crYOLO · AIxSci

What it is

To see the shape of a protein, researchers freeze a watery sample so fast that it turns to glass, then photograph it with a beam of electrons. Each photograph, called a micrograph, holds many copies of the molecule, scattered at random angles and buried in speckle. Before the shapes can be averaged into one clear structure, every copy has to be located. crYOLO does that locating. It is a convolutional neural network, a program that learns to recognise patterns by passing the image through a stack of filters, each built from the one before. It returns a box around each thing it believes is a particle.

The network is trained on images where a person has already drawn boxes round the particles. Anything not boxed is treated as empty background, and the training pushes the network to agree with the human on three things at once: where the box goes, how big it is, and how confident it should be that something is there at all. Because it looks at a whole image in one pass rather than sliding a window over it, it is quick. The same machinery can be pointed at long thin objects, tracing a filament as a chain of boxes along its length rather than one isolated blob.

Why scientists use it

Picking particles is the bottleneck between taking pictures and getting a structure. A single project may rest on hundreds of thousands of individual copies, and the sharpness of the final answer depends on how many were found and how clean they are. Done by eye, this is slow, tiring and inconsistent between people and between sessions. Older automatic methods matched each image against a template, a reference picture of what the particle should look like, which works badly when the particle is faint, oddly shaped or not quite what was expected.

A learned picker sidesteps the template. It is shown examples and works out for itself what the particle looks like against this particular background, so it copes with low contrast and with ice of uneven thickness. Once trained on many samples, a general model can often be used on a new one with no new labelling. In practice it sits as one component inside a longer chain of conventional software: correcting for drift during exposure, estimating the microscope's blurring, then sorting and averaging the picked particles. The network supplies coordinates; the rest of the pipeline does the physics.

In the record

structural-biologyNeural network locates single protein particles in cryo-electron microscope imagesTrained a 21-layer convolutional network to draw bounding boxes round particles in cryo-EM micrographs, learning from positively labelled boxes alone, with all other positions counted as background.structural-biologyCryo-tomography maps the myosin filament inside relaxed mouse heart muscleRun on blurred, reduced-resolution tomograms of heart muscle to pick and trace both thick and thin filaments, supplying the coordinates for later averaging and refinement.structural-biologyCryo-EM maps show water shifting in actin filaments as they ageUsed its filament-picking mode to locate actin segments in micrographs, stepping boxes along each filament and keeping only filaments with at least six boxes.

Limits

The network only finds what its training resembles. Label the wrong blobs, or train on one kind of sample and apply it to a very different one, and it will confidently return junk: ice contamination, aggregates, the edges of the support film. Because unlabelled positions count as background, real particles a human missed teach it to miss them too, and such errors are easy to overlook once averaging makes everything look smooth. It gives positions, not identities, and no measure of whether the picked objects are the molecule of interest. Checking that remains a human judgement.

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