Cyclic BindCraft · v1.0
A modified BindCraft pipeline that closes the N- and C-termini of designed binders into a head-to-tail amide bond — porting RFpeptide's cyclization encoding onto AlphaFold2 hallucination.
Authors: Ali Tarkashvand, Amirali Roohi
Built on BindCraft v1.5.3 · Cyclization ported from RFpeptide · BSD-3-Clause
What it is
Cyclic BindCraft keeps the BindCraft pipeline intact and changes five places — the relative-position encoding, the bonded-edge feature, the termini loss, FastRelax, and the filters.
Stock BindCraft designs linear binders of 70–150 residues by
hallucinating an AlphaFold2 structure module through four
backpropagation stages. Cyclic BindCraft keeps that pipeline
but adds a cyclic wrap so residues 1 and L
on the binder chain are seen as sequence-adjacent — mirroring
RFpeptide's cyclize_relpos trick.
A new cyclic_reses mask, a cyclic bonded edge
between residues L and 1, and a hard
harmonic loss on the real amide-bond geometry (1.33 Å N–C,
3.80 Å Cα–Cα, ω = 180°) replace the soft
add_termini_distance_loss hook.
A coordinate-constrained FastRelax with DeclareBond
closes the ring chemically; CONECT records and a
REMARK 470 CYCLIC line make the topology explicit on
disk. Two new JSON presets, a --cyclic CLI flag, and
a best-effort AfCycDesign / RF2-cyclic validator hook ship with
the patch.
Why cyclic binders
Linear peptides are convenient to compute but poor drugs — they are degraded within minutes, fail to cross membranes, and pay a large entropic penalty on binding. Head-to-tail cyclization addresses all four problems at once.
Exopeptidases require a free N- or C-terminus. A cyclic peptide has neither; endopeptidase access is further reduced by ring rigidity. Empirically, cyclic peptides show 10–100× longer serum half-lives than their linear counterparts.
A linear peptide samples an ensemble; on binding it adopts a
single conformer, paying T·ΔS_conf ≈ 1–3 kcal/mol
for a 15-mer. Ring closure sharply reduces the conformational
space, so the entropic cost of binding is smaller.
For ordered binders, ΔG = ΔH − T·ΔS. Reducing
|T·ΔS_unfavorable| directly improves ΔG — the same
principle as preorganization in host–guest chemistry. Apparent
K_D improves by 1–3 orders of magnitude.
Cyclic peptides can be orally bioavailable. Cyclosporin A — an 11-aa cyclic peptide with 7 N-methylations — is the canonical example. Cyclization reduces solvent-exposed polar surface and promotes the chameleonic conformer needed for passive permeation.
The pipeline
Each iteration of the outer loop is one trajectory: sample a binder length → AF2 hallucination with the cyclic relative positional encoding, bonded edge, and bond loss → constrained FastRelax → ProteinMPNN → AF2 re-prediction + cyclic-aware filters.
load_json_settings reads three JSONs;
binder_hallucination builds the AF2 model and the
cyclic_reses mask via
build_cyclic_reses_mask — a 1-D bool tensor
True for every residue on the binder chain.
3-stage design (design_logits → design_soft →
design_pssm_semigreedy) with cyclized relpos,
a cyclic bonded edge between binder residues L and
1, and a hard add_cyclic_bond_loss on
the real amide-bond geometry.
pr_relax_cyclic calls DeclareBond
between res(L):C and res(1):N, adds
bond-length, bond-angle, and dihedral constraints, then runs
FastRelax with the chainbreak term enabled.
mk_mpnn_model(weights="soluble") with
fix_pos on the target + binder interface; MPNN sees
a single chain B — cyclic geometry is implicit in the backbone,
so MPNN needs no modification.
AF2 monomer re-prediction (with cyclic relpos), PyRosetta
InterfaceAnalyzerMover, then
run_cyclic_filters appends
cyclic_NC_distance, cyclic_omega, and
cyclic_CACA_distance to every design's metric dict.
When #Accepted ≥ number_of_final_designs, the loop
exits and writes final_design_stats.csv sorted by
Average_i_pTM descending (tie-broken by closure
distance when cyclic mode is active).
Stock vs. cyclic
The patch touches nine files. When cyclic: false, the
behavior is byte-for-byte identical to upstream BindCraft v1.5.3 —
nothing is removed, only added.
| Aspect | Stock BindCraft | Cyclic BindCraft |
|---|---|---|
| Binder topology | Linear, chain B with free N- and C-termini | Cyclic, chain B with N(L)–C(1) peptide bond |
| Length range | 65–150 aa typical | 25–55 aa recommended |
| AF2 relative positional encoding | Standard (linear residue index) | Cyclic wrap on binder chain (RFpeptide trick) |
| AF2 bonded-edge feature | +1 only for (i, i+1) | Adds +1 for (L, 1) on binder chain |
| Termini loss | add_termini_distance_loss (ELU, 7 Å, off by default) | add_cyclic_bond_loss (harmonic, 1.33 Å target, on) |
| FastRelax | MoveMap(CHI+BB, jump=off), no closure | DeclareBond + bond/dihedral constraints + chainbreak |
| AF2 validation | AF2 monomer models 1, 2 (sees linear sequence) | Same AF2, with cyclic relpos applied; optional AfCycDesign / RF2-cyclic |
| Clash exclusion | abs(res_i − res_j) == 1 | Adds {res_i, res_j} == {1, L} exclusion |
| PDB output | ATOM / HETATM / MODEL / TER / END / LINK | Adds CONECT + REMARK 470 CYCLIC |
| Filter preset | default_filters.json | peptide_cyclic_filters.json with cyclic_NC_distance, cyclic_omega |
| Recommended preset | default_4stage_multimer.json | peptide_cyclic_3stage.json |
| CLI flags | --settings --filters --advanced | Adds --cyclic, --cyclic_validator |
Key features
The whole cyclic machinery fits in ~250 lines of Python on top of
stock BindCraft v1.5.3 — every change is gated by
if advanced_settings.get("cyclic", False):.
install_cyclic_relpos_patch monkey-patches
colabdesign's AF2 model with a "pre" callback that wraps the
binder's relative-position tensor by ±L — the direct port of
RFpeptide's PositionalEncoding2D.forward.
add_cyclic_bond_loss replaces the soft
7 Å ELU pull with a quadratic penalty on N(1)–C(L) distance,
Cα–Cα distance, and ω dihedral — targeting the real trans amide
geometry (1.33 / 3.80 / 180°).
declare_cyclic_bond creates the covalent N–C edge;
add_cyclic_constraints adds
AtomPairConstraint and DihedralConstraint
(HarmonicFunc + PeriodicFunc) to lock the closing amide during
FastRelax.
calculate_clash_score now pre-computes the
{first_res, last_res} frozenset of the binder chain
and skips those atom pairs — otherwise the 1.33 Å N–C amide bond
always triggers a false clash.
clean_pdb_with_conect preserves CONECT/REMARK
records and re-derives the cyclic CONECT line via
cyclic_bond_conect_records — PyMOL, ChimeraX, and
Rosetta now draw the cyclic bond visually.
peptide_cyclic_3stage.json (3-stage, helix bias
0.5, 100/50/20 iters, MPNN num_seqs=20 at temp 0.05)
and peptide_cyclic_filters.json (relaxed RMSD,
tightened dG, three new cyclic closure gates).
Install in ~30 minutes. Design 50 PD-L1 cyclic binders in ~4 hours on a single A100.