DiffTopo

De novo protein design has advanced rapidly with all-atom generative models, but exploring genuinely novel fold space remains difficult: sampling directly at atomic resolution is expensive and tends to revisit familiar topologies. DiffTopo reframes the problem at a coarser scale, generating the overall arrangement of secondary structure elements first and leaving atomic detail to a downstream designer.

DiffTopo is a diffusion model that operates on a coarse-grained representation of protein structure, capturing the topological organization of a fold rather than its full atomic coordinates. Sampling at this low resolution is fast and produces diverse topologies, which are then handed to RFdiffusion to build full backbones and, in turn, designed sequences. The authors also introduce MirrorTopo, a complementary approach that expands known fold space by mirroring the topological organization of native proteins, generating layouts not observed in nature.

Developed in Bruno Correia's laboratory at EPFL (École Polytechnique Fédérale de Lausanne) and released as a preprint in October 2025, the work grounds its generative pipeline in protein representations to reach uncharted regions of fold space.

#Key Features

• Coarse-grained fold diffusion: DiffTopo diffuses over a low-resolution topology representation, making sampling of overall fold organization efficient and diverse.
• Composable with RFdiffusion: Sampled topologies seed RFdiffusion to generate full atomic backbones, combining fast fold exploration with established all-atom design.
• MirrorTopo fold-space expansion: A companion method mirrors the topology of native proteins to produce novel layouts beyond the natural repertoire.
• Experimental validation: The team generated and tested 30 novel topologies plus 6 mirror topologies, moving beyond in-silico metrics to laboratory characterization.

#Technical Details

DiffTopo is a diffusion generative model defined over a coarse-grained protein structure representation that encodes fold topology instead of atomic coordinates. Because sampling occurs in this reduced space, the model can efficiently explore a broad range of secondary-structure arrangements before any atomic detail is committed. Generated topologies are passed to RFdiffusion to produce full backbones, followed by sequence design. The MirrorTopo variant expands the accessible fold repertoire by mirroring the topological organization of native structures. The authors experimentally generated and tested 30 novel topologies together with 6 mirror topologies, demonstrating that low-resolution sampling can yield designable, novel folds.

#Applications

DiffTopo is aimed at protein designers seeking scaffolds outside the space of natural and previously designed folds. By rapidly proposing diverse topologies that feed into existing all-atom pipelines, it supports the creation of new backbones for functional design tasks such as binder and enzyme scaffolding, and provides a tool for studying the principles that make a fold designable.

#Impact

By separating fold exploration from atomic construction, DiffTopo shows that coarse-grained diffusion over topology can efficiently reach uncharted fold space and, when coupled with RFdiffusion, produce backbones that survive experimental testing. The experimental validation of 30 novel and 6 mirror topologies distinguishes it from purely in-silico generative studies and offers a practical strategy for expanding the designable protein universe. As a preprint, its full scope awaits peer review.