AlphaFold 3 has cracked one of biology's hardest problems: predicting how proteins fold into their functional shapes with near-experimental precision. This breakthrough is now powering the first AI-designed cancer drugs entering human clinical trials, marking a historic shift from theoretical research to real-world medicine. What was once a decades-long puzzle is now being solved in days, fundamentally reshaping how new treatments are discovered and developed.

What Is the Protein-Folding Problem, and Why Did It Stump Scientists for 50 Years?

Every cell in your body contains proteins, which are tiny molecular machines made from chains of amino acids. The shape these chains fold into determines what the protein does and whether it functions correctly. A misfolded protein can cause serious disease, including Alzheimer's, Parkinson's, type 2 diabetes, and cystic fibrosis.

For decades, scientists needed to know a protein's precise three-dimensional structure to design drugs that could interact with it. But determining that structure was brutally slow. Researchers used X-ray crystallography, a technique that involved growing protein crystals and bombarding them with X-rays to create diffraction patterns. By the 1990s, this process took three to five years per protein. By the 2010s, despite all that effort, scientists had experimentally determined only about 170,000 protein structures. Yet there are an estimated 200 million distinct proteins in nature.

The math was sobering. A simple peptide chain with just 35 molecules could fold in up to 3.4 x 10^45 different ways, each theoretically possible but most biologically useless. Many experts argued that even with millions of years of computing power, machines could never solve this problem.

How Did AlphaFold Change Everything?

In November 2020, DeepMind's AlphaFold 2 entered CASP (Critical Assessment of Protein Structure Prediction), a biennial scientific competition that has been running since 1994. Researchers submit their best attempts to predict protein structures, and results are compared against experimentally verified answers. AlphaFold 2 didn't just win; it obliterated the competition.

The system predicted protein structures with accuracy that rivaled and often matched experimental determination. Scientists who had spent their careers wrestling with this problem used language rarely seen in academic papers. John Moult, the founder of CASP, called it "a stunning advance." Andriy Kryshtafovych, one of the assessors, described it as "a once in a generation advance." Eric Topol, one of the most cited medical researchers in the world, simply called it "a scientific earthquake".

The recognition was profound. Demis Hassabis and John Jumper, the two DeepMind researchers most responsible for AlphaFold's development, were awarded the Nobel Prize in Chemistry in 2024. By 2024, the AlphaFold Protein Structure Database had grown to contain over 214 million predicted protein structures, covering nearly every catalogued protein known to science. In three years, AI had mapped more biological territory than humanity had managed in the previous five decades.

AlphaFold 3, the latest version, went further. It can now predict not just protein structures but how proteins interact with drugs, DNA, RNA, and other molecules simultaneously.

How Is AlphaFold Being Used to Design New Drugs?

Knowing a protein's shape is enormously useful. But designing a molecule that fits perfectly into that shape, like a key designed for a lock that has only just been revealed, is far harder. That challenge is what Isomorphic Labs, DeepMind's commercial spin-off founded in 2021 and based in King's Cross, London, was built to solve.

Isomorphic's stated ambition is to reimagine the entire drug discovery process from first principles, not just to accelerate the existing process but to replace it with something fundamentally different. The traditional drug discovery pipeline is nightmarishly slow. From identifying a biological target to getting an approved medicine typically takes 10 to 15 years and costs between one and three billion pounds per successful drug. For every drug that makes it to market, thousands of candidates are abandoned.

Isomorphic's drug design engine, called IsoDDE (Isomorphic Drug Design Engine), announced in February 2026, generates candidate molecules with what researchers describe as near-perfect binding accuracy. It outperforms every publicly available tool by a wide margin. In April 2025, Isomorphic Labs raised 600 million pounds in its first external funding round, led by Thrive Capital, and announced partnerships with pharmaceutical giants Eli Lilly, Novartis, and Johnson & Johnson.

"There are people sitting in our office in King's Cross, London, working and collaborating with AI to design drugs for cancer. That's happening right now," said Colin Murdoch, the company's president.

Colin Murdoch, President at Isomorphic Labs

What Does AI-Driven Drug Discovery Actually Look Like in Practice?

