When a small molecule docks into a protein, the protein is not a frozen lock: sidechains rotate, loops shift, and sometimes an entire hidden pocket opens only once the ligand is in place. Deep learning models now handle this "induced fit" with generative transport — diffusion models that denoise random noise into samples, or flow matching that learns a vector field carrying a broad prior distribution onto the distribution of bound complexes. But every current model treats the competing conformational pathways that lead to the same bound state as probabilities to be added. Collider physicists learned long ago that this bookkeeping is wrong whenever two routes produce indistinguishable final states: there one must add complex amplitudes and square afterwards, and the resulting cross term suppresses probability in some regions while piling it up in others, shifting peaks and carving dips into measured spectra.
This proposal borrows that bookkeeping — the mathematics, not a claim of quantum mechanics governing proteins. We give a flexible docking model a complex amplitude with two branches: a "continuum" branch describing the equilibrium fluctuations of the empty pocket, and a "resonance" branch describing the ligand-induced bound state, joined by a learned phase field. Squaring the sum lets the two branches cancel in configurations where atoms would sterically overlap and reinforce where contacts are correct — something an additive mixture of probabilities cannot express. The physical motivation is direct: in $e^+e^-$ annihilation, neglecting resonance–continuum interference biases extracted resonance parameters at a level rivaling total experimental uncertainty Dielectron Widths of the $\Upsilon(1S, 2S, [3S)$ Resonances](/work/dielectron-widths-of-the-mml-math-xmlns-mml), and the interference term produces measurable dips in cross sections just below resonance peaks Measurement of the total hadronic cross section in $e^+e^-$ annihilation below 10.56 GeV. In docking, the analogous sin is the incoherent additive background: models that average over conformational pathways produce overlapping atoms and distorted poses, and the failures are well documented PoseBusters, Fast and Accurate Blind Flexible Docking.
Core Claim
A single architectural change — replacing the real-valued score of a flow-matching docking model with a two-branch complex amplitude linked by a learned relative phase — will raise the rate of poses that are both geometrically accurate and physically valid, especially when docking into unbound (apo) structures and cryptic pockets, where additive models fail most. The claim is deliberately narrow and falsifiable: the interference cross term is the only difference from a matched real-valued baseline, so a zero-phase ablation with equal parameter count either confirms or refutes it.
Research Program
The model, which we call a coherent amplitude docking network, operates on the joint configuration $q$ of ligand atoms, pocket residue frames, and sidechain torsion angles (the torus-space parameterization that FlexSBDD showed preserves bond geometry and matters more than backbone motion). It is an $\mathrm{E}(3)$-equivariant network with real and imaginary output channels, evolved for roughly twenty Schrödinger-like update steps rather than the real ODE steps of standard flow matching. Its amplitude is the sum of two branches:
$$\psi(q) \;=\; \psi_{\mathrm{cont}}(q) \;+\; e^{\,i\theta(q)}\,\psi_{\mathrm{res}}(q,\ell), \qquad p(q) \;\propto\; |\psi(q)|^2,$$
where $\ell$ is the ligand, $\theta(q)$ is a learned phase field over configurations, and the induced probability $p(q)$ contains the interference term $2\,\mathrm{Re}\!\left(\psi_{\mathrm{cont}}\,e^{-i\theta}\psi_{\mathrm{res}}^*\right)$. Training has two parts. First, samples drawn from $|\psi|^2$ must match the distribution of holo complexes. Second, an interference loss sets the sign of the cross term: destructive where any atom pair sits inside van der Waals overlap, constructive at correct contact distances. We constrain $\theta$ to vary smoothly across conformation space, mirroring the physics assumption that the continuum phase is slowly varying across a narrow resonance window — the assumption that lets precision scans disentangle the two contributions Dielectron Widths of the $\Upsilon(1S, 2S, [3S)$ Resonances](/work/dielectron-widths-of-the-mml-math-xmlns-mml).
A practical payoff of the two-branch split addresses the scarcity of paired apo/holo structures, which currently forces reliance on synthetic relaxation or predicted apo structures FlexSBDD, DynamicBind, FlowDock. The continuum branch can be trained on abundant unbound pocket ensembles (in the spirit of PackDock's empty-pocket sampler PackDock), the resonance branch on holo complexes, and only the phase field needs the scarce paired data — the same division of labor physicists use when resonance parameters and continuum shape are constrained separately before the interference term is fit on scan data. This also lets us test the tension PackDock leaves unresolved: whether conformational selection and induced fit are genuinely sequential stages, or must be superposed coherently.
