When a generic drug seeks approval, it must show in a small crossover study of healthy volunteers that its concentration-time curve matches the brand-name drug's: the 90% confidence interval of the ratio of geometric means for peak concentration ($C_{\max}$) and total exposure (AUC, the area under the concentration-time curve) must fall between 80.00% and 125.00%. This is an average over the study cohort. Yet we know that drug clearance is not a unimodal trait: roughly a third of Chinese pediatric patients carry loss-of-function alleles in the enzyme CYP2C19 that cripple activation of clopidogrel under standard dosing CYP2C19 polymorphisms and lipoproteins associated with clopidogrel resistance in children with Kawasaki disease in China: A prospective study, and 43% of 368 previously unannotated human variants of that same enzyme severely reduce protein abundance Deep mutational scanning of CYP2C19 reveals a substrate specificity-abundance tradeoff. The regulatory premise — that average equivalence in a mixed cohort implies interchangeability for every patient in it — has never been tested against this phenotypic heterogeneity.
This proposal claims that aggregate bioequivalence can mask subgroup-specific failure: a generic can pass the 80–125% criterion overall while a slow-metabolizer stratum accumulates toxic trough concentrations or an extensive-metabolizer stratum is left underexposed. The plan is to build a bioequivalence test that works inside the latent strata themselves. Concretely, we extend the existing model-based bioequivalence machinery — non-linear mixed effects models fitted by stochastic approximation expectation maximization, with the equivalence decision made by a model-based two one-sided test Model‐based bioequivalence approach for sparse pharmacokinetic bioequivalence studies: Model selection or model averaging? — by replacing its single Gaussian random effect with a finite mixture, so that the data, optionally guided by machine-learned variant-function maps, discover the metabolizer subtypes that the current framework smooths away.
Core Claim
Every sparse-sampling bioequivalence method in the literature models inter-individual variability as one Gaussian distribution around a population mean. This is a modeling choice, not a biological finding, and it is the exact point where phenotypic heterogeneity enters and is discarded. Our claim has two parts. First, a mechanistic one: when clearance is multimodal (as pharmacogenomics directly demonstrates), even a modest formulation-dependent difference in dissolution or enzyme saturability produces a subject-by-formulation interaction — the test/reference ratio itself differs across strata — that aggregate testing cannot see and that the crossover design's within-subject control cannot cancel. Second, a methodological one: a mixture-of-phenotypes extension of model-based bioequivalence can recover these strata from sparse crossover data, keep the false-positive rate at the regulatory level when the drug is truly equivalent everywhere, and flag the cases where equivalence holds on average but not within a stratum.
The model replaces the standard assumption $\eta_i \sim \mathcal{N}(0, \omega^2)$, where $\eta_i$ is subject $i$'s deviation in log-clearance, with
$$\eta_i \sim \sum_{k=1}^{K} \pi_k \, \mathcal{N}(\mu_k, \omega_k^2), \qquad \Delta_k = \ln \mu_{T,k} - \ln \mu_{R,k},$$
where $\pi_k$ is the fraction of subjects in stratum $k$ and $\Delta_k$ is the stratum-specific log test/reference ratio. Bioequivalence is claimed only if, for every stratum with non-negligible weight, the 90% confidence interval of $\exp(\Delta_k)$ lies within $[0.80, 1.25]$ — the same window as current regulation, applied where it actually matters. Where genotype or variant-function data exist, they enter not as hard strata but as priors on $\pi_k$ and $\mu_k$: functional scores from Hubble.2D6, a transfer-learned predictor of CYP2D6 haplotype function that reached 88% accuracy on held-out star alleles Transfer learning enables prediction of CYP2D6 haplotype function, and abundance measurements from the CYP2C19 deep mutational scan of 7,660 missense variants Deep mutational scanning of CYP2C19 reveals a substrate specificity-abundance tradeoff. This is the piece no prior work has: functional genomics has produced variant-to-phenotype maps, and virtual bioequivalence has produced trial simulators, but the two have never been connected (a limitation the ipratropium PBPK study states explicitly Bioequivalence study of ipratropium bromide inhalation aerosol using PBPK modelling).
