The Quiet Revolution: How AI Is Rewriting Drug Discovery

Most of the AI industry is arguing about chatbots. Meanwhile, the most consequential application of artificial intelligence is happening in biology labs where nobody on Twitter is paying attention. AI-driven drug discovery has moved from "promising research" to "drugs in human trials." Not in theory. Not in simulations. Real molecules, designed or discovered by AI systems, being injected into real patients in real clinical trials. And the results are starting to come in. This is the story of how AI is rewriting the fundamental economics of pharmaceutical development. It's quieter than the model wars. It's going to matter a lot more. ## The Scale of the Problem Developing a new drug is a nightmare by design. The average cost to bring a single new drug from concept to market is between $1.3 billion and $2.8 billion, depending on who's counting and what they include. The timeline averages 10-15 years. And the failure rate is staggering: roughly 90% of drugs that enter clinical trials never make it to market. Phase III trials — the most expensive stage — fail about 50% of the time. These numbers haven't improved meaningfully in decades. The pharmaceutical industry has a name for this: Eroom's Law (Moore's Law spelled backwards). Since the 1950s, the number of new drugs approved per billion dollars of R&D spending has roughly halved every nine years. We keep spending more and getting less. The bottleneck isn't biology. We understand disease pathways better than ever. The bottleneck is the search space. The universe of possible drug molecules is estimated at 10^60 — that's a number with 60 zeros. Finding the right molecule for the right target is, almost literally, finding a needle in a haystack the size of the observable universe. This is exactly the kind of problem AI was built to solve. ## AlphaFold: The Foundation The story starts with proteins, and the story of AI and proteins starts with AlphaFold. Proteins are the workhorses of biology. They catalyze reactions, transmit signals, provide structural support, and do most of the actual work in living cells. Understanding what a protein looks like — its three-dimensional structure — is essential for understanding what it does and how drugs can interact with it. For sixty years, determining protein structures required painstaking experimental work: X-ray crystallography, cryo-electron microscopy, nuclear magnetic resonance. Each structure took months or years and cost hundreds of thousands of dollars. By 2020, after six decades of effort, scientists had determined the structures of roughly 170,000 proteins. Then AlphaFold arrived. DeepMind's AlphaFold 2, in November 2020, scored above 90 on the Global Distance Test at the CASP14 competition — the biennial challenge that benchmarks protein structure prediction. The results were, in the words of the competition organizers, "astounding." AlphaFold could predict protein structures with accuracy comparable to experimental methods, in minutes instead of months. In July 2021, DeepMind released the AlphaFold database: predicted structures for over 200 million proteins, covering nearly every known protein across all species. The entire history of structural biology — 170,000 structures in 60 years — was dwarfed in a single release. Demis Hassabis and John Jumper won the 2024 Nobel Prize in Chemistry for AlphaFold. Deservedly so. It's one of the clearest examples of AI solving a real scientific problem that matters to human health. AlphaFold 3, announced in May 2024, went further. It doesn't just predict protein structures — it predicts how proteins interact with DNA, RNA, ligands, antibodies, and ions. For drug discovery, this is transformative. Drugs work by binding to proteins and changing their behavior. Predicting how a potential drug molecule interacts with its target protein is the core challenge of rational drug design. AlphaFold 3 made that prediction dramatically faster and cheaper. ## Isomorphic Labs: From Structure to Design Isomorphic Labs is where AlphaFold's research capabilities get turned into pharmaceutical products. Founded in 2021 by Demis Hassabis as an Alphabet subsidiary, Isomorphic Labs is explicitly designed to bridge the gap between DeepMind's AI research and real drug development. The company sits in London, with a second office in Lausanne, and draws directly on AlphaFold's technology. In January 2024, Isomorphic struck partnerships with two pharmaceutical giants: Novartis and Eli Lilly. These aren't exploratory research agreements. They're drug development deals, with Isomorphic using AI to identify drug candidates that the pharma companies will then develop through clinical trials. In April 2025, Isomorphic raised $600 million in its first external funding round, led by Thrive Capital. For a company that had operated entirely on Alphabet's money since founding, taking outside investment signals confidence in the commercial pipeline. Then, in February 2026, came the big one. Isomorphic announced its Drug Design Engine — a platform that doubles the performance of AlphaFold 3 on protein-ligand structure prediction. It predicts small molecule binding affinities with higher accuracy than physics-based methods at a tiny fraction of the time and cost. And it can identify previously unknown binding pockets on target proteins using only the amino acid sequence. That last capability is a big deal. Traditional drug design starts with a known binding site — a pocket on the protein surface where a drug molecule can latch on. If you don't know the pocket, you can't design the drug. Isomorphic's engine can find pockets that nobody knew existed, opening up targets that were previously considered "undruggable." ## Recursion Pharmaceuticals: The Data Machine If Isomorphic Labs represents the structure-first approach, Recursion Pharmaceuticals represents the data-first approach. And it's working. Based in Salt Lake City, Recursion has built what might be the largest proprietary biological dataset in the world. Their approach: systematically test how cells respond to different drugs, gene knockdowns, and perturbations, then use AI to find patterns in