Baker's Team Unveils All-Atom Protein Generation from Text

David Baker's group at the University of Washington and researchers at the National University of Singapore (NUS) have proposed a model that generates a protein structure with all its atoms from a text description. Previously, each type of property required its own input format. Now a single natural-language interface replaces them. The work was reported by Jiqizhixin.

Where de novo protein design gets stuck

Models have been able to generate plausible protein structures for several years. The harder part is getting a protein with predefined properties: a specific shape, binding to a target, a particular function, stability. Existing systems handle this through specialized input representations. One task uses a fixed-motif mask, another a distance map between residues, a third ligand coordinates. Each new type of target requires a new format and, as a rule, additional training.

As a result, the range of tasks is set by what the model's authors managed to build into the architecture. The new work targets exactly this limitation.

How the approach works

The idea follows the logic of text-to-image: the user describes the desired result in words, and the model turns the description into a structure. Generation happens in an all-atom representation. The model works not only with the chain backbone but also with side groups, which determine a protein's chemical contacts and function.

Some context helps with scale. Baker won the 2024 Nobel Prize in Chemistry for computational protein design, and his lab is behind RFdiffusion, ProteinMPNN and the all-atom extensions of that family. Text control looks like the next step in the same direction: removing manual condition preparation from the pipeline.

Key parameters of the work, as known from the available excerpt:

• Authors: David Baker's lab (University of Washington) together with NUS and other contributors.
• Model type: a text-conditioned generative model of protein structure.
• Representation: all-atom, meaning it accounts for all atoms, not just the backbone.
• What it replaces: separate input representations for each target property.
• What is not disclosed: the available text gives no metrics, model size, training data or results of laboratory validation.

What to check after publication

For now, this is a claim of a new interface, not a confirmed jump in quality. Practitioners should look at four questions.

First: how accurately the model follows the wording, especially when the description includes quantitative constraints such as size, number of contacts or fold type. Text is flexible but ambiguous.

Second: whether there is experimental validation. In protein design, the deciding argument is not a good-looking structure on screen but expression, folding and measured activity in the lab.

Third: whether the code and weights are open. That determines whether the community can reproduce the result and build the model into its own pipelines.

The fourth is safety. The easier it is to describe a protein in words, the lower the barrier to entry and the more important filters for dangerous functions become. The industry has so far addressed this problem poorly.

If the text interface can match the quality of specialized models, biotech teams without deep ML expertise will find it easier to test hypotheses. We will wait for the full paper and independent tests.