Designing a cyclic binder for PD-L1.

Download 5JDS, extract chain A, strip waters and ligands:

mkdir -p ~/pdl1_cyclic && cd ~/pdl1_cyclic
wget -q https://files.rcsb.org/download/5JDS.pdb

# Keep only chain A, ATOM records; renumber from 1
python - <<'PY'
from Bio.PDB import PDBParser, PDBIO, Select
class ChainA(Select):
    def accept_chain(self, chain): return chain.id == 'A'
    def accept_model(self, model): return model.id == 0
    def accept_residue(self, r):
        return r.id[0] == ' ' and r.resname in {
            'ALA','ARG','ASN','ASP','CYS','GLN','GLU','GLY','HIS','ILE',
            'LEU','LYS','MET','PHE','PRO','SER','THR','TRP','TYR','VAL'}
io = PDBIO(); io.set_structure(PDBParser(QUIET=True).get_structure('p','5JDS.pdb'))
io.save('PDL1_clean.pdb', ChainA())
PY

Inspect the cleaned structure in PyMOL or py3dmol. The PD-1-binding epitope on PD-L1 spans the C-C' loop and the C strand. Choose three hotspots that flank the binding site:

• A48 — Tyr at the C-C' loop tip
• A53 — Asp in the C' strand (PD-1 contact)
• A66 — Tyr at the start of the C'' strand

These residues receive extra AF2 attention via ColabDesign's hotspot mask. The binder is not required to contact all three, but most passing designs will.

Cyclic BindCraft ships a new preset at settings_advanced/peptide_cyclic_3stage.json. It is the closest analog to stock BindCraft's peptide_3stage_multimer.json with cyclic-specific tweaks. The most important keys:

The filter preset at settings_filters/peptide_cyclic_filters.json extends stock peptide_filters.json with cyclic-specific columns:

The cyclic columns are computed by compute_cyclic_closure_metrics() in functions/cyclic_utils.py and appended to every design's metric dict before the JSON filter engine runs.

Create a target-settings JSON specific to PD-L1 (settings_target/PDL1_cyclic.json):

{
  "design_path": "/home/USER/pdl1_cyclic/",
  "binder_name": "PDL1cyc",
  "starting_pdb": "/home/USER/pdl1_cyclic/PDL1_clean.pdb",
  "chains": "A",
  "target_hotspot_residues": "A48,A53,A66",
  "lengths": [25, 45],
  "number_of_final_designs": 50
}

Note lengths: [25, 45] — we cap at 45 because PD-L1's epitope is small; a 25–45 aa cyclic binder covers it efficiently. Launch:

cd ~/cyclic-bindcraft/repo
python -u bindcraft.py \
  --cyclic \
  --target   /home/USER/pdl1_cyclic/PDL1_clean.pdb \
  --hotspots "A48,A53,A66" \
  --settings settings_target/PDL1_cyclic.json \
  --filters  settings_filters/peptide_cyclic_filters.json \
  --advanced settings_advanced/peptide_cyclic_3stage.json \
  --output   /home/USER/pdl1_cyclic/

Stock BindCraft takes --settings / --filters / --advanced only; the cyclic patch adds the optional --cyclic, --target, --hotspots, and --output shortcuts that override the JSON.

In a separate terminal:

cd /home/USER/pdl1_cyclic/
watch -n 30 'ls -1 Trajectory/ | wc -l; ls -1 Accepted/ | wc -l; tail -n 5 failure_csv.csv'
tail -F /home/USER/pdl1_cyclic/*.log 2>/dev/null

What appears, in order:

1. Trajectory/<binder>_l<len>_s<seed>.pdb — raw AF2 hallucination (chain A target, chain B binder, no closure yet)
2. Trajectory/Relaxed/<...>.pdb — PyRosetta-relaxed trajectory (cyclic bond declared; CONECT record present)
3. Trajectory/LowConfidence/<...>.pdb — terminated early (pLDDT < 0.7 or <3 hotspot contacts)
4. Trajectory/Clashing/<...>.pdb — Cα clash detected
5. MPNN/Binder/<...>_model1.pdb, _model2.pdb — AF2 re-prediction per MPNN sequence
6. MPNN/Relaxed/<...>_model1.pdb — Rosetta-relaxed complex
7. Accepted/<...>_model<N>.pdb — design that passes all filters
8. Rejected/<...>_model<N>.pdb — design that failed filters
9. trajectory_stats.csv, mpnn_design_stats.csv, final_design_stats.csv, failure_csv.csv

Each trajectory takes ~3 minutes on an A100. With acceptance_rate: 0.1 and start_monitoring: 1000 you should see the first accepted design around trajectory #50 (≈2.5 hours). To reach 50 accepted designs, expect ~4 hours total.

