When a small molecule binds a protein, the protein is not a statue. It bends: side chains rotate out of the way, loops shift, and sometimes whole domains swing open to expose a cavity that did not exist a moment earlier. This is called induced fit, and the newest AI docking and molecule-generation models try to capture it by letting both the molecule and the protein move together. But there is a structural blind spot built into every one of these models: they only simulate a cut-out window of the protein around the binding site, typically residues within 15–20 Å, and treat everything outside that window as if it did not exist — like studying waves in a swimming pool with the far half of the pool replaced by a concrete wall. Mechanics says this is wrong. Analyses of 34 allosteric proteins show that binding-induced deformations travel through the entire protein along soft elastic modes whose stiffness shrinks with protein size, roughly as $k_a \sim 1/N$ for $N$ residues Mechanics of Allostery: Contrasting the Induced Fit and Population Shift Scenarios — precisely the motions a fixed window cannot contain.
This proposal takes the partitioning logic of multi-scale quantum-continuum simulation — a fine-grained quantum core, an atomistic sheath, a coarse continuum bulk, and carefully designed interfaces between them — and rebuilds induced-fit modeling on it, using equivariant graph neural networks as the common representation at every scale. The payoff is twofold: receptor deformation predictions that respect long-range elasticity, and a physically sane flexible evaluator that chemical-space search can actually afford to call millions of times.
Core Claim
Every state-of-the-art flexible design model is pocket-truncated. FlexSBDD co-generates ligands and flexible pockets with Riemannian flow matching over backbone frames and sidechain torsions, and famously opened a cryptic pocket on KRAS$^{G12C}$ by rotating a blocking histidine FlexSBDD: Structure-Based Drug Design with Flexible Protein Modeling; DynamicFlow drives an atom-level equivariant network and a residue-level transformer through a stochastic flow from apo to holo states Integrating Protein Dynamics into Structure-Based Drug Design via Full-Atom Stochastic Flows; FABFlex iteratively regresses ligand and pocket coordinates with three equivariant modules Fast and Accurate Blind Flexible Docking. All three cut the protein at a fixed radius, which imposes an artificial infinitely stiff boundary exactly where physics says the soft, system-spanning modes live. We claim that replacing the hard cut with a learned continuum-elastic coupling — and explicitly controlling the spurious "ghost forces" that atomistic–continuum interfaces are known to produce — will improve apo-to-holo deformation prediction and flexible docking, at close to pocket-only compute cost.
The ghost-force problem is not hypothetical. In quasicontinuum coupling, mismatch between atomistic energy sums and local continuum stress creates non-zero forces at perfect equilibrium, and in two dimensions the polluted boundary layer widens to $\mathcal{O}(\sqrt{\epsilon})$ with an $\mathcal{O}(1)$ error in the displacement gradient, where $\epsilon$ is the lattice spacing Ghost Force Influence of a Quasicontinuum Method in Two Dimension. Any neural architecture that grafts a discrete pocket onto a continuum bulk inherits this failure mode unless it is trained against it. No existing work does this: rotation-equivariant networks have been used to simulate continuum mechanics on meshes REMuS-GNN: A Rotation-Equivariant Model for Simulating Continuum Dynamics, and equivariant networks have been used for atomistic pockets Integrating Protein Dynamics into Structure-Based Drug Design via Full-Atom Stochastic Flows, but nobody has coupled the two in one equivariant model with an interface-error objective.
