Partial diffusion and inverse folding#

How to start a run from a structure you already have instead of from noise. See Partial diffusion and inverse folding first if you haven’t: it explains what these do and what partial_t means, which you’ll want before picking a value below.

Partial diffusion#

Pass your structure with inputs= like any other run, then tell the model how much of it to keep versus redesign with a Hydra override under generation.args. There are two ways to set that, and you only need one.

Simple — pick this unless you have a reason not to. generation.args.partial_trajectory_fraction_steps=<f> runs that fraction of the total sampling steps (nsteps). The count is rounded to the nearest integer (at least one step is required). 0.3 runs the last 30% of the trajectory, so the model does less work and stays closer to your input; 0.8 runs most of it, so the output drifts further. Corruption and denoising use the same tail of the configured full schedule, including its late refinement. For example, nsteps=60 with f=0.3333 runs the final 20 intervals, starting at index 40 of that schedule. A restart index is not a diffusion time: changing the schedule or its step count can change the corruption level at a given fraction.
Advanced — reach for this when the backbone and sequence need to change at different rates. Set generation.args.partial_t_backbone=<t> and generation.args.partial_t_latent=<t> separately, each a float in (0, 1) (not a fraction of steps — see the concept page). Only setting one and adding generation.args.couple_partial_t=true derives the other from the default schedule instead of requiring both.

For example, to keep both the fold and the sequence close to the input, regenerating only the last 20% of the trajectory:

rfd4-generate \
    inputs=my_structure.cif \
    generation.args.partial_t_backbone=0.8 generation.args.partial_t_latent=0.8 \
    num_replicates=8 out_dir=./runs/partial

Inverse folding#

Use this when you want to keep a backbone exactly as given and only get new sequences for it — no backbone noise at all, just resampled sequence and side chains.

Simple — start here. Add the flag generation=inverse_folding and pass your backbone with inputs=. That’s the whole setup.
Advanced. Tune sequence temperature with generation.model.local_latents.simulation_step_params.tsr_k (default 4). Lower k raises temperature; tsr_sigma (default 12) provides secondary control over when sharpening takes effect.

rfd4-generate \
    inputs=my_backbone.cif \
    generation=inverse_folding recycle_strategy=medium/flat \
    generation.model.local_latents.simulation_step_params.tsr_k=2 \
    num_replicates=8 out_dir=./runs/inverse_folding