decaf

DeCAF-Pearl: Denoiser Cofolding All-atom Flowmap Model

Distilling Pearl: Flow Maps for Fast All-Atom Cofolding

This repository hosts the DeCAF-Pearl checkpoint (decaf_ckpt.ckpt). The inference/evaluation code lives in the genesistherapeutics/decaf GitHub repository.

Overview

DeCAF is the first flow map model for all-atom cofolding. Instead of taking many steps along the denoising trajectory, a flow map learns to jump directly from one point on the trajectory to another, potentially traversing the entire generation process in just a handful of steps.

DeCAF-Pearl distills the Pearl cofolding foundation model into a fast few-step generator, achieving a 5x inference speedup with near-parity in structure prediction quality. Using over 5x fewer compute steps, DeCAF-Pearl exceeds AlphaFold 3, Chai-1, Boltz-1x, and Boltz-2 on the Runs N' Poses benchmark success rate by 3 to 15 percentage points.

Runs N' Poses (post-2023) success rate, best@5: DeCAF-Pearl nearly matches its full-budget teacher while outperforming AF3, Boltz-1x, Chai-1, and Boltz-2.

Usage

Download the checkpoint and run the bundled end-to-end example from the code repository:

# 1. get the checkpoint (requires `pip install huggingface_hub`)
hf download gianscarpe/decaf decaf_ckpt.ckpt --local-dir .

# 2. run few-step DeCAF inference on the bundled example
bash scripts/run_decaf_example.sh ./decaf_ckpt.ckpt

Or call the predictor directly:

python -m boltz.main predict <input.yaml> \
    --checkpoint ./decaf_ckpt.ckpt \
    --model boltz1 \
    --sampling_steps 10 \
    --diffusion_samples 5 \
    --recycling_steps 3 \
    --accelerator gpu \
    --out_dir ./out \
    --no_kernels

The code auto-detects this checkpoint and switches to the few-step DecafSampler (Detected Decaf checkpoint — using DecafSampler for inference.). See the repository's docs/decaf_prediction.md for full prediction and evaluation instructions.

Citation

@misc{scarpellini2026fewstepcofoldingallatomflow,
      title={Few-step Cofolding with All-Atom Flow Maps},
      author={Gianluca Scarpellini and Ron Shprints and Peter Holderrieth and Juno Nam and Pranav Murugan and Rafael Gómez-Bombarelli and Tommi Jaakola and Maruan Al-Shedivat and Nicholas Matthew Boffi and Avishek Joey Bose},
      year={2026},
      eprint={2606.08375},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2606.08375},
}

License

Released under the MIT License (Copyright (c) 2026 Genesis Molecular AI). The code extends the open-source Boltz project (Apache 2.0).

Acknowledgments

The research team at Genesis is grateful to our collaborators from Massachusetts Institute of Technology: Ron Shprints, Peter Holderrieth, Juno Nam, Rafael Gomez-Bombarelli and Tommi Jaakola; Carnegie Mellon University: Nicholas Matthew Boffi, and Joey Bose from Imperial College London and Mila.