Designing a small molecule that binds a protein is, thermodynamically, a bargain: the molecule must pay for stripping away water, freezing its own rotations, and — when the binding pocket must change shape to accommodate it — for the work of bending the protein itself. Current generative models that co-create molecules and deformed pockets, such as FlexSBDD, let the protein bend for free. They optimize geometric fit and an empirical docking score, with no term asking whether the deformed pocket is a high-energy conformation few proteins would ever visit. At the same time, the neural scoring functions that come closest to rigorous physics — LumiNet rivals free energy perturbation accuracy on one benchmark — only ever see static, already-bound crystal structures, so they cannot tell a designer which new pocket shapes a ligand could plausibly induce.
This proposal builds a single differentiable binding free energy that prices induced fit, and uses it to steer molecular search. The core claim is that the missing quantity — the receptor's reorganization free energy — can be learned without affinity labels, by training an equivariant graph neural network (a network whose 3D outputs rotate correctly when its 3D inputs rotate) as an unsupervised potential on protein conformations, in the spirit of NERE, which learned ligand-binding energies from unlabeled structures by denoising. Attached to an induced-fit generator, this potential turns "the pocket can be bent into place" into "the pocket can be bent into place at this cost," which is exactly the information chemical space exploration currently lacks.
Core Claim
Free energy is a state function, so binding can be written as a thermodynamic cycle:
$$\Delta G_{\text{bind}} \;=\; \Delta G_{\text{inter}} \;+\; \Delta G_{\text{prot}}^{\text{strain}} \;+\; \Delta G_{\text{lig}}^{\text{strain}} \;-\; T\Delta S$$
where $\Delta G_{\text{inter}}$ is the favorable contact energy in the bound state and the strain terms are the costs of deforming the protein and ligand from their unbound ensembles. Every generative induced-fit model in the literature computes, at best, a proxy for the first term. We claim that $\Delta G_{\text{prot}}^{\text{strain}}$ can be predicted by a protein-only equivariant neural potential $E_\theta$, trained by denoising score matching on unbound protein conformations (no affinity labels needed), so that
$$\Delta G_{\text{prot}}^{\text{strain}} = E_\theta(\text{holo pocket}) - \min_{c \,\in\, \text{apo ensemble}} E_\theta(c)$$
measures how far a generated pocket conformation sits above the unbound ensemble's energy floor. Combining this with physics-parameterized interaction terms in the style of PIGNet and LumiNet yields a full, differentiable $\Delta G_{\text{bind}}$ over co-generated ligand–pocket pairs, which then guides generation: sampled molecules that require pocket shapes costing more interaction energy than they gain are suppressed at generation time, not discovered later by an expensive simulation.
Research Program
Three components, each reusing a technique that already works somewhere in the literature, joined for the first time.
First, the protein strain potential. NERE showed that an $SE(3)$-equivariant energy-based model trained purely by denoising unlabeled PDB structures reaches a Pearson correlation of 0.656 on PDBbind v2020 — above MM/GBSA at 0.647 — at ~100 ms per complex versus ~1 hour. But NERE treats proteins and ligands as rigid bodies, and its authors name torsional flexibility as their open challenge. We take up that challenge on the protein side: parameterize sidechains and local backbone by torsion angles on a torus, as FlexSBDD does for generation, and train $E_\theta$ by denoising on unbound conformational ensembles. Paired apo–holo crystal structures are scarce — the central data problem FlexSBDD faced — so we follow its remedy of augmenting with molecular-mechanics relaxations (OpenMM/Rosetta), and we validate $E_\theta$ directly: predicted strain should correlate with the relaxation work of deforming apo structures toward holo geometries, a check no cited method has performed.
Second, the composite scoring function. Interaction energy comes from physics-informed potential terms (van der Waals, directional hydrogen bonds, rotor entropy — the PIGNet recipe), plus a small supervised affinity head calibrated on PDBbind. Ligand strain is the same denoising-potential idea applied to the ligand alone. The composite is trained and evaluated on the FEP congeneric-series benchmark, where the field's most instructive result lives: AEV-PLIG showed that plain supervised scorers reach only weighted-mean PCC 0.41 there, that 78,000 augmented structures lift this to 0.59, and that FEP+ still leads at 0.68. Our bet is that the remaining gap is partly conformational: AEV-PLIG scores rigid docked poses, so a ligand whose binding requires pocket rearrangement is scored in a geometry the rigid pipeline never produced. Giving the scorer induced-fit geometries to evaluate should close part of that gap without another 10× data.
