Which Release Experiments Can Reveal a Nanocarrier's Internal Physics?

Drug carriers that respond to acidity are built to stay intact in blood (pH 7.4) and to fall apart in tumors or inside cells (pH 4.5 to 6.8). These include metal-organic frameworks (crystalline cages of metal ions and organic linkers), polymer micelles, and layered capsules. Their release behavior is still summarized by fitting one empirical curve per experiment, usually a power law $M_t/M_\infty = k t^n$, and reading the exponent $n$ as a mechanism class: "Fickian," "anomalous," or "Case II." Supernova astronomy had a similar habit and recently dropped it. Classes such as plateau versus linearly declining Type II, or faint 1991bg-like versus normal Type Ia, turned out to be cuts through continuous physical parameters such as residual hydrogen envelope mass or outer-ejecta abundances. That field reached this conclusion by pairing a physical forward simulator with likelihood-free Bayesian inference, and by learning which observations constrain which parameters.

This note proposes the same move for pH-responsive release. We will build a pH-gated, multi-layer diffusion–erosion simulator. We will train a neural posterior estimator on it that integrates out experimental nuisance factors. We will then use it to answer a question the drug-delivery literature has not posed: does a standard constant-pH release curve contain enough information to pin down a carrier's internal parameters at all, and what pH protocol would? Our hypothesis has two parts. First, single-pH curves leave the trigger parameters (the pH at which the carrier opens, and how sharply) almost unconstrained. Second, the power-law exponent is a many-to-one summary that merges physically distinct carriers. Third, a short stepped protocol (7.4, then 6.5, then 5.0) makes the parameters identifiable.

Core Claim

The open problem is identifiability, not fitting accuracy. Classical kinetic fits already reach $R^2$ close to 1. In the ivermectin bolus study, Korsmeyer-Peppas gave the best fit and lowest information criterion, but its exponent fell between 0.45 and 0.89, which only says "diffusion plus erosion" (Ruan et al.). A good fit with an uninformative parameter is what non-identifiability looks like.

We claim that simulation-based inference makes the problem measurable. Simulation-based inference means training a neural network on simulated (parameter, data) pairs so that it outputs a posterior over parameters directly, without writing down a likelihood. Applied to a mechanistic release simulator, it turns the question "which mechanism is this?" into the question "how far does the posterior over physical parameters contract, and is it calibrated?" Experimental design choices, chiefly the pH schedule and the sampling times, can then be ranked by how much they shrink that posterior.

What would be new is threefold. The first is the identifiability map itself. The second is a demonstration that the Korsmeyer-Peppas exponent is degenerate across distinct physical parameter regions. The third is a validated protocol recommendation, tested on real metal-organic framework carriers.

Prior Work and Its Role

The motivating evidence comes from astronomy. Raya Dastidar et al. traced the plateau/linear split in Type II supernovae to one continuous parameter, envelope mass. Li et al. used the transitional SN 2012ij to bridge faint and normal Type Ia events. O'Brien et al. used abundance tomography over 40 objects to show that 1991T-like events extend the normal population. They made this tractable with DALEK, a neural emulator of the TARDIS radiative-transfer code with roughly 1% mean fractional error. These papers motivate the hypothesis that discrete labels hide continuous drivers. They do not tell us whether release curves carry enough information to recover those drivers. That is the gap we target.

The inference method comes from SIDE-real. It applied truncated marginal neural ratio estimation (TMNRE) to 86 real Carnegie Supernova Project light curves through a differentiable simulator. TMNRE learns the ratio of posterior to prior for a few parameters of interest at a time and progressively narrows the prior region it simulates from. SIDE-real used it to infer dust properties while integrating out many per-object nuisance parameters. We reuse this estimator family and the truncation idea. Our parameters of interest are carrier physics. Our nuisances are lab conditions: stirring, sink volume, buffer ionic strength, pH drift, and assay baseline drift.

Shen and Gagliano supply a calibration standard. Their conditional diffusion model, DiTSNe-Ia, reported 0.845 coverage for nominal 90% intervals on the Goldstein & Kasen grid, and they warn that models trained on idealized synthetic grids may understate real diversity. We adopt that warning as a constraint: every posterior we report must pass coverage tests on held-out simulations, and our model must be stress-tested against real data the simulator did not generate.

