foundry/models/rfd3/docs

De novo Design of Biomolecular Interactions with RFdiffusion3

RFdiffusion3 (RFD3) is a diffusion method that can design protein structures under complex constraints.

This repository contains both the training and inference code, and both are described in more detail below.

Note

Looking for config documentation? See here

Getting Started

1. Install RFdiffusion3. See Main README for instructions how to install all models to run full pipeline (recommended). If you have already installed all the models skip here.

pip install rc-foundry[rfd3]

2. Download checkpoint to your desired checkpoint location.

foundry install rfd3 --checkpoint-dir <path/to/ckpt/dir>

This sets FOUNDRY_CHECKPOINT_DIRS and will in future look for checkpoints in that directory (alongside the default ~/.foundry/checkpoints location), allowing you to run inference without supplying the checkpoint path. The checkpoint directory is optional, defaulting to ~/.foundry/checkpoints if unset.

Running Inference

To run inference (with foundry installed in your environment, or RFD3 & Foundry src in PYTHONPATH):

rfd3 design out_dir=logs/inference_outs/demo/0 inputs=models/rfd3/docs/demo.json skip_existing=False dump_trajectories=True prevalidate_inputs=True

Additional unnecessary args here are added:

• Including dumping and aligning trajectory structures can be useful for debugging your setup or making cool gifs.
• Printing the config and dumping trajectories are turned off by default, but turned on here for verbosity
• prevalidate_inputs will check that your inputs are valid before running inference. Helpful if your json has a number of different configs you want to debug / double check are valid before loading the checkpoints.
• Only out_dir and inputs are required. The output directory will automatically be created.

There are various interesting ways you can use RFD3 beyond Atom14 design as it's trained on a large array of different tasks. For example, you can fix sequence and not structure (prediction-type task), fix the backbone and unfix the sequence (MPNN-type inverse folding) or unfix the sidechains only (PLACER/ChemNet-style):

For full details on how to specify inputs, see the input specification documentation. You can also see foundry/models/rfd3/configs/inference_engine/rfdiffusion3.yaml for even more options.

Further example jsons for different applications

Additional examples are broken up by use case. If you have cloned the repository, matching .json files are in foundry/models/rfd3/docs that can be run directly, similar to the previous example.

In the examples the paths to the input files are specified assuming that you are running the examples from the foundry/models/rfd3/docs directory. If you would like to run RFD3 from a different location, you will need to change the path in the .json file(s) before running.

Nucleic acid binder design	Small molecule binder design	Protein binder design
Enzyme design	Symmetric design

Training and Fine-Tuning

We make available to the community not only the weights to run RFdiffusion3 but also the complete training code, easily extendable to additional use cases. Any AtomWorks-compatible dataset (and thus, any collection of structure files) can be readily incorporated and used for training or fine-tuning.

Dataset Configuration

PDB Training

To train on the PDB:

1. Set up PDB and CCD mirrors as described in the AtomWorks documentation
2. Update the path configs to point to the correct base directories for the metadata parquets
3. Set the PDB_MIRROR and CCD_PATH variables in your .env file

Custom Datasets

RFdiffusion3 supports arbitrary datasets of structure files for training and fine-tuning via AtomWorks. See the AtomWorks dataset documentation for details on creating custom datasets.

Running Training

After setting up Hydra configs, launch a training run:

uv run python models/rfd3/src/rfd3/train.py experiment=pretrain ckpt_path=<path/to/ckpt>

Supplying ckpt_path=null (default) will start with fresh weights. See the path configs to customize data input and log output directories.

Logging Configuration

Training runs support logging via Weights & Biases. To enable wandb logging:

uv run python models/rfd3/src/rfd3/train.py experiment=pretrain logger=wandb

To run training without wandb (default):

uv run python models/rfd3/src/rfd3/train.py experiment=pretrain logger=csv

Install HBPLUS for training with hydrogen bond conditioning:

1. Download hbplus from here: https://www.ebi.ac.uk/thornton-srv/software/HBPLUS/download.html (available for free)
2. Follow the installation instruction here: https://www.ebi.ac.uk/thornton-srv/software/HBPLUS/install.html
3. Update HBPLUS_PATH in foundry/.env file with the path to your hbplus executable.

