Generative molecular design asks a model to invent new molecules — as chemical graphs or 3D structures — that score well on some property we care about, such as binding to a disease protein. Because real evaluation (lab assays, physics simulations, even docking) is expensive, most practical pipelines score candidate molecules with a surrogate: a neural network trained on ordinary molecules to predict the property. The recurring failure is that the generator, by construction, searches for the highest predicted score it can find — and in doing so finds the surrogate's blind spots rather than real chemistry. The result is well documented: "greasy" oversized hydrophobic molecules that game docking scores DOCKSTRING: easy molecular docking yields better benchmarks for ligand design, and unstable ring systems that fail basic structural filters at rates as high as 88% Tartarus: A Benchmarking Platform for Realistic And Practical Inverse Molecular Design. This is precisely the adversarial-example phenomenon discovered for image classifiers in Intriguing properties of neural networks: an optimizer that queries a model will find inputs that break it.
The proposal: stop treating this as a bug to filter after the fact, and treat it as what it is — an attack. The generative optimizer is the adversary, so we should train the surrogate with the same minimax logic used in adversarial training for robustness, where the "perturbations" are the molecules the generator itself proposes, restricted to chemically valid edits. Concretely, the surrogate is periodically hardened against the worst of the generator's discovered over-scorings, in a closed loop where the generator attacks, the surrogate patches its worst blind spots first, and the next generation of molecules must earn its scores. The payoff we test for is sample efficiency: under the strict 10,000-oracle-call budget of the PMO benchmark, model-based optimization with a hardened surrogate should finally overtake the simple non-model baselines that currently dominate that benchmark.
Core Claim
The claim has two parts. First, mechanistic: surrogate exploitation in molecular optimization is an adversarial-robustness problem, and Madry-style worst-case training — where the attacker is the generative policy itself rather than an $L_p$-bounded pixel shift — measurably shrinks the gap between predicted and true property scores for top-ranked generated molecules, without collapsing the surrogate's clean accuracy. Second, practical: because that gap is what wastes oracle budget, closing it lets a surrogate-guided generator beat direct optimization under a fixed oracle budget.
The training objective makes the mechanism concrete. Let $f_\theta$ be the surrogate, $\mathcal{D}$ the natural-molecule training set, $\pi$ the current generator, and $\mathcal{B}_t$ a batch of generator-proposed molecules that have been re-scored by the true oracle (a deliberate spend of the overall oracle budget). We update the surrogate by
$$\min_\theta \;\; \mathbb{E}_{M \sim \mathcal{D}}\big[\ell(f_\theta(M), y_M)\big] \;+\; \lambda \sum_{M \in \mathcal{B}_t} w(M)\,\ell(f_\theta(M), y_M), \qquad w(M) \propto \exp\!\big(\alpha\,(f_{\theta_{\text{old}}}(M) - y_M)\big)$$
where $y_M$ is the oracle score and the weight $w(M)$ grows exponentially with the overprediction gap — how much the surrogate rewarded a molecule the oracle did not. The exponential worst-case weighting is the adversarial ingredient: the generator's most successful attacks against the old surrogate receive the largest gradient, exactly as the inner maximization in standard adversarial training concentrates effort on the worst perturbation found. This is the discrete, molecular analogue of GANTouch's insight that synthesizing hard near-boundary cases during training pushes the model's decision surface away from the region an attacker can reach GANTouch: An Attack-Resilient Framework for Touch-Based Continuous Authentication System.
The second ingredient defines the perturbation set. Standard adversarial training bounds $\delta$ by a norm; in chemistry the analogue must be chemically valid edits — changing an atom or bond changes valency and synthesizability, and the open problem of defining $\Delta$ in non-Euclidean chemical spaces is explicitly unsolved in the literature summarized by Molecule generation for drug design: A graph learning perspective. We define $\Delta$ through modular reaction edits, reusing the reaction grammar of ClickGen: Directed exploration of synthesizable chemical space via modular reactions and reinforcement learning as the method supplier: the surrogate is additionally smoothed so that its prediction on a generated molecule cannot spike relative to its predictions on the molecule's one-reaction-edit neighbors. This forces the generator out of narrow prediction spikes and onto plateaus of agreement, which are far more likely to be real. The scheduling matters too: following the finding in Robust and Accurate Object Detection via Adversarial Learning that hardening applied during fine-tuning survives downstream use while pre-hardening erodes, we harden the surrogate online inside the optimization loop rather than once up front.
