ConforFlux

ConforFlux is an inference-time procedure for Boltz-2, introduced in Particle-Guided Trunk Repulsion for Diverse Protein Conformations. M parallel particles are coupled through a pairwise Cα-RMSD repulsion gradient, back-propagated to the trunk's single and pair embeddings. Each updated trunk is then decoded by Boltz-2's structure module as usual, so one input yields a diverse set of conformations instead of a single dominant prediction.

Installation

git clone https://github.com/suzuki-2001/conforflux
cd conforflux
pip install -e ./boltz
pip install -e .

The bundled boltz/ is a copy of Boltz-2 v2.2.1 with one addition, a guidance_hooks attribute on the model that the diffusion module consults during inference. Everything else upstream is unchanged.

Command line

conforflux predict forwards every flag to boltz predict, and adds a few options of its own.

conforflux predict input.yaml \
    --out_dir ./out \
    --num_particles 5 \
    --sigma 2.0 \
    --recycling_steps 3 --sampling_steps 200 \
    --output_format pdb

Flag	Default	Meaning
--num_particles	5	Coupled particles M. Replaces Boltz's --diffusion_samples.
--sigma	2.0	RBF kernel bandwidth on Cα RMSD (Å).
--alpha_s	0.02	RMS-normalised step size for the s_trunk update.
--alpha_z	0.02	RMS-normalised step size for the z_trunk update.
--start_frac	0.0	Diffusion-trajectory fraction at which guidance starts.
--stop_frac	0.8	Diffusion-trajectory fraction at which guidance stops.
--update_interval	3	Fire the gradient every K diffusion steps.
--gradient_checkpointing	off	Enable gradient checkpointing on the per-particle structure-module forward pass to reduce peak GPU memory. Off by default.

For broader conformational exploration, raise --sigma (e.g. 2.5–4 Å); for tighter clustering around the dominant fold, lower it.

Container

See container/README.md for Docker and Apptainer/Singularity usage.

Citation

If you use ConforFlux in your work, please cite our preprint.

@article{suzuki2026conforflux,
  title   = {ConforFlux: Particle-Guided Trunk Repulsion for Diverse Protein Conformations},
  author  = {Suzuki, Shosuke and Amagasa, Toshiyuki},
  journal = {bioRxiv},
  year    = {2026},
  doi     = {10.64898/2026.05.16.725138},
  url     = {https://www.biorxiv.org/content/10.64898/2026.05.16.725138v1},
  publisher = {Cold Spring Harbor Laboratory}
}