When a machine-learning model proposes a new synthetic route for a molecule, we now know that grading it against the single route recorded in a patent database is a broken way to evaluate: it fails valid alternatives and rewards memorization. The same failure shape exists, almost unnoticed, one step downstream in drug delivery. A pH-responsive nanocarrier — a particle that stays intact in blood at pH 7.4 and dumps its drug when it hits acidified tumor tissue around pH 6.5 or endosomes near pH 5.0 — is judged by comparing its dissolution curve against a single reference curve measured in a static beaker assay at two fixed pH values. That reference is one arbitrary realization among many chemically valid ones, measured in conditions that don't resemble a tumor. Meanwhile, generative modeling in drug discovery is stuck at static scalars: the best existing diffusion model for pharmacokinetics outputs single numbers like half-life, not the time series of drug concentration that actually determines whether a carrier leaks, switches, or kills the tumor. And when such carriers finally reach human trials, they bring a clinical-trial problem nobody has modeled for nanomedicine: their effects are often perceptible (a warm patch of skin under near-infrared illumination, sudden absence of chemo side effects), so participants and clinicians can guess who got the active carrier, contaminating every subjective endpoint.
This proposal builds one object that serves all three problems: a generative simulator of pH-dependent drug-release trajectories, evaluated by chemistry-consistency checks rather than curve matching, and used as a trial engine in which causal methods for unblinded studies can be validated. The core mechanism comes from trajectory generation; pH-responsive release is the application; exact-match evaluation bias supplies the evaluation lens; unblinded intervention bias supplies the downstream evaluation regime the simulator must support.
Core Claim
The claim has three parts. First, exact-match evaluation bias in computational chemistry has a direct, unexplored analog in drug-release evaluation: scoring a predicted release curve by its distance to one archived reference curve — or by release fractions at two fixed assay timepoints — penalizes chemically valid alternative formulations and cannot distinguish a benign timing shift from a mechanistically impossible curve, exactly as Top-k exact match cannot distinguish a stereochemical variant from a forbidden transformation. Second, whole-trajectory diffusion — a generator that denoises random noise into an entire state-action sequence at once, rather than predicting one step at a time — can be adapted to generate full release time series conditioned on formulation descriptors and non-stationary pH profiles, where step-by-step predictors accumulate compounding error. Third, this simulator closes a stated gap in causal inference for unblinded trials: estimating the Controlled Direct Effect — the effect of treatment if participant expectancy were held fixed at a common reference level — requires dense longitudinal measurement of expectancy and side-effect indicators that real trials rarely collect, and a validated simulator can supply them in unlimited quantity.
The closest existing work is distinct in kind. Diverse and feasible retrosynthesis using GFlowNets (RetroGFN) generates discrete three-step retrosynthesis decisions and scores them with a forward reaction model's round-trip check; it never touches continuous physical dynamics like disassembly kinetics. When Single Answer Is Not Enough: Rethinking Single-Step Retrosynthesis Benchmarks for LLMs replaces single-answer patent supervision with an exhaustively enumerated dataset (CREED, 22.7 million reactions for 1.5 million products) and a constraint checker (ChemCensor); we borrow both ideas — enumerated supervision and constraint-based scoring — but apply them to continuous release curves, not discrete reaction templates. Drug Discovery SMILES-to-Pharmacokinetics Diffusion Models with Deep Molecular Understanding (Imagand) is the strongest generative model bridging molecular structure to pharmacokinetics, but it generates static scalar properties from small-molecule SMILES and its authors state it cannot simulate temporal processes or address the in vitro-to-in vivo gap; we generate temporal release trajectories conditioned on formulation physics it does not represent. Finally, Causal Inference in Studies with Functional Unmasking: Psychedelics and Beyond provides the causal machinery (the Controlled Direct Effect, estimated with the lmtp package) for trials where the active drug unmasks itself — but only for psychiatric interventions, with no carrier, no release dynamics, and no longitudinal simulator. No paper in the corpus connects any of these pieces.
Research Program
The program has four components.
