AtomPaint

AtomPaint is a full-atom diffusion model for designing proteins that bind to non-protein partners — including DNA, RNA, and small molecules — developed in George Church's group at Harvard Medical School and posted to bioRxiv in early 2026. Most recent generative protein-design tools, such as RFdiffusion and related backbone generators, operate primarily at the level of protein backbones and have their strongest results on protein-protein interfaces. AtomPaint instead represents the full atomic content of a binding complex, allowing it to reason about contacts with ligands whose chemistry is not captured by a protein-only representation.

The method reframes interface design as an image-inpainting problem. A binding model is converted into a 3D image, the region to be redesigned is masked, and a diffusion model "paints in" the missing atoms; a separate classification model then reads out the corresponding amino acid identities. This decomposition lets the same machinery handle a wide range of binding-partner chemistries within a single full-atom framework.

As an early preprint, AtomPaint is presented as a proof of concept: the authors report that sequence recovery exceeds random baselines, indicating the model has learned meaningful structural signal, while noting clear room for refinement.

#Key Features

• Full-atom representation: Models all atoms in a binding complex rather than protein backbones alone, enabling design against DNA, RNA, and small-molecule partners.
• Inpainting formulation: Treats interface redesign as masked 3D-image inpainting, a flexible framing that generalizes across binding-partner types.
• SE(3)-equivariant diffusion: Uses SE(3)-equivariant ResNets so that generated structures respect 3D rotational and translational symmetry.
• Two-stage design: A diffusion model generates atomic structure and a separate classifier assigns amino acid identities to the redesigned region.

#Technical Details

AtomPaint converts a binding model into a 3D image, masks the region requiring redesign, and uses an SE(3)-equivariant ResNet diffusion model to inpaint the masked atoms; a second SE(3)-equivariant ResNet classifies the resulting positions into amino acid types. Both models are trained on curated structures from the Protein Data Bank. The preprint reports sequence-recovery benchmarks in which AtomPaint outperforms random baselines, which the authors interpret as evidence that the model captures genuine structural information, while emphasizing that performance has headroom for improvement. The paper is released under a CC BY license. As a recent preprint, released weights and code availability should be confirmed against the manuscript.

#Applications

AtomPaint targets protein engineers and computational designers who need binders for non-protein targets — for example DNA- or RNA-binding proteins and small-molecule binding pockets — applications that are awkward for protein-backbone-only generators. By operating at full-atom resolution, it can in principle propose interface sequences and geometries for nucleic-acid and small-molecule recognition that are then filtered and validated experimentally.

#Impact

AtomPaint extends generative protein design toward the chemically diverse, non-protein binding problems that dominate much of molecular biology and drug discovery, using a unified full-atom inpainting framework. As an early-stage preprint that the authors frame as a proof of concept with above-baseline but not yet refined performance, its practical impact will depend on further development and independent experimental validation.