Boltz

What it is

Proteins are long chains of smaller units called amino acids. The order of those units is easy to read, but what a protein does depends on the shape the chain folds into, and on how it grips its partners. Boltz is a set of models trained on experimentally measured shapes. Given the sequence of units, a model predicts where every atom sits, including in a complex, meaning two or more molecules bound together. The family has several members, released one after another, and the later ones also work the other way round: given a target, they suggest a sequence that should bind to it. Being openly released means anyone can download and run them.

Alongside each prediction, these models report their own confidence, a numerical guess at how much of the answer to trust. Researchers can supply extra information to guide the work: related sequences from other species lined up against each other, called a multiple sequence alignment, or a known structure used as a template. They can also run the same job many times with different random starting points and compare the answers. Because the models are released rather than retrained by each user, most studies simply run them as they come, as a tool rather than something to be rebuilt.

Why scientists use it

Working out a protein's shape in the laboratory is slow and painstaking, and some molecules refuse to cooperate at all. A prediction takes a fraction of the effort, so researchers can screen many candidates and keep only the promising ones for experiment. Short protein fragments, called peptides, are of particular interest because they are small enough to make and alter easily, and they often act as the messages that one protein passes to another.

Because several such models now exist, the models themselves have become objects of study. Researchers collect structures already solved by experiment, hide the answer, ask each model to predict it, and then measure how close it came. That tells them which model to trust for which kind of problem, and whether the confidence scores mean anything. The same exercise checks whether a model has merely memorised shapes resembling those it was trained on.

In the record

Limits

A prediction is an informed guess, not a measurement. These models are trained on structures already solved, so they do best on molecules resembling those, and can fall short on unusual or flexible ones, where several shapes may be equally real. Short, loosely bound fragments are harder than compact folded chains. The confidence a model reports is its own opinion and may be high when the answer is wrong. Running a model many times can give different answers. Only experiment can settle which, if any, is right.

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

Model family

• Diffusion model
• Transformer

Every article