Clash-Free SMOTE on Learned Pose Manifolds for Cross-Docking

Generative docking models — diffusion or flow-matching networks that predict where a small molecule sits inside a protein pocket — now beat classical physics tools when the pocket is pre-arranged around the ligand. But in the realistic setting, the protein comes from an unbound or differently-liganded conformation, and the models fall apart: predicted poses physically overlap with receptor atoms, and performance on genuinely novel proteins drops far below classical tools (PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences reports DiffDock at 12% physically valid and accurate poses versus 58% for AutoDock Vina). A second, less discussed cause is imbalance: training databases are dominated by a few rigid, well-populated binding-mode families, so rare receptor conformations and unusual pocket geometries are severely underrepresented — exactly the tail the cross-docked, "in-the-wild" test cases come from.

The standard fix for imbalance, SMOTE, synthesizes new minority samples by interpolating between neighboring minority points along straight line segments (SMOTE: Synthetic Minority Over-sampling Technique). That is precisely the wrong operation on molecular structures: two clash-free poses connected by a straight chord in coordinate space almost always pass through configurations where atoms occupy the same volume, and interatomic repulsion grows as $r^{-12}$, so the chord crosses an effectively infinite energy barrier. This proposal builds an oversampling method that interpolates in a learned non-linear feature space instead of raw coordinates, and enforces steric non-overlap on every synthesized pose before it enters training.

Core Claim

SMOTE can be made safe and useful for 3D molecular data by (i) mapping poses through a non-linear encoder into a latent space where rare binding modes are locally linear, (ii) performing the SMOTE interpolation there, and (iii) projecting each decoded synthetic pose onto the steric-feasible set before it is used as training data. The theoretical hook comes from work showing SMOTE is a form of kernel-based noise injection with excess-risk guarantees for continuous densities (Touqeer Ahmad et al.); those guarantees implicitly assume the interpolated segment stays inside the data support. On a molecular energy landscape the straight segment leaves the support almost immediately, so the smoothing kernel places probability mass on impossible structures. Our claim is that changing where interpolation happens — a learned kernel-induced latent space, decoded and steric-projected — restores the on-support property, and that populating the rare binding-mode tail this way measurably improves cross-docking generalization.

Concretely, with $\phi$ the learned encoder and $\psi$ its decoder, a synthetic sample is generated as

$$z_{\text{syn}} = (1-\lambda)\,\phi(x_i) + \lambda\,\phi(x_j), \qquad x_{\text{syn}} = \Pi_{\text{clash}}\big(\psi(z_{\text{syn}})\big),$$

where $\lambda \in [0,1]$, $x_i, x_j$ are neighboring minority poses, and $\Pi_{\text{clash}}$ projects onto the steric-feasible set by minimizing a van der Waals energy plus a quadratic tether to the decoded structure. The key ingredients are the non-linear map and the projection; the SMOTE mechanics are unchanged.

Research Program

Three components get built. First, the pose encoder $\phi$: a contrastive encoder over (pocket, ligand-pose) pairs, aligned with the closed-form spectral kernel solution of UniCon: Unified Framework for Efficient Contrastive Alignment via Kernels, which replaces iterative gradient updates on the alignment objective with an SVD in a kernel-induced space and reports up to a 461× speedup over SGD. We adopt its low-rank Nyström approximation to keep the Gram matrix tractable, and we follow its own acknowledged limitation — spectral optimality holds only over frozen features — by alternating between closed-form alignment and encoder updates. Second, the sampler: minority strata are defined by binding-mode clusters in latent space (rare torsional signatures, uncommon pocket families), and SMOTE interpolation runs between latent neighbors within a stratum, following From SMOTE to Mixup for Deep Imbalanced Classification, which showed that naive SMOTE in raw input space is near-useless for deep models while controlled interpolation in representation space is not. Third, the steric projection $\Pi_{\text{clash}}$: a short force-guided correction during decoding, in the spirit of ForceFM: Enhancing Protein-Ligand Predictions through Force-Guided Flow Matching, which doubled PoseBusters pass rates from roughly 17% to roughly 31% by steering sampling with physics-based gradients rather than repairing outputs afterward. The synthetic poses then augment the training set of a torsional diffusion docking model of the kind built by Harmonic Torsional Diffusion for Protein-Ligand Flexible Docking, which already respects periodic torsion geometry.

