Quantum Interaction Descriptors as Fold-Invariant Supervision for Binding Models

Deep learning models that predict how a small molecule binds a protein look excellent on academic benchmarks and then collapse in practice. The reason, documented repeatedly across binding affinity prediction, docking, and ligand generation, is that these models memorize: they learn which chemical scaffolds go with which protein families in the training set rather than the physical chemistry of non-covalent binding. When a test protein has a fold absent from training, performance falls to near chance Evan N. Feinberg et al.; when a test ligand combines chemical building blocks never seen together, the same thing happens Benson Chen et al.. This proposal attacks the memorization problem at its source: the training signal itself. Instead of supervising binding models on dataset affinity labels — the very thing that encodes dataset-specific pairings — we will supervise them on local quantum interaction physics computed from electron densities, quantities that depend only on the chemistry of the contact and are therefore blind to which protein family the pocket belongs to.

The obstacle has always been cost. Density functional theory (DFT), the workhorse method that determines molecular properties from the 3D electron density $\rho(\mathbf{r})$ instead of an exponentially large wavefunction, scales cubically with electron count and is hopeless for a 5,000-atom binding pocket. But two developments change the calculus. Subsystem DFT, which freezes the environment's density and embeds only the active fragment, scales linearly to quadratically and already handles solutes with 180 explicit solvent molecules Moritz Bensberg et al.. And Molecular DFT computes full 3D water density and polarization maps around a drug-like solute in about a minute, $1{,}000\times$ to $100{,}000\times$ faster than explicit simulation, with a mean unsigned error of 5.9 kJ/mol on FreeSolv that matches explicit Monte Carlo Sohvi Luukkonen et al.. Neither method has ever been coupled to a deep learning model for binding prediction. We propose to do exactly that, and to evaluate the result not on random splits but on the hard structural splits where memorization is exposed.

Core Claim

The claim is specific and falsifiable. If a geometry-based binding scorer is trained on quantum interaction descriptors — the many-body interaction energy between the ligand density and a frozen pocket density, decomposed into its exchange-correlation, electrostatic, and kinetic components, plus explicit MDFT water maps — then its performance should transfer across structural splits, shrinking the gap between in-domain and out-of-distribution evaluation that defines the memorization problem. The existing reference point for that gap is AEV-PLIG, which achieves a Pearson correlation of 0.86 on CASF-2016 but only 0.73 on its out-of-distribution test set; a memorizing model shows exactly this kind of drop, and our claim is that quantum supervision should narrow it. The secondary claim is even cheaper to test: because the raw descriptor contains no trained, fold-specific parameters, it can be used directly as a target-agnostic confidence score for rescoring generative docking poses on entirely novel folds, with zero learning and therefore zero opportunity to memorize.

The central quantity is the embedding interaction energy between the ligand density $\rho_L$ and the frozen pocket density $\rho_P$:

$$E_{\mathrm{int}}[\rho_L, \rho_P] \;=\; E[\rho_L + \rho_P] \;-\; E[\rho_L] \;-\; E[\rho_P],$$

where $E$ is the Kohn-Sham energy functional evaluated at the summed density. Because $E_{\mathrm{xc}}$ at the summed density is manifestly non-pairwise, this quantity captures many-body exchange, polarization, and charge-transfer effects that pairwise classical potentials — the Lennard-Jones and Morse terms inside PIGNet and PIGNet2 — cannot express by construction.

How This Differs From Existing Intersection Work

Three prior lines bring electron density near deep learning, and each stops short of this proposal. EDG distills DFT electron densities into geometry-only student networks with real gains (up to 33.7% average improvement across QM9 properties), but it operates exclusively on isolated gas-phase molecules of at most nine heavy atoms, uses random splits, and requires two million up-front DFT calculations; we extend the distillation idea to macromolecular pockets — which is only possible because subsystem DFT makes the descriptor affordable at that scale — and we evaluate on structural out-of-distribution splits that EDG never touches. EDMolGPT does condition generative design on electron density, recovering bioactive-like ligands for 41% of DUD-E targets versus 8% for Pocket2Mol, but it uses crystallographic density as a static spatial envelope for autoregressive generation, never computing Kohn-Sham interaction energies; we use the density as physics, not as a bounding box. PIGNet2 shares our goal of physics-informed generalization but restricts its physics to empirical pairwise functions, and its own authors attribute its worst failures, such as the water-mediated Eg5 system, to exactly the missing many-body and solvent physics our descriptors provide. The hardware-accelerated DFT line — PySCFIPU, which generated $10^8$ conformers in 13 hours and cut off-equilibrium error from 0.348 eV to 0.0486 eV by data scaling alone, and Self-Refining Amortized DFT — motivates the feasibility of quantum data generation but is confined to isolated molecules of at most 12 heavy atoms; we inherit its lesson that data scale matters while escaping its size ceiling via subsystem decomposition.