Research Program
We build a three-zone model that mirrors the quantum-core / atomistic-sheath / continuum-bulk partition of hybrid Car-Parrinello QM/MM simulation, which embedded DFT active sites inside classical AMBER or GROMOS environments for systems up to 24,000 atoms Hybrid QM/MM Car-Parrinello Simulations of Catalytic and Enzymatic Reactions. The inner zone is the binding pocket, represented at full atom resolution by an equivariant scalar-vector network of the DynamicFlow type, driving a flow-matching trajectory from apo to holo. The outer zone is the distal protein, represented not as atoms but as a coarse elastic body: residues become nodes of a mesh carrying displacement fields, processed by a rotation-equivariant continuum network in the spirit of REMuS-GNN, which has already shown that equivariant message passing can reproduce continuum-mechanics fields on irregular meshes. Between them sits the contribution we care most about: a learned coupling layer that passes boundary displacements and tractions in both directions, trained with an explicit ghost-force penalty. Writing $\hat{F}_i(X)$ for the model's predicted force on atom $i$ and $X^*$ for a reference equilibrium configuration (equilibrated snapshots from molecular dynamics), the interface residual is
$$G \;=\; \frac{1}{|B|}\sum_{i \in B} \big\lVert \hat{F}_i(X^*) \big\rVert^2,$$
where $B$ is the band of atoms and mesh nodes near the atomistic–continuum boundary. At a true equilibrium every force vanishes; $G$ measures exactly the spurious force that theory says a naive interface produces Ghost Force Influence of a Quasicontinuum Method in Two Dimension. Penalizing $G$ during training, on configurations where the correct answer is zero, is the key ingredient the proposal isolates. The physical prior comes from the elastic-mode analysis: because the soft mode stiffness scales as $1/N$, the continuum trunk must represent a body whose size actually enters the mechanics, not a decorative backdrop Mechanics of Allostery: Contrasting the Induced Fit and Population Shift Scenarios.
Training data comes from MISATO, which pairs 5,692 protein-ligand complexes with over 46,000 molecular-dynamics conformational snapshots — real, physically equilibrated trajectories rather than the OpenMM- or Rosetta-relaxed synthetic apos that bias FlexSBDD and DynamicFlow training FlexSBDD: Structure-Based Drug Design with Flexible Protein Modeling — supplemented by Apobind's experimentally resolved apo–holo pairs. A second phase upgrades the scoring core: current flexible models score with empirical AutoDock Vina energies, which miss electronic effects entirely — classical force fields could not explain why synthetic mimics of galactose oxidase had a 5 kcal/mol higher barrier than the natural active site, because the difference came from unpaired-electron delocalization Hybrid QM/MM Car-Parrinello Simulations of Catalytic and Enzymatic Reactions. Equivariant machine-learned electronic-structure models have already delivered 10,000× speedups over DFT in a device-interface setting Ab initio Modeling of MoS2/Oxide Device Interfaces with Machine Learned Electronic Structures; we will anchor the pocket-ligand interaction energy to DFT-calculated training points the same way.
Finally, the trained model becomes the flexible oracle inside a closed molecular-design loop. Chemical-space search today is decoupled from receptor motion in both of its main styles: SCENT assembles molecules from purchasable building blocks with a Generative Flow Network but scores them with static empirical surrogates Scalable and Cost-Efficient de Novo Template-Based Molecular Generation, and ChemFlow navigates a continuous latent space of 4.2 million molecules with PDE-regularized flows but with no 3D receptor at all Navigating Chemical Space with Latent Flows. We plug our co-deformation model in as the reward: each candidate fragment assembly is scored against the holo pocket the model predicts for it, so search can find molecules that bind only to conformations a rigid grid never shows.
Baselines
The naive reference point is rigid-receptor docking. DiffDock, a diffusion model over ligand translation, rotation, and torsions with an equivariant score network, reports 38.2% top-1 success (ligand RMSD under 2 Å) on PDBBind, but collapses to 21.7% on computationally folded ESMFold receptors and to 12% physically valid and accurate poses on the PoseBusters set of 308 post-2021 complexes, where classical AutoDock Vina holds 58% validity DiffDock: Diffusion Steps, Twists, and Turns for Molecular Docking; PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences. It anchors the floor.