Research Program
Three components. First, the statistical core: a latent-strata model-based bioequivalence estimator (LS-MB-BE) — the mixture model above, fitted with the same SAEM engine used in the existing sparse-sampling framework, with the number of strata selected by information criteria and the equivalence decision made per stratum with family-wise error controlled across strata. Second, the functional-genomic bridge: a mapping from variant scores to clearance-class priors, calibrated on the CYP2C19 abundance data (where 58% of variants reduce abundance) and the Hubble.2D6 activity-score logic, so that virtual populations can be built with realistic genotype-frequency-weighted mixtures rather than the generic physiological perturbations current simulators use. Third, the evaluation layer: a masked-failure metric — the probability that a method declares bioequivalence while at least one stratum's true ratio falls outside the window — which is the quantity aggregate testing was never designed to measure but which patients experience.
The closest existing work at this intersection is Global Deep Forecasting with Patient-Specific Pharmacokinetics, which uses a hybrid network with a patient-specific pharmacokinetic encoder (a learnable scalar modulating a log-normal absorption curve) and outperformed local-only models by 15.8% on the UVa/Padova type-1 diabetes simulator. It shares our population-plus-individual philosophy but answers a different question: it forecasts one patient's response over time, assumes complete bioavailability, and never compares two formulations, so it cannot speak to bioequivalence at all. The individual-bioequivalence literature of the 1990s gestured at subject-by-formulation variance ($\sigma_{WT}^2$ versus $\sigma_{WR}^2$), but treated it as a continuous variance component to be shrunk, not as discrete strata to be recovered; our proposal is, in effect, the data-driven modernization of that abandoned program. And SuStaIn Uncovering the heterogeneity and temporal complexity of neurodegenerative diseases with Subtype and Stage Inference supplies the proof of concept that unsupervised subtype recovery from cross-sectional data is feasible — it identified genetic mutation carriers at 95% accuracy from MRI alone — but in a neuroimaging context with no pharmacokinetic or regulatory content; we borrow the mixture-recovery logic, not the application.
Baselines
The do-nothing reference point is standard average bioequivalence by non-compartmental analysis on rich sampling: compute the geometric mean ratio of test to reference for $C_{\max}$ and AUC after log transformation and check the 90% confidence interval against 80.00–125.00%, exactly as done in the febuxostat crossover trial Bioequivalence evaluation of generic febuxostat versus Feburic® in healthy Chinese subjects: a randomized crossover study (no reported result found on Lacuna for that trial's numeric point estimates) and the sitagliptin/metformin comparison Pharmacokinetic comparison of sitagliptin and metformin HCl extended-release tablets versus JANUMET® XR in healthy volunteers under fasting and fed conditions. By construction this method "succeeds" whenever the aggregate criterion is met; the informative comparison is therefore not pass/fail agreement but the masked-failure rate: how often the aggregate method passes while a planted stratum-level failure goes undetected. We expect the aggregate method to have a masked-failure rate equal to the planted rate by construction, which is the point.
The strongest existing competitor is the model-based bioequivalence framework with Akaike-information-criterion model selection and model averaging in non-linear mixed effects models Model‐based bioequivalence approach for sparse pharmacokinetic bioequivalence studies: Model selection or model averaging?, which reports Type I error $\le 0.05$ for both $C_{\max}$ and AUC across 800 synthetic two-way crossover datasets, and which reached the same bioequivalence conclusion from a 6-sample sparse subset as from the rich 22-sample design in a 36-subject amlodipine crossover trial. This is the method to beat because it already solves the sparse-data problem our clinical settings force on us, and it is the model whose random-effect structure we are directly modifying. We will run it on exactly the same simulated and real datasets, same sampling schedules, same SAEM estimation settings and compute, with the number of candidate structural models matched, so the only difference is unimodal versus mixture random effects. Beating it means: at matched Type I error $\le 0.05$ under true equivalence, strictly higher probability of detecting planted stratum-specific failures, with the gap reported as the decisive number.
