Segmentation in structural biology and protein science

How it is used

Structural biology works out the shapes of proteins and other large molecules, because shape largely governs what they do. Segmentation here means dividing something up and labelling the parts. Two things get divided. One is the three-dimensional image: maps from cryo-electron microscopy, where a sample is frozen and imaged by electron beam, and from cryo-electron tomography, which builds a volume from tilted views. Each small cube of the volume, a voxel, is labelled as a particular kind of structure or as empty space. The other is the chain itself: each amino acid, or residue, in a sequence is labelled as belonging to a functional region or not.

Models are chosen to match the thing being divided. Image volumes are handled by networks that learn local patterns from neighbouring voxels, and by attention-based networks that weigh many parts of a patch against each other, sometimes reading the same cube from several directions at once. Sequence regions are handled by networks that scan along the chain, and by protein language models, which are trained on vast numbers of sequences and turn each residue into a list of numbers that encodes its context. Protein surfaces are instead turned into meshes, with each mesh point carrying shape and chemistry.

Why researchers use it

The underlying problem is that these signals are noisy, crowded and hard to read by eye. A density map at middling resolution shows where matter is, but not what the matter is: a flat sheet and a coiled helix can look similar. A tomogram of a cell contains many objects packed together. A sequence carries no visible marks saying where one functional region starts. Deciding these things by hand is slow and depends on expert judgement, which does not scale to the volume of data now collected.

Segmentation stands in for that manual judgement, and often for a later step too. Labelled regions guide the fitting of atomic models into a map, pick out objects worth measuring in a tomogram, or point to the residues that bind a partner. It also gives a way to use information learned elsewhere: a language model trained on sequence alone can supply the context that a short, variable region does not reveal on its own, and expert annotations or existing structures provide the labels to learn from.

In the record

structural-biologyChanging the training penalty improves how a network spots beta-sheets in blurry protein mapsLabelled each voxel of cryo-EM maps as helix, beta-sheet or background, using atomic structures and secondary-structure assignments as the training labels.structural-biologyA transformer model segments cryo-electron tomograms from three viewing directions at onceSplit tomograms into voxel patches, read each from three perpendicular views as tokens, and combined attention with a convolutional encoder clustered into graph nodes.structural-biologyNeural networks trained to find the cutting site on serpin proteinsMarked the reactive centre loop within serpin sequences, with residues encoded one-hot, by substitution matrix, and by a protein language model.structural-biologyNeural network reads protein surfaces to predict DNA and RNA bindingTurned protein structures into surface meshes carrying electrostatic, geometric and chemical values at each point, to mark nucleic-acid binding regions.

Limits

Segmentation depends on labels that are themselves inferred, from fitted atomic models, from simulated masks, or from expert annotation, so errors in the labels become errors in the output. Thin or poorly resolved features are easily missed or confused with neighbouring ones, and performance usually falls on structures unlike those seen in training. A label is not a measurement: a voxel marked as a sheet, or a residue marked as part of a loop, still needs checking against independent evidence before any biological claim rests on it.

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