The forward model builds on the drug-delivery side. The multi-stratum capsule model solves nonlinear diffusion across concentric layers with finite volumes. It includes radial anisotropy, coefficients that jump in space and time, and surface erosion, and it was validated against a tannic-acid digestion dataset. It has no pH-gated chemistry and no inverse inference. We add both.

The chemistry that determines the gating comes from three sources. Li et al.'s Bi@ZIF-8 releases cargo when its imidazolate linkers become protonated in acid. Ma et al.'s pyrazole ZnMOF stayed crystalline for 20 days in neutral buffer, against roughly 3 days for ZIF-8, yet still released doxorubicin in acid. Hong et al.'s pH-sensitive micelles disassemble in acid. The ZIF-8/ZnMOF pair matters most to us. It is a natural "two classes" case that our method should be able to place on a single continuum of trigger midpoint and neutral-pH erosion rate.

The file also points to an unsupervised alternative. DRACULA (Ishida et al.) compressed peak-light spectra into four nonlinear dimensions where linear PCA needed ten or more. We treat this as a competing approach and test it directly, as described below.

Research Program

The forward model is a radially symmetric, multi-layer reaction–diffusion model of the drug concentration $c(r,t)$ inside a carrier of radius $R$:

$$ \frac{\partial c}{\partial t} = \frac{1}{r^2}\frac{\partial}{\partial r}\!\left(r^2 D_\ell(\mathrm{pH}(t))\,\frac{\partial c}{\partial r}\right), \qquad D_\ell(\mathrm{pH}) = D_\ell^{0} + \frac{D_\ell^{\max}-D_\ell^{0}}{1+10^{\,h(\mathrm{pH}-\mathrm{pH}_{50})}} . $$

Here $D_\ell$ is the diffusivity of layer $\ell$. It moves from a closed value $D_\ell^0$ to an open value $D_\ell^{\max}$ around a trigger midpoint $\mathrm{pH}_{50}$, with sharpness $h$, mimicking linker or bond protonation. The outer boundary recedes at an erosion rate gated by the same pH curve, and an initial burst fraction represents surface-adsorbed drug. These choices follow the layer-and-erosion structure of the multi-stratum model.

The physical parameter vector $\theta$ therefore has roughly seven to nine components. They are the per-layer closed and open diffusivities, $\mathrm{pH}_{50}$, $h$, the open and closed erosion rates, the shell thickness fraction, and the burst fraction. Nuisance parameters $\eta$ cover sink-volume replacement at each sample, stirring-dependent boundary-layer resistance, pH drift of about ±0.1, multiplicative and additive assay noise, and baseline drift. The simulator outputs cumulative release sampled at realistic times under a given pH schedule $s(t)$.

Next we train TMNRE marginal estimators for single parameters and pairs of parameters, separately for each candidate pH protocol. If a one-dimensional radial finite-volume solve costs well under a second, we simulate directly. A neural emulator in the style of DALEK is built only if repeated simulation becomes the bottleneck, and we report how much error the emulator itself introduces.

The analysis has three parts. First, we compute posterior contraction (prior width divided by posterior width) for each parameter, under each protocol, across the prior. This gives the identifiability map. Second, we take simulated carriers with matched fitted Korsmeyer-Peppas exponents and ask whether their posteriors over $\theta$ fall in well-separated regions. Such a result would show the exponent merges distinct physics. Third, we search a small family of stepped pH schedules for the one that maximizes expected contraction of $\mathrm{pH}_{50}$, $h$ and the erosion rates. The family varies the number of steps (one to three), the step times, and the step levels.

Baselines

The naive reference point fits Korsmeyer-Peppas, Higuchi, first-order and Weibull models to each constant-pH curve and assigns a mechanism class from the exponent, following Ruan et al.. Their quality measure is $R^2$ close to 1 with lowest information criterion, and no stronger numeric result is given. Because these models do not output physical parameters, we compare them on two things. One is held-out curve prediction under a new pH schedule. The other is mechanism discrimination, meaning whether the fitted exponent separates carriers whose true $\theta$ differ. Beating this baseline means correctly predicting release under an unseen pH step, where a power law fitted at one pH has nothing to say.