Third, guided generation and search. The composite energy plugs into an induced-fit flow-matching generator of the FlexSBDD/FlowDock family, via gradient guidance during sampling and as a reranker in latent or fragment-based exploration of chemical space, in the mode ChemFlow uses for latent traversal but with 3D receptor feedback at every step rather than post-hoc docking. One design constraint throughout: the network must be $SE(3)$-equivariant (rotations and translations, but not reflections), because DiffSBDD showed that $E(3)$-equivariant networks treat mirror-image molecules as identical and thereby lose the stereochemistry that real binding depends on. FlexSBDD, though $E(3)$-based, demonstrates the co-generation machinery we need; we rebuild it under $SE(3)$ with the DiffSBDD cross-product trick.
Baselines
The naive reference point is unguided generation scored post hoc by AutoDock Vina — the de facto standard pipeline behind DiffSBDD, FlexSBDD, and ChemFlow's docking evaluations, with no thermodynamic input of any kind. It is the "do nothing" comparator: any claim that energy-aware guidance matters must beat it on molecule quality and on predicted-strain sanity.
The strongest generation-stage competitor is FlexSBDD itself, the current best co-generator of ligands and flexible pockets: on CrossDocked it reports an average Vina score of −9.12 versus −8.39 for its strongest baseline, and it cut steric clashes from 6.43 (DecompDiff) to 1.39 while rediscovering the KRAS$^{G12C}$ cryptic pocket. To be fair, we run the same generator backbone, training data (CrossDocked plus its synthetic apo relaxations), sampling budget, and Vina evaluation, changing only the guidance signal; beating FlexSBDD means generating molecules with equal or better Vina scores and clashes, but with a materially lower distribution of predicted $\Delta G_{\text{prot}}^{\text{strain}}$ — and, decisively, higher ranking accuracy against experiment on held-out congeneric series where Vina is known to mislead.
The strongest ranking-stage competitor is augmented AEV-PLIG, which reports weighted-mean PCC 0.59 and Kendall's $\tau$ 0.42 on the FEP benchmark (versus 0.41/0.26 unaugmented), with FEP+ itself at 0.68/0.49 as the physical gold standard. We give AEV-PLIG the same input complexes our method sees, including the induced-fit geometries our generator produces for it — its architecture accepts docked poses — and the same tuning budget. Beating it means PCC above 0.59 on the same series, ideally closing toward the FEP+ ceiling.
Finally, the key ablation removes the proposal's central ingredient: the identical model, identically trained, but with the strain term zeroed so scoring reduces to interaction energy on co-generated poses. If the ablation matches the full model on congeneric-series ranking and on strain validation, the core claim is wrong and the strain potential is decoration; the proposal stands or falls on this comparison.
Why not just…?
Why not just dock into rigid structures and rescore the top hits with FEP+? For a known holo pocket and a short candidate list, this genuinely works — FEP+ reports PCC 0.68 on the congeneric benchmark, above every neural method, and AEV-PLIG's authors rightly treat it as the gold standard. It fails in exactly the regime this proposal targets: docking into apo or non-cognate structures collapses because of steric clashes and invisible cryptic pockets, which is why rigid baselines produce 6.43 clashes per complex where flexible co-generation produces 1.39, and why the KRAS$^{G12C}$ cryptic pocket is unreachable without a His flip that rigid docking cannot perform. And FEP+ costs hours to days per compound, so it cannot sit inside a search loop over millions of candidates; it can only referee the finalists. The decisive experiment is stratification: on benchmarks split by apo-input and pocket-flexibility level, rigid-then-FEP+ should lose badly on the flexible/apo strata while remaining strong on rigid holo strata — and we will report both, because on the latter it may well win.
