Entropy-Budgeted Watermarking for In-Pocket 3D Molecular Generation

Generative models now design drug-like molecules directly inside a protein binding pocket — a diffusion model or flow-matching model that denoises random noise into a 3D ligand whose shape and chemistry complement the target. These models are expensive to build, and their outputs are valuable intellectual property, so it is natural to want to trace where a generated molecule came from. Text and image models solve this with watermarking: a hidden statistical signature embedded at generation time and recovered later. But two existing molecular watermarking systems, FoldMark (protein backbones) and MolMark (isolated small molecules), both perturb the structure uniformly. That is exactly wrong inside a binding pocket, where some atoms sit in a tight physical corridor — one tenth of an ångström of movement creates a steric clash with the receptor — while solvent-exposed tails have ångströms of freedom.

This proposal imports a lesson from text watermarking, where SWEET showed that a watermark survives and the output stays valid only if injection is skipped wherever the model's next-token entropy is low. The molecular analog of a low-entropy token is an atom whose position is pinned by pocket geometry; the analog of a high-entropy token is an atom free to roam. We propose an entropy-budgeted watermark for pocket-conditioned 3D generation: measure each atom's local spatial freedom during denoising, inject watermark signal only where that budget is positive, and gate injection with a granular receptor–ligand interaction check so the watermark never manufactures a "physical hallucination" — a chemically fluent molecule that violates physics, in the sense standardized by PoseBusters. The third input direction, abstractive hallucination, supplies both the diagnosis and the guard: just as a fluent summary can contradict its source, a fluent-looking ligand can contradict its pocket, and the granular pair-level verification style of SummaC becomes our inference-time quality gate.

Core Claim

A watermark for pocket-conditioned molecular generation can be both decodable and physically harmless if and only if it respects a local, per-atom entropy budget. We claim three things. First, that the entropy-budget theory of Thickstun et al. and Catch-22 — detectability is bounded by the entropy of the distribution you perturb — transfers from discrete token distributions to continuous 3D denoising trajectories, with the per-atom predicted coordinate spread playing the role of token entropy. Second, that a uniform watermark of the MolMark kind, moved from isolated molecules into a pocket, will measurably increase PoseBusters-detectable physical violations, because perturbations land indiscriminately on pharmacophoric atoms and solvent-exposed atoms alike. Third, that gating injection by spatial freedom preserves both watermark detectability (the budget is reallocated, not destroyed, since solvent-exposed regions carry most of the geometric entropy) and binding quality.

Research Program

The first work item is defining the spatial entropy budget. For each denoising step $t$ and atom $i$, we estimate a freedom score combining two quantities the generator already has access to: the predicted per-coordinate variance of atom $i$ (available in any diffusion or flow-matching model that predicts a distribution, and approximable otherwise by short trajectory rollouts), and the geometric clearance $d_i$ — the distance from atom $i$ to the nearest receptor atom or molecular surface. Watermark injection at atom $i$ scales as $\delta_i \propto \max(0,\, \sigma_i(t) - \tau)$, where $\sigma_i(t)$ is the freedom score, $\tau$ a threshold chosen per-pocket by the same calibration logic as power-calibrated watermarking, which replaces heuristic tuning with a formal detectability–distortion trade-off. Atoms deep in the pocket interior with near-zero $\sigma_i$ are left untouched, exactly as SWEET leaves syntactically forced tokens untouched; MolPilot shows such positions exist and matter, since pocket geometry rigidly constrains them.

The second work item is the injection mechanism itself. Following MolMark, we use an SE(3)-invariant decoder over the Euclidean distance matrix, so the signature survives rotation and translation of the whole complex; this is a method we reuse, not reinvent. But we replace its post-hoc, global coordinate modulation with in-trajectory, per-atom gating: the watermark loss is applied as a low-weight auxiliary objective during fine-tuning of a pocket-conditioned generator, weighted by $\max(0, \sigma_i(t) - \tau)$. This differs from FoldMark's WaterLoRA in that the modulation is spatially selective and pocket-aware rather than a global adapter, because FoldMark's regime — perturbations diluted across 60–512 residues of a monomeric protein — does not exist for a 20–50 heavy-atom ligand, where MolMark itself notes severe sparsity and sensitivity.

