Learning Saturable Drug Kinetics from Sparse, Clinician-Timed Blood Samples

When a drug's elimination machinery saturates, doubling the dose can more than double the concentration in the blood. Which patients are in that regime, and at what dose they enter it, are the questions dose escalation and therapeutic drug monitoring exist to answer. The data available for answering them are a handful of blood draws per patient, taken at irregular times. Those times are rarely random: clinicians order a level when they suspect toxicity. We propose an amortized inference model, meaning a network trained once on simulated patients that then maps a new patient's few samples directly to predictions without per-patient fitting. Its training simulator randomizes two things: the nonlinear form of the patient's kinetics, and the rule that decided when blood was drawn. The model's central job is dose extrapolation, predicting a patient's concentration curve at a dose they have not yet received. That task is exactly where linear dose-stacking assumptions fail and where a wrong answer causes harm.

The approach rests on a simple observation. A network trained to predict outcomes from simulated data approximates the posterior under the simulator's joint distribution over dynamics and observation times. If the simulator draws samples on fixed schedules, the network implicitly treats real sample timing as uninformative. Clinician-triggered sampling then gets misread as dynamics: draws cluster at high concentrations, which mimics slow clearance and saturation. If the simulator instead includes state-dependent sampling, the network learns to discount that selection effect. The theory behind continuous-time filters such as Neural Jump ODEs explicitly assumes timing is independent of the state, so it cannot provide this correction. In the simulator approach, the correction comes from the training distribution.

Core Claim

Our claim is that an in-context model with three properties will extrapolate to unseen doses more accurately and with better-calibrated uncertainty than current deep pharmacokinetic (PK) models, generic irregular-time networks, and population models of fixed structure. The model is pretrained on a mixture of linear, saturable, and time-varying kinetics. Its training data are paired with a randomized family of sampling-time policies, including policies that depend on the hidden concentration. It decodes predictions through an explicit compartment ODE whose elimination term can saturate. We expect the advantage to be largest under clinician-timed sampling, and we expect it to vanish when kinetics are truly linear. The second expectation acts as a built-in negative control.

The elimination term the decoder can represent is Michaelis–Menten:

$$\frac{dC}{dt} = \frac{\text{input}(t)}{V} - \frac{V_{max}(t)\,C}{V\,(K_m + C)},$$

Here $C$ is plasma concentration, $V$ the volume of distribution, $V_{max}$ the maximal elimination rate, and $K_m$ the concentration at half-maximal elimination. Kinetics are linear when $C \ll K_m$. Letting $V_{max}(t)$ change over time allows autoinduction and inflammation-driven suppression. Sampling times are generated by a point process with intensity $\lambda(t) = g(C(t), t)$. The simulator draws $g$ from a family that includes fixed protocols ($g$ independent of $C$), uniform random draws, and threshold-triggered draws.

Relation to Prior Work

The closest work is AICMET. It pretrains a Transformer on synthetic one- and two-compartment PK curves and predicts zero-shot from sparse, irregular Phase I data in PK-DB. It outperforms traditional nonlinear mixed-effects models (the standard pharmacometric population models, abbreviated NLME) on 7 of 10 parent compounds. AICMET supplies our method: synthetic pretraining with population-level and individual-level latent codes. It also motivates the gap, because its simulator is predominantly linear and its authors say complex multi-dose schedules and dynamic covariates need fine-tuning. We differ in two ways. First, our simulator includes saturable and time-varying elimination. Second, and more distinctively, it randomizes the sampling-time process. We also evaluate on dose extrapolation, not interpolation within a regimen.

The PK encoder of Challu et al. and the related Forecasting Response to Treatment with Deep Learning and Pharmacokinetic Priors turns insulin doses into log-normal concentration curves, with a learnable rate per patient. It sums overlapping doses linearly and works on a regular 5-minute grid. Its authors name nonlinear PK as future work. That is part of our motivation. We reuse their glucose setting only as the linear control.

On the irregular-time side, Neural Jump ODEs prove convergence to the conditional expectation only when observation times are independent of the state. Multi-Time Attention Networks and Interpolation-Prediction Networks handle arbitrary timestamps without physical structure. IPN in particular shows that sampling intensity is itself predictive. These papers constrain our design: timing must be treated as data, not noise. They also serve as black-box baselines.

Hur et al. found that a generic network pretrained on realistic biokinetic simulations nearly matches an architecture with an embedded ODE ($R^2 \approx 0.515$ vs $0.554$). Pretraining on generic Gaussian-process curves failed. This partly contradicts our choice of an explicit ODE decoder, so we test that choice directly instead of assuming it.

The pharmacology literature grounds the simulator's families. Zhang et al. and the LXI-15029 escalation study document saturable clearance. Cahn et al. document absorption ceilings. Koike et al. and Touchi et al. document autoinduction with threshold dynamics. Jiang et al. report that, in their voriconazole cohort, clearance was suppressed by inflammation measured by C-reactive protein (CRP) and that sparse sampling caused high shrinkage in absorption rate and volume.

Research Program

We will build the simulator first. It draws a structural family: linear, Michaelis–Menten elimination, saturable absorption, or time-varying $V_{max}$ driven either by an autoinduction state or by a CRP-like covariate that is itself sparsely observed. It then draws population and individual parameters, a dosing history with escalations, and a sampling policy $g$. The policy is drawn from protocol schedules, random draws, threshold-triggered draws, and "symptom-triggered" draws with delay and noise.

