Structure determination in structural biology and protein science

How it is used

Proteins are chains of amino acids that fold into particular shapes, and the shape largely decides what the protein does. Working out that shape has traditionally meant growing crystals, freezing samples for electron microscopy, or reading nuclear magnetic resonance spectra, which record how atomic nuclei respond to radio waves in a magnetic field. All of these give indirect signals that someone must interpret. AI enters at two points. It reads the experimental signal: picking out which specks in a microscope image are protein particles, or deciding which peaks in a spectrum are real rather than noise. And it predicts shape from sequence directly, as in the examples where AlphaFold models were generated for large sets of sequences.

Related work runs the problem backwards. Instead of going from sequence to shape, models are given a protein backbone, the chain of connected atoms stripped of its side groups, and asked which amino acid sequence would fold into it. One example encodes that backbone purely as distances between a few atoms per position, plus markers for how far apart two positions are in the chain and whether they sit on the same chain. Another line of work predates deep learning: statistical models fitted to large collections of related sequences, scoring which pairs of positions vary together, on the reasoning that positions in contact tend to change in step.

Why researchers use it

Experimental structure determination is slow, and much of the slowness is human. Spectra and micrographs contain thousands of features that someone must label by hand before anything can be built, and that labour is the bottleneck rather than the measurement. Learned components replace the labelling, having been trained on collections of examples that people annotated previously. One example trained a network on hundreds of thousands of hand-annotated spectrum fragments; another trained a particle picker from a few thousand hand-picked particles and then let it work through the full datasets.

Prediction from sequence is reached for when no experiment exists. It supplies a shape for proteins nobody has solved, so that work can proceed on an estimate: trimming loose ends from a construct before building an assay, or reading off which positions are buried, how many neighbours each has, and what local shape they sit in. It is also used to guess whether two proteins bind, by modelling them as a pair and judging the result.

In the record

Limits

A predicted structure is an estimate, not a measurement, and the models carry confidence scores that are themselves only estimates of their own reliability. They do best on proteins resembling what they were trained on, and training sets are drawn from solved structures, so poorly studied families, flexible regions and unusual folds are weaker ground. Predicting whether two proteins bind is harder still, and a plausible-looking model of a pair is not evidence that the pair meets in a cell. Learned readers of experimental data inherit the habits, and the mistakes, of the people who annotated their training examples.

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

Related

• Structure determination: the idea, across every field
• AI in structural biology and protein science

Every article