Screening Chemical Space Against a Moving Target

Drug discovery searches a space of perhaps $10^{23}$ plausible molecules for ones that bind a particular protein. The search is usually guided by docking: a program that places each candidate into a frozen snapshot of the protein's shape and scores the fit. But proteins are not frozen. A binding pocket opens, closes, and reshapes itself, and some of the best drug targets — the KRAS cancer target, for instance — hide their real binding site behind a side chain that only swings aside when a ligand is already pushing on it. The honest way to score a candidate is therefore molecular dynamics (MD), a physical simulation of the protein and ligand jostling together. That is computationally hopeless at library scale: MD must take femtosecond steps, so simulating the microsecond-to-millisecond events that matter means billions of sequential force calculations, and running that per candidate across millions of molecules is out of the question.

Recent work has built fast generative stand-ins for MD — models that sample a protein's conformational ensemble in seconds instead of hours — but these surrogates are only ever run on one molecule at a time, after someone has already picked it. Meanwhile the frameworks that search huge chemical libraries efficiently still score every molecule against a rigid, static receptor. This proposal is about closing that loop. The core object that makes the closure affordable is an old one given a new job: the residue proximity graph, a network in which amino-acid residues are nodes and edges connect residues that are physically near each other in the folded protein. As the protein moves, edges form and break. I propose to use the pattern of that edge rewiring, conditioned on the ligand, as the compact currency exchanged between an MD surrogate and a large-scale search over chemical space.

Core Claim

The claim is threefold. First, a ligand-conditioned generative surrogate can be trained to output not all-atom trajectories but distributions over residue contact graphs — soft proximity networks whose edge probabilities say which residue pairs come together when this particular molecule binds — while retaining all-atom detail only inside the pocket where the chemistry happens. Second, this contact-graph distribution is a sufficient scoring signal: a molecule's ensemble-aware binding quality can be predicted from its induced contact signature at a fraction of the cost of generating the full conformational ensemble. Third, because even the cheap signature model becomes a bottleneck across $10^8$–$10^{12}$ molecules, the signature should serve as the partition key for bandit-style exploration — directly extending the hierarchical allocation that BOBA introduced for static docking, but now over dynamics-aware scores. No framework in the retrieved literature couples any MD surrogate into closed-loop chemical-space search; the pieces exist separately and the coupling is the contribution.

Concretely, the score for a candidate molecule $m$ is

$$S(m) \;=\; \mathbb{E}_{G \sim p_\theta(G \mid m)}\big[\, f(m, G) \,\big],$$

where $p_\theta(G \mid m)$ is the surrogate's distribution over pocket contact graphs $G$ induced by $m$, and $f$ scores the molecule against each sampled graph (contact complementarity plus pocket-geometry terms). A distilled "signature" model $\phi(m)$ approximates $S(m)$ without sampling, and the search loop spends its expensive surrogate calls where the signature says the uncertainty-times-payoff is highest.

Why Contact Graphs Are the Right Compression

Three strands of evidence in the file motivate representing receptor dynamics as proximity graphs rather than coordinates. LD-FPG, which generates full-atom protein ensembles by latent diffusion over graph embeddings, found that residue-based pooling — organizing the latent space around residue-level graph structure — best captured side-chain torsion distributions, with a Jensen-Shannon divergence of 0.0224 against MD truth on the dopamine D2 receptor, and it reproduced the active/inactive A100 activation index from a 2-µs trajectory (LD-FPG). HeteroBindNet showed that a network mapping drug atoms to residue contacts recovers about 92% of true physical protein-ligand contacts with no 3D supervision at all, purely from cross-modal attention between a molecular graph and a residue contact map (HeteroBindNet) — evidence that the atom-to-residue contact interface carries most of the binding-relevant information. And the broader graph-representation literature establishes that explicit topology is a strong inductive bias: topology-aware graph encodings consistently beat unordered point clouds in solid-state modeling (MatSciML), and propagating features over residue contact edges improves protein function prediction over sequence baselines (DeepFRI).

The design also answers a documented failure mode of coarse compression. GPCRLMD argues that compressing structures into low-dimensional latent bottlenecks destroys atomic detail, and instead anchors an isometric latent space at full coordinate dimensionality to preserve physical topology (GPCRLMD). My proposal adopts a split answer rather than choosing one granularity: all-atom fidelity is kept exactly where small-molecule chemistry happens (the pocket, at atom-to-residue resolution in the spirit of HeteroBindNet's interaction matrix), while the distal scaffold's slow motions are compressed to contact edges. This also addresses the multi-scale mismatch noted across coarse-grained surrogates like TEMPO (TEMPO) and coarse-grained force-field work (Charron et al.).