The third work item is the hallucination gate, and it is where abstractive hallucination research enters as a method, not a metaphor. SummaC fixed a granularity mismatch by checking entailment sentence-pair-by-pair-pair instead of document-to-summary; the analogous move here is a per-interaction check: before a watermark perturbation is accepted at atom $i$, we verify that no receptor–ligand contact pair involving $i$ moves across a physical threshold (van der Waals overlap, hydrogen-bond geometry, aromatic planarity). We build this as a differentiable or quasi-differentiable screening pass over the local contact matrix during denoising, applied only at the perturbed atoms, so it costs far less than a full post-hoc PoseBusters run. The motivation comes from the hallucination literature twice over: Guo showed that training objectives themselves teach models to hallucinate when they reward surface plausibility over grounding, and Ben Abacha et al. showed that a single polarity flip preserves surface overlap while inverting meaning — the molecular version being a sub-ångström shift that preserves RMSD while breaking a hydrogen bond.

Proposed Experiments

We use the two benchmarks in the file that directly test pocket-conditioned generation: CrossDocked2020 for in-distribution generation, where MolPilot reports a 95.9% physical validity rate and superior Vina docking scores, and the PoseBusters Benchmark Set (308 complexes) for out-of-distribution generalization, where MolPilot maintained a 79.1% passing rate against MolCRAFT's 57.8%. We fine-tune a pocket-conditioned generator (MolPilot-class or a comparable non-autoregressive baseline, retrained or used via released weights with identical compute and tuning budget across all watermarked variants) on CrossDocked2020, then measure four things on both benchmarks: watermark bit-extraction accuracy under rotation, translation, and modest conformational relaxation; structural distortion in RMSD; PoseBusters pass rate across its chemical, intramolecular, and intermolecular checks; and Vina docking score relative to unwatermarked generations from the same model and seeds. For OOD robustness we additionally evaluate on targets below 30% sequence identity to training, the regime PoseBusters showed is where deep methods collapse (DiffDock fell to 12% valid-and-accurate while classical AutoDock Vina held 58%).

The decisive checks are three. First, the capacity curve: bit accuracy at 8, 16, and 32 bits as a function of the threshold $\tau$, testing whether the solvent-exposed budget alone carries enough signal — FoldMark reports nearly 100% bit accuracy at 4–16 bits and 85.0% at 32 bits on proteins, and our hypothesis is that small ligands saturate earlier, with the entropy budget explaining where. Second, the hallucination dose–response: PoseBusters pass rate as watermark strength scales up, for uniform versus gated injection. If gated injection keeps pass rate within roughly two points of the unwatermarked baseline at matched detection power — the standard SWEET set, which kept Pass@1 within 2–4% — the core claim holds. Third, the strip test: rename-level molecular analogues of paraphrase attacks, namely re-docking the watermarked ligand into the pocket and re-generating its conformer, since SWEET showed concentration in high-entropy positions is exactly where signature stripping attacks, and our design must survive re-docking because that is how any downstream user would actually process the molecule.

Baselines

The "do nothing" reference point is the unwatermarked pocket-conditioned generator itself, which anchors quality: on CrossDocked2020, the strongest reported number in the file is MolPilot's 95.9% physical validity rate, and on the PoseBusters Benchmark Set MolPilot reports a 79.1% passing rate (MolPilot). Any watermarked variant must stay within a small margin of these figures while adding a decodable signature; this baseline receives identical compute and seeds so the comparison isolates the watermark's cost.

The strongest existing competitor is MolMark, the only published watermark for small 3D molecules, which reports over 95% extraction accuracy at under 0.03 Å RMSD distortion with over 90% property retention on QM9 and GEOM-DRUG. It was never evaluated inside a pocket or against PoseBusters checks, so the fair comparison requires us to retrain its encoder–decoder on CrossDocked2020 pocket–ligand pairs with the same tuning budget as our method, and evaluate it under identical conditions. Beating MolMark means holding comparable bit accuracy at 16 bits while sustaining a materially higher PoseBusters pass rate and Vina score in-pocket — specifically, if uniform MolMark injection in-pocket drops physical validity by more than a few points while gated injection does not, the entropy budget is doing real work.