the resulting data. Their lab runs 24/7. Automated microscopy systems capture high-resolution images of cells under different conditions. Computer vision models analyze the images to quantify cellular responses. And machine learning models identify connections between diseases, targets, and compounds that humans would never spot in the raw data. Recursion went public in 2021 and has been pouring resources into its AI platform ever since. They've built custom foundation models trained on their proprietary data — models that understand cellular biology in the way that GPT understands language. The results are showing up where it matters: clinical trials. Recursion has multiple AI-discovered drug candidates in human testing. Their pipeline includes treatments for rare diseases, oncology, and inflammatory conditions. Several candidates have reached Phase II trials — meaning they've already demonstrated safety in humans and are being tested for efficacy. In 2024, Recursion partnered with NVIDIA to build BioHive-2, a supercomputer designed specifically for biological AI research. They've also struck deals with Roche and Bayer to apply their platform to the pharma giants' pipelines. When companies that spend billions on R&D annually start paying a startup for AI drug discovery capabilities, the technology has proven itself. ## Insilico Medicine: The Speed Record Insilico Medicine, founded in 2014 and headquartered in Hong Kong, holds a record that captures the promise of AI drug discovery more than any benchmark. In 2023, Insilico advanced a drug candidate from target discovery to Phase II clinical trials in approximately 30 months. Traditionally, that journey takes 5-7 years. They didn't just speed up one step — they used AI across the entire pipeline: identifying the target, designing the molecule, predicting its properties, and optimizing it for clinical testing. The drug, INS018_055, targets idiopathic pulmonary fibrosis — a serious lung disease with limited treatment options. The molecule itself was designed by Insilico's Chemistry42 platform, which uses generative AI to design novel molecules optimized for specific properties: binding affinity, selectivity, toxicity, metabolic stability. Insilico's approach is end-to-end AI. Their Pharma.AI platform covers target identification (PandaOmics), molecule generation (Chemistry42), and clinical trial prediction. The company claims to have spent under $3 million on the discovery and early development of INS018_055 — compared to the industry average of $500 million to reach the same stage. Whether the drug works is still an open question. Phase II trials are ongoing. But the speed and cost metrics are real, and they suggest that AI can collapse the timeline and economics of drug development by an order of magnitude. ## What the Skeptics Get Right It's worth noting what AI drug discovery hasn't done yet: it hasn't produced a blockbuster drug. No AI-discovered drug has completed Phase III trials and received regulatory approval. The drugs in the pipeline are promising, but the pharmaceutical industry is littered with promising drugs that failed in late-stage trials. The 50% Phase III failure rate applies to AI-discovered drugs too — we just don't have enough data yet to know whether AI improves those odds. There's also a selection bias problem. The drugs that make it to clinical trials are the ones the AI platforms identified as most likely to work. But the AI's confidence doesn't guarantee success in humans. Biology is messy, complicated, and full of surprises that no model — no matter how sophisticated — can fully anticipate. The CASP16 results from November 2024 provided a reality check for the structure-prediction approach. When tested on predicting protein-ligand interactions for pharmaceutical targets, AlphaFold 3-based models didn't significantly outperform older methods. The top performers actually used AlphaFold 2 with human visual inspection and manual adjustments. AI augmenting human experts beat AI alone. ## What the Skeptics Miss Despite the caveats, the trajectory is undeniable. The cost of generating drug candidates has dropped by 10-100x. The timeline from target to clinical candidate has compressed from years to months. The ability to explore chemical space has expanded by orders of magnitude. And the data keeps getting better — every trial, every experiment, every cellular image feeds back into models that become more accurate over time. Even if AI doesn't improve Phase III success rates — which I think it eventually will, as the models learn from clinical outcomes — the economics still transform the industry. If you can generate 10x more candidates at 10x lower cost, you can afford more shots on goal. The probability of any individual drug succeeding stays the same, but the probability of your portfolio producing winners goes up dramatically. The pharmaceutical industry spends roughly $250 billion annually on R&D globally. If AI cuts the cost of drug development in half — a conservative estimate based on current trajectory — that's $125 billion in freed resources that can be redirected to more programs, rarer diseases, and markets that traditional pharma economics made unviable. ## Where This Goes The next five years will be definitive. If several AI-discovered drugs complete Phase III trials and reach patients, the floodgates open. Every pharmaceutical company in the world will restructure around AI-driven discovery. The ones that already have — Recursion, Isomorphic Labs, Insilico — will have a head start measured in billions of data points. If the drugs fail — if Phase III success rates for AI-discovered candidates turn out no better than traditional methods — the industry recalibrates but doesn't abandon AI. The cost and speed advantages are real regardless of success rates. AI won't replace drug discovery. But it's already transforming it. The Nobel Prize for AlphaFold wasn't just recognition of a technical achievement. It was recognition that AI has crossed a threshold in biological science. The drugs in the pipeline are the first evidence of what comes next. Pay attention. This is the part of the AI revolution that actually saves lives.