Research Program

The program has three components. First, we build a quantum descriptor dataset over the PDBbind refined set, the same 4,514 experimentally measured complexes used to train PIGNet, so that any improvement is attributable to the supervision signal and not to better data. For each complex we compute the frozen pocket density once, then embed the ligand and near-native pose variants, recording $E_{\mathrm{int}}$, its decomposition, and local density-overlap integrals over the contact region; we add MDFT water density and polarization maps for the pocket, following the solvent treatment validated by Sohvi Luukkonen et al.. The per-pocket frozen density is reused across all its ligands, which is what makes virtual-screening-scale amortization possible. We expect roughly the 5% convergence-failure rate on highly charged solutes that the MDFT authors report, and we will report our own rate rather than hide it.

Second, we train a geometry-based student network — a graph network over atomic coordinates, in the SchNet/EGNN family that EDG showed can absorb distilled electronic knowledge — with a dual objective: match the experimental affinity labels and match the quantum descriptor fields via cross-modal distillation from a teacher that sees the descriptors. At inference the student needs only 3D coordinates, so the DFT cost is paid once at training time. We retain PIGNet's derivative regularizer, forcing the learned energy to be stationary at crystallographic poses, since it is architecture-agnostic and orthogonal to our contribution.

Third, we use the raw descriptor, undistilled and untrained, as a physics-only rescoring signal for poses produced by generative docking models on structurally novel targets. This is the purest test of the memorization hypothesis: a quantity with no trainable parameters cannot have memorized anything, so any signal it carries is, by construction, target-agnostic physics.

Baselines

The do-nothing reference is the same student architecture trained on affinity labels alone, evaluated on both a random split and a structural split of PDBbind. This reproduces the "performance mirage" that Alissa M. Hummer et al. describe — high random-split numbers that evaporate under honest splitting — and establishes that any transfer we observe comes from the quantum supervision, not from architecture. Alongside it we include the pure-physics reference: the raw subsystem-DFT interaction energy used directly as a score with no learning, which tells us how much signal lives in the descriptor itself.

The strongest existing competitor is PIGNet2 PIGNet2: a versatile deep learning-based protein–ligand interaction prediction model for binding affinity scoring and virtual screening, because it is the prior model built for exactly our problem — physics-informed generalization beyond memorized pairings — and it holds the numbers to beat: Pearson correlation R = 0.747 on CASF-2016, a DUD-E 1% enrichment factor of 31.2, and an activity-cliff ranking correlation of R = 0.64. We will run it from its released code on the identical PDBbind training data, with the same parameter budget for our student and the same tuning compute (identical random-search budget over hyperparameters, same GPU hours), and we will evaluate both models on identical structural splits that we construct and release. Beating PIGNet2 means three things jointly: matching or exceeding its CASF-2016 correlation, exceeding its out-of-distribution correlation on our structural split by a margin larger than its own in-domain-to-OOD drop, and beating it specifically on water-mediated targets of the Eg5 type where its authors concede failure.

On docking generalization, the competitor is the scaled DiffDock line from Deep Confident Steps to New Pockets: Strategies for Docking Generalization: DiffDock's original 7.1% success rate on DOCKGEN, DiffDock-L's 22.6%, and confidence bootstrapping reaching 9.8% to 24.0% depending on domain, with the best traditional search-based methods around 17.5%. Our rescoring operates on DiffDock's own sampled poses, so the comparison is exactly fair: same poses, same DOCKGEN ECOD-based fold splits, only the confidence signal differs.

Finally, the key ablation isolates our ingredient: the identical student network, identical data and compute, with the quantum distillation term removed (labels only), and a second variant where the quantum descriptors are replaced by pairwise classical descriptors of the PIGNet type computed on the same complexes. This ablation is the experiment — it separates "quantum physics helps" from "any physics-flavored auxiliary loss helps" and from "extra supervision of any kind helps."

Why not just…?