The design is constrained by known negative results. PoseBusters established that RMSD-trained coordinate regression treats molecules as unconstrained point clouds, so we never train on raw RMSD alone; and PoseX: AI Defeats Physics-based Methods on Protein Ligand Cross-Docking showed that "self-docking" is an artificially easy, lab-arranged task, so all headline evaluations use the cross-docking setting where the receptor conformation does not match the ligand.

Proposed Experiments

Training uses PDBbind; the imbalance treatment targets the rare tail of binding-mode clusters, and every compared variant gets the identical base model, training steps, and number of synthetic samples. Evaluation uses three suites. The PoseBusters benchmark set (308 post-2021 complexes, under 30% sequence identity to training data) measures generalization under physical-validity checks; PoseX-CD (1,312 cross-docking pairs from 2022–2025 PDB entries) measures the realistic unconstrained setting; and a stratified analysis splits PoseX-CD by binding-mode frequency in training to isolate where gains come from. Metrics are the PoseBusters physical-validity pass rate, the combined valid-and-accurate rate (RMSD < 2.0 Å with all checks passed), and pose diversity within each pocket (to detect diversity collapse in synthetic data).

The decisive predictions are, first, that clash-free latent SMOTE raises the valid-and-accurate rate on PoseX-CD above what the same model trained without oversampling achieves, with the largest gains on rare-pocket strata and no regression on common pockets; second, that synthetic poses generated this way pass PoseBusters-style checks at a far higher rate pre-relaxation than Euclidean-coordinate SMOTE poses; and third, that the improvement survives an identical OpenMM relaxation applied to all methods' final outputs, which is the test that separates genuinely better training data from merely repairable training data.

As a generality probe rather than a headline result, the same recipe — oversample rare classes inside a disentangled structural-prior space rather than in pixels — will be tried on rare facial expressions in Aff-Wild2 using a FLAME-style 3D face model, the approach that let 3DPE-Gaze: Unlocking the Potential of 3D Facial Priors for Generalized Gaze Estimation cut cross-domain gaze error by over 27%. The motivating evidence is that Expression, Affect, Action Unit Recognition: Aff-Wild2, Multi-Task Learning and ArcFace identifies rare emotions like fear and disgust as an unsolved imbalance barrier, while The 8th ABAW Workshop and Competition reports winning expression-recognition F1 of only 0.365.

Baselines

The naive reference point is the base torsional diffusion docking model trained on PDBbind with no imbalance treatment at all; this is the "do nothing" case, and ForceFM's unguided model gives its expected scale, reporting a PoseBusters pass rate of roughly 17% (ForceFM). It matters because any oversampling method must beat simply ignoring the tail.

The strongest existing competitor is a two-way tie depending on the suite. On the PoseBusters benchmark, the strongest reported number is AutoDock Vina at 58% valid-and-accurate, with CCDC Gold at 55% and DiffDock at 12% (PoseBusters); a learned method must beat 58% to claim physics-free docking has closed the generalization gap. On PoseX-CD, the strongest reported system is SurfDock with OpenMM relaxation at roughly 77% success, having risen from a 41.7% raw physicochemical pass rate on self-docking to 73.6% after relaxation (PoseX). We will run AutoDock Vina and the SurfDock-plus-relaxation pipeline ourselves on our splits under the same evaluation code, and apply identical OpenMM relaxation to our method's outputs, so any difference is attributable to training data, not to post-processing. Tuning budgets and sampling step counts are matched across all learned methods, and every oversampling variant receives the same number of synthetic samples.

