When a small molecule is computationally docked — placed into a protein binding pocket in a candidate 3D orientation — downstream models almost always treat that single guessed geometry as truth. They shouldn't. Docking engines routinely return several plausible, mutually incompatible orientations for the same ligand and protein, and the physics is unforgiving: non-covalent interactions fall off as steep powers of interatomic distance, so a one-ångström shift can turn a hydrogen bond into a steric clash. Every affinity predictor downstream of docking inherits this ambiguity, yet the standard pipeline compresses it away immediately, regressing one scalar binding affinity from one chosen pose. A mixture density network — a neural net whose final layer outputs the parameters of a Gaussian mixture, i.e., a small weighted sum of bell curves rather than a single number — is built for exactly this situation: it can represent several distinct outcomes as separate modes instead of averaging them into a physically meaningless middle value.
The intersection of these two directions has been explored, but only halfway. CarsiDock, Interformer, GenScore, and BioLM-Score all use mixture density networks, but exclusively as likelihoods over interatomic distances — a learned statistical potential for generating or scoring 3D coordinates. Not one of them propagates multi-modal pose uncertainty into the affinity output itself. Meanwhile EGNN-MIL does ingest bags of 10–20 candidate poses, but its attention pooling deliberately collapses the bag into one convex-weighted scalar. This proposal closes that loop.
Core Claim
The claim is that the conditional distribution of binding affinity given a protein–ligand pair, $P(\Delta G \mid \text{protein, ligand})$, is genuinely multi-modal whenever the docking pose is ambiguous, and that a mixture density network conditioned on the whole pose bag — rather than a point regressor on one pose, an attention-pooled bag, or a distance likelihood — both predicts affinity more robustly and, as a byproduct, produces calibrated uncertainty that can decide where expensive physics-based refinement is worth spending compute. The central object is a pose-marginalized density: with a docking pose bag $B$ of $N$ candidate poses, we predict
$$P(\Delta G \mid B) \;=\; \sum_{p \in B} w(p)\sum_{k=1}^{K}\pi_k(p)\,\mathcal{N}\!\big(\Delta G \mid \mu_k(p),\,\sigma_k^2(p)\big),$$
where each pose $p$ gets its own small mixture sub-density and $w(p)$ is either uniform or a learned softmax over docking scores. Training minimizes the negative log-likelihood of the experimental affinity under this bag-level mixture, with a variance floor on $\sigma_k$ to prevent the singular-likelihood blow-ups known to afflict mixture density networks. The defining structural choice is that the mixture components are anchored to poses: each sub-density is computed from a single pose's embedding, so distinct binding orientations produce distinct modes rather than being averaged, and each mode can be traced back to the pose (and its interaction pattern) that generated it. This is the multi-hypothesis trick that Li et al. used to resolve self-occlusion ambiguity in 3D human pose estimation — competing poses play the role of occluded joints — transplanted into molecular regression, with the decoupled-branch stability lessons of Y. Wang et al. and the graph-mixture architecture of Oikarinen et al. informing the head design.
Two secondary hypotheses ride along. First, the anchoring should soften the documented trade-off in GenScore, where raising the affinity weight $\alpha$ in a joint loss monotonically degrades docking and screening power: because our affinity density is assembled per-pose rather than entangled with a shared geometric potential, we expect affinity emphasis to shift mixture weights and widths without corrupting pose-level geometry. Second, the model gives a measurable handle on the unresolved contradiction between Feng et al. (pose error is destructive; equilibration cannot rescue bad poses) and Wu et al. (noisy docked poses act as a regularizer, improving generalization by 32% on DAVIS): predictive width should grow gracefully with pose dispersion, keeping the likelihood calibrated across the range where point regressors collapse, and the crossover point where noise stops helping should be visible in the width-versus-error curve.
Research Program
We will build one system, PoseMDN, and study it. The input is a protein structure, a ligand, and a bag of 10–20 poses sampled from a standard docking engine (DiffDock and AutoDock Vina, plus Glide-score-ranked poses where available). An equivariant graph encoder embeds each pose independently — per-pose embedding, not pooled — and a decoupled mixture head maps each pose embedding to $K$ Gaussian components over affinity (we will sweep $K \in \{1, 2, 3\}$ per pose, since the bag-level mixture count is $N \times K$ and grows for free with pose count, avoiding the fixed-$K$ hyperparameter pain of monolithic mixture density networks). A lightweight bag-level network predicts $w(p)$ from docking scores and pose embeddings. Three training regimes will be compared: bag-level joint training on the marginal likelihood; stochastic single-pose training (each minibatch samples one pose, forcing every sub-density to be individually competent); and free-bag training where an unanchored mixture density network sees only a pooled bag representation — this last is the key ablation isolating whether anchoring, or merely having a density output, does the work. We will reuse the FABind codebase for fast pose generation ($0.12$ s per inference) to keep pose-bag construction cheap, in contrast to the simulation-heavy alternatives documented below.