First, a release-trajectory corpus, built the way CREED replaced single-answer supervision: rather than digitizing a handful of published curves as unique ground truths, we exhaustively enumerate formulation-by-environment combinations through a mechanistic simulator of acid-triggered disassembly — protonation cleaving metal–nitrogen coordination bonds in zeolitic imidazolate frameworks, micellar collapse, and imine hydrolysis — anchored to reported systems in the corpus: the Bi@ZIF-8 photothermal composite, whose framework dissolves in the pH 5.0–6.5 window and whose in vivo behavior is quantified in Rapid synthesis of a Bi@ZIF-8 composite nanomaterial (photothermal conversion efficiency 24.4%, mild hyperthermia at about 43°C, complete tumor disappearance over 14 days), and the pH-sensitive mixed micelles of pH-sensitive micelles for the intracellular co-delivery of curcumin and Pluronic L61 unimers (hydrodynamic diameter 221.3 ± 5.3 nm, doxorubicin encapsulation efficiency about 80.6%, 87.89% late apoptosis in resistant MCF-7/ADR cells). Each entry is a trajectory $\tau = (c_0, c_1, \dots, c_T)$ of released-drug concentration over time under a pH profile that may be non-stationary (7.4 → 6.5 → 5.0), plus a few thousand digitized published curves for external validation. The design directly attacks the in vitro-to-in vivo extrapolation gap that Therapeutics Data Commons flags as a reason preclinical benchmarks fail to predict clinical outcomes.
Second, the generator: a diffusion model that denoises an entire release trajectory in one pass, conditioned on formulation descriptors and a pH profile, with the initial state fixed by inpainting in the style of World Models via Policy-Guided Trajectory Diffusion, which showed that whole-trajectory diffusion avoids the compounding drift of step-by-step prediction and gave a 10× to 50× speedup over autoregressive diffusion. Autoregressive rollouts are a known liability here: release curves span long horizons, and Bootstrapped Transformer for Offline Reinforcement Learning documents how autoregressive self-generation during training is both slow and prone to physically impossible outputs.
Third, the evaluation metric, which is the conceptual heart of the proposal. Instead of distance to a reference curve, we score a generated trajectory by chemical and pharmacokinetic consistency: mass balance (cumulative release never exceeds payload and never decreases), pH-switch consistency (negligible release above the carrier's threshold pH, rapid release below it — the property the whole field is engineered around), and a round-trip check in the spirit of RetroGFN, where a forward pharmacokinetic compartment model must map the release curve to a plausible plasma-and-tumor concentration profile. This is ChemCensor's reaction-center logic and UniCon's learned preference scoring from Learning Reaction-Condition Plausibility for Evaluation and Self-Curation Under Noisy and Non-Unique Supervision, translated from bond transformations to disassembly kinetics. The translation matters because UniCon's own limitations — it cannot account for concentration, stoichiometry, or assembly conditions, and inherits popularity bias from the literature it trains on — are precisely the parameters that govern nanoparticle self-assembly, so the release-domain scorer must be built on enumerated mechanistic ground truth rather than historical literature.
Fourth, the trial engine and causal analysis. We build a simulated unblinded nanomedicine trial in which functional unmasking arises from the release dynamics themselves: perceptible local warmth at the tumor site, or visible absence of chemotherapy side effects when pH shielding works. Expectancy responds to these cues and feeds back into subjective endpoints — the snowball mechanism Loewinger et al. formalize. The simulator injects a known true pharmacological effect, and we test which analysis recovers it: naive comparison, post-hoc stratification on participants' treatment guesses, dose-response comparison, and estimation of the Controlled Direct Effect,
$$\mathrm{CDE} = \mathbb{E}[Y(a, e^\ast)] - \mathbb{E}[Y(a', e^\ast)],$$
where $Y$ is the clinical endpoint, $a$ versus $a'$ denotes assignment to the pH-responsive carrier versus control, and $e^\ast$ is expectancy fixed at one common reference level for every participant. The release-trajectory model supplies the dense longitudinal measurements of unmasking indicators that lmtp-style estimation needs and real trials lack — Loewinger et al. identify this as the main practical obstacle to their method, and no reported result found on Lacuna for the numerical recovery error of their estimator, so quantifying it in simulation is itself a contribution.
One study cuts across all components and addresses the deepest tension in the corpus, flagged as an open gap: model-based filtering of generated data is itself conditioning on a post-treatment variable. When we bootstrap the generator on its own outputs and keep only trajectories passing the plausibility filter — exactly what BooT does with log-probability confidence filtering — the filter's verdict is influenced both by the trajectory's true validity and by the filter's training-distribution bias, making it a collider in exactly the sense Loewinger et al. prove dangerous for participant belief. We will measure whether plausibility-filtered bootstrapping silently truncates valid rare release mechanisms (an analog of UniCon's documented popularity bias toward common catalysts), using held-out mechanistic regimes as ground truth.
Baselines
The naive reference point is the field's standing practice: the static two-point dissolution assay — fraction of drug released at pH 7.4 and pH 5.5 at fixed times in a dialysis bag — optionally upgraded to root-mean-squared error against a single archived reference curve. This is how the carriers in the pH-sensitive micelles and Bi@ZIF-8 studies were characterized, and no reported result found on Lacuna for any machine-learning baseline on this assay. It is the strongest competitor on the question "is anything beyond current practice needed," because it is cheap and regulatory-entrenched. We will show it fails in both directions: it gives full credit to a curve that matches the reference at the two assay points while violating mass balance or switch consistency between them, and zero credit to a chemically valid formulation whose curve differs from the reference everywhere except those points. Beating it means a metric that separates mechanistically consistent curves from inconsistent ones at rates well above chance, verified against enumerated ground truth and, where possible, expert chemist judgment in the style of UniCon's blind trials.