What Is Genuinely New Here
The closest works establish the gap but do not close it. GPMO: Gradient Perturbation-Based Contrastive Learning for Molecule Optimization injects gradient perturbations into the generator's hidden states to fix exposure bias in sequence decoding — it improves generation fidelity and reports 96.72% success on the simultaneous multi-property "All" metric, but never touches the surrogate or defends a property predictor against exploitation. Towards Docking-oriented De Novo Ligand Design via Gradient Inversion (MagicDock) is, in a sense, the mirror image: it uses projected gradient descent as the generator to invert a pre-trained docking evaluator, reporting 7.4% average binding-affinity improvement in molecule design — but because the generator optimizes directly against the evaluator's gradients, the evaluator is exactly the component most exposed to gaming, and MagicDock leaves it unhardened and the receptor rigid. We harden the evaluator side that MagicDock, and every surrogate-based pipeline in the PMO benchmark, leaves static. And a large literature uses "adversarial" to mean the Goodfellow distribution-matching game between a generator and a discriminator MolGAN: An implicit generative model for small molecular graphs; Adversarial Learned Molecular Graph Inference and Generation; Cross-Adversarial Learning for Molecular Generation in Drug Design] — none of these implements minimax worst-case robustness, and all remain exposed to the failure mode we address. Finally, Sample Efficiency Matters: A Benchmark for Practical Molecular Optimization documents that surrogate-based methods do not consistently beat model-free baselines under realistic budgets — this proposal's diagnosis is why (the generator exploits the surrogate, so its scores stop meaning anything) and its fix is targeted at exactly that diagnosis.
Research Program
We build three components. First, the surrogate family: graph- and string-based property predictors trained on the DOCKSTRING matrix (260,000 molecules scored across 58 human targets), where prior work reports Attentive FP regression accuracies of $R^2 = 0.91$ on PARP1 and $R^2 = 0.68$ on PGR DOCKSTRING — these give the reference scale for clean accuracy that hardening must not degrade by more than a stated margin. Second, the generator: we deliberately use the strongest simple generators, SMILES-based policy-gradient REINVENT and Graph GA, because the PMO benchmark showed these dominate under budget constraints; the proposal's value should come from the surrogate, not from architectural novelty in the generator. Third, the closed loop: alternate generation phases (unlimited cheap surrogate queries) with hardening phases (a slice of the oracle budget re-scores the adversarially-most-suspicious generated molecules — those with the largest surrogate-versus-neighborhood disagreement — and the update above reweights them by realized overprediction). We will characterize the loop's dynamics explicitly: how the overprediction gap and the oracle-spend fraction evolve, whether hardening exhibits the robustness–accuracy trade-off known from image classifiers, and whether the generator's successive "best" molecules migrate toward regions where surrogate and oracle agree.
A secondary application, kept deliberately small, extends the same minimax machinery to the target side: replace the molecule-adversary with a pocket-mutation adversary, so the outer generator seeks ligands whose binding survives worst-case amino-acid substitutions, $\min_M \max_{P' \in \mathcal{B}_\epsilon(P)} \Delta G(M, P')$, using the mutation-induced affinity data of A mutation-induced drug resistance database (MdrDB) (100,537 samples across 2,503 mutations; a reference ExtraTrees model there achieves Pearson correlation 0.607 and RMSE 0.656 kcal/mol on a TKI test split). This extension is scoped as a feasibility study because the structural oracle is treacherous: Structural Memorization in AlphaFold: Adversarial Mutations Reveal Template Reliance, Confidence Failures, and Implications for Protein Design shows AlphaFold 3 preserves the wild-type fold (TM-score $\ge 0.5$) even under 40% sequence mutation and its confidence metric picks the correct structure only ~25% of the time, so mutated-pocket structures cannot be trusted to AF3; the MdrDB regression setting avoids that trap. Existing pocket-conditioned generators — DiffBP, which reports 41.07% overall ligand efficiency on CrossDocked2020 DiffBP: generative diffusion of 3D molecules for target protein binding, and DeepICL, which designed EGFR L858R/T790M ligands at −7.67 kcal/mol versus a −6.52 baseline 3D molecular generative framework for interaction-guided drug design — are both conditioned on a single static structure and do mutation handling only post hoc, so the minimax formulation would be new there too.