Research Program

The work has three components. The first is the surrogate itself: a conditional generative model that takes an all-atom ligand graph plus a dual-granularity receptor graph (all-atom pocket, residue-level scaffold with proximity edges) and generates distributions over pocket contact graphs and pocket-side-chain geometry. Training data exists: MISATO provides ~20,000 protein-ligand complexes with 10-nanosecond all-atom trajectories in explicit water, totaling ~170 µs of simulation (MISATO), and GPCRmd provides 703 trajectories of 500 ns each for ligand-bound GPCRs, the same resource GPCRLMD trained on (GPCRLMD). Architecturally this is a residual or flow-based generator in the family of GPCRLMD and BioMD (BioMD); the difference is the output space — contact-edge distributions rather than (or in addition to) coordinates — chosen so that the output is directly usable as a search signal. The second component is the signature model $\phi(m)$: a small graph network distilled from the surrogate that maps a ligand's 2D molecular graph directly to a predicted induced-contact embedding, fast enough for millisecond-scale inference. The third component is the search loop: library molecules are clustered by signature, clusters are treated as bandit arms with upper-confidence-bound allocation in the style of BOBA, and molecules pulled from winning clusters receive full surrogate sampling (or short confirmatory MD) with results fed back to update both the bandit and, periodically, the signature model.

Cost accounting makes the hierarchy concrete. BioMD generates a full 100-nanosecond trajectory in roughly 56 seconds (BioMD) and GPCRLMD runs roughly 180× faster than BioEmu on 500-nanosecond GPCR trajectories (GPCRLMD) — fast per molecule, but 56 seconds times $10^8$ molecules is still ~177 years. Even the millisecond signature model, as BOBA demonstrated, becomes the bottleneck at that scale, which is precisely why allocation must happen above the per-molecule tier.

Baselines

The naive reference point is uniform random selection from the library, which calibrates how much of the apparent win comes from any sensible prioritization at all. It will be run on the same library and evaluated against the same ground-truth ranking.

The strongest existing competitor is BOBA, which partitions chemical space into clusters via K-means on foundation-model molecular embeddings, treats clusters as bandit arms under UCB1, and fits a Bayesian neural network surrogate locally within selected clusters (BOBA). On a 96-million-compound library it reports a 100-fold reduction in surrogate evaluations at $K=100$ while recovering over 90% of the top-100 molecules, and cut optimization-loop runtime from 2,720 seconds to roughly 2,200 seconds. Its oracle, however, is static docking against a rigid receptor. The fair comparison runs BOBA unchanged (its code path, its tuning budget, its evaluation counts) on the same library and the same compute, but with the ground truth defined by ensemble-aware scores rather than static docking — because the claim under test is exactly that static-dock-ranked hits and dynamics-ranked hits diverge on flexible targets. Beating BOBA means recovering a higher fraction of the ensemble-truth top 100 at an equal surrogate-call budget, and specifically recovering cryptic-pocket binders that its static oracle ranks low.

A second serious baseline is the cascade: dock the whole library with a static scorer, then run the expensive MD surrogate only on the top-$N$ docking hits. This is what a pragmatic practitioner would do today, so it gets the same total surrogate budget, with $N$ tuned on validation targets. Beating it means the closed loop surfaces ensemble-good molecules that the cascade's first stage discards.

The key ablation isolates the dynamic ingredient: the identical bandit loop and budget, but with the partition key replaced by (a) molecule-embedding clustering only, i.e. BOBA's key, and (b) a static contact-map signature in the style of HeteroBindNet's predicted 2D residue contacts. If the dynamic induced-contact signature does not beat both at equal budget, the central claim fails.

Why not just…?

Why not just dock everything into a static crystal structure? Because the file contains direct evidence of where that fails. On MISATO's binding-affinity benchmark, models using only static structural data achieved a Spearman correlation of 0.50, while adding MD-derived pocket-adaptability features raised it to 0.64 (MISATO). FlexSBDD showed that rigid-receptor design is blind to cryptic pockets: in KRAS$^{G12C}$, opening the known cryptic site requires rotating a single histidine side chain, and rigid baselines produced severe steric clashes (6.43 per molecule for DecompDiff, versus 1.39 for the flexible model) (FlexSBDD). Static docking remains a fine first stage, which is why it appears as a baseline rather than being dismissed.

