AlphaFold2 at CASP14: protein folding meets deep learning (2020)
At the CASP14 structure-prediction meeting in November 2020, DeepMind’s AlphaFold2 reached accuracy competitive with slow laboratory methods, opening the path to the AlphaFold database and later Nobel recognition.
On 30 November 2020, organisers of CASP14—the community’s blind test of protein structure prediction (CASP)—announced results that stunned structural biologists. DeepMind’s AlphaFold2 system, led by John Jumper under Demis Hassabis, scored at or near laboratory quality on many targets. A problem that had resisted progress since the 1970s suddenly looked tractable.
What CASP measures
Proteins fold into three-dimensional shapes that largely determine what they do. Given only an amino-acid sequence, predict that shape. Every two years, CASP releases unpublished sequences; groups submit models; independent assessors compare predictions to structures determined experimentally but kept secret until scoring finishes.
AlphaFold had competed before, but version two combined deep learning, evolutionary sequence information, and geometric constraints in a way that closed much of the remaining gap. Median scores on hard targets moved from “useful hint” to “trust for many biological questions,” even before peer-reviewed papers and public code arrived in 2021.
From challenge to infrastructure
DeepMind framed the result as a grand challenge solved, not a one-off contest win. Within months it released predicted structures for large swaths of the known proteome, turning AlphaFold from a research prototype into shared infrastructure. Labs that once waited months for crystallography could start from a high-quality model overnight.
The same lineage later anchored the 2024 Nobel Prize in Chemistry shared by Hassabis, Jumper, and David Baker. CASP14 was the public moment when the committee’s later rationale became plausible: AI had crossed from “interesting benchmark” to “default starting point” in molecular biology.
Why it matters
AlphaFold2 at CASP14 showed that deep learning could deliver scientific instruments, not only better classifiers or chat interfaces. It also clarified division of labour: evolutionary couplings supply global constraints, neural networks refine local geometry, and open databases multiply impact.
For the history of AI, the episode sits beside AlphaGo as evidence that the same DeepMind toolkit—large models, heavy compute, careful evaluation—could rewrite an entire field. Protein folding was the clearest case yet that benchmarks, when passed decisively, change how other sciences work.