Blind Source-Filter Deconvolution for Long-Term Nanoparticle Clearance

When a drug nanoparticle is injected into a patient, its fate in each organ is the product of two things nobody can observe separately: the time-varying concentration of particles circulating in the blood, and each organ's own tendency to trap, hold, and slowly release them. Speech processing solved a structurally identical problem decades ago: a recorded voice is the convolution of a glottal excitation (the source) with the vocal tract's resonance profile (the filter), and inverse-filtering techniques can recover both from the recording alone. This proposal argues that the same mathematics, rebuilt with modern differentiable filters, can decompose long-term nanoparticle biodistribution curves into a systemic "source" and organ-specific "filter" — without serial blood sampling, which in animal studies requires sacrificing separate sub-cohorts at each time point.

The motivation is an acute-chronic gap that the nanomedicine literature documents but cannot currently close. Therapeutic platforms such as bismuth nanodots in a degradable zinc framework 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 and hollow mesoporous silicon/carbon particles Degradable Hollow Mesoporous Silicon/Carbon Nanoparticles for Photoacoustic Imaging-Guided Highly Effective Chemo-Thermal Tumor Therapy in Vitro and in Vivo both achieve complete tumor elimination over 14–18 days, yet the ultimate clearance of their inorganic cores and carbon shells is uncharacterized — the silicon carrier dissolves at only ~8.4 wt% over 30 days, and no reported result found on Lacuna for long-term retention of either the bismuth core or the carbon shell. Meanwhile, carbon-based nanomaterials that look benign in acute windows can trigger delayed ferroptosis in microglia during chronic retention Nitrogen-doped graphene quantum dots induce ferroptosis through disrupting calcium homeostasis in microglia. What is missing is a method that predicts multi-week organ retention curves from sparse measurements, and the source-filter view supplies it.

Core Claim

An organ's nanoparticle concentration curve $C_{\text{organ}}(t)$ can be written as a convolution of an unobserved systemic input $e(t)$ — the blood-borne particle concentration shaped by injection, dispersion, and renal elimination — with an organ-specific, time-varying retention filter $h_{\text{organ}}(t)$:

$$C_{\text{organ}}(t) = (e * h_{\text{organ}})(t)$$

A self-supervised analysis-by-synthesis network can jointly recover $e(t)$ and $h_{\text{organ}}(t)$ from measured multi-organ curves, using a differentiable all-pole filter in the reconstruction loop, and this blind decomposition will predict held-out long-term retention trajectories better than regression models that map particle properties directly to scalar kinetic constants. The all-pole form is not an arbitrary import from speech: a linear compartmental mass-balance model with constant rate constants yields a sum of exponentials, which is exactly the impulse response of an all-pole system. Letting the poles vary slowly in time extends this to the regime that matters here — saturation of liver Kupffer cells and other reticuloendothelial trapping, which the Hill-function uptake model of AI-Powered Prediction of Nanoparticle Pharmacokinetics: A Multi-View Learning Approach shows is strongly non-linear.

The key technical ingredient is the differentiable, sample-wise time-varying linear prediction framework of Differentiable Time-Varying Linear Prediction in the Context of End-to-End Analysis-by-Synthesis, which provides GPU backpropagation kernels for time-varying all-pole filters inside a gradient-optimized synthesis loop. That paper developed the machinery for speech vocoders on the CSTR VCTK corpus; here the "sample index" becomes chronological time in days, the source is a network-conditioned estimate of systemic input, and the filter is conditioned on organ identity and particle physicochemistry.

What Is Genuinely New Here

No published work couples source-filter architectures with nanoparticle clearance, but three papers come close enough that the distinction must be sharp. The PDNN system for dynamic PET Self-Supervised Learning for Physiologically-Based Pharmacokinetic Modeling in Dynamic PET also runs a neural network through a forward kinetic model in a self-supervised reconstruction loop, and it is the closest methodological ancestor; but it presumes a known, pre-extracted arterial input function — the source is measured, not inferred — and it evaluates small-molecule tracers over 65 minutes, not nanoparticles over weeks. The multi-view learning approach of Khakpour et al. predicts reticuloendothelial uptake from particle properties, but only as four scalar kinetic constants through tabular regression, never as continuous trajectories and never with a driving input at all. The physical-formula constrained learning of PEMAL contributes a reusable idea — enforcing conservation identities such as $AUC \times CL = K_1$ as training-time constraints — but applies it to static endpoints of small molecules, not to deconvolved time courses. This proposal combines PDNN's self-supervised forward-backward loop, Yu and Fazekas's differentiable time-varying filter, and PEMAL-style conservation constraints, and adds the one thing none of them do: blind estimation of the systemic source.