Baselines
The naive reference point is doing without a surrogate entirely: REINVENT run directly on the true oracle under the 10,000-call budget. This is the "do nothing" baseline and, alarmingly, also the strongest existing method: in the PMO benchmark, REINVENT ranked first overall across the 23 tasks and Graph GA ranked second, beating all modern deep generative models Sample Efficiency Matters. No reported result found on Lacuna for per-task AUC values in PMO, so our primary comparison reports our own reproduction of REINVENT and Graph GA on identical hardware, and the proposal succeeds only if surrogate-hardened generation beats our reproduced REINVENT-on-oracle on mean Top-10 AUC across all 23 tasks in a paired test. Fairness rules: every method gets exactly the same 10,000 oracle calls (re-scoring for hardening counts against the budget), the same generator, the same surrogate architecture and pretraining data, and the same hyperparameter tuning budget via a fixed grid applied uniformly. The strongest competitor among existing surrogate-based methods is the ensemble-plus-uncertainty-penalty setup, because it is the field's default answer to overoptimization; we run it with the same surrogate architecture and five members, spending identical compute-adjusted oracle budget. The rule-penalty baseline — REINVENT with a QED-based filter penalty, the mitigation DOCKSTRING itself demonstrated for the greasy-molecule trap DOCKSTRING — covers the cheap engineering fix. Finally, the ablation that isolates the proposal's key ingredient: identical generator, surrogate, and oracle spend, but with uniform instead of worst-case-exponential reweighting and with random instead of reaction-grammar neighborhoods — i.e., plain self-training on generated data. If uniform self-training matches exponential worst-case weighting, the adversarial structure is doing nothing and the claim collapses to "retraining helps," which is already known.
Why not just…?
Why not just run REINVENT or a genetic algorithm on the true oracle and skip surrogates? This is the most serious objection, and the honest answer is that the proposal's headline experiment is designed to settle it rather than argue around it: if hardened-surrogate REINVENT does not beat direct REINVENT at 10,000 calls on PMO, the proposal fails on its own terms. Two things suggest it can. PMO itself found that model-based methods underperform partly because surrogates get exploited — the failure this proposal targets directly. And the 10,000-call regime understates where surrogates matter: in Tartarus, oracles are physics simulations and TD-DFT calculations Tartarus, and in real programs the oracle is a wet lab, where the cheap-query regime is the only scalable one. A hardened surrogate that stays honest under billions of cheap queries is the principled way to spend a scarce true oracle.
Why not just add rule-based filters (Lipinski, QED, SAscore, PAINS)? DOCKSTRING shows a QED penalty does mitigate the greasy-molecule symptom DOCKSTRING, so this baseline is included head-to-head. But filters mask one symptom rather than fix the score–reality gap: a generator can still park inside the rule boundaries and exploit the surrogate there, producing molecules that pass every filter yet do not bind. DeepICL's experience is a cautionary precedent for rule-based guidance generally — its hand-coded SMARTS interaction patterns limit it when novel binding motifs appear DeepICL. Our decisive metric is the overprediction gap on held-out oracle scores for top-ranked molecules; filters do not act on that quantity at all, and if the filter baseline matches the hardened surrogate on it, we will report that plainly.
