Specification for Model Based TDP

Testing whether integrated modeling improves a biomarker-based dose decision

A staged comparison of responder counts, continuous and longitudinal dose-response models, and integrated PK/PD predictions for a population biomarker target.

Question and Current Status

Does an integrated model improve selection of a dose that produces at least an 80% biomarker decline in at least 90% of the target population, compared with requiring at least nine observed responders among ten patients? Start with three cohorts of ten patients at increasing doses.

The literature review and single-assessment comparison are complete initial deliverables. The longitudinal and integrated PK/PD simulation described below is planned work. This activity assesses a biomarker decision; clinical benefit, safety optimization and adaptive allocation are outside its first stage.

TDP is retained as the activity’s label for the stated decision criterion; its expanded name has not been specified. The operational definition below does not depend on that terminology.

Document	Reader	Kind	Length
Index	Finding the activity’s documents	How-to	Short
Specification	Implementing and reviewing the comparison	Reference	As needed
Literature review	Understanding the prior art and its applicability	Explanation	Focused synthesis
Worked comparison	Checking the calculations and assumptions	Worked example	Reproducible results with interpretation
References	Checking sources and choosing what to read	Reference	Annotated source list

Response Definition and Population

Define response for patient \(i\) at dose \(d\) by

\[ R_i(d)=\mathbb 1\!\left\{\frac{B_i(t^*;d)}{B_i(0)}\le0.20\right\}, \qquad p(d)=P\{R_i(d)=1\}. \]

The first stage uses one protocol-defined assessment \(t^*\), a positive biomarker and a target population shared across doses. The simulation has complete follow-up and no rescue. The biological thresholds, 80% and 90%, are inputs from the motivating question, not findings from the literature.

Before applying the method to a program, specify the assessment window, baseline definition, target population and how intercurrent events affect response. A reduction at one visit, a trough reduction and sustained reduction throughout an interval are different endpoints. If discontinuation or rescue counts as failure, the model must include it in \(R_i\).

Use the observed-assay endpoint for the primary comparison. A latent error-free biomarker threshold is a separate estimand. Removing assay error from model predictions while retaining it in the observed-count comparator would change the target. Values below quantification require a censoring model; substitution by zero or half the limit is a sensitivity analysis, not the default likelihood.

Decisions and Measures of Performance

The dose-selection target is

\[ d^*=\min\{d\in\{d_1,d_2,d_3\}:p(d)\ge0.90\}, \]

with an explicit no-dose outcome when the set is empty. Restrict the initial decision to tested doses. Interpolation can be evaluated later, with its own error criterion and no extrapolation beyond the studied range.

Decision	Rule	Performance to report
Descriptive dose selection	Lowest dose with estimated \(p(d)\ge0.90\)	Correct lowest dose; inadequate dose; higher adequate dose; no dose
Evidence of population attainment	Lowest dose passing a calibrated test of \(p(d)\le0.90\), or a posterior criterion	False attainment declarations across all doses; probability of selecting an adequate dose; lowest-dose selection
Predicting a future cohort	Predict \(P(K_{\rm next}\ge9\mid\mathcal D)\)	Predictive calibration and scoring on independent cohorts

For Bayesian evidence, calculate \(q_d=P\{p(d)\ge0.90\mid\mathcal D\}\) and select when \(q_d\ge\gamma\), with \(\gamma\) set before evaluation. A nominal 95% posterior probability is not a guarantee of a 5% frequentist false-go rate. Calibrate and validate the entire selection procedure across null scenarios, including partial nulls where some doses qualify and others do not.

Statistical power must name an alternative, for example a true response rate of 95% at an adequate dose. High probability of rejecting \(p\le0.90\) when the truth is exactly 90% is incompatible with controlling the boundary type I error. Detecting any dose effect, selecting the lowest adequate dose, and establishing population attainment should not share one power label.

Analysis Comparators

ID	Analysis	Status
M0	Independent responder counts; observed 9-of-10 rule; exact-binomial evidence rule	Implemented
M1	Binary response pooled with an order-restricted dose relationship	Point-estimate rule implemented; calibrated binary dose-model inference planned
M2	Separate continuous log-ratio distribution at each dose	Implemented
M3	Separate means with shared residual variability	Implemented
M4	Continuous dose-response model sharing mean-shape and variance information	Simple log-dose model implemented; Emax/model uncertainty planned
M5	Longitudinal biomarker model with dose, without PK	Planned
M6	Joint population PK/PD model	Planned

Compare successive models and also compare M6 against the best simpler calibrated comparator. Add a shrinkage or bias-reduced binary regression to M1 before attributing gains to mechanistic PK. Evaluate weakly informative and externally informed priors separately, using the same external evidence where it is applicable to competing analyses.