AI-driven drug discovery works fundamentally differently from traditional methods. Instead of the slow, sequential process of target identification, then screening, then synthesis, then testing, AI allows many of these steps to run in parallel. Machine learning models trained on prior experimental data help scientists dodge dead ends early or shortcut them sooner. The result is that fewer resources get spent on candidates heading for failure.

The process includes several key stages:

• Target Identification: AI models scan genomics, proteomics, and clinical datasets to spot which biological targets like genes, proteins, or pathways are connected to a disease, a task that used to take researchers months but now takes days.
• Molecule Generation: Generative AI models can design entirely new molecular structures with desired properties instead of just screening compounds already on hand, dramatically widening the chemical universe researchers can explore.
• Lead Optimization: Once a promising molecule is identified, AI helps refine it by making it more potent, more stable, and more soluble while reducing unwanted side effects, using iterative predictive modeling instead of repeated physical synthesis.
• Toxicity Prediction: Machine learning models trained on historical toxicity records can catch potentially harmful compounds early, before they reach animal or human trials, which is one of the main reasons for the 70 percent reduction in preclinical costs.
• Clinical Trial Optimization: Predictive models help identify ideal trial candidates, forecast enrollment timelines, and catch early safety signals before delays occur, leading to more reliable overall success rates.

These steps reflect a broader automation principle: using intelligent systems to filter out low-value work early so human experts focus only on the highest-potential opportunities.

What Are the Real-World Results So Far?

The economic impact is staggering. Bringing a new drug to market traditionally costs over 2.6 billion dollars and takes 10 to 15 years. Companies using AI for preclinical research are reporting upwards of a 70 percent decrease in preclinical expenditures, according to industry reports analyzing AI-native biotech firms.

But the most significant milestone is clinical. In January 2025, DeepMind CEO Demis Hassabis announced from the World Economic Forum in Davos that the first AI-designed drugs, developed by Isomorphic Labs, would enter Phase 1 clinical trials by the end of the year. By early 2026, that plan had advanced. Isomorphic Labs' AI-designed cancer drug candidates were progressing toward Phase 1 trials, with the first expected to begin in early 2026.

"AI applied to science is a lot richer than just the language models. We and others are working on trying to design drugs with AI and with our spin-out company Isomorphic. I think we will hopefully have some AI-designed drugs in clinical trials by the end of the year," said Demis Hassabis, CEO of DeepMind.

Demis Hassabis, CEO at DeepMind

This represents a historic transition. These are not drugs that AI helped to identify. These are drugs that AI designed from scratch, based on computational predictions of how molecules would interact with disease targets. Human trials will now test whether those predictions hold up in real patients.

How Did AlphaFold Get Built in the First Place?

AlphaFold's development began in the 2010s as AI deep learning techniques improved. The biggest bottleneck in deep learning systems is typically a lack of data, but DeepMind had access to the Protein Data Bank (PDB), a globally connected database of all known proteins and their verified structures. Using this data, DeepMind trained models to look at a simple two-dimensional nucleotide sequence and infer how DNA would be translated into three-dimensional geometry and function.

The initial inspiration came from Foldit, a game where players try to assemble the most accurate protein model possible. Demis Hassabis wanted to see if his team could train a deep learning system to play the game even better than humans could. To increase the model's effectiveness, the team also created a separate algorithm trained specifically on amino acid grouping dynamics and the ways they attract and bond. By training a system to understand the fundamental physical dynamics that make peptide bonds stable, they eliminated many mistakes that an AI could make by not properly accounting for those dynamics.

The result was a system that could take a simple text input and generate a fully interactive three-dimensional protein model, a capability that had seemed impossible just years before.

What Happens Next?

The field is moving fast. Demis Hassabis described the current era as the dawn of a "Golden Age of scientific discovery," a period in which AI doesn't just accelerate research but fundamentally alters the economics and timeline of medical breakthroughs. The first human trials of AI-designed drugs will provide the ultimate test: whether computational predictions translate into safe and effective treatments.

Demis Hassabis

If they do, the implications are profound. A process that has taken 10 to 15 years and billions of dollars could be compressed to years and millions. Diseases that have no treatments today might have candidates in clinical trials within a few years. The bottleneck that has limited drug discovery for decades is finally breaking open.