The Blueprint Layer: UCSF Team Programs Protein Shape the Way Engineers Draw Parts

For a decade, computational protein design has treated the atom as the unit of control. Side chains in binding pockets, catalytic residues in active sites, interface patches between two partners — these local edits have delivered binders, enzymes, and vaccines at a pace that would have read like science fiction in 2015. Yet biology does not operate only at the scale of angstroms. Membrane curvature, flagellar superhelices, and the global envelope of a repeat protein all depend on shape as a first-class variable — and until this week, generative AI had no reliable way to specify that variable directly.

A preprint posted to bioRxiv on July 23, 2026, from Tanja Kortemme’s group at the University of California, San Francisco, proposes a fix that borrows from a discipline older than molecular biology: computer-aided design. The manuscript, led by Yilun Qi and co-authored with Guoxun Zhang, Klaus Yserentant, Sonya Lee, Artem Lyubimov, Kyrellos Ibrahim, Giuseppe Cimicata, James S. Fraser, and Bo Huang, describes a diffusion framework that continuously guides protein backbones toward CAD-defined shape envelopes — and then validates the results in the wet lab at a scale rarely attempted in shape-conditioned design.

From soft hints to hard envelopes

Most protein generative models today — including the widely used RFdiffusion backbone sampler — excel at local topology: helices, sheets, loops, symmetry. Global shape can emerge as a side effect, but it is rarely prescribed the way an industrial designer prescribes the outer contour of a turbine blade before worrying about bolt holes.

Kortemme’s team closes that gap by importing blueprints from standard CAD software or public 3D model repositories and embedding them as a shape potential inside the diffusion loop. During denoising, the evolving Cα backbone is represented as a point cloud and pulled toward the target envelope through two complementary terms: a boundary constraint that keeps atoms inside the blueprint volume, and a global shape-matching term that rewards agreement with the overall form. The implementation uses signed distance functions and point-cloud similarity metrics — mathematical objects familiar to graphics engineers, now repurposed for amino-acid polymers.

The distinction matters. Prior approaches often treat shape as a soft post-hoc filter: generate many candidates, discard misfits, refine survivors. Here, shape is enforced continuously during sampling, interoperable with existing RFdiffusion workflows rather than requiring a parallel generative stack.

Four modes, one design language

The paper organizes its experimental program around four progressively ambitious modes of shape control.

Mode 1 — de novo envelopes. The group generated backbones from Gaussian noise constrained to five CAD targets: cone, sphere, ellipsoid, bracket, and cube. For each blueprint they sampled 100–200 backbones, selected 20–30 with the strongest shape agreement, and ran eight ProteinMPNN sequences per backbone. Fifteen designs meeting strict AlphaFold2 self-consistency thresholds (Cα RMSD below 1.5 Å, pLDDT above 90) entered bacterial expression. Eleven expressed solubly; ten showed size-exclusion peaks consistent with monomeric products. Crystal structures of Bracket SC4 (1.72 Å) and Cone SC1 (2.70 Å) confirmed that the external contours matched the blueprints — including a conical envelope at 1.32 Å backbone RMSD to the design model.

Mode 2 — reshaping existing folds. Natural β-sheet proteins overwhelmingly adopt right-handed twist; left-handed twist is rare. Starting from α/β scaffolds, the authors imposed CAD blueprints encoding left-handed global twist while preserving secondary-structure connectivity. Cryo-EM reconstructions at roughly 4.0 Å resolution supported the intended curved architecture — evidence that topology and global shape can be decoupled as design variables.

Mode 3 — programmable assemblies. Shared building blocks were reshaped into distinct curvature states (270°, 180°, and 90° bends) and docked into closed rings, open helical filaments, and handedness-inverted superhelical coils. Negative-stain TEM and 3D reconstructions matched the intended geometries across the series — closed elliptical rings, progressively open helices, and right-handed counterparts (hh2-270, hh2-180, hh2-90) generated from the same scaffold via reversed blueprints.

