The Machine That Designs Medicines
How AI is compressing drug discovery from decades to years — and what happens when it gets it wrong
EMERGING SCIENCE & TECHNOLOGY BRIEF • ARTICLE 2
Topic: Health & Human Longevity | Domain: AI & Machine Learning | Status: Hot 🔥
A new drug typically takes 12 to 15 years to travel from a scientist’s hypothesis to a pharmacy shelf. Roughly nine out of ten candidates fail somewhere along the way. The cost of bringing a single approved medicine to market, when you account for all the failures, runs into millions and sometimes billons. For most of the history of modern medicine, this was simply the price of doing business.
Something is changing. In laboratories across the United States, the United Kingdom, and Europe, artificial intelligence systems are beginning to reroute that pipeline identifying promising molecular targets faster, predicting how drug candidates will behave in the body before a single test tube is involved, and generating entirely new molecular structures that no human chemist would have thought to try. Several of those AI-designed compounds are now being tested in human clinical trials. The results so far are preliminary. But they are real.
We are shifting from AI as a research assistant to AI as a co-inventor. The question is no longer whether this works it is how far it goes.
Why drug discovery is so hard
Before exploring what AI changes, it helps to understand why the traditional process is so brutally inefficient. The human body contains roughly 20,000 protein-coding genes, and disease often arises when one or more of those proteins misfolds, overactivates, or stops working altogether. Finding a small molecule, a drug, that can specifically bind to the right protein and correct its behaviour, without causing serious side effects elsewhere, is an exercise in navigating an almost incomprehensibly large chemical space.
Estimates suggest there are somewhere between 1023 and 1060 possible drug-like molecules. For context, the observable universe contains around 1080 atoms. Traditional pharmaceutical research has had to sample this space by hand: synthesising candidate compounds, testing them in cells, then in animals, then in humans, each step filtering out failures but also consuming years and vast sums. High-throughput screening helped. Combinatorial chemistry helped. But the fundamental bottleneck, the difficulty of predicting how a molecule will behave, remained.
What AI actually does differently
Modern AI approaches to drug discovery operate on several fronts simultaneously. The most widely discussed is structure prediction. In 2020, DeepMind’s AlphaFold2 cracked one of biology’s grand challenges: predicting the three-dimensional structure of a protein from its amino acid sequence alone. A problem that had resisted 50 years of effort was effectively solved in a single research cycle. AlphaFold’s database now covers virtually all known proteins, giving researchers a structural map that was simply not available before.
Knowing the shape of a target protein is the first step. The next is finding or designing a molecule that fits into it precisely like a key into a lock. This is where generative AI models come in. Systems like those developed by Insilico Medicine, Recursion Pharmaceuticals, and Isomorphic Labs (a spinout of DeepMind) can generate candidate molecules optimised against a set of desired properties: potency, selectivity, stability, and an acceptable safety profile. What used to take teams of medicinal chemists several years can now take months.
Beyond generation, machine learning models are also being applied to predict ADMET properties — absorption, distribution, metabolism, excretion, and toxicity — the suite of factors that determine whether a drug is safe and practical to administer. Predicting these properties before synthesis dramatically reduces the number of dead ends that make it into expensive laboratory and clinical work.
Where it stands right now
The clearest evidence that AI drug discovery has moved beyond proof-of-concept is the pipeline. As of early 2026, multiple AI-generated drug candidates are in Phase II clinical trials — the stage that tests efficacy and safety in larger patient groups. Insilico Medicine’s INS018_055, targeting a rare and fatal lung disease called idiopathic pulmonary fibrosis, is among the most closely watched. Recursion has built a platform that industrialises the discovery process, running millions of biological experiments and feeding the results back into its models. Eli Lilly has partnered with NVIDIA to apply AI to its own internal pipelines.
None of these drugs are approved yet. Phase II is not Phase III, and Phase III is not approval. But the compression of the early-stage pipeline from target identification through lead optimisation is already measurable. Insilico Medicine reported taking INS018_055 from target identification to preclinical candidate in roughly 18 months. The industry average for that same journey is closer to four to five years.
Key Players to Watch
Insilico Medicine: AI-designed drug in clinical trials for idiopathic pulmonary fibrosis (Phase II).
Recursion Pharmaceuticals: high-throughput biology platform; NYSE-listed, partnership with Roche.
Isomorphic Labs (DeepMind spinout): applying AlphaFold-era protein science to drug design.
Eli Lilly / NVIDIA: major pharma + compute partnership focused on AI-accelerated internal R&D.
BenevolentAI: UK-based, applying AI to target identification and rare diseases.
The questions nobody has answered yet
The excitement is warranted. So is the scepticism. Drug discovery has a long history of technologies that seemed transformative in the laboratory and struggled in the clinic. High-throughput screening was supposed to accelerate the industry in the 1990s; it generated enormous amounts of data but not a proportional surge in approvals. The failure rate of clinical trials remains stubbornly high around 90 percent, and most of those failures happen not because the wrong molecule was chosen, but because of unforeseen biological complexity in humans.
AI does not solve that problem. It helps earlier in the process, which is valuable. But it cannot predict human immune responses, rare adverse events that only emerge in large populations, or the complex social and economic factors that determine whether an approved drug actually reaches patients. There is also a harder question about what happens when an AI-designed drug fails. Traditional drug development produces a rich record of experimental data even from failures, data that feeds the next attempt. AI pipelines that are more opaque may generate fewer such lessons.
Then there is access. If AI dramatically reduces the cost and time of drug discovery, who benefits? The conditions most likely to attract AI attention are those with large addressable markets — cancer, metabolic disease, neurodegeneration. Rare diseases in low-income countries, where there is limited financial return, may remain as neglected as ever, unless deliberate effort is made to direct the technology there.
A faster pipeline means little if what gets discovered reflects the same commercial priorities as before. The question of what gets discovered is as important as how quickly.
What comes next
The next few years will be a test of whether the clinical pipeline delivers. If several AI-designed drugs win regulatory approval in the late 2020s, the transformation of pharmaceutical R&D will be essentially confirmed. The major drug companies are already repositioning: almost every large pharma firm now has an AI strategy, internal capability, or external partnership. The question is shifting from whether to use AI to how to use it well.
A second wave of capability is coming from multimodal AI systems that can integrate genomics, proteomics, electronic health records, and imaging data simultaneously building a richer picture of disease biology than any single data type allows. Combined with the rapid decline in the cost of biological experiments, the toolset available to drug developers over the next decade will look genuinely different from what came before.
What is harder to predict is where the surprises will come from. The history of medicine is full of discoveries that upended confident assumptions. AI may accelerate the discovery of things we are already looking for. It may also stumble into things nobody was expecting.
You’re reading The Next Evolution by Neil Catton, articles that explore the human world and the intersection of technology, they try and ask difficult questions - not to scare - but to inform. This is part of the Emerging Science & Technology series.
If someone forwarded this to you, you can subscribe free at neilcatton.substack.com.
Neil Catton is the author of The Next Evolution, The Cognitive Crucible and The Shadow System - available on Amazon, and writes at the intersection of technology, ethics, and human purpose.