How RFdiffusion3 Achieves a Generative Breakthrough in Full‑Atom Protein Design

Recent generative deep‑learning methods have dramatically advanced protein design, but most—including the original RFdiffusion (RFD1) and BindCraft—operate at the amino‑acid‑residue level, limiting their ability to model precise side‑chain interactions with non‑protein components such as small‑molecule ligands or nucleic acids.

RFdiffusion2 (RFD2) partially addressed this limitation but still confined diffusion to the residue plane, preventing effective modeling of additional side‑chain contacts with external molecules.

To overcome these constraints, David Baker’s team released RFdiffusion3 (RFD3), a full‑atom diffusion model that explicitly represents every atom in polymers, enabling the generation of protein backbones and side chains within environments containing ligands, nucleic acids, or other non‑protein atoms. This explicit atom‑level representation simplifies the specification of atomic constraints and provides precise control over hydrogen bonds, ligand contacts, and nucleic‑acid interactions.

Unlike AlphaFold3, which relies on a computationally intensive 48‑layer Pairformer module to extract distance information from sequences, RFD3 redesigns the information‑extraction stage to be far lighter: the Pairformer depth is reduced to just two layers, cutting compute overhead dramatically while keeping the model compact at 168 million trainable parameters.

The authors validated RFD3 by designing a DNA‑binding protein and a cysteine hydrolase, demonstrating that the model can rapidly generate structures guided by complex atomic‑level constraints and thereby broaden the functional scope of protein design.

A step‑by‑step tutorial is hosted on the HyperAI website, allowing users to clone the tutorial repository, select an NVIDIA GeForce RTX 5090 GPU and a PyTorch environment, and launch a Jupyter workspace with a single click to reproduce the demonstrated designs.