Drug-discovery companies increasingly deploy generative models that do not just draw molecules but propose recipes for them: which purchasable chemical building blocks to combine, through which known reactions, to obtain the desired structure. The commercial value of such a system sits less in the neural network than in the curated inventory of building blocks and reaction pathways behind it — precisely the asset that is easiest to steal. An adversary who queries the API can collect input–output pairs and train an unauthorized copy, a theft known as model extraction, and the owner, who sees only the outputs, currently has no way to prove it happened. Meanwhile, the field of molecular generation has spent years closing a different gap: generative models freely emit "chemical hallucinations" — molecules that satisfy valence rules on paper but cannot be made in a flask. The accepted fix is to constrain generation to the forward execution of real reactions over a real catalog of purchasable blocks, as in Generating Molecules via Chemical Reactions, SynGA, and ReaSyn.
This proposal's starting observation is that these two threads collide productively. A generator forced to justify every molecule with a synthesis route over a specific inventory must, in its output distribution, reveal something about that inventory — the recipe is only as good as the stockroom it draws from. Nobody has asked whether this leak is statistically exploitable. We propose to turn the synthesizability constraint itself into an ownership signal: the defender plants a small set of chemically inert "canary" building blocks in the proprietary catalog, trains the generator to prefer routes through them by a controlled margin, and later audits any suspect model with a calibrated hypothesis test on canary usage. The question is not merely practical but scientifically open, because Wahdany et al. found that foundation models trained on purely synthetic, rule-generated data leave no dataset-inference trace at all (0.50 ROCAUC, chance level) — and synthesis-constrained generators are, in exactly their sense, trained on rule-generated data. Whether rule-based chemical generation behaves like their synthetic regime or like their real-data regime is unknown and is the empirical heart of this work.
Core Claim
A synthesis-aware generator trained or fine-tuned over a restricted building-block catalog carries a detectable fingerprint in the distribution of leaf building blocks appearing in its generated routes, and this fingerprint can be made both sticky under model extraction and formally testable. The defender's test is simple: under the null hypothesis that the suspect model was developed independently, a canary block should be used about as often as a matched set of ordinary control blocks from the same catalog, at rate $p_0$. The auditor samples $n$ routes from the suspect model, counts canary occurrences $k$, and computes
$$p\text{-value} = \Pr\big[\mathrm{Binomial}(n, p_0) \ge k\big],$$
rejecting independence when this tail probability falls below a preset significance level. Training biases the generator so that canary usage exceeds $p_0$ by a margin $\Delta$ large enough that a distilled surrogate — trained on the defender's API outputs — inherits the elevated rate, adapting the margin-stickiness insight of Kim et al., who showed that watermarks with large decision margins survive extraction, to the sequence-level setting of reaction-pathway generators. Two access regimes are covered: rich access, where the API returns full routes (audit the leaves directly), and product-only access, where the API returns just the molecule (audit the frequency of the canary's characteristic substructure in products, since forward reactions carry the block's atoms into the product).
The design must respect three constraints that have no analogue in text or image watermarking, and these constraints are where the synthetic accessibility literature does real work for us. First, validity: an arbitrary trigger token spliced into a SMILES string — the text format for molecules — breaks chemical parsers and drops validity to zero, which is why the invisible-Unicode canaries of Li et al. cannot be ported; our canaries are instead real, purchasable compounds that existing reaction templates accept, so validity is preserved by construction, in the spirit of the by-construction generators above. Second, inertness: canary blocks must not distort the generator's property optimization, which we verify empirically on quantitative estimate of drug-likeness (QED) and docking scores. Third, statistical identifiability: canaries are drawn from the long tail of catalog usage frequency and paired with matched control blocks of similar price, reactivity, and structural role, so that anomalous usage is attributable to the defender's choice, not to chemistry.
Research Program
Three components will be built. The first is a canary selection procedure: given a catalog (Enamine-derived subsets, or a synthetic in-house inventory modeled on the 5,955-block Led3 library of Hassen et al.), rank blocks by baseline usage frequency under an unmodified generator, select low-frequency blocks that are accepted by the catalog's reaction templates and carry a distinctive, easily matched substructure, and assemble matched control sets for calibrated nulls. The second is margin training: during reinforcement-learning fine-tuning for property optimization, following the reward-shaping pattern of SynGA's Bayesian optimization loop and Hassen et al.'s DrugEx v3 pipeline, add a bounded bonus for routes that use canary blocks, tuned so the usage margin $\Delta$ is large while property scores and validity are unchanged. The third is the audit itself: the binomial test above for leaf-level access, a substructure-frequency variant for product-only access, and permutation calibration over control blocks to certify false-positive rates, in the direction of the certified-rate ambitions articulated by Kallas et al.. We also model a realistic adversary who distills a surrogate from API outputs, re-plans routes to avoid suspicious blocks, or quantizes the generator.