The key ablation is our own method with the gate removed: the same injection mechanism, the same bit capacity, but a uniform perturbation applied to all atoms with no spatial-freedom weighting and no interaction-pair verification. This isolates the proposal's central ingredient — the entropy budget and its hallucination gate — from everything else (the SE(3)-invariant decoder, the fine-tuning recipe, the data). A second ablation removes only the interaction gate while keeping entropy weighting, separating the contribution of the two components; no reported result exists on Lacuna for either ablation, which is itself evidence the question is open.

Why not just…?

Why not just run force field minimization after watermarking? This is the cheapest objection and the file answers it directly. PoseBusters tested post-prediction molecular mechanics minimization and found it improves plausibility scores but does not fix the underlying generalization failure — DiffDock remained at 12% valid-and-accurate on novel targets against Vina's 58%. Worse for this objection, minimization is a coordinate perturbation attack on the watermark itself: MolMark decodes through the Euclidean distance matrix at 0.03 Å precision, and energy-gradient relaxation moves atoms along exactly those coordinates. The strip test in our experiments measures this directly, but the prior is that post-hoc repair and watermark survival pull in opposite directions, which is an argument for building validity in at injection time rather than repairing afterward.

Why not just watermark the SMILES string instead of the 3D coordinates? Standard token-level watermarking is far simpler, and SWEET already handles rigid grammars. But the file gives two concrete reasons it falls short here. FORGE documents that BPE tokenization of SMILES splits rings and functional groups into chemically invalid fragments, so token bias in this grammar breaks validity even before entropy is considered — the same collapse WLLM showed on code before SWEET's thresholding, but with less recoverable entropy to exploit. More fundamentally, a SMILES watermark says nothing about the 3D pose, torsions, or pocket complementarity, which is the asset being protected in structure-based design; the signature we need must survive re-docking, which a 1D string cannot carry. This is a real limitation we accept: our watermark protects the generated geometry, not the molecular identity, and the experiments are scoped accordingly.

Why not just use a WaterLoRA-style global adapter like FoldMark? FoldMark is the closest existing system, achieving near-100% bit accuracy at 4–16 bits with sub-0.1 Å RMSD, so a reviewer could reasonably ask why we do not simply port it to ligands. The answer is arithmetic, not taste: FoldMark dilutes perturbations across thousands of backbone atoms in proteins of 60–512 residues, while a ligand has 20–50 heavy atoms, a regime MolMark explicitly flags for sparsity and sensitivity, and each perturbed atom sits microns of chemical significance from a receptor surface. Porting WaterLoRA without spatial gating is precisely our uniform-injection ablation, so this alternative is not argued around — it is built and tested as the ablation that isolates our contribution, and the hallucination dose–response experiment is what settles whether the difference matters in practice.

Closing Checks and Risks

Two honest risks remain. If pocket-conditioned generators concentrate almost all coordinate entropy in a few terminal atoms, the budget may be too small for a robust signature at 16 bits; the capacity curve tests this first, and the fallback is a hybrid scheme that watermarks discrete chemistry choices (atom types, bond orders at solvent-exposed positions) alongside coordinates, extending SWEET's thresholding across modalities in the spirit of MolPilot's coordinated discrete–continuous treatment. Second, verification could be attacked by stripping solvent-exposed groups entirely; we will report bit accuracy under simulated truncation of the top-budget atoms, since SWEET already showed high-entropy concentration is where watermark stripping bites. The proposal succeeds if, on CrossDocked2020 and the PoseBusters Benchmark Set, gated injection matches MolMark-class detection power at 16 bits while keeping physical validity within roughly two points of the unwatermarked MolPilot-anchored baseline (95.9% in-distribution, 79.1% out-of-distribution) and preserving Vina scores — the first demonstration that provenance tracing and physical validity are compatible inside a binding pocket, rather than trading one for the other.