Why not just use an ensemble or an uncertainty penalty? This is a genuine threat and gets the most careful comparison. PMO reports that adding surrogates or ensembles did not consistently improve sample efficiency without exceptional tuning Sample Efficiency Matters, and ensembles cost compute linearly while sharing inductive biases — a generator that shifts its policy can fool members jointly, since the ensemble average smooths but does not concentrate training on worst-case failures. The exponential worst-case weighting is precisely the mechanism ensembles lack: it allocates gradient mass to the generator's most damaging discoveries rather than averaging over all queries. We run the ensemble baseline at matched compute and identical oracle budget; if it matches the hardened surrogate, the simpler method wins and we will say so.
Why not adversarially train the generator instead, as GPMO does? GPMO's perturbations live in the generator's own hidden space and serve a contrastive objective that fixes exposure bias in sequence decoding GPMO; it never evaluates, let alone hardens, a property surrogate. The two mechanisms are complementary rather than competing, and our design keeps the generator deliberately simple so any gain is attributable to the surrogate — but nothing in the loop prevents combining them later.
Proposed Experiments
The primary experiment is the full PMO protocol: 23 tasks, Top-10 AUC, hard 10,000-oracle-call cap, with all baselines above run under identical conditions; success requires beating reproduced REINVENT-on-oracle on mean AUC and beating the ensemble baseline by a margin that survives a paired test across tasks. The second experiment measures the mechanism directly on DOCKSTRING: run generation against per-target surrogates across a target subset, and track (a) the overprediction gap between surrogate score and Vina re-score for the top-100 molecules per run, (b) the molecular-weight and logP inflation that defines the greasy pathology, and (c) surrogate clean accuracy per target, with the $R^2 = 0.91$ (PARP1) and $R^2 = 0.68$ (PGR) figures as the pre-hardening reference points, accepting at most a pre-registered drop. The third experiment applies Tartarus's structural-filter pass rates as the synthesizability check, against the documented failure floor of deep models passing as few as 12% of structures on the protein-ligand task Tartarus; reaction-grammar neighborhoods should push pass rates toward the genetic-algorithm end of that spectrum, and if they do not, the perturbation-set design is wrong. Distributional sanity uses the MOSES suite molecularsets/moses to confirm the hardened surrogate has not simply collapsed to conservative predictions. The mutation-resilience extension trains the MdrDB minimax variant and reports worst-case $\Delta\Delta G$ under held-out mutations, comparing against the post hoc alternative of conditioning on known mutant structures à la DeepICL.
The decisive checks, in order: the overprediction gap for top-ranked molecules must shrink monotonically with hardening rounds (mechanism), the PMO headline must beat direct REINVENT (practical value), the uniform-reweighting ablation must underperform exponential worst-case weighting (ingredient necessity), and clean surrogate accuracy must stay within the pre-registered margin (no robustness–accuracy collapse). Any one of these failing names the specific assumption that was wrong, which is what a proposal this size owes its skeptics.
References
- DOCKSTRING: Easy Molecular Docking Yields Better Benchmarks for Ligand Design
- Tartarus: A Benchmarking Platform for Realistic And Practical Inverse Molecular Design
- Intriguing properties of neural networks
- GANTouch: An Attack-Resilient Framework for Touch-Based Continuous Authentication System
- Molecule generation for drug design: A graph learning perspective
- ClickGen: Directed exploration of synthesizable chemical space via modular reactions and reinforcement learning
- Robust and Accurate Object Detection via Adversarial Learning
- GPMO: Gradient Perturbation-Based Contrastive Learning for Molecule Optimization
- Towards Docking-oriented De Novo Ligand Design via Gradient Inversion
- MolGAN: An implicit generative model for small molecular graphs
- Adversarial Learned Molecular Graph Inference and Generation
- Cross-Adversarial Learning for Molecular Generation in Drug Design
- Sample Efficiency Matters: A Benchmark for Practical Molecular Optimization
- A mutation-induced drug resistance database (MdrDB)
- Structural Memorization in AlphaFold: Adversarial Mutations Reveal Template Reliance, Confidence Failures, and Implications for Protein Design
- DiffBP: generative diffusion of 3D molecules for target protein binding
- 3D molecular generative framework for interaction-guided drug design