Interface Guide
What interface prediction (docking) is, every metric's definition and how to read it, and the command line.
Interface prediction generates a 3D antibody–antigen complex pose on internal GPU. From that model it extracts where the chains touch (epitope / paratope + contact pairs) and describes the pose with confidence and geometry metrics. It is a structural hypothesis for review, complementary to sequence-based models and experiments.
Read the metrics together, never in isolation. Agreement among confidence and geometry signals makes a predicted pose more internally consistent, but does not establish binding, affinity, or specificity without orthogonal evidence.
For CD19-BCMA or other dual-target CARs, use the dedicated `/car-interface-assessment` pilot module instead of submitting the full CAR and both antigens as one unconstrained complex. Phase 1 runs a 2 x 2 matrix: arm 1 vs antigen 1, arm 1 vs antigen 2, arm 2 vs antigen 1, and arm 2 vs antigen 2.
This compares intended-pair and counter-screen pose signals; it does not by itself prove intended binding or cross-binding. Full tandem-CAR geometry, linker reach, hinge/TM constraints, and membrane orientation remain separate modeling problems.
ipTM
Model confidence (0–1) in the predicted relative arrangement of the chains.
How: Interface predicted TM-score from the complex model, restricted to inter-chain residue pairs.
Read: Useful for ranking predicted poses, but not a probability of binding or specificity. A non-binder can still receive a confident modeled interface.
pTM
Overall fold confidence (0–1) for the whole predicted structure.
How: Predicted TM-score over all residue pairs.
Read: Higher values indicate greater predicted global confidence, but interpretation is model- and context-dependent. It is less interface-specific than ipTM.
pLDDT
Mean per-residue confidence (0–100).
How: Averaged from the model's per-residue confidence (stored in the B-factor column).
Read: ≥90 very high and 70–90 confident for the modeled local structure; <50 is low model confidence and can be compatible with disorder. See interface pLDDT for the interface specifically.
Interface pLDDT
Mean pLDDT over just the interface residues.
How: For pDockQ, averaged over residues participating in a cross-chain Cβ contact within 8 Å (Cα for glycine).
Read: More local than global pLDDT, but still describes model confidence rather than experimental binding.
pDockQ
A fitted predicted-DockQ score for a generated two-chain pose; higher is better within its calibration.
How: 0.724 / (1 + e^(−0.052·(x−152.6))) + 0.018, where x = interface pLDDT × log10(unique cross-chain Cβ/Cα residue contacts within 8 Å) (Bryant et al. 2022).
Read: Predicts the quality of a generated pose relative to DockQ; it is not an affinity, binding, or specificity score.
Buried area (BSA / ΔSASA)
Interface size in Ų — how much surface the two chains bury against each other.
How: (SASA of antibody alone + SASA of antigen alone − SASA of the complex) / 2, via Shrake-Rupley solvent-accessible surface area.
Read: Describes the size of the generated interface. A larger modeled BSA does not by itself imply tighter or real binding.
Hydrogen bonds
Cross-chain donor–acceptor pairs identified by distance geometry in the generated pose.
How: A no-explicit-hydrogen distance heuristic: a donor candidate (N or hydroxyl O) and an N/O acceptor candidate 2.4–3.6 Å apart.
Read: Use the list to inspect pose geometry. The count alone does not establish affinity or specificity; donor/acceptor candidates are listed by `hbonds`.
Salt bridges
Charged-pair contacts (Arg/Lys/His ↔ Asp/Glu) in the generated pose.
How: Cationic N ↔ anionic O within 4.0 Å.
Read: Inspect as distance-geometry candidates. Their count alone does not quantify electrostatic energy, specificity, or pH sensitivity.
Clashes
Steric overlaps between heavy atoms across the interface.
How: Heavy-atom pairs closer than (rᵥ𝒹ᵂ,A + rᵥ𝒹ᵂ,B − 0.6 Å).
Read: A few are normal; many flag a poor/strained model — discount that prediction.
Epitope / Paratope residues
Antigen residues (epitope) and antibody residues (paratope) at the interface.
How: Residues with ≥1 heavy atom within the contact cutoff (default 4.5 Å) of the other chain.
Read: Compare the modeled footprint with an explicit CDR annotation; a framework-heavy footprint is a manual-review flag, not an automatic rejection.
The command line under a result runs over the computed metrics and contacts, echoing ChimeraX-style output to the log. Command history with ↑/↓.
| metrics | all interface metrics at a glance |
| hbonds | hydrogen bonds (donor, acceptor, distance) |
| saltbridges | salt | salt bridges |
| contacts [N] | top-N contact pairs (default 20) |
| buriedarea | bsa | buried surface area breakdown |
| clashes | steric clashes |
| epitope | paratope | interface residues on each side |
| help / clear | list commands / clear the log |