The strongest existing generative method is Imagand, the SMILES-to-pharmacokinetics diffusion model of Drug Discovery SMILES-to-Pharmacokinetics Diffusion Models with Deep Molecular Understanding. It is the closest published system: a diffusion model conditioned on molecular structure that generates pharmacokinetic quantities, with reported distributional fidelity of an average Hellinger distance of 0.15 between its synthetic pharmacokinetic data and real distributions from eleven Therapeutics Data Commons datasets, and reported downstream gains when its synthetic data augments real data for regression models. We will run its discrete-noise transformer architecture on our corpus, flattened to predict whole release curves as long property vectors, with the same training data, compute budget, and tuning effort as our trajectory diffusion model, so the comparison isolates architecture and conditioning design rather than resources. Beating it means lower trajectory-level error on held-out mechanistic ground truth — especially under out-of-distribution, non-stationary pH profiles, where its static-property design has no mechanism for temporal coherence — and better downstream utility, measured as calibration of predicted tumor exposure against the forward pharmacokinetic model.
The strongest generation-side competitor at matched compute is an autoregressive sequence transformer with self-generation bootstrapping, the design of Bootstrapped Transformer for Offline Reinforcement Learning, which improved average normalized return by 16.8% over a Trajectory Transformer baseline on D4RL locomotion and reached 23.7 versus 20.1 on Adroit manipulation — evidence that bootstrapped generation works when trajectories are dynamically coherent. It will receive the same corpus, the same filtering budget, and the same hyperparameter search. Beating it means lower compounding error as a function of horizon and lower wall-clock cost per generated trajectory, with PolyGRAD's reported 10× to 50× speedup over autoregressive diffusion as the reference scale for what whole-trajectory diffusion should deliver.
Two ablations isolate the proposal's key ingredients. Removing the plausibility filter from the bootstrapping loop isolates the contribution of chemistry-consistency scoring to output validity, and directly measures the collider-bias cost of filtering on rare mechanisms. Swapping the plausibility metric for reference-curve root-mean-squared error inside the same generator isolates the evaluation-metric claim from the generation claim.
On the causal side, the competitors are naive stratification on participants' treatment guesses and dose-response comparison, both of which Loewinger et al. formally show are biased (collider bias and a placebo dose-response, respectively), and the lmtp Controlled Direct Effect estimator run on sparse measurement schedules resembling real trials. Beating them means the Controlled Direct Effect estimated on the simulator's dense measurement stream recovers the injected true pharmacological effect within a stated tolerance, while the alternatives show quantifiable bias.
Why not just…?
Why not just write down the mechanistic differential equations and skip the learned generator? This is the most serious alternative, and we embrace half of it: the mechanistic simulator is how we build ground truth, exactly as CREED used expert-coded templates to build enumerated supervision. But a mechanistic model needs known parameters for every carrier and environment, and the environments that matter — non-uniform tumor acidification microgradients, protein corona formation, shear stress, the very factors the corpus identifies as why static assays fail — are precisely those without clean mechanistic descriptions. The learned model's job is to interpolate and quantify uncertainty where mechanism is unknown. The decisive experiment is a held-out regime where mechanistic parameters must be extrapolated, not interpolated: if the mechanistic simulator matches the learned model everywhere within measurement noise, the generative component collapses to a simulation study and the proposal should shrink accordingly. We will run that check first.
Why not just extend Imagand to output more static scalars? Imagand's authors state their model is limited to static continuous properties and cannot simulate temporal in vivo processes or close the in vitro-to-in vivo gap, and its Hellinger distance of 0.15 is measured against static distributions, not trajectories. A release curve cannot be summarized into scalars without discarding exactly the information this proposal depends on: the timing of the pH switch determines whether a carrier leaks during circulation, and the slope of release determines the intensity and timing of the somatic cues that drive functional unmasking in trials. Two formulations with identical aggregate half-life can have opposite leakage and unmasking profiles. The baseline run of Imagand-on-curves will test this directly; if flattened-curve prediction matches trajectory diffusion, the temporal machinery is not earning its keep.
