Detection in structural biology and protein science

How it is used

Structural biology works out the three-dimensional shapes of proteins, nucleic acids and the assemblies they form, because shape largely determines what a molecule does. The main experimental methods produce very noisy raw data. In electron cryo-microscopy, thousands of copies of a molecule are frozen in thin ice and photographed; each image is faint and grainy, and the first step is simply to find where the copies are. In nuclear magnetic resonance, the raw output is a spectrum of peaks whose positions carry chemical information. Detection here means marking every real instance in that noise: a particle box in a micrograph, a filament running through a reconstructed volume, a genuine peak rather than an artefact.

The models used are mostly convolutional neural networks, that is, networks that scan an image with learned filters and respond where a familiar pattern appears. In the examples, one such network predicts boxes around particles in cryo-microscope images, and the same tool is run on reconstructed volumes to trace long filaments. Another study trained a picker on a few thousand hand-marked particles and then let it work through whole datasets. In spectroscopy, one network sorts every peak-like feature into real signal or artefact, while a second estimates the positions of overlapping components hidden under a single feature. Newer work uses transformer models, which weigh relationships across a whole patch at once.

Why researchers use it

Detection is the bottleneck at the front of the pipeline. A single session can produce more images than a person could mark by hand in a reasonable time, and everything downstream, averaging, classification and refinement, depends on having found enough correct instances. Hand-picking is also inconsistent: judgements drift between operators and across a long sitting. A trained network applies one standard everywhere, quickly, and can be taught from a modest set of hand-marked examples, as one study did by training a picker on roughly three thousand manually picked particles before turning it loose on the full data.

The method also stands in for template matching and other rule-based approaches, which need you to say in advance what the thing looks like. That is awkward when the structure is unknown, which is usually the point of the experiment, and it fails on objects without a fixed outline, such as filaments of variable length. Learned detectors infer the appearance from examples instead, and in one case were trained on synthetic data generated from the physics of the measurement, which sidesteps the need for hand annotation altogether.

In the record

Limits

A detector learns what it was shown. Trained on one specimen, ice thickness or microscope setting, it may miss instances or confidently mark contamination and ice edges as particles. Because only marked positions count as positive during training, anything the annotators overlooked teaches the network to ignore it, so human bias is reproduced rather than removed. False positives survive into later averaging, where they can blur or bias the final structure, and rare orientations or states are the ones most likely to be dropped. Picks therefore still need checking, and downstream classification does much of that work.

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

Related

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

Every article