We predicted how 184 small molecules sit in the PXR protein, then the real crystal structures were released. This page asks, for the {{potent.tier.n}} strongest binders ([[pec50|pEC50]] ≥ 6.0): where did our submission land, how close could the best predictions have been, and what would a perfect score require? Scores are [[lddt|LDDT-PLI]] (0 = wrong, 1 = every contact distance reproduced).
- The belief that potent compounds can “statistically” reach 0.90–0.95 is not supported by these data. We scored {{potent.final}}. Only {{potent.or09.cnt}} of {{potent.tier.n}} ({{potent.or09.pcti}}) potent ligands have any pool pose ≥ 0.9, and we submitted one for {{potent.f09.cnt}}.
- 0.90–0.95 is also at or above the crystal’s own reproducibility. Comparing the two protein copies in the crystal gives {{potent.rep}}; only {{potent.rep09.pct}} of ligands reproduce at ≥ 0.9 and {{potent.rep095.pct}} at ≥ 0.95.
- Even the most generous ceiling stops short of 0.9 on average ({{potent.union}}); {{potent.union09.pct}} of potent ligands reach ≥ 0.9 there.
- Where the shortfall sits: [[selection|picking the right pose from the pool]] is about 31% of it; the pool simply lacking a good pose is 69%, mostly [[clash|pocket side chains]] and conformer. The [[site2|second site]] costs nothing for potent ligands.
- Weak binders’ crystals look informative, not like soak noise, by every proxy available without density maps (see the weak-binder section).
Caveats (please read before trusting any number)
- Post-hoc and exploratory. Nothing was pre-registered except the tier cut-offs. About 30 weak-binder metrics were compared with no multiplicity correction.
- Truth quality is a proxy. Without electron-density maps there is no RSCC/EDIAm; “informative” rests on B-factors, occupancy, contacts and copy agreement. “Cross-copy LDDT-PLI” is an ad hoc reproducibility measure available for 130 of 184 ligands (26 of 29 potent). The leaderboard takes the better of the two copies, so a model can in principle beat that number by matching one copy.
- The analog-transfer ceiling is optimistic. It uses the target’s own crystal protein (so side-chain clashes the models suffer are absent), one crystal form, conformer choice by clash against the crystal pocket, and “best of 3 templates” chosen by the true score. Those steps are not available at prediction time. Same-series templates (13 of 29 potent ligands) do most of the work.
- Series clustering. Ligands fall into a few chemical series, so bootstrap intervals that treat ligands as independent are probably too narrow.
- 90 transfer ligands were built before a timeout fix. An embedding timeout (6 s per conformer embed) was added partway through; about the first 90 ligands ran without it. It can only affect slow cases, but the run is not perfectly uniform.
- The pool is one pose per model (13 models, so ~9–12 poses per ligand), not the 20-sample pools; the oracle is therefore a lower bound on what generation could supply, and partial-coverage engines make per-ligand oracles unequal.
- The gap split is attribution, not cause. It uses ad hoc geometric thresholds on the single best pool pose; the pieces overlap physically.
- Tiers rest on an un-re-audited join. pEC50 comes from a pre-joined activity table; 76 ligands have no record (73 fragments + 3 drug-like) and are never called weak.
- Small n. 29 potent ligands; 26 with both copies.
source: docs/FINDING_050_pxr_potency_ceiling.md § Caveats
Glossary: what every technical term means click any dotted term anywhere on the page
Every dotted-underlined term on this page is a button that opens a plain-language card right where you are. They are all listed here too.
Where did we land? Scores by potency tier chart + table
A [[tier|tier]] groups ligands by [[pec50|pEC50]]: potent ≥ 6.0, moderate 5.5–6.0, weak < 5.5, and no record (mostly [[fragment|fragments]] with no activity curve; never called weak). Each row shows the mean score at five levels with its 95% [[bootstrap|bootstrap]] interval (whisker). The dashed line is 0.9.