Across these three modes, the authors structurally validated 31 of 59 selected designs by X-ray crystallography or electron microscopy — a hit rate that, for global shape programming, is the paper’s central empirical claim.

When shape becomes physics

Mode 4 asks whether programmed architecture can encode emergent physical behavior, not just static structure.

Low-Reynolds-number hydrodynamic simulations on the designed helical propellers showed that symmetric ring controls produced negligible axial pumping, while left- and right-handed helical variants drove coherent flows in opposite directions under identical rotational forcing. Pumping magnitude scaled with an effective helical slope coordinate (pitch divided by π times diameter) — qualitatively consistent with resistive-force theory for microscopic helical transport.

The capstone experiment couples design to motor chemistry. The team attached an hh2-90-4 helical propeller module to the rotary PS3 F₁-ATPase through a bridge protein engineered with BindCraft rotor binders and RFdiffusion motif scaffolding. Negative-stain TEM supported the intended Rotor–Bridge–Propeller organization.

Single-particle tracking then tested whether the assembly responded to fuel. Fluorescently labeled propellers were imaged before and after adding 2 mM ATP. Across seven independent experiments with the complete three-component assembly, mean apparent diffusion increased after ATP addition; a Bridge–Propeller control lacking the rotor showed the opposite trend, consistent with surface adsorption rather than motor activity. The effect required all three components — propeller, bridge, and rotor.

The authors are careful about what this proves. The assay does not measure propulsion force or demonstrate directional swimming in bulk fluid. What it does establish is selective, ATP-dependent mechanical coupling between a biological rotary motor and a CAD-shaped propeller — a prototype molecular swimmer in the narrow sense of an ATP-responsive nanoscale assembly, not yet a cargo-carrying microbot.

Why the field should watch a preprint

This work sits at a convergence point the AI-biology community has been circling for years. Structure prediction — AlphaFold and its successors — solved the forward problem of inferring shape from sequence. Inverse folding and diffusion models solved much of the sequence-from-structure problem. What remained awkward was the industrial-design layer: specifying an outer envelope independent of fold class and enforcing it through generation rather than selection.

Kortemme’s group, with affiliations spanning UCSF’s Quantitative Biosciences Institute and Biohub San Francisco, is not a neutral observer of that history. Fraser and Kortemme have published de novo fold-family design for years; this manuscript explicitly positions CAD blueprinting as a complement to local interface engineering, not a replacement.

Because the paper is a bioRxiv preprint — posted July 23, 2026, not yet certified by peer review — the structural statistics and swimmer tracking data should be read as strong early evidence, not settled doctrine. Independent replication, peer review, and release of code (not yet bundled with the preprint at publication time) will determine how quickly CAD-guided diffusion becomes a standard module in design pipelines.

Still, the conceptual shift is already legible. If global shape becomes as programmable as a catalytic triad, the design space for synthetic biology expands beyond binders and enzymes into architectures: membrane remodelers, tunable filaments, and rotor-coupled nanomachines whose geometry is drawn before a single residue is chosen.

The practical read

For practitioners, the immediate takeaway is methodological. Blueprints from off-the-shelf CAD tools can plug into RFdiffusion sampling through a continuously enforced shape potential — no exotic model retraining required for the proof-of-concept implementation described in the manuscript.

For the broader AI-biotech narrative, the takeaway is strategic. The hottest headlines in July 2026 have tracked CRISPR specificity, enzyme evolution starting points, and frontier-model policy. This preprint addresses a quieter bottleneck: you cannot engineer what you cannot specify. By making the outer contour an explicit input, Kortemme’s team turns protein design a degree closer to mechanical engineering — draw the part, diffuse the backbone, validate in the microscope, and only then ask what it can do.

The molecular swimmer is a teaser, not a product. But teasers with crystal structures and ATP-dependent tracking data are how fields announce that a new design layer has arrived.

Sources

• bioRxiv — Programming protein shape as an explicit design layer via CAD blueprint-guided diffusion (July 23, 2026)
• bioRxiv — Full preprint PDF (July 23, 2026)