Two-sided partial diffusion with sequence gaps?

Start at the run_inference.py entry point and trace how contigmap.contigs and contigmap.provide_seq are parsed for partial diffusion. Reproduce the attached 9sps_binderA_targetBgaps.pdb example with the continuous and breaks configurations. Done means two-sided partial diffusion preserves both the input sequence gap and the original target chain label without post-processing.

Hello RFdiffusion devs! I'm trying to do something pretty specific in RFdiffusion, and I am hoping for some guidance. I want to do two-sided partial diffusion on a binder-target complex, where the target contains a gap in the sequence (missing residues). I've constructed a simple minimal example using a modified version of PDB ID 9SPS (attached) where I set the 'binder' as chain A, 'target' as chain B, and I deleted residues 80-100 from chain B.

$INPUT_PDB="9sps_binderA_targetBgaps.pdb"

9sps_binderA_targetBgaps.pdb

One-sided partial diffusion works fine and preserves the sequence gap. Of course, it helps here that the contig includes the chain label, so we can just specify chain B twice!

run_inference.py \
    "inference.input_pdb=$INPUT_PDB" \
    'inference.num_designs=3' \
    'diffuser.partial_T=15' \
    'contigmap.contigs=["124-124/0 B1-79/0 B101-195/0"]' \
    "inference.schedule_directory_path=$OUTPUT_DIR" \
    "inference.output_prefix=$OUTPUT_DIR/onesided"

onesided_0.pdb

I tried two different versions of the syntax for two-sided partial diffusion:

# version one - pretend the target is continuous
run_inference.py \
    "inference.input_pdb=$INPUT_PDB" \
    'inference.num_designs=3' \
    'diffuser.partial_T=15' \
    'contigmap.contigs=["124-124/0 174-174/0"]' \
    'contigmap.provide_seq=["124-297"]' \
    "inference.schedule_directory_path=$OUTPUT_DIR" \
    "inference.output_prefix=$OUTPUT_DIR/continuous"

# version two - include the sequence gap in the target
run_inference.py \
    "inference.input_pdb=$INPUT_PDB" \
    'inference.num_designs=3' \
    'diffuser.partial_T=15' \
    'contigmap.contigs=["124-124/0 79-79/0 95-95/0"]' \
    'contigmap.provide_seq=["124-297"]' \
    "inference.schedule_directory_path=$OUTPUT_DIR" \
    "inference.output_prefix=$OUTPUT_DIR/breaks"

continuous_0.pdb
breaks_0.pdb

Unsurprisingly, version one ("continuous") stitches together the gap in the target sequence, creating an unhelpful franken-target. Version two ("breaks") is more promising, except that it interprets the discontinuity as a chain break too, and so the target gets split into chain B before the break and chain C after the break.

Breaking up the provide_seq flag into two parts, e.g. 'contigmap.provide_seq=["124-202,203-297"]' \ produced the same output as version 2.

Other than the chain labels, version 2 gives me exactly what I'm looking for, so I could just do some post-processing to combine chains B and C back into chain B and call it a day. But, it would certainly be preferable to have this handled by RFdiffusion itself! Perhaps the syntax could support something like 'contigmap.provide_seq=["124-202/0 203-297"]' \.

To summarize, is there any way to do two-sided partial diffusion, while preserving the gaps in the input sequence and preserving the chain labels of the input sequence? Thank you!