AlphaFoldの5年: タンパク質構造予測がもたらした影響 AlphaFold: Five years of impact

Five years after DeepMind first released AlphaFold, the company has taken stock of how the protein-structure prediction system has reshaped biology, and the numbers are striking. AlphaFold is now used by more than 2.5 million researchers across over 190 countries, and the freely accessible AlphaFold Protein Structure Database covers more than 200 million predicted structures — essentially the known protein universe.

Before AlphaFold, determining a single protein structure by X-ray crystallography or cryo-EM could take months or years and cost significant lab resources. By collapsing that bottleneck, AlphaFold has shifted structural biology from a craft practiced by specialists to an everyday tool available to virtually any life-science lab. The impact was formally recognised in 2024 when Demis Hassabis and John Jumper shared the Nobel Prize in Chemistry for their work on the system.

The range of downstream applications is broad. Researchers have used AlphaFold to study surface proteins of the malaria parasite as vaccine targets, to design enzymes that break down plastics more efficiently, to investigate misfolded proteins implicated in Parkinson's disease, to search for compounds active against antibiotic-resistant bacteria, and even to develop new strategies against agricultural pests. Universities are increasingly weaving AlphaFold into structural-biology curricula, lowering the barrier for students entering the field.

In 2024, DeepMind released AlphaFold 3, extending predictions beyond single proteins to complexes that include DNA, RNA, small-molecule ligands and post-translational modifications. That shift is particularly important for drug discovery, where the interaction between a target protein and a candidate molecule is the central question. Isomorphic Labs, the DeepMind spinout commercialising the technology, has signed partnerships with Novartis and Eli Lilly and has signalled ambitions to take its own programs into clinical trials.

AlphaFold does not exist in isolation. Meta's ESMFold offered a fast language-model-based alternative, while the Baker Lab at the University of Washington has produced RoseTTAFold and the diffusion-based de novo design tool RFdiffusion, which is now widely used to invent proteins that do not exist in nature. Together these systems form a rapidly maturing ecosystem in which prediction, generation and experimental validation feed each other.

Limitations remain. AlphaFold predicts static structures well but struggles with conformational dynamics, the effects of point mutations, intrinsically disordered regions and rare folds underrepresented in training data. Several benchmarking studies have suggested that confidence scores can be over-optimistic in such regimes, and experimental verification is still essential for high-stakes decisions like clinical drug candidates. The next phase of progress is likely to come from tighter integration with wet-lab pipelines, better modelling of dynamics and ensembles, and closer coupling with generative design tools.

Five years in, AlphaFold's most lasting contribution may be cultural as much as technical: it normalised the idea that a deep-learning model could become foundational scientific infrastructure, used daily and cited in tens of thousands of papers. Whether the same pattern repeats in genomics, chemistry or materials science is one of the open questions that the next five years are likely to answer.