A molecule's 3D shape — its conformation — determines whether it binds a protein or absorbs light, so generative models that produce realistic conformer ensembles are workhorses in drug discovery. The best training data for these models comes from expensive quantum calculations: GEOM compiled 37 million conformations for over 450,000 molecules, each annotated with semi-empirical density functional theory (DFT) energies. Such datasets are slow to build and, when they come from a pharmaceutical company's proprietary library, deeply sensitive: an attacker who can tell whether a specific molecule's geometry was in the training set has learned something the company intended to keep secret. Meanwhile, the dominant conformer generators, such as GeoDiff — a diffusion model, i.e. a generator that learns by gradually denoising random noise into samples — pay a heavy latency price, integrating a trajectory step by step for every sample.
This proposal starts from a simple observation nobody has tested: in a one-step conformation generator, the scale of the random noise injected at generation time acts like a temperature knob, and that knob should control both sample diversity and privacy leakage at once. At low temperature the model regurgitates sharp, memorized geometries; at higher temperature it spreads probability mass across plausible alternatives. If the perturbation is applied only to the soft degrees of freedom (rotatable bonds) rather than to raw coordinates, chemical validity survives the flattening. The result would be a single dial that lets a data owner publish a fast, useful, synthetic conformer generator while measurably suppressing training-set re-identification — a molecular analogue of what DP-Prompt achieved for text with temperature-scaled paraphrasing, but in continuous 3D geometry, with quantum-accurate evaluation, and at one function evaluation per sample.
Core Claim
We claim three things, none of which has been measured. First, fast conformer generators — both one-step flow models in the spirit of Mean Flows for One-step Generative Modeling and direct regressors like Direct Molecular Conformation Generation (DMCG) — memorize enough of their training geometries that a membership inference attack can separate members from non-members. Second, this leakage is monotone in the temperature of the injected randomness: scaling the latent noise upward drives attack success toward chance while costing coverage only gradually. Third, because the perturbation that provides privacy can be routed through torsional rather than Cartesian degrees of freedom, the privacy gain does not have to destroy chemical validity. The dial is:
$$c = G_\theta\!\left(T_{\mathrm{ens}}\,\varepsilon\right), \qquad \tau_j \leftarrow \tau_j + T_{\mathrm{tor}}\,\eta_j,\ \ \eta_j \sim \mathcal{N}(0,1),$$
where $G_\theta$ is a one-step generator mapping latent noise $\varepsilon$ to atomic coordinates $c$, $T_{\mathrm{ens}}$ scales the latent (selecting which conformer basin is reached), and $T_{\mathrm{tor}}$ perturbs only the torsion angles $\tau_j$ of rotatable bonds after generation — never bond lengths or angles, which is what keeps the molecule chemically intact.
The nearest existing work stops short of this in different ways. DP-Prompt proved that logit clipping plus temperature scaling can certify local differential privacy, but only for autoregressive text on IMDB and Yelp; it never touches continuous coordinates, quantum benchmarks, or fast samplers. SecMI showed that diffusion models leak through step-wise reconstruction error concentrated at small timesteps, but only for image diffusion, and its attack requires a multi-step trajectory — which one-step models deliberately eliminate. A Tale of Two Temperatures showed that a second temperature recovers diversity in low-latency sampling, but for discrete tokens on math and code, with no privacy analysis. 3D MeanFlow gives the one-step 3D machinery (about 5 ms per shape, over 8× faster than prior one-step baselines) but works on macroscopic point clouds, admits its one-step mode collapses diversity, and says nothing about atoms, energy surfaces, or memorization. This proposal sits exactly in that empty intersection, and one result in the file makes the setting genuinely tense: SecMI found that Weight Averaging, a standard fidelity trick, amplifies leakage — so we must also ask whether fine-tuning a generator to match DFT energy minima, as the evaluation protocol of GEOM-Drugs Revisited rewards, quietly does the same thing.