The model is an AICMET-style encoder over (time, dose, concentration, covariate) tokens. Its decoder outputs parameters of the ODE above, integrated with a fixed-step solver. This is cheap for a three-state system, so the ODE-cost concerns raised by mTAN largely do not apply. Training uses a proper predictive loss on held-out query times and held-out future doses.

The analysis component asks when saturation is identifiable from $n$ samples. We expect identifiability to depend on whether any sample falls near or above $K_m$. Clinician-triggered sampling may therefore help identify nonlinearity even as it biases naive estimates. We will measure this trade-off directly.

Baselines

The naive reference is a population one-compartment linear model with linear dose stacking, individualized by maximum a posteriori fitting. This is roughly what routine TDM software does. Any useful model must beat it at unseen doses, and we expect it to under-predict concentrations badly once escalation enters the saturated regime.

The strongest classical competitor is an NLME model given the correct structural family: Michaelis–Menten elimination where that is the truth. It is fit with standard mixed-effects estimation on each synthetic study or PK-DB compound. Because it is told the true structure, it is an oracle on structure. Matching its accuracy without manual specification, and with better calibration under triggered sampling, would count as success. We will also run it with a misspecified linear structure, to show what manual choices cost.

The strongest deep competitor is AICMET's recipe as published, with a linear prior. No numeric error values for it on PK-DB were found on Lacuna. The only reported result is that it beats NLME on 7 of 10 parent compounds. We will therefore reimplement it with the same encoder, compute, and number of pretraining curves as our model. Beating it means lower error and better interval coverage at held-out doses on PK-DB compounds with documented nonlinearity, with no loss on linear compounds.

Black-box irregular-time baselines are mTAN and Neural Jump ODEs. For context, mTAN reports a PhysioNet mortality AUC of 0.858 and NJ-ODE a PhysioNet extrapolation MSE of 1.945, but neither result is on a PK task. We will train each on the same simulated corpus with matched tuning budgets, so any gap reflects structure rather than data.

Two ablations isolate the key ingredients. The first keeps the nonlinear simulator but uses only fixed-schedule sampling, isolating the timing randomization. The second keeps the full simulator but replaces the ODE decoder with a generic MLP decoder, isolating the mechanism. All models share the same hyperparameter-search budget and the same test patients.

Why not just…?

Why not just add Michaelis–Menten curves to AICMET's simulator? This is the obvious extension, and it is the threat we take most seriously. It is exactly our fixed-schedule ablation. If that ablation matches the full model under clinician-triggered test sampling, our timing contribution is void and the paper reduces to "extend the prior." The experiment comparing models across train and test sampling policies settles this. Our prediction is that the fixed-schedule model will overestimate saturation and exposure when draws cluster at peaks.

Why not just fit a nonlinear NLME model in standard pharmacometric software? When a pharmacometrician knows the right structure and samples cover the saturating range, NLME works well. The oracle baseline acknowledges this. Its weaknesses are the weeks of manual specification noted by AICMET and the parameter shrinkage under sparse sampling reported by Jiang et al.. Standard NLME likelihoods also treat sampling times as fixed by design. We will report where the oracle wins, which we expect to be dense sampling with known structure, rather than claim uniform superiority.

Why not just pretrain a generic network on the simulations, as Hur et al. suggest? Their result says the mechanism may be redundant once the simulator is realistic. Our MLP-decoder ablation tests this directly. If it matches the ODE decoder on in-range doses but not on extrapolated doses, we have evidence that the mechanism matters specifically outside the training range. That would refine their claim rather than contradict it. If it matches everywhere, we will report the simpler model as the recommendation.

Why not just use an irregular-time network like mTAN on the raw draws? These models handle timestamps natively. However, nothing in them encodes mass balance, so we expect unphysical extrapolation from five to fifteen samples. That expectation is consistent with the collapse of an unconstrained MLP reported by Hur et al. ($R^2 \approx 0.168$) and with AICMET beating neural ODE variants. The matched black-box baselines test it directly.

Proposed Experiments

On simulated cohorts with known ground truth, we will cross three training sampling policies (fixed, random, mixed including triggered) with four test policies. The fourth test policy is a triggered family held out from training, to probe misspecification of the timing process. For each cell we will measure error at held-out escalated doses, accuracy in classifying supratherapeutic exposure, and 90% interval coverage. The supratherapeutic threshold mirrors voriconazole's 5.5 mg/L ceiling. The decisive check is the off-diagonal cells. Fixed-trained models should show systematic exposure bias under triggered test sampling that the mixed-trained model does not. That bias should grow with the strength of the trigger.

On PK-DB we will follow AICMET's compound set and splits. We will add a dose-extrapolation split: context drawn from a patient's lower dose levels, queries at higher ones. We will report separately compounds the literature describes as saturable and those it describes as linear.

As a negative control we will use UVA/Padova simulated insulin data, where Challu et al.'s linear encoder reports up to 16.4% gains on critical events. There, our model should match a linear prior and not hallucinate saturation. A false-positive saturation rate is reported explicitly.

If a cohort like the Guangxi voriconazole data (69 patients, 233 samples) can be obtained under agreement, we will test whether sparse CRP observations let the time-varying-clearance family predict concentrations better than the CRP-stratified population model of Jiang et al. We treat this as optional, since we cannot confirm access.

The project fails cleanly if the fixed-schedule ablation equals the full model on triggered test data, or if the linear AICMET reimplementation matches us on the dose-extrapolation split. Either result would be informative and publishable as a negative finding.