Two ablations isolate the proposal's ingredients. Euclidean-coordinate SMOTE inside the identical pipeline — same encoder capacity, same sample count, same relaxation — isolates the contribution of the non-linear latent interpolation. Clash-free latent SMOTE with the steric projection $\Pi_{\text{clash}}$ removed isolates the contribution of the physical feasibility step. Beating the baseline means, concretely: on PoseX-CD, exceeding the 77% SurfDock-plus-relaxation success rate; on the PoseBusters benchmark, exceeding Vina's 58%; and on rare-pocket strata, exceeding the un-oversampled ablation while remaining within noise of it on common pockets.

Why not just…?

Why not just run OpenMM relaxation on everything? This is the most serious alternative, and the evidence for it is strong: PoseX shows relaxation lifts SurfDock's physicochemical pass rate from 41.7% to 73.6% and that relaxation applied to synthetic samples would also repair local clashes. But relaxation repairs geometry, not distribution: a synthetic pose synthesized across an energy barrier relaxes back toward one of its two parents, so the minority tail gains duplicated near-copies rather than new binding modes, and diversity shrinks — exactly the overfitting-by-duplication failure documented for naive oversampling in Synthetic Oversampling for Imbalanced Classification. The proposal's experiments settle this directly: we report pre-relaxation clash rates and post-relaxation pose diversity for Euclidean-SMOTE-plus-relaxation versus clash-free latent SMOTE, plus downstream rare-pocket success with relaxation applied identically to both. If relaxation alone closes the rare-pocket gap, the simpler method wins and we will say so.

Why not just use cost-sensitive or margin losses? Cheap and proven in-domain: ArcFace-style margins lifted AffectNet F1 from 0.54 to 0.63 without synthesizing anything (Expression, Affect, Action Unit Recognition). But From SMOTE to Mixup shows margin losses alone plateau under severe imbalance — LDAM-DRW reached minority-class accuracies of 0.63 and 0.66 on CIFAR-10 at imbalance ratio 100, while margin-aware interpolation reached 0.79 and 0.82 — and A Comprehensive Survey on Imbalanced Data Learning stresses that loss reweighting cannot create structural diversity that is absent from the tail. Still, we run a class-balanced-loss variant at matched budget; if it matches our method on rare-pocket strata, the extra machinery is not justified.

Why not just train a bigger model on more data? Capacity without physical or structural priors has already been tried and has already failed where it matters: PoseBusters shows large learned docking models collapsing to 12% on novel proteins while small classical physics tools hold 58%, and 3DPE-Gaze shows the same pattern across domains — scaling 2D CNN encoders overfits environmental nuisance rather than absorbing it. The problem here is not model capacity but a thin, skewed training distribution of binding modes, which more parameters do not address.

Why not just use vanilla SMOTE or duplication? Vanilla SMOTE in coordinate space is the ablation we run, and the theory says why it fails: its kernel-noise-injection guarantees (Touqeer Ahmad et al.) hold for on-support smoothing, while PoseBusters demonstrates what happens when linear coordinate assumptions meet 3D atomic structures — validity collapses. Duplication, meanwhile, is documented to cause memorization rather than minority learning (Synthetic Oversampling for Imbalanced Classification).

Decisive Checks

The single most decisive experiment is the pre-relaxation validity rate of synthetic poses: if latent interpolation plus steric projection does not dramatically outperform Euclidean-coordinate SMOTE on clash-free generation before any force field touches the samples, the core mechanism is wrong and the proposal should not proceed. The second check is the rare-pocket strata analysis on PoseX-CD: the entire motivation rests on gains concentrating in the underrepresented tail, so gains that appear only on common pockets would mean the imbalance story, and the method, missed. The third is the relaxation-equality control: our method and the strongest baseline must both receive identical post-processing, so that any reported advantage is a statement about training data rather than about who ran OpenMM.