Same numbers as a table
Take-away: for every tier the submission sits well below the pool oracle, and the oracle sits well below the crystal’s own copy-to-copy agreement (~0.88–0.89). Moderate ligands scored lowest ({{moderate.final}}); potent {{potent.final}}. Pose accuracy is not related to potency itself (see the weak-binder section).
source: data/processed/pxr_post/potency_belief.csv (values and CIs), potency_tier_summary.csv (n). FINDING 050 T1/T3.
How close could we get? The ladder for potent ligands 0.634 → 0.9–0.95
Each step relaxes a rule. Read it bottom-up as “what do I have to assume to get there”. The bar axis starts at 0.50. Whiskers are 95% CIs. The pink dashed lines mark 0.90 and 0.95; the pink band is the believed range.
Not a staircase you can climb. The steps are not nested: the crystal-copy line is a property of the truth (how well the structure reproduces itself), not of any predictor. A prediction that matched one copy exactly would still score 1.0 under the best-of-A/B leaderboard metric, so the 0.888 is an expectation for an independent re-determination, not a hard cap.
source: data/processed/pxr_post/potency_belief.csv (potent, transfer variant C, n = 29; copy ceiling n = 26 two-copy ligands). FINDING 050 T2/T3.
Where does the gap to 0.9 sit? selection, side chains, conformer, orientation, second site
For each ligand the shortfall is max(0, 0.9 − score). Averaged over the tier, it is split by what the best pool pose looks like. Pick a tier:
These pieces overlap physically (a clash-limited pose is also a shifted one): the split is attribution by the best pool pose, not causation. Thresholds are ad hoc, from FINDING 045.
source: data/processed/pxr_post/potency_gap.csv. FINDING 050 T4.
3D structures: stable vs unstable, side by side crystal vs ours vs best pool vs analog transfer
Four exemplars chosen by a rule fixed in code before structures were inspected (see “Sources”). Thick sticks are ligands, coloured by source; thin lines are pocket side chains. Drag to rotate (rotation is synced across panels when the box is ticked), scroll to zoom. You can change any panel’s ligand, or click a row in the ligand table (only the 16 exemplars have 3D data).
Scores are LDDT-PLI against the crystal (the leaderboard takes the better of the two copies, so the crystal copy B chip is its agreement with copy A). Morph blends atom i of the crystal with atom i of the prediction; the atom order is shared by construction (crystal display-chain order) and midpoints are not chemically valid geometries. Only side-chain atoms of residues within 10 Å are stored; the page draws those within 6.5 Å of copy A. Everything is in the crystal frame of the displayed copy.
source: data/processed/pxr_post/viewer/structures.json (16 exemplars), tiers.json. Viewer data are 2-decimal coordinates.
Weak binders: are their crystals informative or noise? 47 of 48 in the main site
A weak binder’s crystal pose might be a soak artefact (the ligand just stuck there) and so a poor yardstick. Without electron-density maps we can only use proxies. They say: informative, with two differences that fit “smaller, less anchored” ligands.
Evidence the pose is informative
- 47 of 48 weak ligands sit in the main site (potent 29/29).
- They reproduce the Gln285 and Phe288 [[anchor|anchor contacts]] at 83% / 98% (potent 86% / 100%).
- Ligand-to-pocket B-factor ratio 1.03 vs 0.96: ordered about like the potent ones.
- Copy-to-copy agreement 0.891 vs 0.888.
- The 8 weakest ligands (pEC50 < 4.5) are all on the main site, ordered, reproduce across copies (0.93) – the opposite of a random soak artefact, but n = 8.
Real differences, and a different population
- Weak ligands occupy both [[copies|copies]] less often (62% vs 90%).
- Fewer peripheral hydrophobic contacts (Met323 50% vs 79%), consistent with smaller ligands.
- Fragments with no activity record are different: 13 of 73 (18%) sit off the main site, less ordered (B ratio 1.20), far fewer anchor contacts – though those in the site and in both copies still reproduce (0.885).
- “Never tested” does not mean “does not bind”.
The evidence table
Cells are the means/proportions as written in potency_q2_table.csv (n per group inside the cell). “vs potent” is Hedges g (continuous) or a difference (proportions) with a 95% bootstrap CI. ~30 metrics, no multiplicity correction.