Why not just generate a ligand-blind protein ensemble and dock into each conformer? This is the most natural cheap alternative and I take it seriously. The counter-evidence is GPCRLMD's direct comparison: ligand-blind models such as BioEmu failed to reproduce induced-fit motions, including the transmembrane TM2–TM6 transitions that define GPCR activation, and failed on ligand-receptor contact accuracy, whereas the ligand-aware model reproduced them (GPCRLMD). The physical reason is that the pockets a drug cares about are often stabilized by the drug. FuncBind's experimental results point the same way: 45% of its de novo antibodies bound rigid targets in vitro, but only 2–4% bound flexible targets (FuncBind). Still, ensemble docking will be run as an additional baseline on the screening benchmark, because on targets whose apo ensemble genuinely covers the holo-like states it could match the ligand-conditioned surrogate at lower cost.

Why not just run the surrogate on the top-N static docking hits (the cascade)? This is a real threat, and for targets with rigid, well-formed pockets it will likely work as well as my loop — I am not going to pretend otherwise. The proposal's bet is that the population of targets where static docking and dynamics disagree (cryptic pockets, allosteric GPCR sites) is exactly where lead discovery currently stalls, and there the cascade's first stage has already thrown away the winners. The decisive experiment below is built around this: on such targets, I will measure how many ensemble-truth binders sit in the cascade's discard pile and whether the closed loop recovers them. If the cascade matches the loop even on cryptic-pocket targets, the dynamic signature is not earning its cost and the claim should be abandoned.

Why not just fine-tune an affinity predictor on MISATO's MD features? MISATO's own benchmark shows dynamic features help static predictors (Spearman 0.50 to 0.64), but those features are per-complex summaries of pre-computed trajectories; they cannot answer the counterfactual question the search loop needs — what would the pocket do if this new molecule bound — and MISATO's 10-nanosecond trajectories are too short for large-scale conformational change (MISATO). A regression over fixed features is a baseline component, not a replacement for a conditional generative surrogate.

Experiments and Decisive Checks

Surrogate fidelity comes first, on held-out trajectories. On MISATO, the pocket-adaptability prediction benchmark (the original paper reports a mean Pearson correlation of 0.66) is the target to match or exceed; on GPCRmd, the checks are reproduction of the TM2–TM6 distance fluctuation, following GPCRLMD, and contact-edge rewiring statistics against MD truth, measured as Jensen-Shannon divergence over edge probabilities in the manner of LD-FPG's dihedral evaluation. Physical validity is checked via steric-clash counts, where LD-FPG's reported inflation (1145.6 generated clashes versus roughly 1023 in ground-truth MD on D2R-MD) sets the known failure level to improve on; JAMUN's PoseBusters pass rate of 94.7% (JAMUN) is the reference for validity checks. Because no surrogate can be fully trusted, a few hundred real short MD runs on stratified samples validate the surrogate's rankings before any search experiment.

The search experiment uses the Enamine REAL 5-million-molecule subset and the ZINC subsets used by BOBA, on three target classes: rigid pockets (where the cascade should win and is expected to), conventional flexible pockets, and cryptic-pocket targets in the KRAS$^{G12C}$ mold. Ground truth for ranking is defined by high-sample-count surrogate scoring, validated on the real-MD subsample; metrics are top-100 hit recovery and wall-clock cost at matched surrogate-call budgets, against random selection, BOBA, the cascade, ensemble docking, and the two ablation keys. The single most decisive number is the recovery of ensemble-truth binders from the cascade's discard pile on cryptic-pocket targets: that is the result that would establish, or refute, the value of putting receptor dynamics inside the search loop rather than after it.

Two risks deserve honesty. The signature model may be too weak to steer the bandit — in which case the ablations will show it, and the fallback is to use the surrogate's own cheap samples as the key, accepting a smaller library scale. And the MISATO trajectories are short; the GPCRmd data mitigates this for the GPCR target class, but fidelity claims must be scoped to the timescales the training data actually covers. The broader bet, though, is uncluttered: the field has fast generators of protein motion and fast searchers of chemical space, and nothing joins them. The residue proximity graph — cheap, ligand-conditionable, and directly interpretable as the geometry a drug actually perturbs — is a plausible joint currency, and the experiments above are built to find out whether it pays its way.