Why not just keep FlexSBDD and run FEP+ post hoc on its outputs? This two-stage pipeline — flexible generation, then physical rescoring of the top few hundred — is a real threat, and we will run it as a baseline rather than argue around it. Its weakness is that it corrects only after the search budget is spent: the generator, steered by Vina, fills its candidate pool with molecules optimized for the wrong objective, and FEP+ can only re-rank what survived. Our claim is that differentiable guidance during search concentrates the pool on thermodynamically viable molecules in the first place, so the same number of FEP+ referee calls yields more confirmed binders. The measurable prediction is enrichment: at fixed FEP+ call budget (say 200 per target), guided generation should place more true sub-micromolar ligands inside the called set than Vina-guided FlexSBDD does.
Why not just train a bigger supervised scorer on more augmented structures? AEV-PLIG's own results cap this route: even with 78,000 extra docked structures it reaches 0.59 versus FEP+'s 0.68 on congeneric series, and on the out-of-distribution test its PCC falls from 0.86 to 0.73. PIGNet shows the same memorization trap from the other side — deep scorers without physical inductive bias collapse (KDEEP's 29.1% docking success versus 87.0% with embedded potentials on CASF-2016). Scaling supervised data attacks the symptom, not the cause; the cause is that the scorer never sees the conformational physics. Our strain potential is trained unsupervised on protein conformations alone, so it cannot memorize ligand chemotypes by construction, and the OOD test is where we expect the gap to show.
Proposed Experiments
Experiment one validates the strain potential in isolation. On apo–holo pairs (real where available, OpenMM/Rosetta-relaxed synthetic pairs otherwise, the FlexSBDD protocol), we test whether predicted $\Delta G_{\text{prot}}^{\text{strain}}$ correlates with computed deformation work, and whether it separates true holo-like deformations from the clash-resolving but high-strain deformations a geometric generator prefers. No cited method reports a strain-prediction metric, so there is no reported result found on Lacuna for receptor reorganization energy prediction against which to calibrate; establishing this baseline is itself a contribution.
Experiment two tests ranking on the FEP congeneric-series benchmark, scored as weighted-mean PCC and Kendall's $\tau$. The published ladder is AEV-PLIG augmented at 0.59/0.42, FEP+ at 0.68/0.49, and, on the related FEP1 absolute-energy set, LumiNet at RMSE 1.13 kcal/mol versus FEP+ at 1.08. Success is PCC significantly above 0.59 with the strain term, and the ablation (interaction-only) landing at or below it — which would directly implicate conformational geometry, not data volume, in the residual gap. On PDBbind v2020 core, the comparison is NERE's 0.656 and MM/GBSA's 0.647, evaluated in an apo-input regime where the rigid-pose methods are handicapped by construction.
Experiment three tests generation on CrossDocked under matched conditions against FlexSBDD's reported −9.12 Vina average and 1.39 average clashes. The honest expectation is a slightly worse mean Vina score — we are deliberately refusing free pocket deformation that Vina rewards — traded for lower strain and better downstream ranking. The KRAS$^{G12C}$ cryptic-pocket case from FlexSBDD serves as the qualitative test: can energy-guided generation still reach the cryptic site (it requires one rotamer flip, a cheap deformation), while refusing geometrically convenient but thermodynamically implausible large loop motions?
Experiment four runs the closed search loop: guided and Vina-guided generators explore the same pocket with the same sample budget; the top 200 molecules from each go to FEP+ as referee. The decisive check is enrichment of confirmed strong binders per referee call, plus a PoseBusters-style physical-plausibility audit (FlowDock reports 51% blind docking success at RMSD ≤ 2 Å there, a reference point for pose quality under apo inputs). If Vina-guided search matches our enrichment at equal referee budget, the differentiable energy is not earning its cost, and the two-stage pipeline wins — we will say so.
The program is deliberately falsifiable at three joints: strain prediction validity, ranking gains over augmented rigid-pose scorers, and search enrichment over post-hoc rescoring. Each has a named number to beat drawn from published results, and each failure mode points to a specific simpler method that should then be preferred.
References
- FlexSBDD: Structure-Based Drug Design with Flexible Protein Modeling
- Robust protein–ligand interaction modeling through integrating physical laws and geometric knowledge for absolute binding free energy calculation
- Unsupervised Protein-Ligand Binding Energy Prediction via Neural Euler's Rotation Equation
- PIGNet: a physics-informed deep learning model toward generalized drug–target interaction predictions
- Narrowing the gap between machine learning scoring functions and free energy perturbation using augmented data
- Structure-based Drug Design with Equivariant Diffusion Models