Open final_design_stats.csv. The most important columns:

The cyclic_* columns are the ones that distinguish this from a stock BindCraft CSV.

Sort by Average_i_pTM descending, then re-rank by cyclic_NC_distance ascending. Apply these selection criteria:

1. Top quartile by Average_i_pTM.
2. Cyclic closure: cyclic_NC_distance < 1.4 Å and cyclic_omega ∈ [170°, 190°].
3. AF2 agreement: Average_Binder_RMSD < 1.5 Å.
4. Rosetta quality: Average_dG < -25 kcal/mol and Average_SAP < 30.
5. Interface size: Average_n_InterfaceResidues >= 6.
6. Diversity: Pick at most one design per binder-length bucket (e.g. one 30-mer, one 35-mer, etc.) to maximize sequence diversity.

import pandas as pd
df = pd.read_csv('/home/USER/pdl1_cyclic/final_design_stats.csv')
df = df[(df.cyclic_NC_distance < 1.4) &
        (df.cyclic_omega.between(170, 190)) &
        (df.Average_Binder_RMSD < 1.5) &
        (df.Average_dG < -25) &
        (df.Average_SAP < 30)]
df = df.sort_values(['Average_i_pTM','cyclic_NC_distance'],
                    ascending=[False, True])
# keep at most one per length bucket
top = df.groupby('Length').head(1).head(5)
top.to_csv('top5.csv', index=False)
print(top[['Design','Length','Average_i_pTM','Average_dG',
           'cyclic_NC_distance','cyclic_omega']])

A realistic output after a 4-hour run on 1×A100:

These are your top 5. Five is the right number for a first synthesis round.

AF2-with-cyclic-relpos is an approximation of a true cyclic-prediction network. For designs you intend to synthesize, run an independent cyclic-aware predictor:

AfCycDesign

# Install once
pip install git+https://github.com/bhardwajlab/AfCycDesign.git

# Re-predict each top-5 design's binder+target complex
for design in top5_*.pdb; do
  python -m afcycdesign.predict \
    --target PDL1_clean.pdb \
    --binder $design \
    --cyclic_chain B \
    --output ${design%.pdb}_afcyc.pdb
done

Acceptance criteria: AfCycDesign iPAE < 0.3 (PD-L1 is moderate-difficulty; for harder targets use <0.4) and RMSD < 1.5 Å between the BindCraft design and the AfCycDesign prediction.

RF2-cyclic (orthogonal)

If both AfCycDesign and RF2-cyclic agree, you have high confidence. Disagreement is a strong signal to drop the design. Typically 3 of 5 survive this cross-validation.

Send the surviving 2–3 designs to a peptide-synthesis core. The chemistry is standard:

1. Fmoc SPPS of the linear protected peptide on 2-chlorotrityl chloride resin (so the C-terminus is cleaved as a free acid under mild acid).
2. Cleavage from resin with 1% TFA in DCM (preserves side-chain protecting groups).
3. Head-to-tail cyclization in dilute solution (1 mM peptide in DMF) with PyAOP / DIEA (1.5 eq) for 2–4 h at RT — this forms the N→C amide bond.
4. Global deprotection with TFA/TIS/water (95/2.5/2.5) for 2 h.
5. Purification by reverse-phase HPLC; confirm by MALDI-TOF (mass should equal sum(residue_masses) − 18.02 for the lost water in the new amide bond).
6. Affinity by SPR against immobilized PD-L1 (target: K_D < 1 µM for first round; <100 nM after one round of affinity maturation).
7. Stability in human serum at 37°C (cyclic peptides should retain >80% after 24 h, vs. minutes for linear controls).

Expect ~1 in 3 designs to show measurable binding in the first round. Cyclic peptides that pass the design filters and the AfCycDesign cross-validation typically have 1–10 µM K_D; one round of point-mutation affinity maturation gets them to <100 nM.