Distributed Training

To use distributed training, you could use a command such as this (we use Lightning Fabric to handle ddp)

EFFECTIVE_BATCH_SIZE=16
DEVICES_PER_NODE= #INSERT NUMBER OF DEVICES PER NODE
NNODES = # INSERT NUMBER OF NODES
GRAD_ACCUM_STEPS=$((EFFECTIVE_BATCH_SIZE / (DEVICES_PER_NODE * NNODES)))
uv run python models/rfd3/src/rfd3/train.py \
    experiment=pretrain \
    trainer.devices_per_node=$DEVICES_PER_NODE \
    trainer.num_nodes=$SLURM_NNODES \
    trainer.grad_accum_steps=$GRAD_ACCUM_STEPS"

Notably, fabric must receive devices_per_node and the number of nodes (num_nodes) you're training on.

Dataset Paths: See the paths configs to customize the paths where data is read from and where logs are written. There is also a wandb config that can be enabled if you want to log training through wandb.

Hydra configs and experiments: In the example above, the experiment argument is a hydra-native argument. For RFD3, it will look for config overrides in /models/rfd3/configs/experiment/<experiment-name>.yaml and apply them on top of the base configs

Conditioning during training: RFD3 is trained on a multitude of conditioning tasks, and does so by randomly 'creating problems' for it to solve during training. For example, for a random training example it gets a random set of tokens to be 'motif tokens', then subsets those to whether specific atoms should be fixed, and further subsets the information to whether, say, sequence, coordinates or the sequence index should be fixed. It's pretty complicated to evaluate and how it was put together was more of an art than a science. There's likely still room for further optimization!

In models/rfd3/configs/datasets/design_base.yaml there's the shared configs for all datasets under global_transform_args. The dials that control the conditioning described above go under training_conditions, where for example tipatom - a specific preset conditioning sampler which more frequently fixes few tokens with few atoms - and others can be found.

Training with WandB: We strongly recommend tracking your runs via wandb. To use it, simply have your WANDB_API_KEY set and use the wandb logger. For more details see here

Appendix

Install HBPLUS for hydrogen bond conditioning:

One of the examples shows how to incorporate hydrogen bond conditioning into your designs. To make use of this feature, you will need to additionally complete the following steps:

1. Download hbplus from here: https://www.ebi.ac.uk/thornton-srv/software/HBPLUS/download.html (available for free)
2. Follow the installation instruction here: https://www.ebi.ac.uk/thornton-srv/software/HBPLUS/install.html
3. Update HBPLUS_PATH in foundry/.env file with the path to your hbplus executable.

Citation

If you use this code or data in your work, please consider citing:

@article {butcher2025_rfdiffusion3,
	author = {Butcher, Jasper and Krishna, Rohith and Mitra, Raktim and Brent, Rafael Isaac and Li, Yanjing and Corley, Nathaniel and Kim, Paul T and Funk, Jonathan and Mathis, Simon Valentin and Salike, Saman and Muraishi, Aiko and Eisenach, Helen and Thompson, Tuscan Rock and Chen, Jie and Politanska, Yuliya and Sehgal, Enisha and Coventry, Brian and Zhang, Odin and Qiang, Bo and Didi, Kieran and Kazman, Maxwell and DiMaio, Frank and Baker, David},
	title = {De novo Design of All-atom Biomolecular Interactions with RFdiffusion3},
	elocation-id = {2025.09.18.676967},
	year = {2025},
	doi = {10.1101/2025.09.18.676967},
	publisher = {Cold Spring Harbor Laboratory},
	URL = {https://www.biorxiv.org/content/early/2025/11/19/2025.09.18.676967},
	eprint = {https://www.biorxiv.org/content/early/2025/11/19/2025.09.18.676967.full.pdf},
	journal = {bioRxiv}
}