AlphaFold 2 is published in Nature and open-sourced

DeepMind published “Highly accurate protein structure prediction with AlphaFold” in Nature and released the system’s source code on GitHub under an open licence, alongside a second Nature paper giving predicted structures for the human proteome. This was the disclosure that followed the CASP14 result seven months earlier, in November and December 2020, when AlphaFold had produced structure predictions accurate enough that competition organisers described the fifty-year-old protein-folding problem as substantially solved. At the time of that result the method itself had not been published; the July 2021 papers and code release were what let outside laboratories actually use it.

The Nature paper described a deep learning architecture that predicted a protein’s three-dimensional structure directly from its amino-acid sequence, reaching accuracy comparable to experimental methods such as X-ray crystallography for a large share of targets, without requiring a database search for a structurally similar known protein. Alongside the code, DeepMind and the European Bioinformatics Institute launched the AlphaFold Protein Structure Database, publishing an initial set of over 350,000 predicted structures covering the human proteome and twenty other organisms important to research, with a stated plan to expand it to nearly every catalogued protein.

The open-source release, rather than the CASP result alone, was what let the method be verified, adapted and built upon outside DeepMind: it was cited more than 40,000 times in the years that followed, and became a standard tool in structural biology and drug discovery research. It stands as one of the more widely cited examples of a frontier AI lab publishing full working code for a system with obvious commercial value, at a moment when other labs, including DeepMind’s own parent Google and OpenAI, were moving toward keeping comparable systems closed.