Beyond the Argmin

Every ranking item needs (target, pose, symmetry-corrected RMSD) triples. Run the F3 arm (512 orientations, 0.375 Å, 1.42 s median) over train, retaining the top 32 poses per target rather than 8 — the extra poses cost nothing and give the ranker negatives near the decision boundary.

First

Check whether scratch/ranker-v2-guarded-seed42.json and the re_export_corpus_depth.py pipeline already carry pools with RMSDs. If so, reuse and skip the regeneration entirely.

Store

Poses plus the landscape summary from W1 — recomputing the field later would cost as much as the docking run itself.

Budget

35,349 × 1.42 s ÷ 6 cores ≈ 2.3 h, plus RMSD evaluation.

Every item below emits features. Give them one home so the Rust scorer and the Python trainer cannot drift apart — the train/inference mismatch that killed EquiField (relu(2.5−d)²·5.0 in the trainer against 4.0·(3.4−d)² in equifield.rs:577) came from exactly this having no single owner.

pub struct PoseFeatures { /* ~60 named f32 fields */ }

pub fn extract(
    grids:     &ReceptorGrids,
    landscape: &LandscapeSummary,   // W1
    hydration: &HydrationField,     // W2
    pose:      &Pose,
) -> PoseFeatures;

Serialise as a stable named schema, never a bare float array. Emit the same struct from the Rust path and the Python export, and unit-test that they agree bit-for-bit on a fixed complex.

The flagship. Vina's entire entropy model is −0.0585 × N_rot: a torsion count. Replace it with the actual configurational integral over the field you already have.

Z(R) = Σ_T exp(−β·E(T,R)) · h³
F    = −kT · ln Σ_R w_R · Z(R)

One exp-and-sum pass over arrays correlate.rs already produced — negligible beside the FFT itself. Then watershed the field by steepest descent into basins, integrate Z per basin, and rank basins by F rather than by E_min. A broad basin beats a sharp one of equal depth, which is the right physics: crystal poses sit in redundant, robust basins while empirical false minima are knife-edge artefacts of one lucky contact.

New file

src/landscape.rs, consuming correlate::TranslationField

Emits

F_basin, basin volume, occupancy p_i = Z_i/ΣZ_j; pocket-normalised z (E_min − μ)/σ against the landscape's own distribution; landscape entropy −Σp ln p; rank gap between basin 1 and 2.

Params

β alone. Grid-search on train CV against top-1 accuracy.

Basins across rotations

Watershed is per-rotation. Merge basins whose minima fall within 1 Å RMSD — the retained-pose clustering already does this, so reuse it rather than inventing a second convention.

The pocket-normalised z may be the highest-value line here and is nearly free: Vina scores scale with ligand size, which is why absolute scores are near-useless for cross-ligand ranking. Normalising against this ligand's own landscape in this pocket cancels both ligand-size and pocket-size bias. Expect the effect in screening enrichment (vs_enrichment.py) more than in single-target top-1.

Where Vina is weakest. Its hydrophobic term is a distance-based contact count with no notion of enclosure or water structure, yet the hydrophobic effect is driven by the entropy of displaced water, not by carbon proximity.

Check first: ReceptorGrids already holds a per-atom-type map, so the O/OA map may already be W(x), the water-oxygen affinity at every voxel. If so this item needs no new correlation at all — only the enclosure term and the combination rule.

Enclosure B(x) comes from erosion.rs: the sliding-window minima machinery already computes how deeply a voxel sits inside the free-space mask. Then classify:

• high B, low W → trapped, poorly-coordinated water → displacement is strongly favourable. This is the hydrophobic hotspot, properly derived.
• high B, high W → conserved water → penalise displacement, unless the displacing ligand atom is a donor/acceptor that can replace its H-bonds.

New file

src/hydration.rs

Params

Two thresholds and two weights. Fit on train CV.

Cost

Receptor-side and built once — amortised across an entire screen, which is where it matters most.

Validation

Before touching docking: do predicted conserved-water sites coincide with crystallographic waters in the train receptors? That is a direct, honest check of the physics with no scoring involved.