We will also port a diagnostic habit from spectroscopy. Plotting the model's induced density along a reaction coordinate (say, ligand–pocket separation or a key torsion) should show an asymmetric lineshape if the mechanism is real: enhancement on one side of the induced-fit barrier, a localized dip on the other, exactly as the $\pi$-radian phase rotation across a Breit-Wigner resonance flips constructive interference below the peak to destructive above it Study of the $\rho'(1600)$ Mass Region Using $\gamma p$ collisions. And we will check for spurious attractors — regions of high $|\psi|^2$ absent from data — the docking analog of the spurious poles that dispersive unitary analysis filters out in three-body amplitudes Pole position of the $a_1(1260)$ from $\tau$-decay.
Proposed Experiments
Evaluation uses the standard suite from the file. Primary benchmark: the PoseBusters Benchmark Set of 308 post-2021 complexes, reporting the valid-and-accurate rate — ligand RMSD $\le 2.0$ Å together with RDKit chemical and steric checks — the metric under which DiffDock fell to "12%" physical validity for accurate poses while AutoDock Vina held "58%" PoseBusters. Secondary: the DISCO cross-docking set, docking into apo and non-native structures, where PackDock reports "48.21%" success at RMSD $< 2.0$ Å and reduced AlphaFold2-model ligand RMSD "from 3.68 Å to 1.77 Å" PackDock; the PDBbind test set, where FABFlex reports "40.59%" success Fast and Accurate Blind Flexible Docking; and the CrossDocked generative setting, where FlexSBDD reports an average Vina Dock score of "-9.12" and reduced average steric clashes from "6.43" to "1.39" FlexSBDD. We add an unseen-receptor split, using FABFlex's "32.46%" unseen-receptor success as the reference point, a cryptic-pocket case study on $\mathrm{KRAS}^{\mathrm{G12C}}$ following FlexSBDD's histidine-rotation example, and the Major Drug Target dataset as a secondary evaluation (no reported result found on Lacuna for a per-family MDT breakdown, so we will establish our own reference numbers there). Runtime is reported throughout, since fast regression baselines average "0.49 seconds" per pair.
Baselines
The naive reference point is incoherent additive docking: sample equilibrium pocket conformers, dock the ligand into each independently, and combine scores by addition — effectively PackDock's two stages run with no joint conditioning, close to classical AutoDock Vina on apo structures, which achieves "21.07%" on DISCO PackDock. This sets the floor and demonstrates the cost of ignoring coupling; beating it is necessary but not interesting.
The strongest existing competitor on our primary benchmark is FlowDock, a conditional flow-matching model that transports ESMFold apo structures to holo complexes and reports a "51% blind docking success rate" at RMSD $\le 2$ Å on the PoseBusters benchmark FlowDock. On the apo cross-docking question specifically, the strongest competitor is PackDock at "48.21%" on DISCO PackDock. Both will be run from their released configurations on identical inputs (same AlphaFold2/ESMFold apo structures where they use them), with our model trained on the same PDBbind v2020 and Binding MOAD data (FlowDock used 46,000 Binding MOAD structures), matched parameter count, the same ~20-step sampling budget, and the same validation split for tuning. Beating them means a higher RMSD $\le 2$ Å success rate and a higher valid-and-accurate rate on PoseBusters, and more than 48.21% success on DISCO.
The ablation that isolates the proposal's key ingredient is the zero-phase variant: the identical two-branch network with $\theta \equiv 0$ and the amplitudes constrained real and non-negative, which reduces exactly to an additive mixture with the interference loss replaced by an equivalent additive clash penalty. Any gain of the full model over this ablation is attributable solely to the interference cross term.
Finally, we apply the PoseBusters post-prediction force-field energy minimization protocol to every system — ours and all baselines — so that the comparison is never confounded by post-processing.
Why not just…?
Why not just run force-field energy minimization after prediction? This is the most serious alternative, because PoseBusters explicitly showed that molecular-mechanics minimization after prediction restores much of the physical validity that deep learning models lose. But minimization is a patch applied to a pose already chosen: it repairs geometry without changing which pose the model prefers, and the same paper shows deep models still fail to generalize to novel sequences. It is also expensive at search time — classical induced-fit docking workflows are "25 times slower" than deep flexible docking PackDock. Our decisive experiment runs minimization on both sides: if minimization equalizes valid-and-accurate rates between our model and the additive ablation, the interference mechanism's value collapses to pose ranking rather than pose geometry, and we will report exactly that.