Research Program
We will build one system and three analyses around it. The system is a permutation- and rotation-equivariant one-step conformer generator, trained with a DMCG-style roto-translationally invariant loss on GEOM-QM9 and on the kekulized GEOM-Drugs split, using the training and evaluation scripts released as the DirectMolecularConfGen/DMCG codebase as the scaffolding. The generator will support the two-temperature sampling above as an inference-time option, requiring no retraining. Crucially — and we want to be honest about this — the dial does not depend on our particular backbone. The latent of any conditional generator, including DMCG's own VAE prior, can be tempered. The backbone is an instantiation; the durable contributions are the attack construction, the temperature–leakage curves, and the quantum-alignment study.
The first analysis is the attack. SecMI's step-wise error cannot be computed for a one-step model, so we construct its natural analogue: given a candidate conformer, approximately invert the generator by optimizing the latent $\varepsilon$ to minimize reconstruction error, and use the residual — a "latent error" playing the role of SecMI's small-$t$ error — as the membership signal, validated alongside the Likelihood Ratio Attack used in Privacy for Free. No reported result found on Lacuna for membership inference on any molecular conformation generator, so these are first measurements, and the multi-step GeoDiff gives the natural comparison object: does collapsing the trajectory concentrate or dissipate leakage?
The second analysis is the temperature sweep, tracing utility (coverage, matching, quantum energy fidelity) and leakage (attack AUC, true positive rate at 1% false positive rate) jointly along $T_{\mathrm{ens}}$ and $T_{\mathrm{tor}}$.
The third analysis tests the Weight-Averaging analogy: after fine-tuning the generator's outputs against GFN2-xTB relaxed geometries — the quantum-accurate protocol that GEOM-Drugs Revisited shows is mandatory for honest evaluation (classical MMFF94 force fields misrepresent structural precision) — does the attack AUC rise? If quantum alignment behaves like Weight Averaging, the field's preferred accuracy fix is a privacy hazard, which would be a finding worth knowing regardless of the dial.
Baselines
The "do nothing" reference is the untempered one-step generator: the same model at $T_{\mathrm{ens}} = 1$, $T_{\mathrm{tor}} = 0$, evaluated for both utility and leakage. It anchors both ends of the trade-off: it is the fastest, sharpest setting, and we expect it to be the leakiest. Beating it means finding an operating point with materially lower attack AUC at a small, quantified coverage cost.
The strongest utility competitor is GeoDiff, the multi-step diffusion generator that set the standard on these benchmarks; it reported recall coverage of roughly 89% on GEOM-Drugs against about 67% for GeoMol. We will retrain it on our splits under the same tuning budget and compare coverage, matching score, HOMO-LUMO gap error, and per-conformer wall-clock time. Beating it means matching its coverage and matching score within a few points while generating at one function evaluation instead of a full diffusion trajectory.
The strongest latency competitor is DMCG itself, which reported a 200× speedup over ConfGF and a 36% matching-score improvement on large-scale GEOM-Drugs. Here the comparison is deliberately about what DMCG cannot do: its randomness is a fixed standard-normal VAE prior with no entropy control and no privacy evaluation anywhere in the paper. Beating it means matching its speed and accuracy while demonstrating an operating point it cannot reach — high ensemble diversity with attack AUC near chance. If tempering DMCG's own latent achieves everything our backbone does, the backbone was never the contribution, and we will say so in the paper; that experiment is included precisely to find out.
The privacy baseline is the attack suite: our latent-error attack (calibrated against SecMI's reported severity on images — attack success rate of 0.889 on a DDPM trained on CIFAR-10, versus 0.560 for GAN-Leaks, and a Stable Diffusion v1-5 training-set detection AUC of 0.701) plus LiRA. Success is an AUC at or below roughly 0.55 at matched coverage. Two ablations isolate the key ingredient: a single uniform temperature scaling both latent and coordinates with one knob, versus our split dial; and plain Cartesian Gaussian jitter at matched noise magnitude, versus torsion-only perturbation, scored on chemical validity and energy after GFN2-xTB relaxation.
Why not just…?
Why not just train with DP-SGD, or add noise to the outputs? Formal DP is the stronger guarantee, and we do not provide one — that is a real limitation of this proposal. But the evidence says naive DP pays too much: Privacy for Free found condensed data beat the best DP-trained generators by 33.4% test accuracy at matched privacy on FashionMNIST, and SecMI notes DP degrades diffusion quality. As for output noise, unstructured Cartesian jitter at privacy-meaningful magnitudes will break bond lengths and aromatic valence — the same failure mode GEOM-Drugs Revisited documents when geometries are evaluated against corrected quantum valence tables. We run the Cartesian-jitter arm anyway, and its expected collapse into chemical invalidity is itself one of the results.
