being without causing collateral damage. That makes total sense. And even if you do manage to solve both the chemistry problem and the biology problem, especially in the realm of drug repurposing, you run headfirst into a massive infuriating legal maze. You have to survive the intellectual property gauntlet. Ah, the IP paradox. Yeah. This is where the economics of repurposing get really strange. Let's say I use co-scientist. I run a massive simulation and I discovered that a cheap generic arthritis drug that costs pennies to make actually cures a rare devastating brain disease. Okay, hypothetically. I have solved the biology problem. Yeah. But how do I patent that? It's an old drug. The patent expired 20 years ago. Anyone can manufacture it. You have just hit on the central commercial bottleneck of the entire repurposing industry. In the pharmaceutical world, without a patent, you have no market exclusivity. And without exclusivity. Without exclusivity, no venture capitalist or pharmaceutical giant is going to give you the $50 million required to run the massive phase two and phase three products. human clinical trials needed to prove it works for the brain disease. But wait, doesn't the FDA have a fast track for this? The FDA does have a pathway to accelerate repurposed drugs. It's called the 505B2 regulatory pathway. It is designed to let you skip the phase-i-safe trials since the generic drug is already proven safe in humans. Oh, that's helpful. It is. But getting regulatory approval from the FDA is a completely different universe from getting a legal patent from the U.S. Patent and Trademark Office. Because basic patent law requires an invention to be novel and non-obvious. But the physical molecule already exists. It isn't novel. Exactly. Since you can't patent the physical molecule itself, your only option is to file a method of use patent. You are trying to patent the specific concept of using drug X to treat disease Y. But the patent examiners are notoriously strict. They will often reject these applications by claiming the discovery is obvious. Obvious. How is it obvious if an AI had to find it? Well, they'll look at the literature and say, well, given the known anti-inflammatory properties of this arthritis drug, it was obvious to a person of ordinary skill in the art to try it on this inflammatory brain disease. That is infuriating. It is. To beat that rejection and secure the patent, you have to prove unexpected results. What does unexpected results look like in a biology lab? How do you prove you surprise the scientific community? You can't just show that the drug works. You have to show that it works in a way that defies conventional understanding. A classic way to prove this is synergy. You combine the old generic arthritis drug with the second generic drug, and you prove in the lab that the combination is not just additive, but synergistic. Right. 1 plus 1 equals 10. Exactly. It is 10 times more effective than the sum of its parts. Or you discover a completely novel mechanism of action. You prove that the arthritis drug cures the brain disease through a bizarre, hidden biological pathway that no scientists on Earth ever suspected the drug interacted with. Okay, let's play this out. Let's say I am a brilliant tech bio startup. I jump through all those flaming hoops. I find the unexpected synergy using my AI. I fight the patent office. I get the method of use patent. I raise the $50 million. I run the trials. And I bring my shiny, newly branded repurposed drug to market for the brain disease. Congratulations. Thank you. But what is stopping a doctor from just looking at my expensive new pill, realizing the active ingredient is literally just the cheap arthritis generic, and legally prescribing the cheap generic off-label to their patient to save them a fortune? Nothing. That is the nightmare scenario that keeps repurposing executives awake at night. Off-label prescribing can entirely erode a company's market share in a matter of months. And this threat is exactly why pharmaceutical companies build massive, impenetrable legal fortresses around their drugs, known as patent thickets. Patent thickets. The phrase itself sounds defensive and deliberately prickly. Oh, it is a highly aggressive legal strategy. A pharmaceutical company won't just patent the core drug. They will employ armies of lawyers to file dozens, sometimes hundreds, of secondary overlapping patents around every conceivable aspect of the drug. Like what? They patent the specific extended release coating on the pill. They patent the exact crystalline structure of the powder, called a polymorph. They patent the specific manufacturing process used in the factory. And they use something called Marcuse structures, right? Yes. Marcuse structures are the ultimate weapon in a patent thicket. A Marcuse claim is a complex legal and chemical description that doesn't just patent one single molecule. It effectively patents an entire family of millions of theoretical variations of that core molecule. Wow. It's like a chemical Mr. Potato Head. The patent claims the core structure and says, we also own the rights to this molecule if you attach a chlorine atom here or a fluorine atom there or a carbon chain here. They build a fortress of patents so dense, so mathematically massive, that a generic competitor cannot possibly manufacture a similar drug without infringing on something in the thicket. So big pharma uses these sprawling thickets to keep generics out, maintain their monopolies, and keep prices artificially high. I mean, it seems incredibly anti-competitive. It is highly controversial and it is a major driver of health care costs. But here is the fascinating, almost ironic twist for our story. today. Okay, I