Analysis deliverables alongside the system: a multimodality audit identifying complexes whose pose bags contain two or more RMSD-separated clusters (RMSD being the standard measure of pose error in ångströms); a component-to-cluster alignment score measuring whether mixture modes map back to pose basins; and calibration curves (predicted versus empirical coverage of, say, 90% intervals) as pose noise is deliberately injected at controlled RMSD levels from 0.5 to 6 ångströms.
Baselines
The naive reference point is scoring each docked pose with the docking engine's own energy function (AutoDock Vina's score on the top pose), which requires no learning and represents what a screening pipeline gets before any machine learning is added. It is the floor to beat on every discrimination metric. The first learned reference is a single-top-pose deterministic regressor: the same equivariant encoder and training budget as PoseMDN, but taking only the top-ranked pose and regressing a scalar with mean squared error, in the style of DAAP, whose authors document that a wrong pose corrupts the distance bins and disables the attention layers downstream. The file reports that EGNN-MIL outperforms such single-top-pose baselines by 3–5% in Pearson correlation on CASF-2016, so this baseline is expected to lose; its role is to quantify the value of the pose bag itself.
The strongest existing competitor is EGNN-MIL: an equivariant graph neural network that ingests exactly the same kind of pose bags (10–20 poses) we propose to use, and reports Pearson $r = 0.904$ on CASF-2016 — the best pose-bag-affinity number in the file. It is the strongest competitor because it already solves half the problem (pose-ensemble input, equivariant encoder) and its weakness is precisely our claim: its attention pooling flattens the bag into one convex-weighted scalar, so it cannot represent competing affinity modes and offers no calibrated uncertainty. We will run it under identical conditions — same pose bags from the same engines, same training data and split, same tuning budget of sweeps over encoder size and learning rate — and also retrain its pooling layer on our backbone so the comparison isolates the output representation rather than the encoder. Beating it means matching or exceeding its Pearson correlation on CASF-2016 while delivering substantially better negative log-likelihood and continuous ranked probability score on docked-pose inputs, and demonstrably separating modes on the multimodal subset where its point predictions should fall between modes. We also run GenScore (its fine-tuned Graph Transformer configuration, $R_p \approx 0.80$ on CASF-2016 with docking success above 90%) as the representative of the distance-likelihood-plus-scalar-affinity family, since it owns the scoring-versus-docking trade-off we claim to soften.
Two ablations isolate the proposal's key ingredients. The scalar-head ablation keeps the identical backbone, pose bags, and training data but replaces the mixture head with attention pooling to a single number — if it matches PoseMDN everywhere, the density output is not earning its keep and we would report that. The free-bag ablation keeps the mixture head but feeds it a pooled bag representation, isolating whether per-pose anchoring (rather than density outputs in general) is what preserves mode structure and calibration.
Why not just…?
Why not just use EGNN-MIL's attention pooling over the pose bag? This is the most serious alternative, and we treat it as the primary baseline rather than arguing around it. The documented limitation is architectural: attention pooling outputs a convex-weighted scalar, and when a ligand binds in two distinct orientations with genuinely different affinities, any point predictor — however good its inputs — lands on an expectation that may correspond to no physical state, the same failure mode that motivates mixture density networks in inverse problems like Cheng et al.'s nanophotonics work. But pooling could win on pure correlation if multimodal complexes are rare. Our decisive check is the multimodality audit: we identify complexes with RMSD-separated pose clusters and test whether PoseMDN's per-mode predictions beat EGNN-MIL's single prediction against the experimental value, plus a likelihood comparison that a point predictor cannot win by construction but that only matters if the density is honest — which the calibration curves test directly.
Why not just average the scores of the top $N$ poses? This is the cheapest alternative and will be included as a baseline for that reason. The file gives two pieces of evidence against it: Wu et al. found that structured architectural handling of top-5 poses improved correlation by 12.86% and reduced mean squared error by 8.10% relative to unstructured use, and the mode-cancellation argument is structural — averaging two flipped orientations cancels their distinguishing interactions, which is exactly the averaging failure mixture density networks exist to avoid. Still, if score averaging matches PoseMDN on CASF-2016 and DUD-E discrimination, the multimodal and calibration benefits would be the entire contribution, and we would say so plainly.
Why not just run molecular dynamics equilibration or free energy calculations on everything? Physics-based refinement is the quality ceiling: Feng et al. raised average ROC AUC on three DUD-E targets from 0.67 to 0.83 by refining top Glide SP compounds with absolute binding free energy calculations, and crystal-pose calculations in Heinzelmann & Gilson land within 1.5 kcal/mol of experiment. But the same papers show the ceiling is expensive and conditional: equilibration cannot rescue fundamentally wrong poses, and automated free energy workflows cost $12–15 per perturbation on commercial cloud infrastructure (Moore et al.). We do not compete with physics; we use the mixture's width and entropy to select the small subset of ambiguous, high-stakes complexes where refinement is worth its cost — a tiering strategy evaluated directly in the experiments below.
