ProtPainter

ProtPainter generates protein backbones conditioned on a 3D curve, letting a designer sketch the path a chain should take through space and receive designable structures that follow it. Existing conditional backbone generators specify topology parametrically — the count, approximate placement, and orientation of secondary structure elements, the classical "blueprint" formulation. That representation is compact but covers only a thin slice of three-dimensional topology space and offers no natural way to edit a fold once generated. A curve is a far more expressive handle: it can be drawn, dragged, rotated, joined to another curve, or locally re-annotated, and each of those operations maps to a concrete change in the generated backbone.

The work comes from Zhejiang University and was published as a conference paper at ICLR 2025, with a preprint posted in April of that year. It is built on top of RFdiffusion and is deliberately retraining-free with respect to that backbone: rather than fitting a new generative model, ProtPainter injects curve information into RFdiffusion's existing denoising trajectory, so the released structure prior does the structural heavy lifting while the new components supply the control signal.

ProtPainter sits alongside other conditional structure generators — Chroma conditions on secondary structure annotations or point clouds, TopoDiff on a learned topology latent — but targets a control modality none of them expose, and the paper introduces a benchmark to measure it.

#Key Features

• Curve-conditioned generation: A 3D curve, drawn by hand or extracted from an existing structure, is the sole topology condition, supporting draw, drag, rotate, joint, and local secondary-structure editing operations on the resulting fold.
• CurveEncoder: An equivariant graph neural network predicts per-position secondary-structure labels from curve geometry, which parametrizes conversion of the bare curve into a full backbone sketch at matching dimensionality.
• Helix-Gating fusion scheduling: The sampler compares the helix percentage of its current structure prediction against the sketch to switch from a low-guidance confidence phase into a high-guidance controllable phase, rather than applying a fixed guidance weight throughout.
• Retraining-free guidance: Conditioning enters through filtering and fusion at sampling time, so the method reuses pretrained diffusion weights instead of requiring a new training run.
• A topology benchmark and metric: The Protein Restoration Task and self-consistency Topology Fitness (scTF), a Procrustes similarity between the refolded backbone's curve and the guidance curve, give the first quantitative test of topology-conditioned generation.

#Technical Details

Generation runs in two stages. Sketching abstracts a backbone into a curve, annotates it with CurveEncoder — an EGNN with 100 tokens, dimension 32, and depth 3, trained on 15,000 curve samples derived from PDB structures and scaffolds at three sampling granularities — and expands the annotated curve into a parametric sketch. Sketch-guided sampling then runs the DDPM process, upsampling the curve to sketch resolution while low-pass filtering generated backbones to a matching frame, with RoseTTAFold translation estimates providing self-conditioning. Evaluation uses three restoration sets of increasing length and complexity: HHH_ems three-helix bundles of 50–60 residues, a 1a0b cluster of 100–250 residues, and GPCRs of 280–400 residues, each of 50 targets, with designs refolded through ProteinMPNN at temperature 0.1 followed by OmegaFold. On GPCRs, ProtPainter reaches scTF 0.800 and fit designability 0.792, against 0.262/0.000 for RFdiffusion, 0.182/0.000 for Chroma, and 0.179/0.000 for TopoDiff; on the short HHH set it records 0.767 scTF and 0.654 fit designability. Confident designability stays high (0.832–0.936) across all three sets. Restoring CATH Mainly Alpha topologies without refolding, over 60% of targets exceed scTF 0.7, rising to near 0.8 for up-down bundles and falling to 0.249 for the more tangled DNA polymerase topology.

#Applications

The intended workflow is interactive structural design. A ChimeraX plugin lets a user draw curves directly onto a protein surface to define binder conditions and assign secondary structure to the intended output, so a designer can sketch a scaffold that wraps a target epitope, drag an existing three-helix bundle into a two-helix arrangement, or splice two folds together, and inspect designability immediately. That suits de novo binder and scaffold design, membrane-protein and GPCR-like scaffold construction, and hinge or multi-state design where the topology transition is the object of interest.

#Impact

ProtPainter reframes topology control as a drawing problem and supplies the benchmark that makes the claim measurable, with the GPCR results showing that unconditional and loosely conditioned diffusion samplers essentially never hit a specified complex topology while curve guidance does so most of the time. The limits are practical: designability drops relative to unconstrained RFdiffusion on longer targets, self-consistency RMSD grows with chain length, and the approach inherits whatever biases the underlying structure prior carries. No training code or model weights for ProtPainter itself have been released — the only public code is the ChimeraX plugin — so reproduction currently requires reimplementing the method from the paper.