A third baseline is naive stratification: genotype participants with commercial star-allele panels and run standard two one-sided tests within each genotype stratum. Its known weakness is the uncurated-variant problem — the ordinal logistic regression baseline in the Hubble.2D6 work achieved only 54% accuracy on functional status versus 88% for the transfer-learned network Transfer learning enables prediction of CYP2D6 haplotype function, and 43% of unannotated gnomAD CYP2C19 variants escape commercial panels entirely Deep mutational scanning of CYP2C19 reveals a substrate specificity-abundance tradeoff. We will run it on cohorts where a controlled fraction of variants are unannotated, to quantify how its detection rate decays with annotation coverage. Finally, two ablations isolate our ingredients: the full mixture machinery with $K = 1$ forced (recovering the Gaussian baseline — shows what latent strata add), and the mixture without functional-genomic priors (shows what the variant maps add when the data alone are sparse).
Why not just…?
Why not just rely on the crossover design's within-subject control? The reviewer's argument is that in a two-period crossover each subject receives both formulations, so every phenotypic difference cancels in the within-subject test/reference comparison. This is true only under zero subject-by-formulation interaction: the cancellation requires the test/reference ratio to be the same in every physiological background. When a poor metabolizer's elimination enzyme operates near saturation, a modest dissolution difference between formulations produces a nonlinear exposure divergence that an extensive metabolizer never experiences — the ratio itself becomes phenotype-dependent, and the crossover design's cancellation fails precisely for the patients at greatest risk. Our first simulation experiment makes this quantitative by sweeping the size of the planted interaction, and we will be honest about the boundary: for drugs with linear kinetics and trivial excipient differences, the objection is correct, and our method should (and, if well-calibrated, will) recover no strata and change no conclusions.
Why not just stratify by known star alleles with PCR genotyping? This is a real threat and we treat it as such — for well-characterized enzymes in well-genotyped cohorts it may match our method. The evidence for its limits is concrete: over 70 CYP2D6 star alleles remain uncurated, driving frequent indeterminate clinical calls, and the best conventional predictor reached only 54% accuracy versus 88% for the deep-learning approach Transfer learning enables prediction of CYP2D6 haplotype function; 43% of unannotated gnomAD CYP2C19 variants significantly destabilize the protein and would be invisible to commercial panels Deep mutational scanning of CYP2C19 reveals a substrate specificity-abundance tradeoff; and in voriconazole pharmacokinetics, C-reactive protein overrides genotype as the driver of clearance, which no static genotype stratum can represent. Our design settles this empirically: the simulation benchmark includes a star-allele-oracle arm (perfect annotation of the known alleles only), so the gap between the oracle, naive stratification, and our latent-class method with functional priors is measured rather than asserted.
Why not just handle heterogeneity after approval, with therapeutic drug monitoring and dose titration? Post-hoc monitoring demonstrably works when infrastructure exists: in a 112-patient chronic myeloid leukemia cohort, monitoring guided successful de-escalation from 100 mg to 50 mg daily in 22 patients without loss of deep molecular response, using the thresholds $C_{\max} > 51.85$ ng/mL for major molecular response and $C_{\min} > 2.48$ ng/mL for pleural effusion risk Dasatinib Dose Optimization Based on Therapeutic Drug Monitoring in Patients with Chronic-Phase Chronic Myeloid Leukemia — the file's identifier is Dasatinib Dose Optimization Based on Therapeutic Drug Monitoring in Patients with Chronic-Phase Chronic Myeloid Leukemia. But that same study emphasizes that monitoring requires tandem mass spectrometry, repeated blood draws, and costs that are infeasible outside specialized centers — and for narrow-therapeutic-index drugs, the harm occurs before the first dose adjustment can be made. Monitoring is a remedy after the failure; the proposal aims to prevent the failure at the approval stage, and the two are complements, not substitutes.
Proposed Experiments
The first experiment extends the established synthetic benchmark: 800 two-way crossover datasets generated under the three structural models of the sparse-sampling study Model‐based bioequivalence approach for sparse pharmacokinetic bioequivalence studies: Model selection or model averaging?, but now with two clearance strata whose prevalence is calibrated to observed genetics (a one-third loss-of-function carrier rate, matching the Kawasaki disease cohort and the BOCLA-Plan finding that standardized antiplatelet dosing left roughly 33% of stented patients unprotected BOCLA-Plan study) and with a formulation difference that interacts with stratum through enzyme saturation. The decisive readout is a detection-versus-false-positive curve per method: the aggregate and Gaussian-mixed-effects baselines should sit on the chance line for masked failures (they cannot see them), naive stratification should degrade as annotation coverage drops, and LS-MB-BE should dominate at matched Type I error $\le 0.05$.
