Learn · Intermediate

De novo protein design: how a computer invents a molecule that sticks

De novo protein design is the practice of inventing a protein that has never existed, with a sequence no organism ever evolved, engineered so that it folds into a specific shape and latches onto a chosen target molecule. It works today because the inverse problem got tractable: models can now predict, cheaply and fairly accurately, what shape a given amino acid chain will fold into, which means a computer can grade thousands of candidate designs before anyone spends a week in a laboratory building one.

That capability is why Anthropic could report that Claude designed binders against 14 of 15 protein targets with two contract labs building every design. Understanding what that result does and does not mean requires understanding the pipeline underneath it.

The problem

A protein is a chain of amino acids, twenty possible letters at each position, that spontaneously folds into a three-dimensional shape. The shape is the function. An enzyme works because its folded surface has a pocket that grips one particular molecule; an antibody works because a loop on its surface fits a particular patch on a virus.

For fifty years the hard question was the forward one: given a sequence, what shape? Christian Anfinsen showed in the 1960s that the sequence alone determines the fold, which meant the answer was in principle computable and in practice hopeless. The search space is astronomically large and the energy landscape is rugged.

AlphaFold, from DeepMind's John Jumper, Demis Hassabis and colleagues, largely settled the forward problem in 2021 by predicting structures at accuracy competitive with experimental methods. That mattered enormously for biology. But it also quietly unlocked something else: a fast, reasonably trustworthy referee.

Why a referee changes everything

Design is a search problem, and search problems are gated by how cheaply you can score a candidate. If checking whether a design works costs six weeks of wet-lab time, you can test perhaps a dozen ideas a year. If checking costs thirty seconds of GPU time and is right most of the time, you can test a million and only build the survivors.

The analogy is a lock and key. Designing a key for a lock you cannot see is impossible. Designing one for a lock you can photograph, when you also own a simulator that tells you whether a proposed key will turn, is an engineering problem. Structure prediction is that simulator.

The modern pipeline

Contemporary de novo binder design usually runs three specialised models in sequence, and this division of labour is the thing most coverage skips.

Generate a backbone. RFdiffusion, from David Baker's lab at the University of Washington, is a diffusion model that operates on protein structure rather than pixels. Give it a target surface and it hallucinates a plausible protein backbone shaped to nestle against it, by the same denoising logic that turns static into an image. Independent work by Namrata Anand and Tudor Achim demonstrated the equivalent idea for joint structure and sequence generation.

Choose a sequence for that backbone. A shape is not a molecule. ProteinMPNN, also from the Baker lab, solves the inverse-folding problem: given a desired backbone, which amino acid sequence will actually fold into it? This is where most naive designs die, because many beautiful shapes have no sequence that reliably produces them.

Filter hard. Run the proposed sequence back through a structure predictor and check whether it folds the way you intended and docks the way you hoped. Protein language models such as ESM, trained on evolutionary sequence data with the same next-token objective that powers chatbots, add another filter by scoring whether a sequence looks like something biology would tolerate at all.

Only the survivors get synthesised. In a typical campaign, 10 to 15 percent of synthesised designs turn out to bind, which sounds terrible until you compare it to the alternative of not having a filter.

What a general model adds

The three-model pipeline is powerful and fiddly. It requires choosing targets, setting hotspots, tuning diffusion parameters, ranking outputs, and deciding what to order, which is why campaigns historically consumed weeks of a specialist's attention per target.

What changed recently is orchestration. A general reasoning model can drive the whole stack, deciding what to try next based on what came back, which is closer to running the campaign than to running a tool. The distinction matters for how you read the headlines: the novelty in results like Anthropic's is not a new protein model, it is a competent operator that never gets tired. Our lesson on the harness covers the same phenomenon in software.

The caveat that keeps getting dropped

Almost every published AI protein design result validates binding, measured with a standard assay that reports how tightly two molecules stick. Binding is necessary and nowhere near sufficient. A molecule that grips its target may do nothing useful, may grip fifty other things too, may be destroyed by the body in minutes, or may provoke an immune response.

When a result says "no design was tested for activity and no structure was solved," it means exactly that: the poses in the figures are predictions, and the biology is unproven. That is still a real result. It is not a drug. Keeping those two things separate is most of what informed reading of this field consists of.

Ground Truth. (2026, August 18). De novo protein design: how a computer invents a molecule that sticks. Ground Truth. https://groundtruth.day/learn/de-novo-protein-design.html

@misc{groundtruth:de-novo-protein-design,
  title  = {De novo protein design: how a computer invents a molecule that sticks},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {aug},
  url    = {https://groundtruth.day/learn/de-novo-protein-design.html}
}

Topics: protein-design · biology · ai-for-science · diffusion · drug-discovery · fundamentals