RFdiffusion: When Diffusion Models Meet Protein Design

Original: Watson et al., Nature 2023
DOI: 10.1038/s41586-023-06415-8

1. Background: From Structure Prediction to Protein Design

Protein design is the "inverse problem" of structural biology: while structure prediction infers structure from sequence, protein design starts from functional requirements to create sequences that fold into specific structures. The complexity lies in the vast sequence space (20^n, where n is the number of amino acids), with functional foldable proteins occupying only a tiny fraction.

In the deep learning era, protein structure prediction has made substantial progress. Models like AlphaFold2 and RoseTTAFold can predict protein structures with near-experimental accuracy. These models encode deep understanding of protein structures, opening new possibilities for protein design.

However, adapting structure prediction models for design faces two core challenges:

Diversity Problem

Traditional methods (e.g., Rosetta) use deterministic search, producing limited solutions for given design constraints. Protein design typically requires exploring numerous candidates to find solutions satisfying multiple constraints.

Constraint Satisfaction Problem

Many design tasks require only partial structural information (e.g., functional site coordinates), with overall folding to be inferred by the algorithm. Existing methods often fail on such "under-constrained" problems.

Diffusion models offer a new approach to address these issues. These models have demonstrated remarkable capabilities in image generation, producing high-quality, diverse images through iterative denoising from Gaussian noise. Applying diffusion models to protein design could theoretically solve both diversity and constraint satisfaction: random noise initialization ensures diversity, while the iterative denoising process gradually establishes structural constraints.

2. Method: Diffusion-izing RoseTTAFold

2.1 Core Technical Approach

RFdiffusion's key innovation lies in transforming the RoseTTAFold (RF) structure prediction network into a denoising network for diffusion models. This transformation is based on key observations: RoseTTAFold possesses high-precision structure generation capability, rotational equivariance, and multi-level conditioning mechanisms—properties that make it well-suited as a foundation for diffusion models.

2.2 Training Process

• Data preparation: Sampling structures from the Protein Data Bank (PDB) with up to 200 steps of noise addition.
• Loss function: Using Mean Squared Error (MSE) loss rather than FAPE to promote global coordinate frame continuity across timesteps.
• Self-conditioning: Allowing the model to condition on previous predictions between timesteps, significantly improving performance.
• Transfer learning: Fine-tuning from pre-trained RoseTTAFold weights outperforms training from scratch.

2.3 Generation Pipeline

1. Initialization: Random initialization of residue frames (Cα coordinates and N-Cα-C rigid orientations)
2. Iterative denoising: RFdiffusion performs denoising predictions, updating each residue frame along the predicted direction
3. Sequence design: Using the ProteinMPNN network to design sequences for generated structures
4. Validation: Verifying foldability through AlphaFold2 or ESMFold single-sequence prediction

3. Experimental Results and Performance Analysis

3.1 Unconditional Monomer Design

RFdiffusion can generate complex protein structures from scratch, covering various topologies including α-helices, β-sheets, and α/β mixed structures. Experimentally validated designs show circular dichroism spectra consistent with the designs and exhibit exceptional thermal stability.

3.2 Protein Binder Design

In protein binder design tasks, RFdiffusion achieves approximately two orders of magnitude improvement in success rate compared to traditional Rosetta methods (19% vs ~0.1%).

Key Results:

• For 5 different targets, testing fewer than 100 designs per target, successful binders were obtained for all targets
• Multiple designs showed nanomolar affinity (e.g., HA and IL-7Rα binders at ~30 nM)
• Cryo-EM validation: Designed influenza hemagglutinin binders were determined by cryo-EM, showing near-perfect agreement with design models

3.3 Symmetric Oligomers and Functional Site Scaffolds

RFdiffusion supports symmetry constraints, enabling design of C3, C4, C5, and C6 symmetric oligomer structures. Application cases include SARS-CoV-2 spike protein binder design and metal-binding protein design, with experimental validation showing high consistency with design models.

4. Technical Significance and Impact

4.1 Methodological Contributions

• Architecture reuse: Powerful structure prediction architectures can be directly adapted for generation tasks
• Loss function design: MSE loss is more suitable for training diffusion models compared to FAPE
• Self-conditioning mechanism: Information transfer between timesteps is crucial for generation quality

4.2 Impact on Subsequent Research

RFdiffusion laid the foundation for subsequent protein design models. Chai-2 (2025) adopts a similar diffusion model architecture, achieving 16% experimental success rate in antibody design tasks—more than 100× improvement over previous methods.

4.3 Limitations and Open Problems

• Computational cost: Diffusion models require multiple iterative denoising steps, resulting in higher computational costs
• Joint sequence-structure design: RFdiffusion focuses primarily on backbone generation, with sequence design relying on external tools
• Experimental validation scale: The cost of large-scale experimental validation remains high

5. Conclusion

RFdiffusion represents an important paradigm shift in protein design from "search" to "generation." By introducing diffusion models to protein structure generation, the method achieves substantial progress across multiple dimensions including diversity, constraint satisfaction capability, and experimental success rate. The cryo-EM validation of designed binder structures marks a milestone where computationally designed proteins can achieve atomic-level accuracy. With further breakthroughs by subsequent models like Chai-2 on specific tasks, diffusion-based protein design is becoming an important technical approach in the AIDD field.

References

Watson, J.L., et al. (2023). De novo design of protein structure and function with RFdiffusion. Nature, 620, 1089-1100. https://doi.org/10.1038/s41586-023-06415-8

Code: https://github.com/RosettaCommons/RFdiffusion