Why not just predict a mean and a variance (a heteroscedastic Gaussian, or a deep ensemble)? A single Gaussian with a predicted spread captures uncertainty but is unimodal: it widens around the average of competing orientations rather than separating them, and it cannot attach an interpretation to any component. Interformer already demonstrated the interpretive payoff of attaching Gaussian components to physical interaction classes (its hydrogen-bond and hydrophobic components), and we extend that principle from interaction types to pose hypotheses. Concretely, we will run the $K{=}1$ version of our own head as the heteroscedastic control; if it matches the mixture on every metric including the multimodal subset, the anchoring claim fails and the simpler model should be preferred — an outcome we will report if it occurs.
Proposed Experiments and Decisive Checks
The first experiment establishes parity on the standard benchmark. On CASF-2016 scoring power, with pose bags built from docking engines and crystal poses included as an upper-anchor condition, we report Pearson and Spearman correlations against the file's reference numbers: EGNN-MIL at $r = 0.904$, DAAP at $R = 0.909$ (a crystal-pose method, included to show the cost of moving to docked poses), and GenScore at $R_p \approx 0.80$. PoseMDN passes this stage if it matches EGNN-MIL's correlation on identical inputs while delivering strictly better likelihood and calibration — metrics the point predictors cannot report natively.
The second experiment is the robustness and crossover study, attacking the unresolved regularizer-versus-destructor contradiction. We perturb crystal poses with controlled noise from 0.5 to 6 ångströms RMSD and track correlation, predictive width, and calibration for PoseMDN, the scalar-head ablation, and single-pose baselines. The hypothesis is that point-predictor correlation degrades early (consistent with the fragility documented for DAAP and ASFP) while PoseMDN's width grows to keep negative log-likelihood flat, and that the noise level where width stops compensating — the regularizer-to-destructor crossover — is measurable and consistent with the FDA pipeline's finding that DiffDock-pose noise improved generalization by 32% on DAVIS.
The third experiment moves to structures with no crystal pose at all, on the DAVIS and KIBA kinase panels under the "both-new" split (unseen proteins and unseen ligands), following the FDA protocol of folding targets, docking with DiffDock, and predicting affinity. The file's numbers to beat are $R_p = 0.29$ on DAVIS and $0.51$ on KIBA for the FDA pipeline, against $0.22$ and $0.45$ for the docking-free MGraphDTA — docking-free regressors are included to confirm that bypassing geometry is not the answer. Because these panels are exactly where pose ambiguity is unmeasurable by crystal reference, the mixture's behavior here is the purest test of pose-marginalization.
The fourth experiment tests the refinement-gating payoff. On the three DUD-E targets used by Feng et al. — BACE1, CDK2, and thrombin, where docking alone achieved ROC AUCs of 0.82–0.87 at baseline and ABFE refinement lifted average AUC from 0.67 to 0.83 on selected tiers — we replicate their tiering but replace score-ranked selection with uncertainty-ranked selection using the mixture's entropy. Success is reaching at least the refined AUC (0.83 average) while spending at most half the free energy calculations, since the file prices those at $12–15 each.
The fifth experiment is the interpretability and multimodality audit, the decisive check for the core claim. On complexes whose pose bags contain two or more RMSD-separated clusters, we measure how often the dominant mixture components align with the correct clusters, and whether per-mode predictions bracket the experimental value where EGNN-MIL's single prediction falls between modes. A synthetic variant — deliberately flipping ligands 180 degrees to manufacture bimodal bags — provides ground truth for mode separation that experimental data cannot. If the components do not align with pose basins above chance, anchoring is decorative, and the free-bag ablation should be preferred on simplicity grounds; that falsification would itself resolve how much of the multimodality story is real.
The program is falsifiable at three levels: the density may not beat the scalar head (then only calibration is gained), the anchoring may not align with poses (then only the density output matters), or the uncertainty may not gate refinement efficiently (then the pragmatic payoff fails even if the science holds). Each failure mode has a named experiment above, and the file's reported numbers give every comparison a concrete bar.
References
- CarsiDock: a deep learning paradigm for accurate protein–ligand docking and screening based on large-scale pre-training
- Interformer: an interaction-aware model for protein-ligand docking and affinity prediction
- A generalized protein–ligand scoring framework with balanced scoring, docking, ranking and screening powers
- BioLM-Score: Language-Prior Conditioned Probabilistic Geometric Potentials for Protein-Ligand Scoring
- Protein-ligand binding affinity prediction using multi-instance learning with docking structures
- Generating Multiple Hypotheses for 3D Human Pose Estimation With Mixture Density Network
- Predicting Real-World Penny Auction Durations by Integrating Game Theory and Machine Learning
- GraphMDN: Leveraging graph structure and deep learning to solve inverse problems
- Absolute binding free energy calculations improve enrichment of actives in virtual compound screening
- A Folding-Docking-Affinity framework for protein-ligand binding affinity prediction
- Distance plus attention for binding affinity prediction
- Transfer Learning-Assisted Inverse Modeling in Nanophotonics Based on Mixture Density Networks
- Automation of absolute protein-ligand binding free energy calculations for docking refinement and compound evaluation
- Automated relative binding free energy calculations from SMILES to ΔΔG