love a twist. These defensive patent thickets built to lock down knowledge are actually becoming literal treasure maps for multi-agent AI systems like co-scientists. Wait, how does an AI turn a legal fortress into a treasure map? Well, to get all those hundreds of secondary patents and Marcush claims approved by the patent office, the pharmaceutical company legally has to disclose their data. They have to provide physical experimental examples to prove to the patent examiner that their broad claims actually work. OK. Often they will publish data on hundreds of chemical variations of their drug. Now, many of those variations might have had weak, unhelpful activity for the specific disease the company cared about. Or they might have accidentally shown activity against a completely different secondary target that the company ignored. Because they were focused on just getting the patent for their main drug. Exactly. Now, to the human eye, reading these patents is a nightmare. It is thousands of pages of failed or suboptimal experiments buried in dense legal jargon. But to a system like CoScientist, which has a million token context window and can read 50 massive patents simultaneously. To the AI, it is an absolute goldmine. The AI reads those failed experimental examples hidden in the fine print. It pulls that data into its knowledge graph, connects it to other biological data across the world, and realizes, wait a minute. This specific chemical variation that the original company abandoned in 2012 because was a terrible arthritis drug. Its 3D shape is actually the perfect binder to inhibit the specific cancer pathway. Oh wow. The massive defensive wall built by the original company literally becomes the open blueprint for the AI to discover entirely new repurposing opportunities. That is incredible. It's absolute poetic justice. The very legal mechanism used to hoard and lock down medical knowledge becomes the data set that sets new discoveries free. It's beautiful. So let's step back and look at the massive journey we have taken today. We started with the blind, almost comical serendipity of discovering Viagra and Rogaine. We moved to Don Swanson, manually cross-referencing paper index cards in a Chicago library. We watched the rise of massive computational knowledge graphs and jams imagining new molecules. It's been quite a ride. And now we have arrived at DeepMind's co-scientist. Right. Where does this leave the human being? Are researchers about to be entirely outsourced to algorithms? No, and that is perhaps the most critical point to take away from the DeepMind paper. The integration of AI and biopharma is not about replacing human scientists. Right. It is about fundamentally shifting the bottleneck of discovery. Shifting the bottleneck. Yeah, yeah. For the last 20 years, the bottleneck was data processing. Human brains simply could not read, retain, and connect enough information. AIs decisively solved the data processing problem. Now the bottleneck shifts upward to high-level strategic problem solving. We are entering the era of the augmented human. The augmented human. I really like that framing. It means the definition of what makes a scientist an expert completely changes. It isn't about rote memorization of cellular pathways anymore. It isn't about spending 10,000 hours pipetting clear liquids into plastic tubes in a basement lab. Precisely. If an AI system like CoScientist can read every biomedical paper ever published, synthesize the data, and debate thousands of hypotheses in a single afternoon, the human's value is no longer in knowing all the answers. The human's value lies entirely in asking the right questions. It is in providing the moral, ethical, and strategic oversight. The AI is the ultimate co-pilot, doing the heavy computational lifting of data synthesis. But the human remains the pilot, determining the destination. And the destination matters deeply because this technology is ultimately about addressing those neglected rare diseases that the old economically broken model left behind. So to bring this all home, give you a final thought to mull over as we wrap up. Think about the concept of end of one medicine. Oh, that's the dream. Right now, because of Eroom's law and the $2.6 billion price tag, drugs have to be mass produced for millions of people just to make a profit. But in the near future, powered by TechBio, if you get sick with a complex, mysterious, or incredibly rare disease, your doctor might not reach for a mass-market blockbuster drug off the shelf. No, they won't have to. Instead, the hospital might sequence your specific individual genome... They feed your unique biological profile, along with the specifics of your disease, into a multi-agent AI-like co-scientists. The AI runs a virtual tournament of ideas against the millions of existing already approved compounds. Exactly. It simulates how your specific cells will react to different chemical combinations. And overnight while you sleep, it discovers a custom repurposed treatment combination designed exclusively for you. A completely bespoke cure found in a matter of hours using cheap drugs that already exist. It brings us right back to that wooden chair we talked about at the very beginning. For decades, E-Room's law meant the pharmaceutical chair was getting harder, slower, and exponentially more expensive to build, leaving millions of patients without a seat. But by systematizing serendipity… Yes, by teaching machines to debate each other and by augmenting human brilliance with multi-agent AI, we aren't just figuring out how to build the chair faster. We are realizing we already have a warehouse full of furniture. And we finally have the flashlight to see it in the dark.