This post was originally published on here
AlphaFold, the artificial intelligence system developed by Google DeepMind, has just turned five. Over the past few years, we’ve periodically reported on its successes; last year, it won the Nobel Prize in Chemistry.
Until AlphaFold’s debut in November 2020, DeepMind had been best known for teaching an artificial intelligence to beat human champions at the ancient game of Go. Then it started playing something more serious, aiming its deep learning algorithms at one of the most difficult problems in modern science: protein folding. The result was AlphaFold2, a system capable of predicting the three-dimensional shape of proteins with atomic accuracy.
Its work culminated in the compilation of a database that now contains over 200 million predicted structures, essentially the entire known protein universe, and is used by nearly 3.5 million researchers in 190 countries around the world The Nature article published in 2021 describing the algorithm has been cited 40,000 times to date. Last year, AlphaFold 3 arrived, extending the capabilities of artificial intelligence to DNA, RNA, and drugs. That transition is not without challenges—such as “structural hallucinations” in the disordered regions of proteins—but it marks a step toward the future.
To understand what the next five years holds for AlphaFold, WIRED spoke with Pushmeet Kohli, vice president of research at DeepMind and architect of the AI for Science division.
WIRED: Dr. Kohli, the arrival of AlphaFold 2 five years ago has been called “the iPhone moment” for biology. Tell us about the transition from challenges like the game of Go to a fundamental scientific problem like protein folding, and what was your role in this transition?
Pushmeet Kohli: Science has been central to our mission from day one. Demis Hassabis founded Google DeepMind on the idea that AI could be the best tool ever invented for accelerating scientific discovery. Games were always a testing ground, and a way to develop techniques we knew would eventually tackle real-world problems.
My role has really been about identifying and pursuing scientific problems where AI can make a transformative impact, outlining the key ingredients required to unlock progress, and bringing together a multidisciplinary team to work on these grand challenges. What AlphaGo proved was that neural networks combined with planning and search could master incredibly complex systems. Protein folding had those same characteristics. The crucial difference was that solving it would unlock discoveries across biology and medicine that could genuinely improve people’s lives.
We focus on what I call “root node problems,” areas where the scientific community agrees solutions would be transformative, but where conventional approaches won’t get us there in the next five to 10 years. Think of it like a tree of knowledge—if you solve these root problems, you unlock entire new branches of research. Protein folding was definitely one of those.
Looking ahead, I see three key areas of opportunity: building more powerful models that can truly reason and collaborate with scientists like a research partner, getting these tools into the hands of every scientist on the planet, and tackling even bolder ambitions, like creating the first accurate simulation of a complete human cell.
Let’s talk about hallucinations. You have repeatedly advocated the importance of a “harness” architecture, pairing a creative generative model with a rigorous verifier. How has this philosophy evolved from AlphaFold 2 to AlphaFold 3, specifically now that you are using diffusion models which are inherently more “imaginative” and prone to hallucination?
The core philosophy hasn’t changed—we still pair creative generation with rigorous verification. What’s evolved is how we apply that principle to more ambitious problems.
We’ve always been problem-first in our approach. We don’t look for places to slot in existing techniques; we understand the problem deeply, then build whatever’s needed to solve it. The shift to diffusion models in AlphaFold 3 came from what the science demanded: We needed to predict how proteins, DNA, RNA, and small molecules all interact together, not just individual protein structures.
You’re right to raise the hallucination concern with diffusion models being more generative. This is where verification becomes even more critical. We’ve built in confidence scores that signal when predictions might be less reliable, which is particularly important for intrinsically disordered proteins. But what really validates the approach is that over five years, scientists have tested AlphaFold predictions in their labs again and again. They trust it because it works in practice.
You are launching the “AI co-scientist,” an agentic system built on Gemini 2.0 that generates and debates hypotheses. This sounds like the scientific method in a box. Are we moving toward a future where the “Principal Investigator” of a lab is an AI, and humans are merely the technicians verifying its experiments?
What I see happening is a shift in how scientists spend their time. Scientists have always played dual roles—thinking about what problem needs solving, and then figuring out how to solve it. With AI helping more on the “how” part, scientists will have more freedom to focus on the “what,” or which questions are actually worth asking. AI can accelerate finding solutions, sometimes quite autonomously, but determining which problems deserve attention remains fundamentally human.
Co-scientist is designed with this partnership in mind. It’s a multi-agent system built with Gemini 2.0 that acts as a virtual collaborator: identifying research gaps, generating hypotheses, and suggesting experimental approaches. Recently, Imperial College researchers used it while studying how certain viruses hijack bacteria, which opened up new directions for tackling antimicrobial resistance. But the human scientists designed the validation experiments and grasped the significance for global health.
The critical thing is understanding these tools properly, both their strengths and their limitations. That understanding is what enables scientists to use them responsibly and effectively.
Can you share a concrete example—perhaps from your work on drug repurposing or bacterial evolution—where the AI agents disagreed, and that disagreement led to a better scientific outcome than a human working alone?
The way the system works is quite interesting. We have multiple Gemini models acting as different agents that generate ideas, then debate and critique each other’s hypotheses. The idea is that this internal back-and-forth, exploring different interpretations of the evidence, leads to more refined and creative research proposals.
For example, researchers at Imperial College were investigating how certain “pirate phages”—these fascinating viruses that hijack other viruses—manage to break into bacteria. Understanding these mechanisms could open up entirely new ways of tackling drug-resistant infections, which is obviously a huge global health challenge.
What Co-scientist brought to this work was the ability to rapidly analyze decades of published research and independently arrive at a hypothesis about bacterial gene transfer mechanisms that matched what the Imperial team had spent years developing and validating experimentally.
What we’re really seeing is that the system can dramatically compress the hypothesis generation phase—synthesizing vast amounts of literature quickly—whilst human researchers still design the experiments and understand what the findings actually mean for patients.
Looking ahead to the next five years, besides proteins and materials, what is the “unsolved problem” that keeps you up at night that these tools can help with?
What genuinely excites me is understanding how cells function as complete systems—and deciphering the genome is fundamental to that.
DNA is the recipe book of life, proteins are the ingredients. If we can truly understand what makes us different genetically and what happens when DNA changes, we unlock extraordinary new possibilities. Not just personalized medicine, but potentially designing new enzymes to tackle climate change and other applications that extend well beyond health care.
That said, simulating an entire cell is one of biology’s major goals, but it’s still some way off. As a first step, we need to understand the cell’s innermost structure, its nucleus: precisely when each part of the genetic code is read, how the signaling molecules are produced that ultimately lead to proteins being assembled. Once we’ve explored the nucleus, we can work our way from the inside out. We’re working toward that, but it will take several more years.
If we could reliably simulate cells, we could transform medicine and biology. We could test drug candidates computationally before synthesis, understand disease mechanisms at a fundamental level, and design personalised treatments. That’s really the bridge between biological simulation and clinical reality you’re asking about—moving from computational predictions to actual therapies that help patients.
This story originally appeared in WIRED Italia and has been translated from Italian.