Why not just parameterize sidechains in torsion space and add a clash penalty? Torsion-space parameterization guarantees physical bond geometry PackDock, and FlexSBDD's clash-aware flow matching already cut average steric clashes from "6.43" to "1.39" FlexSBDD. We regard this as a genuine threat to the proposal, which is why the zero-phase ablation is designed to be exactly this baseline with matched capacity: same torus parameterization, same clash penalty, no phase. If the ablation matches the full model on PoseBusters and DISCO, the hypothesis is refuted and the paper becomes a negative result. What the penalty approach cannot do by construction is resolve conflicting pathways — a local distance penalty cannot express that two conformational routes should cancel each other in one region and reinforce in another — and FABFlex still admits "occasional steric clashes where atoms overlap" despite iterative refinement Fast and Accurate Blind Flexible Docking.
Why not just use fast deterministic regression like FABFlex? FABFlex reaches "40.59%" on PDBbind in "0.49 seconds" per pair, a 208× speedup over DynamicBind's 102.12 seconds Fast and Accurate Blind Flexible Docking, using the public QizhiPei/FABind codebase. But deterministic regression averages over multimodal bound states, and its success drops to "32.46%" on unseen receptors. Our target metrics are validity and apo/unseen generalization, not throughput, and we will state plainly if FABFlex plus minimization matches our model on the DISCO and unseen-receptor splits at a fraction of the cost — in that case the coherent-amplitude machinery would not earn its complexity.
Decisive Checks
Three results decide the proposal. First, the zero-phase ablation gap on PoseBusters valid-and-accurate rate and DISCO success: no gap, no paper. Second, the phase diagnostic: the learned $\theta(q)$ must show a structured $\pi$-rotation across the induced-fit transition coordinate, producing the asymmetric lineshape signature familiar from resonance scans Study of the $\rho'(1600)$ Mass Region, Measurement of the total hadronic cross section below 10.56 GeV; if metrics improve while $\theta$ is structureless noise, the gain is ordinary extra capacity and the mechanism claim fails. Third, the minimization-stack check: our advantage must survive when every competitor receives the same post-processing. If all three hold, the interference formalism earns a place in flexible docking; if not, we will have cleanly bounded what additive probability transport can and cannot do.
References
- Dielectron Widths of the<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>Υ</mml:mi><mml:mo stretchy="false">(</mml:mo><mml:mn>1</mml:mn><mml:mi>S</mml:mi><mml:mo>,</mml:mo><mml:mn>2</mml:mn><mml:mi>S</mml:mi><mml:mo>,</mml:mo><mml:mn>3</mml:mn><mml:mi>S</mml:mi><mml:mo stretchy="false">)</mml:mo></mml:math>Resonances
- Measurement of the total hadronic cross section in<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>e</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>e</mml:mi><mml:mo>−</mml:mo></mml:msup></mml:math>annihilation below 10.56 GeV
- PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences
- Fast and Accurate Blind Flexible Docking
- FlexSBDD: Structure-Based Drug Design with Flexible Protein Modeling
- DynamicBind: Predicting ligand-specific protein-ligand complex structure with a deep equivariant generative model
- FlowDock: Geometric Flow Matching for Generative Protein-Ligand Docking and Affinity Prediction
- PackDock: a Diffusion Based Side Chain Packing Model for Flexible Protein-Ligand Docking
- Study of the<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:msup><mml:mrow><mml:mi>ρ</mml:mi></mml:mrow><mml:mrow><mml:mo>′</mml:mo></mml:mrow></mml:msup></mml:mrow><mml:mn/><mml:mo>(</mml:mo><mml:mn>1600</mml:mn><mml:mo>)</mml:mo><mml:mn/></mml:math>Mass Region Using<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>γ</mml:mi><mml:mi>p</mml:mi><mml:mo>→</mml:mo><mml:mrow><mml:msup><mml:mrow><mml:mi>π</mml:mi></mml:mrow><mml:mrow><mml:mo>+</mml:mo></mml:mrow></mml:msup></mml:mrow><mml:mrow><mml:msup><mml:mrow><mml:mi>π</mml:mi></mml:mrow><mml:mrow><mml:mo>−</mml:mo></mml:mrow></mml:msup></mml:mrow><mml:mi>p</mml:mi></mml:math>at 20 GeV
- Pole position of the <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:msub><mml:mrow><mml:mi>a</mml:mi></mml:mrow><mml:mrow><mml:mn>1</mml:mn></mml:mrow></mml:msub><mml:mo stretchy="false">(</mml:mo><mml:mn>1260</mml:mn><mml:mo stretchy="false">)</mml:mo></mml:mrow></mml:math> from <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>τ</mml:mi></mml:math>-decay