Convolutional neural networks in structural biology and protein science

How it is used

Structural biology works out the three-dimensional shape of proteins and other large molecules, because shape largely decides what a molecule does. The raw evidence is rarely a picture of a protein. It is a grid of numbers: a map of electron density from an electron microscope, a spectrum from a nuclear magnetic resonance machine, or a table of distances between the atoms of a backbone. A convolutional network is built for exactly this kind of grid. It slides small learned filters across the data, looking for the same local pattern wherever it occurs, then combines those local findings into larger ones. In this field it is asked to label voxels, find and clean up peaks, and judge the consequences of changing one amino acid.

The examples here show that range. One pipeline uses stacked convolutional blocks to sort genuine signals in a spectrum from artefacts, and a second network to separate overlapping peaks into their components. Another trains on density maps paired with known atomic structures, labelling each small cube of the map as helix, sheet or neither. A third turns a protein pair into maps of backbone distances, within each chain and between them, and learns which amino acids suit a given binding site. In each case the network's input is a grid, and its output is a label or a number attached to a position on that grid.

Why researchers use it

The step these networks replace is usually slow expert judgement. Deciding which bumps in a spectrum are real, or where a sheet lies in a blurred density map, has long been done by hand, one structure at a time, by people who have spent years learning to do it. That work does not scale to the number of maps and spectra now being collected. Convolutional networks are attractive because the judgement is largely local and repeated: the same visual cue means the same thing wherever it appears, which is the assumption the architecture is built on.

In the record

structural-biologyDeep learning turns raw NMR spectra into protein structures without human interventionResidual networks scored every signal peak in NMR spectra as real or artefact, and regressed the positions of overlapping components around each selected peak.structural-biologyChanging the training penalty improves how a network spots beta-sheets in blurry protein mapsCryo-electron-microscopy density maps paired with known structures were cropped into subregions, and each voxel labelled helix, beta-sheet or background for segmentation training.structural-biologyNeural networks pick amino acid sequences for short peptides in protein binding sitesPeptide and binding-site complexes were encoded as backbone distance maps plus secondary structure, and used to learn sequence design for protein interfaces.structural-biologyResearchers measure folding stability for hundreds of thousands of protein variantsPredicted structures supplied burial, contact-count and secondary-structure features, and guided trimming of loosely contacting terminal segments before stability assays were designed.

Limits

These networks learn from structures already solved, so they inherit that collection's biases towards well behaved, well studied, soluble proteins, and they are weakest on the unusual cases that most need help. Their labels come from existing assignment software or modelled structures, so errors in the training labels are reproduced confidently. They report no honest measure of doubt, and a wrong voxel label or a spurious peak can propagate into a published structure. Local pattern matching also misses constraints that span a whole chain.

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