RoseTTAFold

What it is

Proteins are long chains of small chemical units called amino acids. The chain folds up into a particular shape, and that shape decides what the protein does: which molecules it grips, which reactions it speeds up. Reading the sequence of amino acids is now easy; seeing the folded shape is hard, because it has usually meant growing crystals of the protein or freezing it and imaging it with electrons. RoseTTAFold is one of a family of programs that skip that step. It is a neural network, a piece of software whose internal settings are tuned by showing it many examples until it can reproduce known answers, in this case protein shapes already solved in the laboratory.

The network is trained on sequences and on the matching shapes from decades of experimental work. It also leans on evolution. If two positions in a chain change together across the same protein in many different species, they are probably touching in the folded form. Aligning related sequences this way, a step called a multiple sequence alignment, supplies that hint. RoseTTAFold combines several views of the problem at once: the sequence and its relatives, a map of which positions sit near which, and the actual coordinates in space. It passes information between those views repeatedly until they agree, then reports a shape along with a score saying how sure it is.

Why scientists use it

Shape is the bridge between a gene and a function. Knowing the folded form of a protein suggests what it binds, how a mutation might break it, and where a drug molecule could dock. Laboratory structure determination is slow, costly and fails outright for many proteins, so for most known sequences no experimental shape exists. A prediction program fills that gap cheaply and in bulk, and it can be pointed at proteins that have never been purified at all.

The same programs are also used for pairs and groups of molecules, not just single chains. Predicting how two proteins fit together, or how a short fragment settles into a pocket on a larger one, lets researchers test ideas about signalling, immunity or enzyme control before committing to the bench. Because the output comes with a confidence estimate, a laboratory can triage: follow up the predictions the program is sure about, treat the rest as guesses. RoseTTAFold emerged from the Rosetta community of structural modelling tools, and sits alongside other prediction networks that researchers now routinely compare against each other.

In the record

Limits

A prediction is an inference from past examples, not an observation. Shapes unlike anything in the training data, and proteins with few known relatives to align, come out poorly. The programs tend to report one rigid form, while real proteins flex, unfold in places and change shape on binding. Confidence scores help but can be wrong, and they track the shape of each part better than how parts sit together. Predicting how two molecules dock remains harder than folding one chain, so laboratory structures are still the reference against which these tools are judged.

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

Model family

• Transformer

Every article