Research Program

The build has three parts. First, a source network takes dose, hydrodynamic diameter, zeta potential, core material, and shape — the feature set of Khakpour et al. — and outputs a systemic input curve $e(t)$, initialized to a standard biexponential bolus-plus-elimination shape but free to deviate. Second, an organ-filter network outputs time-varying all-pole coefficients per organ, implemented with the sample-wise differentiable linear prediction kernels of Yu & Fazekas so that the convolution is fully differentiable end-to-end. Third, training is purely by reconstruction: the model convolves its estimated source and filter, compares the synthesized organ curves to measured ones, and backpropagates the discrepancy — no ground-truth kinetic parameters required, exactly as PDNN trains without kinetic labels — with PEMAL-style conservation identities as additional soft constraints so that data-rich organs supervise the scarce ones. A deliberate secondary question, following the finding in Yu & Fazekas that analytically derived filter coefficients differ systematically from gradient-learned ones, is whether the same discrepancy appears here: do the learned organ filters deviate from classically fitted PBPK rate constants, and does the deviation carry physiological signal (e.g., progressive RES saturation)?

Because all code will be released against the open SBML standard — motivated by the reproducibility failures documented in Making PBPK models more reproducible in practice, where published compartmental models could not be reproduced across platforms — every experiment below is specified to be re-runnable.

Baselines

The naive reference point is a two-parameter exponential decay fit per organ curve, plus a standalone multilayer perceptron mapping particle features to the four scalar kinetic constants ($KTRES_{max}$, $KTRES_n$, $KTRES_{50}$, $KTRES_{release}$). The perceptron is the documented failure case: Khakpour et al. report $R^2 = -0.16$ for it on $KTRES_{max}$ over 5-fold cross-validation on their 280-mouse dataset, which is worse than predicting the mean and establishes the floor.

The strongest existing competitor is the multi-view cross-attention ensemble of Khakpour et al., which splits particle features into raw and domain-engineered views and reports $R^2 = 0.17$ on $KTRES_{max}$, significantly beating ridge regression, support vector regression, and the perceptron. It is the strongest because it is the only published method that both works on this data and encodes the physiological rule — hydrodynamic diameter dominance — that governs uptake. To make the comparison fair, it will be re-run on the same folds, same features, same compute and tuning budget, and evaluated not only on scalar constants but on full trajectories: its predicted constants will be pushed through its own Hill-function uptake model to synthesize organ curves, so both methods are scored on the same trajectory-level metric (RMSE and $R^2$ on held-out organ concentration curves, and MAE on predicted 90-day residual retention). Beating it means a higher trajectory $R^2$ with non-overlapping error bars across folds, and better extrapolation past the training horizon — a capability the scalar pipeline cannot express at all.

The critical ablation isolates blind source estimation: the identical network with the source fixed to the parametric biexponential bolus rather than learned. If the learned source adds nothing over the fixed one, the proposal's central claim is refuted and the contribution collapses to a differentiable re-implementation of known-input forward fitting, i.e., PDNN transplanted to nanoparticles. A second ablation replaces the sample-wise time-varying filter with a frame-wise time-invariant one, which Yu & Fazekas showed causes training-inference mismatch in speech; the question is whether the mismatch matters at the coarse day-scale sampling of pharmacokinetics, where it plausibly may not.

Why not just…?