Does pose accuracy track potency?
Spearman correlations with 95% CI. Overall, pEC50 vs our score is {{ALL.corr0}} – indistinguishable from none. The potent tier scores higher than weak by +0.112 (CI +0.021 to +0.202), but ligand size, rigidity and series are confounded with potency.
source: potency_q2_table.csv, potency_q2_corr.csv. FINDING 050 T5 and verdict by sub-population.
What was believed vs what the data say the 0.90–0.95 claim
The belief, as stated
“For potent compounds one can statistically reach 0.90–0.95; the rest is luck.”
(Quoted from the framing in FINDING 050 item 1.) It implies that for a strong binder the crystal pose is well defined and a good enough method lands near it most of the time.
What the 29 potent ligands show
- Submission: {{potent.final.ci}}; ≥ 0.9 for {{potent.f09.cnt}} of {{potent.tier.n}} ({{potent.f09.pcti}}).
- Any pool pose ≥ 0.9: {{potent.or09.cnt}} of {{potent.tier.n}} ({{potent.or09.pcti}}); pool-oracle mean {{potent.oracle.ci}}.
- Generous ceiling (transfer best-of-3 chosen with the truth, unioned with the pool oracle, holo protein): mean {{potent.union.ci}}; ≥ 0.9 for {{potent.union09.pcti}}.
- The truth against itself (copy A vs B, {{potent.rep.n}} ligands): mean {{potent.rep.ci}}; ≥ 0.9 for {{potent.rep09.pcti}}; ≥ 0.95 for {{potent.rep095.pct}}.
- Both reproducible at ≥ 0.9 and within the generous ceiling: {{potent.both09.pcti}}.
A neutral reading
- These data do not support 0.90–0.95 as a level that is statistically reachable for the potent set as a whole. The best case examined averages about 0.86.
- They do support “a strong binder is somewhat better predicted than a weak one” (+0.11, exploratory) and that a sizeable minority of potent ligands can reach ≥ 0.9 when a same-series crystal is available.
- 0.95 is above the crystal’s own reproducibility for about 85% of potent ligands, so a mean near 0.95 would mean beating the truth’s own agreement; a prediction that matches one copy would still score near 1.0 under the better-of-two metric.
- What these data cannot rule out: n = 29 with series clustering (intervals are probably too narrow); a different pool (20-sample pools were not used); methods that use a same-series structure in the same crystal form, which is the only lever found that approaches the reproducibility ceiling.
source: potency_belief.csv; FINDING 050 summary item 1–3, T2 verdict.
All 184 ligands sortable, filterable; click an exemplar row to load it in the viewer
Rows with a cyan left edge have 3D data (16 exemplars).
“transfer, MCS” = analog-transfer pose from the largest-MCS template (variant C, all 183 others); “transfer, best of 3” uses the truth to pick among the top 3 templates. “A↔B” = cross-copy LDDT-PLI where both copies hold the ligand. — = not available.
source: viewer/tiers.json (184 records); Emax joined from potency_master.csv (emax_pc, as released).
Which models were in the pool? potent ligands, per engine
One pose per model per ligand. For potent ligands no single engine averaged above 0.586, and no engine put more than 10% of its poses at ≥ 0.9. Partial-coverage engines have fewer ligands (n column).
source: potency_engine_by_tier.csv (potent rows). FINDING 050 T1 notes.
Sources and method every number traces to a file
- Prose: docs/FINDING_050_pxr_potency_ceiling.md and docs/FINDING_045_pxr_postmortem.md (scorer validated: 0.5651 vs 0.5640 on the leaderboard).
- Data embedded as-is: data/processed/pxr_post/viewer/tiers.json, viewer/structures.json; potency_belief.csv, potency_tier_summary.csv, potency_gap.csv, potency_q2_table.csv, potency_q2_corr.csv, potency_engine_by_tier.csv, potency_master.csv (Emax).
- Generator: scripts/viz/build_pxr_page.py (fails the build if headline numbers drift from FINDING 050). If this page and the repo disagree, the repo wins.
- Exemplar rule (fixed in code before inspecting structures):