RFDiffusion — NA-Binding Protein Design

This is a fork of RFDiffusion modified for the design of nucleic acid-binding proteins.

Key modifications

1. Nucleic acid support — added handling for DNA/RNA chains with dedicated auxiliary potentials based on the substrate_contacts framework.

2. Rigid-body averaging — potentials computed per residue can now be averaged across the protein chain and applied as collective impulses for translation and rotation of the entire chain as a rigid body.

3. Px0-based potential calculation — instead of computing potentials on the noisy intermediate structure (xt), potentials are evaluated on px0 (the extrapolated structure used for self-conditioning). Since px0 is closer to a natural structure, this provides more accurate protein-NA interaction estimates.

Usage

# Basic design with NA-binding auxiliary potentials
python ./scripts/run_inference.py \
    inference.output_prefix=examples/example_outputs/design_dsdna \
    inference.input_pdb=input_pdbs/5EDN.pdb \
    "contigmap.contigs=[70-130]" \
    potentials.guiding_potentials=["type:na_contacts,s:2,r_0:8,rep_r_0:7,rep_s:2,d_0:2,rep_r_min:1,smooth=0.6,predicted:1"] \
    potentials.guide_scale=1

For full documentation, see the original RFDiffusion README.

References

• Original RFDiffusion: Watson et al., Nature 2023
• This fork: Geraseva et al., JCIM 2026