RoseTTAFold: an AI that works out the 3D shape of proteins

Why does the shape of a protein matter so much?

Proteins are the tiny machines of the body: they digest, defend, carry and relay signals. Each one is a long chain of building blocks (amino acids) that folds into a precise three-dimensional shape. That shape lets the protein do its job, and it also lets scientists design drugs that can latch onto it.

Finding this shape used to need long and costly experimental methods, such as X-ray crystallography or cryo-electron microscopy. For many proteins, the shape is still unknown.

Working out how a long chain folds into its final shape from its sequence alone has been a major challenge in biology for decades.

How was RoseTTAFold built?

In 2020, at the 14th international assessment known as CASP14, DeepMind presented remarkably accurate structure estimates. The researchers behind this study explored neural network designs that follow related ideas.

Their best performance came from a three-track network. One track works on the sequence (one dimension), a second on a map of distances between the building blocks (two dimensions), a third on coordinates in space (three dimensions). Information is transformed and exchanged between these tracks, step by step.

The three tracks talk to one another, which is the key idea: what the network learns about the sequence helps it refine the distances, and what it learns about the distances helps it refine the 3D coordinates.

What can the tool do?

The network produces structures whose accuracy approaches DeepMind's at CASP14. It makes it possible to quickly solve hard modelling problems in X-ray crystallography and cryo-electron microscopy, two methods that rebuild a protein's shape from experimental measurements.

It also gives clues about the role of proteins whose structure was unknown. And it quickly generates models of complexes, meaning proteins that assemble with one another, from sequences alone. This skips the usual approach, which models each protein separately and then works out how to fit them together.

What could this mean for health?

Knowing a protein's shape sooner helps researchers understand how a disease works and guide the search for new drugs. The authors made their method available to speed up biological research.

This work is a tool for laboratories. It does not treat patients directly and it was not tested on patients.

What are the limits of this work?

RoseTTAFold's accuracy approaches AlphaFold's, it does not exceed it. A calculated shape is still a model: researchers check it against experiments before relying on it for important decisions.

The study mainly reports performance on proteins and cases chosen for the evaluation. The road from a modelled structure to a usable drug is long, and nothing in this study shows a benefit for patients.

The authors made the method available to the scientific community, so other teams can test it and build on it. That openness also makes it easier for others to check its limits.