AlphaFold at Five: A Decade of AI Transforming Science

AlphaFold, the artificial intelligence system developed by Google DeepMind, marks its fifth anniversary with a legacy defined by transformative breakthroughs in molecular biology. Since its debut in November 2020, the system has not only redefined protein structure prediction but also catalyzed a scientific revolution: in 2023, it was awarded the Nobel Prize in Chemistry for its foundational role in advancing understanding of biological macromolecules.

Key Milestones

• AlphaFold2 (2020): This iteration revolutionized protein folding by predicting 3D protein structures with atomic precision, culminating in a database of over 200 million predicted structures—encompassing nearly the entire known protein universe. Used by 3.5 million researchers across 190 countries, it became a cornerstone of global scientific collaboration.

• Citation Impact: The 2021 Nature article detailing AlphaFold2’s algorithm has been cited over 40,000 times, underscoring its immediate and enduring influence.

• AlphaFold3 (2023): Building on this success, AlphaFold3 extended capabilities to predict interactions involving DNA, RNA, and small molecules (e.g., drugs), though this expansion introduced challenges like "structural hallucinations" in disordered protein regions—highlighting the need for rigorous validation.

Interview with Pushmeet Kohli: AI as a Catalyst for Scientific Discovery

Q: Transitioning from Games to Protein Folding—Role and Vision

WIRED: AlphaFold’s evolution from mastering Go to solving protein folding represents a paradigm shift in AI’s role in science. How did this transition unfold, and what guided your strategic focus?

Pushmeet Kohli (DeepMind VP of Research): At DeepMind, our mission has always centered on science. Games like Go were critical testing grounds to develop AI techniques—neural networks, planning, and search—that could later tackle high-stakes problems. Protein folding was such a problem: it demanded computational breakthroughs and held the promise of revolutionizing medicine and biology.

I defined my role as identifying "root node problems"—scientific challenges where conventional methods would fail to deliver transformative progress within a decade. Protein folding was precisely such a case: its solution would unlock new avenues in drug discovery, disease mechanisms, and synthetic biology.

AlphaGo’s success validated that AI could master complexity; protein folding shared this complexity but with urgency: solving it meant improving human health.

Q: The "Harness" Architecture—Balancing Creativity and Rigor

WIRED: AlphaFold3’s adoption of diffusion models introduces risks of "structural hallucinations." How has your "harness" architecture (creative generation + verification) evolved to address this?

Pushmeet Kohli: The core principle—pairing generative creativity with rigorous validation—remains unchanged. What has evolved is how we apply this to increasingly ambitious systems.

Diffusion models in AlphaFold3 were adopted because we needed to model interactions between DNA, RNA, proteins, and small molecules—not just individual protein structures. To mitigate hallucinations, we integrated confidence scores that flag unreliable predictions, critical for disordered proteins. Crucially, five years of lab validation by scientists have established trust: AlphaFold’s predictions reliably align with experimental data.

Q: AI Co-Scientist—Redefining Research Collaboration

WIRED: The "AI co-scientist" (an agentic system on Gemini 2.0) aims to generate and debate hypotheses. Could this shift diminish human agency, with AI replacing Principal Investigators?

Pushmeet Kohli: No—AI will augment, not replace, scientists. The dual role of scientists—framing questions ("what to study") and executing methods ("how to study")—will evolve: AI automates "how," freeing researchers to focus on "what."

The "AI co-scientist" acts as a virtual collaborator: it identifies research gaps (e.g., viral hijacking of bacteria by Imperial College researchers), generates hypotheses, and proposes experiments. Humans then validate these hypotheses and interpret findings for clinical relevance.

Q: The Holy Grail—Simulating a Human Cell

WIRED: What unsolved problem keeps you focused, and how will AI tools bridge simulation to real-world impact?

Pushmeet Kohli: Simulating a complete human cell is the ultimate goal. Cells are integrated systems: DNA (the "recipe") dictates protein (the "ingredient") synthesis, which drives cellular function. To model this, we first decode the nucleus—the cell’s command center—studying how genetic code is transcribed, signaling molecules are produced, and proteins are assembled.

Once achieved, this would transform medicine: computationally test drugs, decode disease mechanisms, and design personalized therapies. This bridges AI simulation and clinical reality, delivering on the promise of predictive biology.

Conclusion

AlphaFold’s five-year journey embodies AI’s potential to rewire scientific discovery. As Kohli’s insights highlight, the future lies not in replacing human curiosity but in empowering it—turning AI from a tool into a collaborative co-scientist.

This story originally appeared in WIRED Italia and has been translated from Italian.