TDFold

TDFold is a single-sequence protein structure prediction method developed by researchers at the School of Artificial Intelligence, Beijing Normal University, and collaborators. It was first posted to bioRxiv in July 2025, updated in February 2026, and subsequently published in Nature Machine Intelligence in 2026. The method addresses the long-standing difficulty of predicting structure for proteins that lack rich evolutionary information: tools such as AlphaFold2 rely on multiple sequence alignments (MSAs), and their accuracy degrades for orphan proteins, designed sequences, and fast-evolving families where deep alignments are unavailable. TDFold instead works directly from a single sequence.

The key conceptual move is to treat inter-residue geometry as an image. TDFold re-purposes a powerful text-to-image diffusion model (Stable Diffusion) to generate high-quality two-dimensional geometric templates—matrices of inter-residue distances and orientations—from a protein sequence. Because pairwise geometry is naturally represented as 2D feature maps, image-generation machinery can be adapted to produce it. These generated templates then guide three-dimensional structure prediction, providing the geometric constraints that an MSA would otherwise supply.

TDFold's pipeline has two stages: a template diffusion stage that produces the inter-residue geometric templates, followed by a Sequence-template Co-evolved Learning (SCL) stage that fuses the templates with the sequence to predict the final 3D structure. The result is a method that the authors report is both more accurate than existing single-sequence approaches on low-homology targets and dramatically more efficient than mainstream folding tools.

#Key Features

• Image-diffusion-based geometry generation: TDFold adapts Stable Diffusion to generate 2D inter-residue distance and orientation templates, reframing geometry prediction as a conditional image-generation task.
• Single-sequence prediction: The method requires no MSA, making it suitable for orphan proteins, de novo designs, and fast-evolving sequences where alignments are shallow or absent.
• Sequence-template Co-evolved Learning (SCL): A lightweight co-evolved graph network fuses a residue-level branch (sequence and inter-residue correlations) with an atom-level branch (side-chain influence on backbone conformation) to recover 3D structure.
• Low compute footprint: TDFold can be trained within about a week on a single consumer NVIDIA 4090 GPU and uses less memory than AlphaFold2 or ESMFold.
• Fast inference: For long proteins, TDFold is roughly 10× to 100× faster than ESMFold, AlphaFold2, and AlphaFold3—about 10 seconds for a 500-residue protein versus ~100 seconds for ESMFold and ~240 seconds for AlphaFold3.

#Technical Details

TDFold is an end-to-end, two-stage architecture. The first stage is a diffusion model adapted from Stable Diffusion that generates high-quality inter-residue geometric templates—2D matrices encoding pairwise distances and orientations—conditioned on the input sequence. The second stage, Sequence-template Co-evolved Learning, is a lightweight co-evolved graph network with two branches: a residue-level branch that models sequence and inter-residue correlations, and an atom-level branch that captures how side-chain atoms influence backbone conformation; fusing the two yields the predicted 3D structure. The authors report superior single-sequence prediction accuracy relative to protein language model and homology-based methods on limited-homology datasets, alongside large efficiency gains—training feasible on a single RTX 4090 within a week, lower GPU memory than AlphaFold2 and ESMFold, and roughly 10–100× faster inference on long sequences. The bioRxiv preprint is released under a CC BY-NC license.

#Applications

TDFold is most useful where evolutionary information is scarce and where compute budgets are limited. It is well suited to predicting structures for orphan proteins and singleton sequences, de novo designed proteins that have no natural homologs, and rapidly evolving families such as some viral and immune proteins. Its low memory footprint and fast inference make high-throughput single-sequence folding practical on a single consumer GPU, lowering the barrier for labs without large compute clusters and enabling rapid screening of designed or mutated sequences during protein engineering campaigns.

#Impact

TDFold demonstrates that image-generation diffusion models can be productively repurposed to produce inter-residue geometry, offering an efficient route to single-sequence folding that sidesteps the MSA dependence of MSA-based predictors. Its progression from bioRxiv preprint to publication in Nature Machine Intelligence reflects peer-reviewed validation of its accuracy and efficiency claims on low-homology targets. The method's emphasis on accessibility—training and inference on a single consumer GPU—broadens who can run accurate single-sequence structure prediction. As with any single-sequence method, performance on targets that do have deep alignments may not surpass strong MSA-based predictors, and independent benchmarking will help delineate where its advantages are largest.