The strongest competitor is likelihood-based Bayesian inference on the same simulator. We run Markov chain Monte Carlo per curve, with an assumed Gaussian noise model and the nuisances either fixed at nominal values or sampled jointly. It gets the same simulator, the same priors, and a matched total simulation budget. It is the strongest because, whenever its noise model is right, it is the exact target that neural inference approximates. Our method beats it if it reaches equal or better coverage when the nuisances are misspecified relative to the Gaussian assumption, and costs less per curve once amortized over many curves. If it ties under correct specification, that is expected and we will report it plainly. No reported result was found on Lacuna for simulation-based inference applied to drug-release curves. The closest calibration figure is DiTSNe-Ia's 0.845 coverage for 90% intervals (Shen and Gagliano), which comes from a different domain and is used only as a reference level.

The unsupervised baseline is an autoencoder with Isomap in the style of DRACULA, plus PCA, trained on the same simulated curves. We then regress the true $\theta$ on the learned coordinates.

Two ablations isolate the key ingredients. The protocol ablation runs the identical estimator on constant-pH data versus stepped-pH data. It tests whether identifiability comes from the experimental design or from the inference machinery. The nuisance ablation trains with the lab nuisances frozen at nominal values and then tests on data where they vary. It measures how much marginalizing over nuisances contributes.

Why not just…?

Why not just fit Korsmeyer-Peppas or Weibull curves as everyone does? Those fits describe a single curve well but cannot extrapolate to a new pH history, and their exponent is the quantity we suspect is degenerate. The matched-exponent experiment settles the question. If carriers with equal fitted $n$ turn out to have overlapping posteriors over $\theta$, then the classical summary loses nothing and our case weakens. We will report that outcome if it occurs.

Why not just learn a low-dimensional manifold of release curves with an autoencoder or PCA, as DRACULA did for spectra? A learned manifold shows that curves vary continuously. It does not say whether its axes are physical, and it is silent about experimental design. Ishida et al. themselves interpreted their four axes only after the fact. We include this baseline. If its coordinates predict $\theta$ as well as our posterior means and need no simulator, that is a real threat. In that case the contribution narrows to the protocol-design result.

Why not just run MCMC on the simulator directly? This is the most serious alternative. If the simulator is cheap and lab noise really is Gaussian, MCMC is exact, and neural inference adds only the saving from reusing one trained network across many curves. The contest is decided under misspecified nuisances and on the real metal-organic framework data below. If MCMC with jointly sampled nuisances matches our coverage there, we will recommend it and present the identifiability map as the main finding, since that map does not depend on which inference engine produced it.

Why not just use response-surface design of experiments to choose the protocol? Response-surface methods, as in Ruan et al., fit quadratics to scalar endpoints. Our design target is the expected posterior contraction of hidden parameters, which is not one of their endpoints. We will check whether a response-surface search over the same schedule family, scored on single-point cumulative release, picks schedules that also maximize contraction.

Decisive Experiments

On simulated data we hold out 10,000 carriers drawn from the prior. We report the coverage of 50%, 90% and 95% intervals, posterior contraction per parameter under each protocol, and prediction error on release under an unseen pH schedule. The central test is the protocol ablation. If stepped schedules do not raise contraction of $\mathrm{pH}_{50}$ and $h$ at least twofold over constant pH, the identifiability claim fails. To probe misspecification in the way Shen and Gagliano warn about, we also test on curves generated with an extra mechanism the simulator lacks, namely carrier swelling.

On real data there are two tests. First, the multi-stratum capsule tannic-acid digestion data, whose gastric-to-intestinal pH transitions form a natural stepped protocol, serves as an out-of-simulator check. Second, a wet-lab arm reproduces doxorubicin-loaded ZIF-8 and pyrazole ZnMOF following Ma et al.. It measures release in triplicate under the constant pH 7.4, constant pH 5.0, and optimized stepped protocols.

Success on real data requires three things. The posteriors from the stepped protocol must place the two frameworks in separated $(\mathrm{pH}_{50},\ \text{neutral erosion rate})$ regions consistent with their reported stability gap of 20 days versus 3. The stepped-protocol posterior must predict a held-out fourth schedule better than both the power-law fits and the MCMC baseline. And the constant-pH posteriors must fail to separate the two frameworks. That last result would be the release-kinetics counterpart of a "class" turning out to be a cut through a continuum.