Intrapatient dose escalation for first-in-human T-cell engagers

Working specification for a trial-design simulation

trial design
simulation
pharmacometrics
oncology
Does escalating dose within a patient reach a confirmable active dose sooner, and does the time it saves outweigh the extra visits and hospital time it costs, or is it a wash? A specification for the simulation that would answer it.
Published

September 6, 2026

This is a working document, not a finished analysis. It specifies a simulation that has not been built yet. Parameter values are simulation assumptions drawn from published models, not clinical estimates, and the conclusions listed in Section 26 are hypotheses to be tested rather than results. The project index lists the other documents in this folder, and References carries the sources.

In brief

The question. In a first-in-human T-cell engager trial, the starting dose may sit orders of magnitude below the biologically active one, and conventional escalation crosses that gap one cohort at a time. Escalating within a patient crosses it faster. Does that actually get the trial to a confirmable active dose sooner, and does the time it saves outweigh what it costs in extra visits, doses and hospital time? It can save more than it costs, cost more than it saves, or come out a wash, and which of those it is depends on the scenario. Section 1 sets out the three axes the answer is reported on.

Why it might work. The value of the strategy does not come from the size of the gap between the starting and active dose. It comes from uncertainty about that gap. If translational prediction is roughly right, conventional escalation already reaches the active range efficiently. The design earns its keep only when the prediction is wrong by a large factor, so prediction error is the primary simulation axis rather than a sensitivity analysis.

The design. One scout patient escalates weekly until peripheral B cells fall below 5 cells/uL, then stops. That patient’s ladder is not an interpretable regimen, so it nominates a dose neighbourhood rather than a dose: three fresh patients then receive a conventional step-up regimen at the nominated target, and only those patients count toward the primary endpoint. The strategy is a scouting device with a pharmacodynamic stopping rule, not a replacement for conventional dose characterisation.

What cannot be quantified. Escalating weekly inside a toxicity process with a longer latency means a normal observation at the next dose decision is incomplete reassurance. For CRS this is tractable and the ladder is arguably protective, since the scout is better primed at any given dose. For ICANS it is not: no published exposure-response model supports simulating that risk, and the document declines to invent one. That limitation belongs to the strategy, not just to the simulation.

flowchart TD
    A["Large uncertainty about the active dose"] --> B["Rapid within-patient traversal"]
    B --> C["First strong pharmacologic signal"]
    C --> D["Stop escalating"]
    D --> E["3-patient conventional backfill<br/>at the nominated target dose"]
    E --> F["Ordinary expansion / dose optimisation"]

How to read the answer. The simulation quantifies a benefit against a risk it explicitly cannot quantify, so no region of the result is a recommendation on its own. What it can say is where the saving is large enough for the safety discussion to be worth having, and where it is small enough that the discussion is unnecessary. A saving under about eight weeks, one conventional cohort cycle plus its accrual, is inside the noise of ordinary trial execution. Section 26b reads the map.

1. High-level objective

Develop a simulation and operational framework to answer:

Under what conditions does intrapatient dose escalation (IPDE) materially shorten the time required in a first-in-human T-cell engager study to reach and conventionally evaluate a biologically active dose/regimen, or reach that point at the same time having treated fewer patients at subtherapeutic exposures, and what safety and operational risks accompany either?

Three axes, not one. The saving can arrive as calendar time, as patients, or as what the trial consumes, and a design can win on one without the others. Catch-up escalation is expected to be exactly that case: little change in \(T_{3,\rm eval}\), substantially less patient-time at clearly subactive doses (Section 6). An objective stated on time alone scores that design as a failure when it is not one.

The three headline quantities are:

  • Time. \(T_{3,\rm eval}\), and \(\Delta T\) against the conventional comparator (Section 4).
  • Patients spent reaching the answer. The number enrolled before the first profound B-cell depletion, and patient-weeks with B cells at or above 5 cells/uL (Section 5).
  • What the trial consumes. Total patient-visits, and total observation and hospitalization time in patient-hours (Section 19).

Report the third axis per patient and per trial, because they point opposite ways. A scout accumulates a visit, a dose, a pharmacy preparation and an observation window at every rung, where a conventional patient sits at one dose level for the same period, so per patient the ladder is heavier. Across the trial it may be lighter: the scouts are few and each rung costs one patient’s visits, where a conventional escalation spreads a smaller per-patient burden across three patients at every rung. A single total hides that, and the two readings answer different questions, one about what a participant undergoes and one about what the trial consumes. Section 19 expects the per-patient size to depend on what the patients were scheduled for anyway: close to free for a QW oncology TCE where the visits are already planned, substantial for an immune-reset programme whose eventual treatment is one or two administrations.

