Baker lab publishes RoseTTAFold and opens it to all

Background

A protein is a chain of amino acids that folds into a three-dimensional shape, and that shape largely determines what the protein does inside a cell. Knowing the structure is central to understanding how a protein works, how a mutation alters it, or where a drug might bind. Determining a structure in the laboratory, however, is slow: the Institute for Protein Design put the cost at years of laboratory work for a single protein^[2]. Computational structure prediction aims to close that gap — to take a protein's sequence and return a plausible three-dimensional structure without doing the experiment.

What happened

On 15 July 2021 the Institute for Protein Design announced RoseTTAFold, its protein structure prediction software, alongside the paper describing it in the journal Science^[1][2]. The work came out of the Baker lab. The lead author was postdoctoral scholar Minkyung Baek^[2]. In the announcement the institute framed the release as accurate protein structure prediction made accessible to all^[2], putting as much weight on who would be able to use the method as on its accuracy.

Key details

Accessibility was built into the release in two ways. The code was published on GitHub, and a public web server was opened^[2]. The code lets groups run the method themselves; the server lets anyone who can submit a sequence get a prediction back without maintaining the software or the hardware it runs on.

The hardware requirement was the second half of the story. The institute said the software can compute a structure in as little as ten minutes on a single gaming computer^[2] — against years of laboratory work for a single protein^[2]. A gaming computer here means an ordinary machine built for graphics-heavy games rather than a specialist computing cluster; the point is that the calculation fits on hardware a normal laboratory might already own.

Demand showed up quickly. Baek said more than 4,500 proteins had been submitted to the web server in the previous month^[2], that is, in the month before the announcement, while the tool had not yet been formally released.

Why it matters

Protein structure prediction had been a research problem for decades; what a release like this changes is availability. Publishing the code and opening a server turns a method from something a specialist group implements into something a student can try, and it lets biologists who work on a particular protein get a structural hypothesis without first becoming experts in the prediction method. The institute's own framing — accessible to all — is the claim the release should be held to^[2].

It is worth being clear about what a prediction is. A computed structure is a model, not a measurement; experimental methods remain the way to confirm what a protein actually looks like, and a prediction is most useful as a guide for which experiments to do.

The work was supported in part by Microsoft, Open Philanthropy, Schmidt Futures, the Washington Research Foundation, the National Science Foundation, the Wellcome Trust and the National Institutes of Health^[2].