The second experiment reanalyzes the 36-subject amlodipine Phase I crossover trial, subsampling from its rich 22-sample design to the 6-sample regime where the baseline method matched the rich-sampling conclusion Model‐based bioequivalence approach for sparse pharmacokinetic bioequivalence studies: Model selection or model averaging?. We ask whether the mixture model recovers latent strata and whether any stratum-specific ratio leaves the window while the aggregate passes. We state plainly that 36 subjects cannot support more than two strata; this experiment is a feasibility and calibration check, not a powered claim.
The third experiment runs virtual bioequivalence trials: physiologically based pharmacokinetic simulation in the style validated for ipratropium aerosol, where the model fit achieved $r^2 \approx 0.98$ with prediction error for $C_{\max}$ and AUC under the 20% acceptance threshold, and predicted test/reference ratios of 99.76% for $C_{\max}$ and 104.3% for AUC in a 30-subject validation study Bioequivalence study of ipratropium bromide inhalation aerosol using PBPK modelling. We build virtual populations for a CYP2C19-metabolized substrate whose clearance classes are drawn from the VAMP-seq abundance map and Hubble-style activity scores — this substrate model is synthetic by construction, and we flag it as such — and sweep plausible excipient-driven dissolution shifts to find the boundary where the aggregate ratio stays within 80–125% while a stratum exits it.
The fourth experiment grounds the toxicity stakes in real clinical data, using the dasatinib cohort's bifurcated exposure structure: efficacy driven by $C_{\max}$ (mean $103.8 \pm 54.0$ ng/mL in major-molecular-response achievers versus $48.6 \pm 13.9$ in non-achievers) and toxicity by $C_{\min}$ (mean $4.14 \pm 1.19$ ng/mL with pleural effusion versus $1.89 \pm 1.01$ without) Dasatinib Dose Optimization Based on Therapeutic Drug Monitoring in Patients with Chronic-Phase Chronic Myeloid Leukemia. We ask the question that cohort could not: could a generic with a modest release-rate shift keep aggregate AUC equivalent while pushing a susceptible (elderly, accumulating) stratum's $C_{\min}$ across the 2.48 ng/mL toxicity threshold — a rate-versus-extent decoupling that aggregate AUC testing is structurally blind to.
The decisive checks are three numbers: masked-failure detection probability at matched false-positive rate against the strongest baseline; Type I error remaining $\le 0.05$ under true all-strata equivalence; and, on real data, whether any recovered stratum changes the regulatory conclusion. If the simulation shows the crossover design's cancellation holds at realistic interaction sizes, or that star-allele stratification matches the latent-class method whenever annotation is available, the honest conclusion is that the problem is narrower than claimed — and the boundary itself becomes the publishable finding.
References
- CYP2C19 polymorphisms and lipoproteins associated with clopidogrel resistance in children with Kawasaki disease in China: A prospective study
- Deep mutational scanning of CYP2C19 reveals a substrate specificity-abundance tradeoff
- Model‐based bioequivalence approach for sparse pharmacokinetic bioequivalence studies: Model selection or model averaging?
- Transfer learning enables prediction of CYP2D6 haplotype function
- Bioequivalence study of ipratropium bromide inhalation aerosol using PBPK modelling
- Global Deep Forecasting with Patient-Specific Pharmacokinetics
- Uncovering the heterogeneity and temporal complexity of neurodegenerative diseases with Subtype and Stage Inference
- Bioequivalence evaluation of generic febuxostat versus Feburic® in healthy Chinese subjects: a randomized crossover study
- Pharmacokinetic comparison of sitagliptin and metformin HCl extended-release tablets versus JANUMET® XR in healthy volunteers under fasting and fed conditions
- Dasatinib Dose Optimization Based on Therapeutic Drug Monitoring in Patients with Chronic-Phase Chronic Myeloid Leukemia
- Tailored antiplatelet therapy can overcome clopidogrel and aspirin resistance - The BOchum CLopidogrel and Aspirin Plan (BOCLA-Plan) to improve antiplatelet therapy