Show your work.

Pose accuracy — 42-target Astex benchmark

The number the field reports: for each target, re-dock the crystal ligand and ask whether the top-ranked pose lands within 2 Å of the experimental one. We run 42 targets of the Astex Diverse Set through one frozen protocol — docking box from the crystal ligand, exhaustiveness 16, single chain, no manual curation — the same fully automated path a user gets. (The full Astex Diverse Set is 85 complexes; we report the 42 our automated pipeline fetches and prepares end-to-end without hand-editing.) Every number below was computed live with the result cache off.

How to read this against the literature: published redocking numbers — GNINA 1.0's own paper (~73%, PDBbind core), GNINA in PoseBusters v2 (~63%), AutoDock Vina (~44–58%) — use different benchmark sets with hand-curated inputs, so they are a directional band, not a head-to-head. Astex is an older, generally easier set. The honest, apples-to-apples comparison is our own automated pipeline against itself: 60 → 69 with the preparation fix.

The confidence badge — which pose to trust

Most co-folding tools report one average accuracy. Almost none tell you, for the pose in front of you, whether to believe it. Boltz-2 emits a per-ligand confidence (pLDDT); we bin it green / amber / red. On a 30-target holdout the thresholds were never fit to — overall base rate 15/30 (50%) correct, point-biserial r = 0.66 between confidence and correctness — the badge separates right from wrong cleanly:

Boltz-2 is run single-sequence (empty MSA, so wild type and mutants are predicted under identical conditions), with pocket conditioning on the residues inside the crystal-ligand box, one diffusion sample. The independent 12-target run reproduced the shape. One red pose (pLDDT 35) did land at 1.8 Å — which is why red means "don't rely on it," not "always wrong." This per-pose signal, not the raw success rate, is what makes an automated docking result actionable.

Affinity ranking — 8-target MM/GBSA benchmark

One rung above docking: an MM/GBSA binding-energy estimate on a reference pose, correlated against experimental affinity across eight congeneric series (184 ligands). The score is a single snapshot; a separate short MD run is a pose-stability trust badge, not averaged into the number. Read it as a ranking signal for ordering compounds within a target — not a calibrated K_d.

We lead with rank correlation — the honest measure of a rescoring signal — and report every target, including the weak ones (HIF2A ρ = 0.46, Thrombin ρ = 0.54). A per-target calibrated RMSE (≈0.8 kcal/mol) exists but is deliberately kept out of the headline: it linearly maps scores onto the experimental scale, so it measures within-target ranking, not absolute affinity, and is easily misread as a K_d. Solute dielectric ε = 1 (tuned for ranking, not a physical constant); no explicit salt. FEP (relative free-energy perturbation) is in the pipeline for the close calls; we are not reporting FEP accuracy until a rigorous apples-to-apples benchmark is complete.

Catalog audit

Every catalog target's pocket centre is independently verified against the chain-A co-crystal ligand centroid in the canonical RCSB PDB. Every promoted mutation is checked for residue identity and reachability inside the docking box. The check runs as a blocking gate before every backend deploy — a regression cannot ship.

Verification script: backend/scripts/verify_catalog.py · runs in CI before every deploy

Approved-drug redocking — 8 complexes

A complementary, drug-focused check alongside the 42-target Astex run above: pull each drug out of its own co-crystal structure, re-dock it into the same pocket, and measure the RMSD to the experimental pose. Under 2 Å is the field's accepted pass mark. This is redocking (self-docking), the standard first check — not blind cross-docking: the search box is the target's pre-defined catalog box or, where none exists, a 22 Å cube auto-centred on the co-crystal ligand, so the pocket location is given. Exhaustiveness 16, one run per cell. Eight crystallographic complexes spanning ABL, KIT, EGFR, BRAF, ERα and trypsin — every number below was computed live with the result cache switched off, so nothing is recycled from an earlier run.

The honest read: fast Vina reproduces the crystal pose to well under 2 Å on 5 of 8 and misplaces the ligand on two EGFR ATP-site boxes (erlotinib 7.8 Å, gefitinib 6.6 Å). GNINA's CNN rescoring pulls both of those back into the pocket (→ 2.0 and 1.7 Å) for 7/8 under 2 Å. Boltz-2 co-folding, scored after pocket superposition, is the strongest on the well-behaved cases (KIT/imatinib 0.35 Å) at 6/8 under 2 Å — its one real miss is ABL/imatinib (30 Å), where it appears to fold ABL in the DFG-in state that imatinib can't occupy (the identical imatinib on KIT lands at 0.35 Å). Lapatinib is the lone case all three read as borderline (~3 Å). Verdict is PASS when any engine places the pose under 2 Å; NOISE when none clear 2.

Scripts: rescore.py (Vina / GNINA) · boltz_aligned.py (Boltz-2 pocket-aligned) · symmetry-corrected RMSD via spyRMSD · cache off, fresh GPU computes 28 Aug 2026

Docking speed

Time to a finished, interaction-annotated pose for one compound on a warm GPU worker (single-compound, no queue). Physics docking returns in well under a minute; Boltz-2 folds the entire protein, so it costs minutes — the price of a co-folded structure. Jobs fan across multiple GPU workers, so a whole screen finishes far faster than these per-pose numbers imply.

Where we sit: our GPU Vina/GNINA are in the seconds-per-compound band with Uni-Dock and Vina-GPU and comfortably ahead of CPU Vina and licensed Glide, while Boltz-2 is in the same minutes-per-complex band as every other co-folding model. The advantage is offering all three — fast screen, CNN-rescored accuracy, and full co-folding — behind one API with no license and automatic multi-GPU fan-out. Competitor figures are indicative ranges from published tooling, not head-to-head runs on identical hardware.

Screening, ADMET & throughput

Three more known-answer checks on the surrounding engines, reported at face value — modest where the method is modest.

Screening enrichment is honest but small-sample (24 compounds on a single ABL target) — read it as a sanity check that actives rank above decoys, not as a DUD-E-scale claim. ADMET is rule/ML-based triage scored against public benchmarks: useful for flagging obvious liabilities (high sensitivity on hERG and BBB), weak on the harder directional splits. Throughput is the reliability work from this sprint — a job now fans its compounds across multiple GPU workers concurrently, and the per-job GPU-worker count is shown on every result and in history.

Scripts: vs_analyze.py · admet/eval.py · benchmarked Aug 2026 on production liganx-api

Positive-control validation

Eleven (target, mutation, drug) pairs whose mutation-driven binding shifts are published in the clinical and pharmacology literature. We submit each to the live Liganx pipeline at exhaustiveness=16 (2× the product default — tighter sampling for a tighter noise band), capture Δ(mutant − WT), and check whether the direction agrees with the published shift. The point is direction, not magnitude — Vina scoring isn't free energy and isn't calibrated to cellular IC50.

Noise floor: ±1 kcal/mol (Vina scoring reproducibility at default exhaustiveness). NOISE means the Δ direction may be correct but magnitude sits below the noise floor — see the per-case rationale. FAIL means the Δ direction explicitly disagrees with the published literature direction; that's a regression and we investigate.

Validation script: backend/scripts/validate_positive_controls.py · each row's case name links to the live Liganx job result on this deployment.

Case-by-case rationale

One paragraph per case explaining the published literature, the number Liganx returned, and why a NOISE result (where applicable) is a method limitation rather than a bug.