主要内容
AlphaFold, the AI system developed by Google DeepMind, marks its fifth anniversary this year. Over the past five years, it has revolutionized biology, with AlphaFold 2 (2021) becoming a milestone: it predicted protein 3D structures with atomic accuracy, compiling a database of over 200 million structures—nearly all known proteins. Used by 3.5 million researchers in 190 countries, its 2021 Nature paper has been cited 40,000 times.
AlphaFold 3 (2023) extended capabilities to DNA, RNA, and drug design, though challenges like "structural hallucinations" in disordered protein regions persist. This transition signals progress toward AI-driven scientific discovery.
In an interview with WIRED, DeepMind VP Pushmeet Kohli (architect of the AI for Science division) explained the shift from Go AI to protein folding. "Science is our core mission," he noted. "Games like Go tested AI techniques, which later tackled real-world problems like protein folding."
Kohli emphasized three future goals: building AI models that collaborate with scientists, making tools accessible globally, and simulating complete human cells. "We focus on 'root node problems' where AI can transform science," he added.
The next five years may see AlphaFold evolve into a true scientific partner, unlocking unprecedented biological insights.