Why not just use RDKit/MMFF94 conformers, which need no training data and therefore cannot leak? This is the most serious alternative, and for any application that only needs generic chemistry on public molecules, it wins outright — we concede that setting. Our setting is different: an organization holding a private library of novel molecules that wants to share a useful generator or synthetic conformer pool without exposing the library. There, heuristic force fields cannot help by construction, and their quality is anyway superseded: DMCG shows learned conformers beat RDKit ensembles in downstream docking and HOMO-LUMO fidelity, and GEOM-Drugs Revisited shows modern generators surpass MMFF94 under GFN2-xTB evaluation. We will quantify this gap directly on relaxed coverage and matching.
Why not just run a multi-step diffusion model with an off-the-shelf fast solver and one global temperature? Two failures, both documented. Truncated solvers degrade sharply: AC-Sampler reports a DPM-v3 model at 5 function evaluations at FID 12.41, recovering to 9.88 only with added MCMC correction steps; on molecules, truncation error surfaces as steric clashes, the exact pathology DMCG and FABFlex were built to avoid. And a single uniform temperature is the wrong control in accelerated regimes: A Tale of Two Temperatures found uniform scaling forces roughly 8,000 function evaluations where a decoupled pair needs about 4,000 at a 70% pass rate. If the uniform-knob ablation nonetheless suffices on molecules — possible, since molecular mode structure differs from text — we will report that and drop the split dial; the privacy-latency audit stands on its own.
Proposed Experiments and Decisive Checks
All training uses GEOM-QM9 and the kekulized GEOM-Drugs split of GEOM-Drugs Revisited (which reports about a 5% validity improvement for models retrained on kekulized data), with utility scored under the corrected GFN2-xTB relaxation protocol — against which the file reports GFN2-xTB itself matches full DFT at a mean RMSD of 0.36 Å on GEOM, so relaxation is a faithful quantum referee. Utility metrics are recall coverage and matching score, HOMO-LUMO gap error, chemical validity, ensemble diversity, and milliseconds per conformer. Privacy metrics are attack AUC and true positive rate at 1% false positive rate (the regime where SecMI reports nearly 38% TPR at 1% FPR on CIFAR-10 DDPMs) on balanced member/non-member splits.
Three checks decide the proposal. First, the temperature–leakage curve: does attack AUC fall monotonically toward 0.5 as $T_{\mathrm{ens}}$ rises, and does the knee sit at a coverage level still competitive with GeoDiff's ~89%? If AUC barely moves, the core claim is dead and the paper becomes a negative result about latent tempering. Second, the trajectory comparison: at matched coverage, does the multi-step GeoDiff leak more than its one-step counterpart under our attack, or does one-step generation concentrate leakage in the single latent? Either outcome is publishable and settles an open question SecMI left behind. Third, the quantum-alignment check: after GFN2-xTB fine-tuning, does AUC rise the way Weight Averaging amplified SecMI? If yes, every group optimizing for quantum fidelity is currently trading privacy invisibly, and this proposal will have measured the exchange rate.
References
- GEOM, energy-annotated molecular conformations for property prediction and molecular generation
- GeoDiff: A Geometric Diffusion Model for Molecular Conformation Generation
- Locally Differentially Private Document Generation Using Zero Shot Prompting
- Mean Flows for One-step Generative Modeling
- Direct Molecular Conformation Generation
- Are Diffusion Models Vulnerable to Membership Inference Attacks?
- A Tale of Two Temperatures: Simple, Efficient, and Diverse Sampling from Diffusion Language Models
- 3D MeanFlow: One-Step Point Cloud Completion and Generation via Average-Velocity Transport
- GEOM-Drugs Revisited: Toward More Chemically Accurate Benchmarks for 3D Molecule Generation
- Privacy for Free: How does Dataset Condensation Help Privacy?
- AC-Sampler: Accelerate and Correct Diffusion Sampling with Metropolis-Hastings Algorithm
- Fast and Accurate Blind Flexible Docking