Why not just fit compartmental ODE models with standard solvers? This is the default in pharmacokinetics, and where the compartment structure and the input function are both known, it is the right tool — our synthetic-benchmark experiments will use exactly such a solver as ground truth, and we expect to match, not beat, its accuracy there. The case for the proposed method rests on documented failures of the default in the target regime: nonlinear least-squares fitting is prohibitively slow at scale (De Benetti et al. report 8.7 minutes per 2D slice, roughly 58 hours for a 400-slice scan, versus about 1 minute for their network); fixed rate constants cannot represent progressive RES saturation; and the literature suffers severe reproducibility failures from proprietary solver pipelines Making PBPK models more reproducible in practice. Still, the honest test is in the plan: if ODE fitting with a correctly specified structure matches our trajectory accuracy even when the input is withheld and crudely approximated, the differentiable machinery is not earning its complexity, and we will report that.

Why not just use off-the-shelf tabular machine learning on particle features? Because the documented evidence says it fails on this data: the standalone perceptron scores $R^2 = -0.16$ on $KTRES_{max}$ Khakpour et al., and unconstrained deep models collapse under label noise, whereas embedding physical constraints reduced clearance MAE by 19.6% and AUC MAE by 15.2% in PEMAL, which also stayed stable when 80% of labels were corrupted. The real threat here is not the perceptron but the tuned multi-view ensemble, which is why it is the primary baseline; if it matches our trajectory accuracy, that settles the question in its favor and we will say so.

Why not just measure the arterial input curve directly instead of estimating it? Where it is measurable, this is genuinely better, and we do not dispute it — on the dynamic PET dataset of De Benetti et al., a measured blood-pool input is available and our blind estimate will be scored against it. But the target regime is rodent nanomaterial studies over weeks, where serial blood sampling is destructive terminal sampling across separate sub-cohorts (the very design that makes the 280-mouse dataset noisy), and image-derived inputs from thoracic regions suffer cardiac and respiratory motion artifacts that systematically underestimate vascular blood volume and contaminate downstream kinetic estimates De Benetti et al.. Blind deconvolution is a claim of necessity in this regime, and the dPET experiment — blind source versus measured source — is precisely the check that would falsify it: if the blind estimate diverges from the measured input on data where the input exists, the method has no business being trusted where it does not.

Proposed Experiments and Decisive Checks

Three testbeds, in increasing order of realism. First, synthetic ground truth: the open SBML adult whole-body PBPK benchmark of Making PBPK models more reproducible in practice will be used to simulate multi-compartment concentration trajectories, then the blood-compartment signal is deleted and the model must recover both it and each organ's transfer behavior from organ curves alone. Recovery is scored by cosine similarity between the recovered and true systemic source, with 0.78 — the mean organ-level cosine similarity PDNN achieved against non-linear least-squares fits De Benetti et al. — as the reference point a known-input forward model attains. Second, the 280-mouse nanoparticle dataset of Khakpour et al.: trajectories are generated by forward-simulating the documented Hill-function uptake model from measured parameters, the model is trained on curves from a subset of organs and particles and evaluated on held-out organ-particle combinations, with the 5-fold protocol preserved for direct comparison against the reported $R^2$ values (0.17 versus -0.16 on $KTRES_{max}$). A data-efficiency test follows PEMAL, which showed a constrained model trained on 20% of data outperforming an unconstrained one trained on 100%: our conservation-constrained model should show the same curve on 280 samples. Third, the 23-patient long axial field-of-view dynamic PET dataset De Benetti et al., used as the falsification test for source recovery against the image-derived aortic input, on the known 65-minute, 62-frame protocol.

The decisive long-horizon check speaks directly to the clinical gap: train on data truncated at day 30 — the timescale over which the silicon carrier of Zhang et al. dissolves only 8.4 wt% — and evaluate predicted residual retention at days 60–90, where no reported result found on Lacuna for any of the cited therapeutic platforms exists, which is exactly why a predictive model rather than a measurement campaign is the practical path. If the source-filter model extrapolates beyond the training horizon substantially better than the fixed-source ablation and the multi-view baseline (which cannot extrapolate at all without re-simulation assumptions), and if its recovered source matches the measured input on dPET, then the case is made: organ biodistribution is a convolved system whose driving input can be recovered blind, and long-term clearance becomes predictable from data that already exists.