Lab Newsletter — September 5, 2026: Designing Life's Missing Parts

For almost all of biology’s history, a protein was something you found. You purified it from a cell, crystallized it, solved its structure, and named it. This week’s digests have lived in that world — predicting a protein’s function from its sequence, folding two proteins together to estimate how tightly they bind, mapping the flexible ensembles a single protein visits. All of it reads molecules that already exist. Today’s story runs the other way. De novo protein design is the art of writing proteins — inventing sequences and structures that evolution never made — and it has quietly become one of the most striking demonstrations that the deep-learning revolution in structural biology can be run in reverse.

🧱 First, proof it was even possible

Long before neural networks, the Baker lab settled the in-principle question (Kuhlman … Baker, Science, 2003). The challenge, as they framed it, is stark: “a major challenge of computational protein design is the creation of novel proteins with arbitrarily chosen three-dimensional structures.” Their answer was Top7, built with “a general computational strategy that iterates between sequence design and structure prediction.” It wasn’t a tweak of a natural protein — it had “a novel sequence and topology,” and when they made it, “Top7 was found experimentally to be folded and extremely stable, and the x-ray crystal structure of Top7 is similar (root mean square deviation equals 1.2 angstroms) to the design model.” The closing line set the agenda for two decades: this makes possible “the exploration of the large regions of the protein universe not yet observed in nature.”

🌀 Then, inverting the predictor

The deep-learning era began with a mischievous idea: if a network can predict structure from sequence, can you run it backwards to design? Hallucination (Anishchenko … Baker, Nature, 2021) did exactly that, asking “whether the information captured by such networks is sufficiently rich to generate new folded proteins with sequences unrelated to those of the naturally occurring proteins used in training.” The recipe reads like a dream: start from random amino acids, then “carry out Monte Carlo sampling in amino acid sequence space, optimizing the contrast … between the inter-residue distance distributions predicted by the network and background distributions” — sharpening noise into a confident, protein-like structure. And it wasn’t just a computer trick: they “obtained synthetic genes encoding 129 of the network-‘hallucinated’ sequences … 27 of the proteins yielded monodisperse species,” with three structures solved by crystallography and NMR. The lesson, in their words: “deep networks trained to predict native protein structures … can be inverted to design new proteins.”

✍️ Writing the sequence for a shape

Design splits neatly into two questions: what shape do I want? and what sequence folds into it? ProteinMPNN (Dauparas … Baker, Science, 2022) nailed the second. Its abstract names the gap it closed: “although deep learning has revolutionized protein structure prediction, almost all experimentally characterized de novo protein designs have been generated using physically based approaches such as Rosetta.” ProteinMPNN changed that, and the numbers are blunt: “on native protein backbones, ProteinMPNN has a sequence recovery of 52.4% compared with 32.9% for Rosetta.” Better still, it plays well with others — the authors demonstrate it “by rescuing previously failed designs, which were made using Rosetta or AlphaFold, of protein monomers, cyclic homo-oligomers, tetrahedral nanoparticles, and target-binding proteins.” It’s the fast, reliable sequence-writer that the rest of the pipeline leans on.

🎯 Generating structure and function, by diffusion

The leap that put design on magazine covers was borrowing the math behind image generators. RFdiffusion (Watson … Baker, Nature, 2023) noted that a truly general design framework — “including de novo binder design and design of higher-order symmetric architectures” — was still missing, then supplied it by “fine-tuning the RoseTTAFold structure prediction network on protein structure denoising tasks.” The result is a generative model of backbones that handles “protein binder design, symmetric oligomer design, enzyme active site scaffolding and symmetric motif scaffolding.” The proof is beautifully concrete: “the accuracy of RFdiffusion is confirmed by the cryogenic electron microscopy structure of a designed binder in complex with influenza haemagglutinin that is nearly identical to the design model.” And the framing is exactly the one we keep meeting — “in a manner analogous to networks that produce images from user-specified inputs, RFdiffusion enables the design of diverse functional proteins from simple molecular specifications.” A designed molecule, made real, and confirmed by cryo-EM.

🧭 Programming protein space

If RFdiffusion is a camera for proteins, Chroma (Ingraham … Nature, 2023) is a programmable one. It opens on the scale of the opportunity — “three billion years of evolution has produced a tremendous diversity of protein molecules, but the full potential of proteins is likely to be much greater” — and offers “a generative model for proteins and protein complexes that can directly sample novel protein structures and sequences, and that can be conditioned to steer the generative process towards desired properties and functions.” The framing is the one that ties design to the agentic thread this digest keeps following: Chroma “achieves protein design as Bayesian inference under external constraints, which can involve symmetries, substructure, shape, semantics and even natural-language prompts.” You can, in a real sense, ask for a protein. And it holds up at the bench — “the experimental characterization of 310 proteins shows that sampling from Chroma results in proteins that are highly expressed, fold and have favourable biophysical properties.”

🔁 The predictor is also the referee

None of this floats free of the tool that started it. AlphaFold (Jumper … Hassabis, Nature, 2021) was “the first computational method that can regularly predict protein structures with atomic accuracy even in cases in which no similar structure is known,” validated “in the challenging 14th Critical Assessment of protein Structure Prediction (CASP14) … competitive with experimental structures.” That accuracy is what design inverts (hallucination, RFdiffusion are literally structure networks turned around) and what it trusts as a judge: the standard in-silico test of a design is to fold it back with a predictor and check that it returns to the intended target. Prediction and design are two directions of a single map — and CASP is the prove-it benchmark underneath both.

🧬 Why it’s our kind of problem

It lands right on the lab’s map. If predicting and annotating proteins is about reading biology, design is about writing it — and a serious virtual cell or Human Cell Simulator will need both: custom sensors to watch a cell, binders to perturb it, enzymes to rewire it. Chroma’s natural-language prompting points at the same agent-driven future we keep sketching — describe the protein you need, let the model draft it, validate in silico, test at the bench. And the way these tools travel is our ethos exactly: RFdiffusion, ProteinMPNN and Chroma are open source, AlphaFold and CASP are the public yardstick — the same publish-the-model-and-the-test spirit behind the BioImage Model Zoo and BioEngine. For most of biology, proteins were life’s given parts. Now they’re parts we can design — and that changes what a laboratory is for.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (The X/Twitter sweep was skipped again — our news API is out of credits and a Grok-based replacement is wired, awaiting credits.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.