The strongest competitor on our primary question — flexible docking against AlphaFold2-predicted apo receptors — is FABFlex, which decomposes blind flexible docking into pocket prediction, ligand regression, and pocket adaptation modules refined over 6 iterative rounds; it reports 40.59% top-1 success (RMSD under 2 Å) on PDBBind v2020 with AlphaFold2 apo structures, 32.46% on unseen receptors, and 0.49 seconds per pair Fast and Accurate Blind Flexible Docking. On generative design against flexible pockets, FlexSBDD reports an average Vina score of −9.12 on CrossDocked against −8.39 for its strongest baseline, with steric clashes cut from 6.43 to 1.39 FlexSBDD: Structure-Based Drug Design with Flexible Protein Modeling, and DynamicFlow reports a Vina score of −7.65 against −7.84 for natural ligands on MISATO/AlphaFold2-apo, with 52.5% of generations beating the native ligand's predicted affinity Integrating Protein Dynamics into Structure-Based Drug Design via Full-Atom Stochastic Flows. We will rerun the strongest of these on identical splits (PDBBind v2020, MISATO, CrossDocked, PoseBusters) under a matched tuning budget — a fixed random-search budget over hyperparameters and a compute envelope anchored to FlowDock's reported one week on 4 GPUs FlowDock: Geometric Flow Matching for Generative Protein-Ligand Docking and Affinity Prediction — rather than quoting their published setups. Beating FABFlex means exceeding 40.59% top-1 success by at least five points on the same PDBBind v2020 / AlphaFold2-apo protocol, exceeding FlowDock's 51% PoseBusters blind-docking rate at equal or better physical validity, and cutting the ghost-force residual $G$ by an order of magnitude relative to the truncation ablation below. The key ablation isolates our ingredient: the identical architecture with the continuum trunk removed and a hard 15–20 Å cutoff restored, plus a second variant with the pocket simply enlarged to 30 and 45 Å. If the enlarged pocket matches the coupled model, our claim fails and we will say so.
Why not just…?
Why not just enlarge the pocket cutoff? This is the most serious objection, and we treat it as the decisive ablation rather than arguing around it. Two things suggest it falls short: the soft elastic modes have system-spanning wavelength with stiffness scaling as $1/N$, so any finite cutoff misestimates their stiffness and cannot represent domain closure beyond the window Mechanics of Allostery: Contrasting the Induced Fit and Population Shift Scenarios; and atom count grows roughly with cutoff volume, so a 45 Å pocket costs many times a 20 Å one while the continuum trunk scales with residue count. But if the 45 Å ablation matches our accuracy at acceptable cost on the deformation benchmarks, the simpler recipe wins and the paper becomes a negative result with a useful boundary. The experiment settles it either way.
Why not just dock rigidly and relax with molecular dynamics? The literature documents the failure at both ends. Rigid docking into apo structures rejects correct scaffolds before any relaxation can run, because scoring functions treat minor sidechain overlaps as prohibitive TANKBind: Trigonometry-Aware Neural NetworKs for Drug-Protein Binding Structure Prediction; Protein-protein binding site identification by enumerating the configurations, and on KRAS$^{G12C}$ rigid baselines could not reach a pocket blocked by a histidine that only co-generation rotates away FlexSBDD: Structure-Based Drug Design with Flexible Protein Modeling. At the other end, running MD to equilibrium on thousands of candidates takes days to weeks per batch, where DynamicFlow's joint relaxation produces 10 molecules in about 36 seconds Integrating Protein Dynamics into Structure-Based Drug Design via Full-Atom Stochastic Flows — a gap of four to five orders of magnitude that no screening campaign can bridge.
Why not just repair AI poses with force-field minimization? PoseBusters showed that classical minimization fixes local bond geometry but cannot rescue a ligand placed in the wrong global binding mode, and that deep-learning methods keep only 12% physically valid accurate poses on novel sequences PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences. It also inherits the blind spot that motivates our QM-anchored scoring: classical force fields missed the 5 kcal/mol electronic-delocalization effect in galactose oxidase entirely Hybrid QM/MM Car-Parrinello Simulations of Catalytic and Enzymatic Reactions. We will nonetheless run a minimization-repaired variant of our own model's truncated ablation, so the comparison is against the best version of this objection, not a straw one.