For M4, consider a limited prespecified set of shapes: a log-dose line, an Emax curve, and a plateau or flexible monotone alternative. Three dose levels provide limited independent information about shape. Fitting an unconstrained four-parameter sigmoid simply because there are 30 observations does not make its parameters identifiable. Parameter restrictions require scientific support and sensitivity analysis. Generalized MCP-Mod is related prior art for shape uncertainty, as described in the review.

Integrated Model

A starting M6 generator can use a population PK model for exposure \(C_i(t;d)\) and an inhibitory Emax effect on biomarker production:

\[ \frac{dB_i}{dt}=k_{{\rm in},i} \left(1-\frac{I_{\max,i}C_i(t;d)}{IC_{50,i}+C_i(t;d)}\right) -k_{{\rm out},i}B_i(t),\qquad B_i(0)=\frac{k_{{\rm in},i}}{k_{{\rm out},i}}. \]

This is a candidate mechanism, not a claim about the unspecified biomarker. Use random effects for biologically supported heterogeneity and an observation model for PK, baseline and repeated biomarker samples. Alternative generators should include stimulation of loss, a delayed effect, incomplete maximum response and a refractory patient subgroup.

For each parameter draw from a fitted model:

1. Sample new patients from the target covariate distribution and population random-effects distribution. Preserve relevant covariance.
2. Simulate each candidate regimen, the biomarker trajectory and the protocol’s observation process, including assay noise for the observed endpoint.
3. Apply the same response rule as M0 and calculate the responder fraction \(p_d\) in a sufficiently large simulated population.
4. Repeat over parameter uncertainty to estimate \(q_d\), credible intervals for \(p_d\), and the distribution of \(d^*\).

The inner simulation represents variation among patients; the outer draws represent uncertainty in the fitted model. Averaging \(p_d\) over draws gives the posterior predictive responder probability, which is different from \(P(p_d\ge0.90\mid\mathcal D)\). Counting successes among the original patients’ empirical Bayes predictions can understate population variability.

M5 should use the same biomarker sampling times and observation model, with a parsimonious dose-dependent trajectory in place of the PK link. Evaluate M6 with jointly estimated PK and PD; a two-stage version must propagate PK estimation uncertainty. Conditioning on perfectly known exposures is an explicit idealized benchmark only.

Simulation Design

Hold the cohort allocation fixed while comparing analyses. The initial example uses relative doses 1, 2 and 4; a program-specific version must choose doses and sampling times based on the plausible biomarker dynamics.

Dimension	Required scenarios
Target attainment	None adequate; lowest, middle or highest dose first adequate; all adequate; rates just below, at and above 90%
Dose shape	Gradual increase, early plateau, late steep rise, nonmonotonic response
Distribution	Normal log ratio, skewness, heavy tails, refractory mixture, dose-dependent variance
PK contribution	Exposure informative versus redundant; high versus low PK variability; sparse versus rich PK sampling
Repeated measures	Few versus many visits; low versus high within-patient correlation; delayed or transient suppression
Observation process	Assay error, noisy baseline, quantification limits, unequal assessment times
Trial conduct	Missingness related to observed history or latent response; rescue; sequential-cohort population drift
External information	None, compatible prior, conflicting prior; uncertain population variability
Model uncertainty	Correct structure, wrong mean shape, wrong variance, wrong PK/PD mechanism
Sample size	5, 10, 20 and 30 patients per dose, with identical candidate doses

First vary one feature at a time to identify its effect, then combine plausible adverse features. Use separate simulation seeds and trials for decision-threshold calibration and performance evaluation. Preserve failed fits in the denominator and report their frequency and fallback decision. Do not discard difficult simulated studies.

The principal performance outputs are correct lowest-dose selection and selection of an inadequate dose. Also report probability bias, root mean squared error, interval coverage and width by dose, selection of a higher adequate dose, no-dose decisions, and failure to converge. For evidence rules, report both selection of an inadequate dose and declaration of any inadequate dose, even when a different dose is selected.

Use paired simulation draws across methods and report Monte Carlo standard errors for differences. Begin with 10,000 replicates per scenario for cheap models; the worst-case standard error for a single proportion is then 0.5 percentage points. Choose the replication count for expensive joint models from the precision required for the false-go comparison.

Define sample savings only at matched performance: the same target correct decision rate, the same acceptable false-go rate, and adequate coverage in the prespecified sensitivity scenarios. More frequent positive decisions alone do not establish a benefit.

Completion Criteria for the Integrated Comparison

• Reproduce M0’s operating characteristics analytically and M2’s normal-model uncertainty calculation independently.
• Show where each source of information changes the decision, using the M0–M6 comparisons and the same simulated studies.
• Verify population predictions against a large independently generated population, including the response tail and all failure components.
• Identify scenarios in which added modeling gives no benefit or selects inadequate doses more often.
• Report whether the gain survives matched false-go rates, model uncertainty and conflicting external information before recommending a smaller study.

Source claims and open reading tasks are maintained in References; calculations and implementation checks live in the worked comparison.