Empirical functions reward nonspecific bulk — a greasy ligand that sticks anywhere scores well everywhere. Ask instead how well this ligand scores elsewhere on the same protein, and score the difference:

specificity = E_site − E_best_elsewhere

Uniquely affordable here because the FFT already sweeps translations exhaustively; a sampling engine would pay for a second full docking run. Decoy-free and training-free.

Cost control

Run the counterfactual sweep on a coarse 1.0 Å grid with a reduced rotation set. You need a crude "does it stick elsewhere" signal, not a second accurate dock — this is what keeps the item from costing 10–50× a normal run.

Emits

Absolute gap, percentile rank of the true site among all surface patches, and the number of competing sites within 2 kcal/mol.

Risk

Large proteins with several genuine pockets will show a small gap for correct poses. Report the percentile alongside the gap so the ranker can learn the distinction rather than being misled by it.

Nothing in src/ looks like a strain library, and this one compounds with the pose-quantisation work directly.

Histogram observed torsion angles by bond type across the 35,349 train natives — no external dataset needed, you already hold it. Define strain(τ) = −ln p(τ | bond type), smoothed as a von Mises mixture so the potential and its derivative are continuous.

The payoff is structural: a per-torsion prior is a unary potential, and unary potentials are exactly what the tropical DP already handles exactly. It costs nothing in torsion_field.rs, it biases enumeration toward conformers that are chemically real, and PoseBusters explicitly checks internal energy — so top-1 and PB validity improve from the same table lookup.

Bond typing

Key on the two central atoms' AutoDock types plus heavy-atom degree and aromaticity. Coarse but sufficient; fall back to a pooled distribution when a type has < 50 observations.

Artifact

release/torsion_prior.bin, hashed into release/model-provenance.toml like every other committed binary.

Validation

Held-out natives must score lower strain than their own docked decoys. If they don't, the typing is too coarse — check before wiring it into the DP.

The cheapest item on the list and one of the strongest known discriminators of real binding modes. Empirical functions reward formed H-bonds and stay silent on the converse: desolvating a polar atom and giving it no partner costs several kcal/mol.

hbond_repair.rs already fixes donor/acceptor assignment — which matters here, since a receptor prepared without polar hydrogens presents every backbone nitrogen as an acceptor and would corrupt this count exactly as it corrupts the hbond term. Run the repair first, then count.

Definition

A ligand polar atom is unsatisfied when B(x) exceeds the burial threshold and no complementary partner sits within standard H-bond distance and angle.

Emits

Separate counts for ligand-side and receptor-side unsatisfied atoms — burying a receptor donor is a distinct and equally real cost.

Reuse

B(x) from W2. Sequence W2 first or share the enclosure computation.

Three jobs, one model. LambdaMART over ~60 features, groups keyed by target.

1. Fusion. Where W1–W5 cash out. Combine them with the existing Vina, Vinardo, CNN and ranker-v2 scores. The new features are orthogonal to what a pose scorer can perceive, and orthogonal signals fuse well — which the existing z-mean fusion already demonstrates in a cruder form.
2. Diagnostic. Train once on physical features alone, without ranker-v2 output. The gap against ranker v2 measures how much its lift comes from the learned representation versus features a tree could read directly. Either answer is useful: a small gap says simplify, a large gap says the GNN is earning its keep.
3. Pocket conditioning. Don't build a separate mixture-of-experts — feed pocket descriptors (buriedness distribution, polar/apolar ratio, metal presence, volume) as features and let the trees learn the conditioning. Same effect, one model, no extra machinery.

Protocol

Grouped 5-fold CV on train for both hyperparameters and early stopping. Val (68) as a smoke check. Test (225) once.

Determinism

Fixed seed, sorted feature order, no hash-map iteration in the funnel — the bit-identical invariant in CLAUDE.md applies to this path too.

Confidence

Emit P(top-1 ≤ 2 Å), calibrated by isotonic regression on CV out-of-folds. Report a reliability diagram and ECE, not just accuracy.

That last deliverable is the one with no competitor: no mainstream docking engine tells you when to distrust its answer, and W1's landscape entropy is a genuinely informative input to it.