There is no cost model and there should not be one yet. Turning visits, doses, hospital hours and patients into a single number requires weights, and a sponsor, a site and a patient weigh them differently. Section 19 reports the components separately for that reason, and Section 5 keeps the list short. The three quantities above are the headline set because they are what a programme argues about, and a programme wanting a currency figure can price its own counts.

What comes out of this

A map of the design space, not a verdict on a programme. The question is which conditions favour IPDE and which do not, so the deliverable is \(\Delta T\), \(\Delta N_{\rm sub}\) and the burden counts reported across the scenario grid in Section 21, with the regions where each is material named.

Report contours, not a pass or fail. A figure such as eight weeks saved is a line drawn on the map for reading it, not a bar the design has to clear. Eight weeks is roughly one conventional cohort cycle at \(W_{\rm cohort}=28\) days plus its accrual, so a saving smaller than that sits inside the noise of ordinary trial execution and is worth marking as such. A programme applying this work later brings its own number. Section 26b works through how to read the map, including the case where the saving is patients rather than time.

The safety side is unquantified wherever the map is read. Section 16 sets out why no defensible quantitative model of ICANS is available, and that is a limitation of the strategy rather than of the simulation. No region of the map is a recommendation on its own. What a region says is whether the benefit there is large enough for the safety discussion to be worth having at all.

Once this stops being an exploration and starts supporting a specific programme decision, that is the point to fix a threshold in advance, so that the team is not negotiating one against results it has already seen. That is a later step and not what this document is for.

Scope

The main use case is a B-cell-depleting TCE, with profound peripheral B-cell depletion as the rapid pharmacodynamic signal.

Raising the starting dose is out of scope. A better MABEL attacks the same gap from the other end, and it is complementary rather than competing, but this project holds \(D_{\rm start}\) fixed and varies what happens above it.

This is primarily a FIH / high-dose-uncertainty strategy. It is not intended as a general later-development strategy. Once the biologically active neighborhood is known, the rationale for traversing a large dose range within individual patients diminishes substantially.

A second conceptual use case is solid-tumor TCE development. A third, useful “negative-control” archetype is another BCMA TCE, where prior class knowledge should substantially improve prediction of the active dose and therefore reduce the expected value of IPDE.

2. Central hypothesis

The value of IPDE comes from rapidly traversing a dose range that is likely to be subtherapeutic.

The key determinant is not simply \(D_{\rm active}/D_{\rm start}\) but our uncertainty about that quantity.

A useful conceptual parameter is:

\[ E_{\rm pred} = \frac{D_{\rm true,active}}{D_{\rm predicted,active}} \]

If translational prediction is accurate, conventional escalation may already reach the relevant range efficiently.

If the predicted active dose is wrong by 10–30-fold or more, IPDE can potentially provide substantial acceleration.

This makes prediction error one of the primary simulation design variables.

3. Important conceptual distinction

IPDE is a scouting strategy, not the ultimate regimen-evaluation strategy.

An IPDE patient might receive:

\[ 0.1\rightarrow0.3\rightarrow1\rightarrow3\rightarrow10\ {\rm mg} \]

and achieve profound B-cell depletion following the 10-mg administration.

That does not establish that a conventional 10-mg regimen has been adequately characterized. The response arose after a sequence of previous TCE exposures.

Therefore:

IPDE identifies the candidate dose neighborhood. A separate conventional backfill cohort tests an actual candidate regimen.

This is essential to the study design.

4. Primary and secondary endpoints

Primary endpoint

Define:

\[ \boxed{T_{3,\rm eval}} \]

as:

Calendar time from the first dose in the FIH study until 3 patients have been fully evaluated on a conventional candidate regimen terminating at the biologically active target dose.

These 3 patients must not be patients who reached the target dose through a large exploratory IPDE sequence.

For IPDE, this means:

  1. Scout patient reaches profound B-cell depletion.
  2. Candidate target dose is nominated.
  3. Backfill cohort opens.
  4. 3 new patients receive a standard candidate regimen.
  5. All 3 complete the required full safety/PD evaluation window.

For conventional dose escalation, the 3 patients in the first conventional active-dose cohort can themselves count, because they received an interpretable regimen from the outset.

This makes the comparison deliberately conservative with respect to IPDE.

Endpoint rules use observables only

The trial never observes \(D_{\rm true}\), so \(T_{3,\rm eval}\) cannot be defined as “3 patients evaluated at the biologically active dose”. The computable definition is:

\(T_{3,\rm eval}\) = calendar time until 3 patients complete the full evaluation window \(W_{\rm full}\) on a conventional candidate regimen whose target dose produced profound B-cell depletion in those same patients.