Baselines
The naive reference point is an uncalibrated eyeball audit: look at a handful of outputs from the suspect model and judge whether the molecules "feel" like yours. This establishes the floor — it has no controlled false-positive rate, and under the null its error rate is unbounded, so any principled test beats it trivially; it is included only to make the value of calibration concrete.
The strongest existing competitor is the multi-signal dataset inference of Wahdany et al., which assembles a fingerprint vector from many input perturbations and output signals and classifies it with XGBoost (internal audit) or Isolation Forests (external audit). It is the strongest because it is the only prior method that proves whole-dataset provenance under black-box access with strong reported numbers: 0.997 ROCAUC on internal audits of a tabular foundation model and 0.851 ROCAUC on external audits, with 0.50 ROCAUC — chance — on purely synthetic pre-training. No reported result found on Lacuna for dataset inference applied to molecular generators of any kind, so we will port it faithfully: same perturbation battery adapted to molecules (stereochemistry flips, salt stripping, scaffold shuffling), same classifier stack, the same query budget and the same hyperparameter search budget as our method, run on identical training corpora. Beating it means higher AUROC on the same audit pairs at matched query cost, and — more decisively — detection power in the regime where it scores at chance, namely models trained on rule-generated data.
Two further operational baselines complete the picture. A retrosynthesis-planner audit runs each suspect output through AiZynthFinder, a tool that searches backward from a molecule to purchasable precursors, and inspects which building blocks appear in the recovered routes; it gets the planner's full commercial catalog, including the canaries, and unlimited compute, so the comparison is generous to it. A distributional baseline computes Maximum Mean Discrepancy on Morgan fingerprints of sampled molecules, following the substitution-detection use of MMD by Gao et al.; no reported result found on Lacuna for MMD-based auditing of molecular generators. Finally, the key ablation is our own passive test: the same leaf-distribution audit run on a generator with no canary margin, detecting only the natural fingerprint of catalog restriction. This isolates the proposal's active ingredient — the deliberately planted, margin-reinforced canaries — from whatever passive leakage exists anyway.
Why not just…?
Why not just run a membership inference attack on the generated molecules? Standard membership inference asks whether a specific record was in the training set, thresholding per-sample loss or confidence, as in the p-value formulation of Hu et al.. It is the wrong question here for a structural reason: forward-reaction generators produce molecules that never existed verbatim in the training data — Molecule Chef reports 87.1% novelty — so per-molecule membership is moot even when the derivation is real. Moreover, Wahdany et al. show that sample-level inference is unreliable in foundation models while dataset-level signals reach 0.997 ROCAUC; our canary test is dataset-level by construction. Their approach will nonetheless be run as the strongest baseline rather than argued away.
Why not just pass the suspect model's outputs through a retrosynthesis planner? This is the most serious alternative, and we will run it head-to-head rather than dismiss it. Two known shortcomings temper it: latency — retrosynthesis planners are slow enough that Hassen et al. had to train an XGBoost surrogate to stand in for AiZynthFinder inside their optimization loop, and SynGA was built expressly to avoid such calls — and insensitivity: shrinking a catalog by a factor of 3,000 changed the fraction of molecules solved by only about 12%, because the planner simply found longer routes (mean steps rose from 4.5 to 6.6). Aggregate solve rates are therefore too coarse for provenance. The sharper version of this alternative — dissecting recovered routes for canary blocks — is a genuine threat to the passive variant of our test, and two design choices answer it: canaries can be private compounds absent from public planner catalogs, so no planner will route through them, and the extraction-surrogate target (a stolen model emitting products, not a route database) offers the planner nothing to dissect. Experiment E5 settles this empirically.
