The encyclopedia · R&D & Science · Technical decision · 2018–2021
AlphaFold predicted a protein's shape from its sequence instead of measuring it
DeepMind's AlphaFold solved the 50-year protein-folding problem by predicting 3D structure from sequence at near-atomic accuracy.
Google DeepMind
the move
A protein's job depends on its 3-dimensional shape, and knowing the sequence alone does not tell you the shape. Determining one structure experimentally can take a research team months to years, which is why only about 100,000 unique protein structures were known despite billions of protein sequences.
DeepMind's AlphaFold attacked the problem computationally. It was trained on the structures scientists had already solved, and it used evolution as a signal: sequences that share a common ancestor must fold similarly, so multi-sequence alignments carry information about which parts of a protein are near each other in space.
The network then produces the 3D coordinates of every atom from a sequence alone. In the blind CASP14 assessment it was competitive with experimental structures for most targets and far outperformed all prior computation, doing in minutes what could otherwise take a lab years.
The win was not a bigger model for its own sake; the key was encoding physical and evolutionary knowledge into the architecture, then letting it generalize across the enormous space of unseen proteins.
why it works
- Experimental structure determination is one of the slowest, most expensive steps in biology, so replacing it with prediction changes the scale of what can be studied.
- Multi-sequence alignments let evolution substitute for measurement, giving the model a large private training signal beyond the few known structures.
- The model solves a general task, so it works even when no similar structure has ever been measured.
- Near-atomic accuracy is enough to be useful, turning structure into a routine input rather than a scarce resource.
what transfers
When the ground truth is expensive, build a model that generalizes it — learn the rule from a small set of examples instead of re-measuring the same thing for every new case.
what came after
AlphaFold's structures were released for nearly the whole of the known protein universe, and the open database became a standard tool in drug discovery, genomics and disease research. Its approach helped end a 50-year open problem and showed that a carefully designed deep network could extract a general physical rule from limited, expensive data.
references
spotted an error? The archive wants to know.