Confirmation is by observed PD in the confirming cohort, for both designs:

  • Conventional: between-patient escalation stops at the first cohort whose patients reach \(B<5/\mu L\); those 3 patients count.
  • IPDE: the backfill cohort counts only if its patients reach \(B<5/\mu L\) on the nominated regimen. A backfill cohort that does not deplete triggers the re-escalation rule in Section 7, and the clock keeps running.

The two failure directions are handled asymmetrically on purpose: a nomination that is too low costs calendar time inside \(T_{3,\rm eval}\) through the backfill regimen rule in Section 7, and a nomination that is too high does not extend \(T_{3,\rm eval}\) but is scored by the dose-selection operating characteristics in Section 5.

Secondary endpoint

\[ \boxed{T_{\rm BCD}} \]

is the time to the first observation of profound B-cell depletion, \(B<5\ {\rm cells/\mu L}\).

Call this “time to first active-dose signal” or “time to first profound B-cell depletion,” rather than claiming that the isolated dose responsible has been identified.

The difference

\[ T_{3,\rm eval}-T_{\rm BCD} \]

is the confirmation penalty after rapid discovery.

5. Operating characteristics and other outputs

Three families of operating characteristics, all computable from what the simulated trial observes plus the simulator’s knowledge of \(D_{\rm true}\) for after-the-fact scoring. Deterministic Phase 1 yields a single value for each; once enrollment stochasticity (Phase 2) and IIV (Phase 4) enter, report the median and 10th–90th percentile across simulation replicates.

Speed

  • \(T_{\rm BCD}\) and \(T_{3,\rm eval}\) (Section 4, endpoints).
  • \(\Delta T = T_{3,\rm conventional} - T_{3,\rm IPDE}\), and once stochastic, \(P(\Delta T>0)\) and \(P(\Delta T>4\ {\rm weeks})\).
  • Probability of confirming an active regimen within the administrative horizon. A scout or backfill patient who never depletes makes this less than 1 once IIV enters.

Dose selection

The trial cannot see \(D_{\rm true}\), so selection is scored after the fact by the simulator:

  • Confirmed target dose relative to \(D_{\rm true}\), in grid steps (below / at / above).
  • Number of nomination-confirmation cycles before a regimen confirms (Section 7, backfill regimen).
  • Number of patients exposed above \(D_{\rm true}\).

Together with the endpoint rule in Section 4 this answers what “right” means for the design’s operating characteristics: too low shows up as calendar time (extra confirmation cycles inside \(T_{3,\rm eval}\)), too high shows up as overshoot here and in the exposure counters below. Neither is hidden inside the endpoint definition.

Patient and operational cost

Keep these relatively simple:

  • Number of patients enrolled before first profound B-cell depletion.
  • Patient-weeks with B cells >=5 cells/uL.
  • Number of patients treated only at subactive exposures.
  • Total doses administered.
  • Number of safety-review decisions.
  • Extra visits relative to conventional treatment.
  • Hospitalization/observation days.
  • Delayed/cumulative toxicity behavior in the neutropenia stress test.

Do not initially make the following major endpoints:

  • EC50 precision.
  • Bias in the B-cell exposure-response model.
  • Dose-response estimation efficiency.
  • Probability of correctly estimating an “active dose range.”
  • Clinical efficacy.

Those are valid later questions but obscure the current decision.

6. Designs to compare

Four designs, varying two things independently: how many patients sit at each low dose level, and whether a patient may move up within their own treatment.

The four designs
Design Low-dose cohort Within-patient escalation Escalation stops when
Conventional (A) 3 patients None A cohort reaches profound B-cell depletion.
Accelerated (A2) 1 patient, reverting to 3 once activity or toxicity appears None As conventional.
Catch-up (B) 3 patients A patient already enrolled may move up to a dose the trial has since cleared As conventional.
Scout-and-backfill (C) 1 scout The scout leads the escalation on a weekly ladder That scout’s B cells fall below 5/uL. A fresh conventional cohort then confirms the nominated dose.

Three contrasts, and each isolates one mechanism:

  • A against A2 is the micro-cohort effect on its own. Fewer patients per rung, nothing else changed.
  • A against B is within-patient escalation on its own, since both keep 3-patient cohorts and the same escalation decisions.
  • A2 and B against C asks whether prospective scouting delivers more than the two mechanisms separately, and what the pharmacodynamic stop adds beyond either.

This is the shape of Simon’s comparison, where escalation scheme and intra-patient escalation are varied independently rather than bundled. Reporting only A against C would credit scouting with a saving that a 1-patient cohort produces by itself.

The letters are kept because the figures and the rest of this document use them. Conventional, accelerated and catch-up escalate the trial; only scout-and-backfill escalates a patient toward an unknown dose, and only it stops on pharmacology rather than on a cohort review.

