ColabFold · AIxSci

What it is

Proteins are long chains of building blocks called amino acids. The chain folds into a particular shape, and the shape decides what the protein does. AlphaFold is a trained computer program that guesses that shape from the sequence alone. One of its slowest steps is looking through enormous databases for related sequences from other species, since comparing relatives hints at which parts of a chain sit close together. ColabFold keeps the same prediction machinery but swaps in a much faster search, and packages everything so it can be run in a free online notebook or installed on a laboratory's own computers. A companion version, AlphaFold-Multimer, predicts how several chains fit together.

ColabFold also exposes the knobs. A user can feed in a single sequence with no relatives at all, limit how many times the program loops back over its own answer to refine it, turn off the use of known experimental structures as templates, ask for many predictions from different random starting points, or hand it short fragments rather than a whole chain. Predictions come with the program's own confidence scores, which flag which regions to trust.

Why scientists use it

Working out a protein's shape in the laboratory takes months of specialist work, and some proteins resist every method. A prediction arrives in minutes, so scientists use it as a first sketch: something to compare with patchy experimental maps, to test an idea about which proteins bind each other, or to generate thousands of candidates in a design loop. ColabFold matters because cost and access shape what gets tried. Running the original program well needs large databases and heavy hardware. A free notebook, or a modest local installation, lets a small group run one protein or an entire screen, and lets method developers modify the code to study how the predictions are produced.

In the record

structural-biologyCryo-EM maps and AlphaFold models explain how two bacterial helper proteins activate an enzymePredicted the shape of a full-length bacterial protein, and several arrangements of it paired with a partner protein, scoring each predicted pairing.structural-biologyBenchmark measures how well AlphaFold predicts antibody-antigen complex structuresRan antibody and antigen pairs through a modified copy that reported a prediction at every refinement loop, and produced many predictions per pair from different random seeds.structural-biologyRepeating-coil proteins designed from random sequences using AlphaFold2 in an evolution loopPredicted each candidate sequence in a design loop, using single sequences with no relatives and one refinement round, to evolve new repeating protein shapes.structural-biologyClassifier sorts true from false AlphaFold predictions of human protein pairsPredicted structures for many pairs of proteins from a local installation, using its fast relative search, to build a database of possible interactions.

Limits

A prediction is a guess, not a measurement. The programs were trained largely on single, stable, well-behaved structures, so they are weaker on floppy regions, on shapes that depend on binding partners or chemical tags, and on chains that adopt more than one form. Predicted pairings are especially unreliable: a confident-looking model of two proteins together is not evidence that they ever meet in a cell. Confidence scores help but can be wrong in both directions. The faster search may find fewer relatives, and results shift with settings, so the same sequence can give different answers.

Written by AIxSci: the background from general knowledge, the examples from checked records.