A single-step generative model for protein structure prediction and binder design that reaches AlphaFold3-level accuracy with a claimed ~15x inference speedup.
DCFold is a pretrained single-step generative model for protein structure prediction and protein binder design. It is built on the AlphaFold3 architecture but replaces AlphaFold3's iterative diffusion sampling—which requires many sequential denoising steps at inference—with a single forward pass. The authors report that this design reaches AlphaFold3-level accuracy while delivering a claimed ~15x speedup in inference, addressing one of the main practical bottlenecks of diffusion-based structure predictors: the cost of generating high-quality structures one denoising step at a time.
The model's central methodological contribution is a Dual Consistency training framework paired with a Temporal Geodesic Matching (TGM) scheduler. Consistency-style training teaches a generative model to map directly from noise to a clean sample, collapsing an iterative trajectory into one step; DCFold applies this idea to the AlphaFold3 structure module so that a single evaluation produces a usable structure rather than a starting point for further refinement. A notable property is that one checkpoint serves both structure prediction and binder design without task-specific retraining.
DCFold was introduced in a May 2026 arXiv preprint (arXiv:2605.17899) and accepted as an Oral at ICLR 2026. The arXiv listing does not state author affiliations; the work is attributed here to Tsinghua University's Institute for AI Industry Research (AIR) based on senior authors Hao Zhou and Wei-Ying Ma, who lead AIR's protein generative-modeling group (e.g., ProfileBFN, ESM-AA), with co-authors including members of that group. This affiliation is inferred from the authors' group, not stated on the arXiv record.
DCFold inherits the AlphaFold3 architecture, including its representation trunk and diffusion-style structure generation module, and changes how that generative module is trained and queried. Rather than learning a step-by-step denoiser sampled iteratively at inference, the Dual Consistency framework trains the model to be self-consistent across the generative trajectory so that a single step yields a high-quality structure, while the Temporal Geodesic Matching scheduler governs how points along that trajectory are matched during training. The headline empirical claims are AlphaFold3-comparable accuracy on structure-prediction and binder-design evaluations together with an approximately 15x reduction in inference cost from the single-pass design. The arXiv abstract does not describe the training datasets, parameter count, or detailed benchmark tables, so those specifics are not characterized here.
DCFold targets workflows where the throughput of AlphaFold3-class predictions matters: large-scale proteome or variant screening, high-throughput binder and interface design campaigns, and any pipeline where iterative diffusion sampling is a latency or compute bottleneck. Because one checkpoint covers both folding and binder design, it can serve as a single backend for groups that would otherwise maintain separate predictive and generative-design models, benefiting protein engineers, structural biologists, and computational drug-discovery teams.
DCFold contributes to a broader effort to make AlphaFold3-quality structure prediction cheaper and faster by distilling iterative diffusion into single-step generation, an approach that, if its accuracy claims hold up under independent evaluation, could lower the cost of running modern structure predictors at scale. Its acceptance as an ICLR 2026 Oral signals strong peer interest in consistency-based acceleration of biomolecular generative models. As of this writing, however, no code, model weights, or HuggingFace release has been linked, and the training data is not described in the preprint, so independent reproduction and downstream adoption remain to be established.