Protein Design
Proteins are programmable machines. We develop models that read structure, sequence, and function, and then write new ones.
Applications
Sequence and structure design under stability and manufacturability constraints.
Binder generation, affinity maturation, and developability scoring in one objective.
Activity, stability, and specificity optimized jointly against the conditions a real process runs in.
De novo binders against structurally defined epitopes.
Take a working protein and move it along the axis a process actually needs.
Explore families with no close natural neighbour.
From predicting proteins to designing them.
Method
Multi-objective generation instead of generate-then-filter, so that the model never spends its budget on regions the constraints already excluded.
Learned scores rank; simulation adjudicates the disagreements that ranking cannot resolve.
Hit rate, failure modes, and the designs that did not work are part of the result.
Why AI
Structure predictors and protein language models are trained on the whole evolutionary record, so they absorb which amino-acid patterns can physically fold and function. They work in the space of paths evolution could have taken, not only the ones it did.
A human pipeline mutates something it already has. A generative model samples a new backbone conditioned on a target surface, an active site, or a fold, and can land most of the way across sequence space in one step.
Accurate prediction lets millions of designs be judged in silico so that only the plausible few are ordered as DNA. In one landmark case, half of the hundredfold gain in success rate came from the generator and half from filtering before any experiment.
The 2024 Nobel Prize in Chemistry went to David Baker for computational protein design, and jointly to Demis Hassabis and John Jumper for protein structure prediction.
Evidence
Peer-reviewed results where the design came out of a model and the protein came out of a flask.
In the blind CASP14 assessment AlphaFold2 predicted backbones to roughly the width of a carbon atom, against 2.8 Å for the next best method. The public database now holds predictions for over 214 million sequences, against about 100,000 structures solved experimentally in the previous half century.
Jumper et al., Nature, 2021 ↗Running a structure predictor backwards as a design engine produced experimentally confirmed binders for twelve hard targets, including CRISPR-Cas9, with no high-throughput screening. The same paper puts the earlier physics-based baseline below one in a thousand.
Pacesa et al., Nature, 2025 ↗Asked to invent proteins that grip five medically relevant targets, RFdiffusion succeeded with fewer than 100 designs per target where earlier campaigns tested thousands. One designed MDM2 binder reached 0.5 nM, about a thousand times tighter than the natural peptide it replaces.
Watson et al., Nature, 2023 ↗A protein language model was asked for something that glows and produced a bright green fluorescent protein sharing 58 percent of its sequence with the closest known fluorescent protein and 36 percent with jellyfish GFP. The authors estimate the gap at more than 500 million years of natural evolution.
Hayes et al., Science, 2025 ↗A designed luciferase built inside an invented fold emits light, survives 95 °C, and matches natural enzymes on catalytic efficiency while being far more selective about its substrate.
Yeh et al., Nature, 2023 ↗The honest other half: the first round of that luciferase campaign found 3 working designs out of 7,648, and even the best binder pipelines fail nine times out of ten on their hardest target. Wet-lab validation, not compute, is still the gate. Every result above ends with protein expressed in E. coli and measured on a real instrument.
