binderqc

Quality control and tag-site scoring for designed protein binders.

Install

pip install binderqc

Python 3.10+. Pulls in biotite, numpy, and pandas. To develop from source, pip install -e ".[test]".

Usage

binderqc --binder-chains A --target-chains B --out out.csv complex.cif some_dir/

from binderqc import score_structure
rows = score_structure("complex.pdb", binder_chains=["A"], target_chains=["B"])

Inputs are PDB/CIF files, globs, or directories. Leave --binder-chains off to guess the binder as the shortest chain (20-250 aa, printed for each file). --target-chains defaults to the remaining chains.

Scoring is CPU-only: no folding, no GPU, no network. Files in a batch are independent, so -j scales near-linearly across cores on large directories.

Example output for the bundled LCB1 minibinder (a few of the columns):

Its warnings field reads "both termini ~equidistant from interface (ambiguous)", a tag-site advisory, so qc_pass stays True.

What it reports

Per binder chain:

• Interface: buried surface area, interface residue count, hydrogen-bond and salt-bridge counts, and a contact-packing density (a lightweight proxy for contact molecular surface).

• Self-association (with --self-fold): given a predicted binder homodimer, how much of the paratope the self-interface buries (self_paratope_overlap_frac) and whether it is more than a patch that size would bury sitting anywhere on the binder at random (self_paratope_enrichment). Interface PAE, ipTM and ipSAE say whether a self-interface is predicted; they cannot say where. Self-association away from the paratope is a formulation problem, self-association through it competes with target binding. self_verdict bands the buried fraction at 0.5 and 0.2 — a stated convention, not a calibrated cutoff, since no dataset of de-novo binders with measured self-association exists to fit one on.

• Pose: approach angle (end-on vs. lying across the surface).

• Grippability: epitope planarity, hydrophobic fraction, aromatic anchors, and a SASA-aware glyco-occlusion check (epitope_glyco_occluded, epitope_glyco_sites). The check flags N-glycosylation sequons (N-X-[S/T]) that are exposed on the free target and sit at or near the epitope, where an installed glycan would mask an otherwise grippable patch. grippability_consensus(row, iara_score) optionally cross-checks this physical read against a learned target-side score (for example an IARA epitope mean, computed by the caller so binderqc stays dependency-clean) and returns grippable, flat, or disagree. Pass --iara-score on the CLI to add it as a column.

• Tag site: recommended terminus (N/C) and the numbers behind it: relative SASA, CA-CA distance to the paratope, orientation, and a terminal cysteine's SG SASA.

• Developability: two complementary aggregation scores, an SAP-style spatial aggregation score (sap_score, sap_total, and sap_per_res = load per residue, since sap_total scales with chain length and the literature SAP < 35 bar was calibrated on 52-65 residue binders) and an Aggrescan3D score (a3d_score, a3d_total_positive, a faithful pure-Python port of Aggrescan3D 1.0.2's a3v scale and algorithm, Pearson r≈0.92 vs the reference tool). Plus TAP-style surface charge patches (Raybould et al. 2019): charge_patch_pos and charge_patch_neg are the strongest exposed clusters of like charge, which net charge and pI miss and which drive viscosity and self-association. Following TAP's actual insight, the same idea is also computed over the binding region: paratope_hydrophobicity and paratope_charge say whether the interface itself is sticky or charge-clumped. These are reported numbers, not gates: TAP's thresholds are calibrated on antibody Fv surfaces and do not transfer to minibinders. It also reports the ProtParam instability index (Guruprasad 1990), GRAVY, pI, MW, and ε₂₈₀. Sequence liabilities cover deamidation (N-[G/S/T]), Asp isomerization (D-[G/S/T/D/H]), unpaired cysteines, N-glycosylation sequons, and Met/Trp oxidation hotspots. The oxidation check is SASA-aware, so only surface-exposed residues count.

A warnings column flags problems: small, flat, anchorless, or glyco-occluded interfaces; buried, ambiguous, or interface-facing tag sites; hydrophobic sequences. qc_pass is true when there are no quality warnings (tag-site advisories like an ambiguous terminus do not count), and --fasta writes those binders.

Tests

pip install -e ".[test]"
pytest

Tests run against a bundled example, PDB 7JZU (the LCB1 minibinder on the SARS-CoV-2 RBD). LCB1 comes from Cao et al. (2020), which frames the problem binderqc targets: the bottleneck is selecting good binders, not designing them.

"…not in the de novo design of proteins with shape and chemical complementarity to the target surface, but in recognizing the best candidates."

In that work LCB1 buries ~1,000 Ų and forms "multiple hydrogen bonds and salt bridges … consistent with the subnanomolar affinities". binderqc on 7JZU agrees: 1021 Ų buried area (Cao reports ~1,000 Ų), plus 2 salt bridges and 17 interface polar contacts. See Cao et al., De novo design of picomolar SARS-CoV-2 miniprotein inhibitors, Science 370, 426-431 (2020), doi:10.1126/science.abd9909.

tests/pisa_correctness.py is a separate script, not part of the unit tests. It downloads 18 public complexes from RCSB and PDBePISA and checks the interface area against PISA (r ~ 1.0, about 1% median error):

pip install -e ".[validation]"
python tests/pisa_correctness.py

Figures

The complex in the banner is a real render, not a drawing. docs/render_structure.py draws a cartoon under a translucent molecular surface (binder blue, target grey) from actual coordinates, using 3Dmol.js in headless Chromium. It works on any PDB/CIF, so you can reuse it for your own figures:

pip install -e ".[docs]" && python -m playwright install chromium
python docs/render_structure.py complex.pdb A B out.png   # structure, binder chain, target chain, output

License

MIT