Why not just clean up unphysical poses with a force-field minimization step? PoseBusters tested exactly this: post-prediction molecular mechanics minimization does remove steric clashes and planar distortions, but it does not touch the fold-generalization gap — on the post-2021 benchmark set with under 30% sequence identity to training, DiffDock still reached only 12% valid-and-accurate poses even with cleanup available, versus 58% for AutoDock Vina and 55% for Gold. Force fields with fixed point charges also cannot rank congeneric series, which is where many-body polarization matters AEV-PLIG. Geometry cleanup and quantum supervision address different failures; we need the latter, and our DOCKGEN rescoring experiment measures it directly.

Why not just scale up parameters and training data? This is the most serious alternative, and the literature gives it real force: PySCFIPU showed that scaling from $10^4$ to $10^8$ conformers cut off-equilibrium error from 0.348 eV to 0.0486 eV. But the same paper shows the limit: models trained on ten-heavy-atom molecules fail completely on other sizes without explicit multi-system augmentation — scale fixes interpolation, not composition. On the docking side, scaling DiffDock from 20M to 30M parameters with added data lifted DOCKGEN success from 7.1% to 22.6% and then plateaued, leaving over three-quarters of novel folds unsolved Deep Confident Steps to New Pockets; and AEV-PLIG's 78,000 augmented poses raised FEP-benchmark Kendall's $\tau$ from 0.26 to 0.42 yet stayed below rigorous physics at 0.49 AEV-PLIG. Scaling and quantum supervision are not mutually exclusive — our descriptors are themselves a form of scaling the right data — and our label-only ablation at matched compute is precisely the experiment that separates the two explanations. If the ablation closes the structural gap too, the skeptics win and we will report it.

Why not just use empirical pairwise physics terms, as PIGNet already does? Because the pairwise ceiling is visible in the evidence. PIGNet reaches 87.0% top-1 docking success on CASF-2016, but its own successor concedes that pairwise empirical potentials fail on targets where many-body polarization or explicit water networks dominate binding, naming Eg5 explicitly PIGNet2. Pairwise terms also cannot express charge transfer or orbital-overlap effects, which the density overlap region in our descriptor captures natively. We treat this honestly as a threat rather than a dismissal: the classical-descriptor ablation above runs the comparison inside one architecture, on one dataset, at one budget, so the outcome is decided by measurement rather than argument.

Proposed Experiments and Decisive Checks

Experiment one validates the descriptors themselves: correlation of $E_{\mathrm{int}}$ with measured affinities across PDBbind, reproduction of known mediating-water networks from the MDFT maps, and a transparent accounting of convergence failures against the ~5% rate reported for MDFT on charged solutes Sohvi Luukkonen et al.. Experiment two is the distillation study: student models trained with and without quantum supervision, evaluated on CASF-2016 (target: at or above PIGNet2's R = 0.747), DUD-E screening (target: above EF$_{1\%}$ = 31.2), activity cliffs (target: above R = 0.64), the FEP congeneric benchmark (target: above AEV-PLIG's augmented $\tau$ = 0.42, approaching FEP+'s 0.49), and — decisively — a structural split of our own construction, where the headline metric is the in-domain-to-OOD drop, to be compared against AEV-PLIG's 0.86-to-0.73 fall. Experiment three is the zero-training test on DOCKGEN: subsystem-DFT rescoring of DiffDock's sampled poses on entirely unseen ECOD folds, where beating DiffDock's original 7.1% with no learned parameters would already demonstrate that fold-agnostic physics carries signal, and approaching or exceeding the 17.5% traditional-method level would be strong evidence that memorization, not insufficient data, is the binding constraint. Experiment four applies the PoseBusters physical-validity suite and its under-30%-identity set, comparing against DiffDock's 12% valid-and-accurate rate and AutoDock Vina's 58% PoseBusters, and probes the water-mediated targets where PIGNet2 fails to see whether the MDFT solvent descriptor specifically repairs them. We state the risk plainly: if frozen-density embedding with a GGA functional, corrected with D3/D4 dispersion, proves too crude — self-interaction error is a known limitation of the standard functionals — the descriptor-affinity correlation in experiment one will expose it early, and the fallback is to use the descriptors as auxiliary supervision alongside labels rather than as the primary signal. Either outcome, positive or negative, would be the first rigorous measurement of whether quantum physics is the ingredient that molecular deep learning has been missing when it fails on structures it has never seen.