Why not just generate freely and filter by retrosynthesis? For the search component, generating unconstrained 3D structures and filtering afterward wastes the oracle budget: SCENT documents that unconstrained generative models mostly yield molecules that fail retrosynthetic planning or need expensive starting materials, while template-based assembly with a dynamic fragment library guarantees synthesizability and cut synthesis costs by 30–50% Scalable and Cost-Efficient de Novo Template-Based Molecular Generation. Our loop therefore generates in fragment space from the start, with the flexible model as reward — and we will measure whether static-grid versus co-adapted rewards actually change the discovered scaffold set.
Decisive Experiments
The first experiment is the ghost-force audit: on MISATO equilibrated snapshots, where true forces vanish, we compare $G$ for the hard-truncation model, the enlarged-pocket variants, and the coupled model. Theory predicts the truncated interface produces a spurious force band of $\mathcal{O}(1)$ gradient error Ghost Force Influence of a Quasicontinuum Method in Two Dimension; we test whether the learned coupling suppresses it by an order of magnitude, and whether the deformation error decays with cutoff radius as slowly as the $1/N$ mode-stiffness scaling implies Mechanics of Allostery: Contrasting the Induced Fit and Population Shift Scenarios. The second experiment is apo-to-holo prediction on MISATO and Apobind, scored on both pocket RMSD and whole-protein displacement — the latter is where we expect the unique signal, and we note that no reported result found on Lacuna for whole-protein apo-to-holo deformation accuracy on Apobind, so we will establish this metric ourselves against the DynamicFlow and FABFlex baselines rerun under matched budgets. The third is flexible docking on PDBBind v2020 with AlphaFold2 apo receptors against FABFlex's 40.59%, plus the PoseBusters 308-complex set against FlowDock's 51% blind-docking rate and DiffDock's 12% validity figure, running all 18 physical checks PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences. The fourth is a cryptic-pocket case study on KRAS$^{G12C}$-style targets, asking whether continuum coupling opens pockets that pocket-only models cannot. The fifth closes the loop: a SCENT-style fragment search over 64,000 Enamine building blocks with our model as flexible reward, measuring how the discovered scaffold set diverges from a static-grid run and whether oracle cost stays within an order of magnitude of DynamicFlow's 36-seconds-per-10-molecules figure. If the enlarged-pocket ablation matches us everywhere, the honest conclusion is that elasticity-aware coupling is unnecessary at current benchmark scale — but the scaling argument, and the ghost-force audit, predict otherwise.
References
- Mechanics of Allostery: Contrasting the Induced Fit and Population Shift Scenarios
- FlexSBDD: Structure-Based Drug Design with Flexible Protein Modeling
- Integrating Protein Dynamics into Structure-Based Drug Design via Full-Atom Stochastic Flows
- Fast and Accurate Blind Flexible Docking
- Ghost Force Influence of a Quasicontinuum Method in Two Dimension
- REMuS-GNN: A Rotation-Equivariant Model for Simulating Continuum Dynamics
- Hybrid QM/MM Car-Parrinello Simulations of Catalytic and Enzymatic Reactions
- Ab initio Modeling of MoS2/Oxide Device Interfaces with Machine Learned Electronic Structures
- Scalable and Cost-Efficient de Novo Template-Based Molecular Generation
- Navigating Chemical Space with Latent Flows
- DiffDock: Diffusion Steps, Twists, and Turns for Molecular Docking
- PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences
- FlowDock: Geometric Flow Matching for Generative Protein-Ligand Docking and Affinity Prediction
- TANKBind: Trigonometry-Aware Neural NetworKs for Drug-Protein Binding Structure Prediction
- Protein-protein binding site identification by enumerating the configurations