Why not just ask trial participants what they think they received, and adjust for it? Because Causal Inference in Studies with Functional Unmasking: Psychedelics and Beyond proves, with directed acyclic graph analysis and simulation, that participant belief is a collider influenced by both the randomized assignment and unmeasured prognostic traits, so conditioning on it opens a spurious back-door path that can manufacture or mask efficacy. The same paper rules out the dose-response escape: higher doses produce stronger somatic cues, so the dose-response curve can simply track an escalating placebo response. The simulated trial will quantify both failure modes for nanocarrier trials specifically — where unmasking cues are tied to release timing, a variable our simulator controls — and verify that the Controlled Direct Effect estimator with dense mediator measurements recovers the truth. The honest residual risk is that CDE identification assumes no unmeasured confounding between expectancy and outcome; our simulator can violate that assumption deliberately and measure how fragile the estimator is, which is a study real trials cannot run.
Proposed Experiments and Decisive Checks
Experiment one evaluates the generators. All three models — whole-trajectory diffusion, bootstrapped autoregressive transformer, and Imagand-on-curves — train on the enumerated corpus with matched compute and tuning. We report trajectory error against held-out mechanistic ground truth on in-distribution and out-of-distribution pH profiles, compounding error as a function of horizon, generation cost per trajectory, and distributional fidelity via Hellinger distance with Imagand's reported 0.15 average as the reference scale.
Experiment two evaluates the metrics. We construct paired test sets in which multiple chemically valid formulations differ from a single archived reference, and in which curves are corrupted to violate mass balance or pH-switch consistency while still passing through the two standard assay timepoints. We measure, for exact-curve matching, the plausibility score, and the round-trip check, the rate of correct credit and correct rejection. The target pattern mirrors what plausibility scoring achieved in chemistry: RetroGFN's round-trip-based evaluation reached 63.6% round-trip accuracy at k = 50 versus roughly 54% for the next best model on USPTO-50k, and UniCon agreed with strict chemist preferences in 94.1% of cases while experts preferred its protocols over patent ground truth in 61.8% of cases where they had a preference — both demonstrations that reference-free scoring captures validity that exact match discards. Where the corpus provides no numbers, we say so: no reported result found on Lacuna for numerical plausibility scores on the URSA out-of-distribution benchmark, which is why our out-of-distribution release profiles are a fresh measurement rather than a comparison.
Experiment three runs the collider study. We bootstrap the diffusion generator with and without plausibility filtering across successive self-generation rounds, tracking coverage of held-out rare mechanism classes. The hypothesis, drawn from the formal collider argument, is that filtering systematically removes valid rare mechanisms at a rate that grows with the filter's own distributional bias; the unfiltered arm quantifies the price of unfiltered hallucination, giving a bias-variance-style trade-off curve rather than a one-sided answer.
Experiment four runs the simulated unblinded trial. With a known pharmacological effect injected, we compare naive analysis, belief stratification, dose-response, and the lmtp Controlled Direct Effect estimator across measurement frequencies and sample sizes, since Loewinger et al. note their estimator's appetite for dense longitudinal measurement without quantifying it, and no reported result found on Lacuna for the numerical recovery error of their estimator on their simulation benchmark. We also test estimator fragility by injecting mediator-outcome confounding. The deliverable is a concrete schedule: how many expectancy probes per participant, and at what trial size, the Controlled Direct Effect recovers the injected effect within a stated tolerance — a number a nanomedicine trialist can act on.
The single decisive check, run first: on held-out mechanistic regimes requiring extrapolation, does the learned trajectory model beat the mechanistic simulator it was trained against? If not, the proposal should drop the generator and keep the metric and the trial engine — still, to our knowledge after searching the corpus, the first application of counterfactual mediation to nanomedicine evaluation, and the first replacement of reference-curve matching with chemistry-consistency scoring for release dynamics.
References
- Diverse and feasible retrosynthesis using GFlowNets
- When Single Answer Is Not Enough: Rethinking Single-Step Retrosynthesis Benchmarks for LLMs
- Drug Discovery SMILES-to-Pharmacokinetics Diffusion Models with Deep Molecular Understanding
- Causal Inference in Studies with Functional Unmasking: Psychedelics and Beyond
- Rapid synthesis of a Bi@ZIF-8 composite nanomaterial as a near-infrared-II (NIR-II) photothermal agent for the low-temperature photothermal therapy of hepatocellular carcinoma
- Author Correction: pH-sensitive micelles for the intracellular co-delivery of curcumin and Pluronic L61 unimers for synergistic reversal effect of multidrug resistance
- Therapeutics Data Commons: Machine Learning Datasets and Tasks for Drug Discovery and Development
- World Models via Policy-Guided Trajectory Diffusion
- Bootstrapped Transformer for Offline Reinforcement Learning
- Learning Reaction-Condition Plausibility for Evaluation and Self-Curation Under Noisy and Non-Unique Supervision