Why not just port the existing text watermark to SMILES? The strongest text canary method, Li et al., hides non-rendering Unicode characters in fine-tuning text and achieves 96.7% true positive rate at 0% false positive rate with as few as 40 marked documents. It cannot transfer: invisible characters corrupt SMILES parsers and alter molecular identity, and the marks are stripped by trivial preprocessing — a sanitization vulnerability the authors themselves flag. Generic backdoor watermarks fail for the adjacent reason that arbitrary trigger substructures violate valence and are scrubbed by standard cheminformatics sanitizers; and on the utility side, CREDIT reports that generic watermarking baselines degraded protected-model accuracy to as low as 60% on CIFAR-10, versus 94.67% for its own protected model against a 95.70% vanilla baseline. Our canaries are ordinary chemistry, so sanitizers have nothing anomalous to remove.
Why not just compare output distributions with MMD on molecular fingerprints? Zhu et al. show that MMD and Kolmogorov–Smirnov tests have substantially lower detection power than rank-based tests when an adversary mixes genuine and substitute outputs, and ReaSyn shows molecular fingerprints are lossy representations that mangle synthetic-pathway information. MMD on fingerprints also cannot certify a false-positive rate. We include it as a baseline; if it matches our power at equal query budget in E5, the simple version of our claim — that leaf-level statistics are needed — would be wrong, and we would report that.
Proposed Experiments
All experiments use ReaSyn-style autoregressive route generators trained on USPTO-derived reaction data with building-block leaves drawn from Enamine subsets, evaluated alongside a SynGA-style tree search; reported reference points for generator quality are 99.0% validity and 87.1% novelty for the reaction-based generator of Molecule Chef, a perfect 1.000 validity rate for SynGA, and a 75.7% analog rate versus 31.4% for the next best model for ReaSyn — our generators must stay within these ranges or the watermark is buying detection at the price of the model's reason to exist.
Experiment E1 asks whether passive catalog fingerprints exist at all: train pairs of generators on disjoint 10,000-block sub-catalogs of a common reaction rulebook, then audit each model against each catalog definition. This is the direct test of whether rule-generated chemical data behaves like the synthetic regime of Wahdany et al. (chance detection, 0.50 ROCAUC) or like their real-data regime (0.997 ROCAUC). Either answer is a finding; the proposal's viability requires at least the active mechanism to work if the passive one does not. Experiment E2 is the core extraction test: an adversary distills a surrogate from 10k–1M API queries against a margin-trained defender; we audit the surrogate with the binomial test and report true positive rate at certified false-positive rates of 1% and 10⁻⁶, taking as the shape of success the 96.7%-TPR-at-0%-FPR level that Li et al. achieve in the text domain, while noting no reported result found on Lacuna for any ownership verification of generative molecular models. The margin ablation (margin $\Delta = 0$ versus tuned $\Delta$) isolates stickiness, testing whether the Kim et al. margin logic transfers from classification backdoors to route usage. Experiment E3 restricts the auditor to product-only access and tests substructure-frequency detection. Experiment E4 runs adaptive adversaries: route re-planning to avoid canaries, substitution of structurally similar non-canary blocks, 4-bit quantization, and output filtering; survival under re-planning is the hardest case, because multiple routes can reach the same product, and it determines the practical ceiling of the method. Experiment E5 is the baseline bake-off on identical query budgets, tuning budgets, and data.
The decisive falsifier is explicit. If E1 shows chance-level passive fingerprints — replicating the Wahdany synthetic-data result in chemistry — and E2 additionally shows that margin-trained canaries do not survive extraction or are erased by route re-planning at realistic query budgets, then the synthesizability constraint genuinely closes the provenance leak, and the field's reaction-driven generators are, from an IP standpoint, unownable by black-box statistical means. That negative result would itself be the first quantified answer to a question no paper in the file has touched.
References
- Generating Molecules via Chemical Reactions
- A Genetic Algorithm for Navigating Synthesizable Molecular Spaces
- Exploring Synthesizable Chemical Space with Iterative Pathway Refinements
- Dataset Inference for Data Provenance and Privacy Auditing in Tabular Foundation Models
- Margin-based Neural Network Watermarking
- Data Provenance Auditing of Fine-Tuned Large Language Models with a Text-Preserving Technique
- Generate what you can make: achieving in-house synthesizability with readily available resources in de novo drug design
- ROSE: A RObust and SEcure DNN Watermarking
- Model Equality Testing: Which Model is this API Serving?
- VeriDIP: Verifying Ownership of Deep Neural Networks Through Privacy Leakage Fingerprints
- CREDIT: Certified Ownership Verification of Deep Neural Networks Against Model Extraction Attacks
- Auditing Black-Box LLM APIs with a Rank-Based Uniformity Test