ModelAngelo

What it is

Cryo-electron microscopy freezes many copies of a molecule, photographs them with a beam of electrons and combines the images into a map: a three-dimensional cloud showing where the molecule's matter is dense. The map is not a structure. Someone still has to decide which chemical chain, with which amino acids in which order, threads through the cloud. Traditionally a scientist does this by hand, sliding fragments around on screen until they match. ModelAngelo belongs to a family of tools that learn this fitting from maps whose answers are already known, then propose a complete chain of atoms for a new map on their own.

Such tools work on the map in small cubes of space. They convert each cube into numbers, pass it through a learned network, and predict where a chain's backbone runs and which of the twenty amino acids sits at each position. The predicted sequence can then be compared with the real genetic sequence of the protein to confirm the identity of what has been seen. Related steps in the same pipeline can also be learned: picking out which specks in the raw micrographs are molecules at all, or using a predicted structure from sequence alone as a starting shape to be pushed into the map.

Why scientists use it

Building a model by hand is slow and depends on the judgement of the person doing it. A large assembly of several proteins can take weeks, and two experts may produce slightly different answers. An automatic builder does the tedious part in a fraction of the time and does it the same way each time, which makes the result easier to check and to reproduce. It also lets scientists look at more maps, including poorer ones they might otherwise set aside.

The structure matters because shape is function. Knowing how a protein assembly is put together shows how its parts touch each other, where a drug might bind, and how a change in one amino acid, inherited as a disease, could break the whole thing. Automatic model building stands in for the hours of expert hand-fitting between a blurry cloud and a usable picture of a molecule.

In the record

Limits

An automatic builder can only be as good as the map. Where the cloud is blurred, because part of the molecule moves or the sample is thin, the program may still place atoms confidently and produce a chain that looks finished but is wrong. It tends to reproduce what it was trained on, so unusual chemistry, bound small molecules and sugars may be missed or misassigned. The output is a proposal, not evidence: it needs checking against the known sequence and against the map itself, and a human still has to judge the doubtful regions.

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

Model family

• Graph neural network

Every article