Prior art

The catch-up and scout-and-backfill designs are not new in structure, and the document should say so before a reviewer does. Simon and colleagues published accelerated titration designs in 1997 (J Natl Cancer Inst 89:1138–1147). Their designs 2, 3 and 4 combine single-patient cohorts in the low-dose region with intrapatient dose escalation for patients who neither respond nor experience toxicity. The catch-up design is that mechanism, and the accelerated single-patient cohorts in the conventional design are the same family. What is new here is the stopping rule: Simon’s designs escalate within a patient until toxicity or lack of response over a fixed schedule, where scout-and-backfill escalates until a pharmacodynamic threshold fires and then hands off to a conventional cohort.

The paper has been read and checked; the reading notes carry the numbers and their page citations. Simon’s designs are numbered 1 to 4 and his intrapatient options lettered A and B, which are not this document’s designs A, B and C.

The finding that matters here is that intrapatient escalation moves no clock. Adding it changes neither the number of patients nor the number of cohorts, in either the standard or the accelerated design. What it buys is patients: those whose worst toxicity was grade 0-1 fall from 23.3 to 19.3 in the standard design. Time is bought by dose step size instead. No table in the paper splits the intrapatient options apart, so the size of the time effect is unpublished, and a literature search found no other paper that quantifies it.

Three consequences:

  1. The catch-up helps patients, but not time to evaluate active dose hypothesis in Section 26 is supported in direction and unquantified in size. Cite Simon for the direction; the simulation still has to produce \(\Delta T\).
  2. The efficient conventional comparator is Simon’s design 3 with option A, which is now Design A2 below rather than a parameter.
  3. The dose escalation factor in Section 13 stops being a parameter and becomes the mechanism. Simon’s accelerated ladder moves about 2-fold per double step and about 3.3-fold at the first step; the scout ladder holds roughly 3.3-fold throughout.

The prior art is not only oncology’s. Immunology finds doses by intrapatient escalation too, and in the autoimmune cytopenias it is ordinary practice. Rilzabrutinib’s phase 1–2 in immune thrombocytopenia is an adaptive dose-finding study whose escalation mechanism is within-patient, stepping every 28 days on platelet response; the same molecule’s pemphigus study gates the step on clinical response and Bruton tyrosine kinase occupancy, which is a target-engagement marker doing what B-cell count does in Section 13. A first-in-human study in Graves’ disease is registered as a within-subject dose-escalating design. The second reading queue on the references page carries these with the question each has to answer, and nothing there has been read, so this paragraph reports a search rather than a set of papers.

Those escalations run the other way from this one. They climb until the patient responds and stop because the patient is treated, where a scout climbs until the marker moves and stops because the programme has found its dose. What the search did not find is a ladder whose purpose is to nominate a dose for a separate confirming cohort, so the stopping rule and the scout-and-backfill split remain what is new here.

None of the accelerated titration literature bears on the rest. Those escalation rules stop on toxicity and seek a tolerated dose, with no pharmacodynamic stopping rule, no scout-and-backfill split, and no requirement that a separate cohort confirm an interpretable regimen, so none of it touches \(T_{3,\rm eval}\), the confirmation penalty, or the prediction-error axis. What remains to be shown is the value of the pharmacodynamic stop and the size of the confirmation penalty.

Model-based escalation designs (continual reassessment method, BOIN and variants) are out of scope. They allocate patients across dose levels more efficiently under a toxicity model, and the binding constraint here is pharmacology in a range believed to be subtherapeutic rather than toxicity statistics. Section 27b carries the argument for deferring them and what would bring them back.

Conventional escalation — Design A

Patients receive a conventional TCE regimen with a step-up dose followed by a target dose.

Base case: one step-up dose. A no-step-up comparator is unrealistic for modern TCE development, and a two-step regimen makes the conventional arm slower than it needs to be, which biases the comparison toward IPDE. Two step-up doses are the sensitivity analysis.

The generic schedule is therefore \(D1:\ SUD,\qquad D4:\ D_{\rm target}\), followed by QW target dosing if needed.

Teclistamab uses 0.06 and 0.3 mg/kg before the 1.5 mg/kg target dose, approximately 4% and 20% of the target. With a single step-up, take the higher of those fractions, \(SUD = 0.20D_{\rm target}\). The two-step sensitivity uses \(SUD1 = 0.04D_{\rm target}\) and \(SUD2 = 0.20D_{\rm target}\) on D1 and D4, with the target on D8.

Do not treat those fractions as biologically universal; they are simply a reasonable reference regimen.

After adequate safety review of dose level \(D_k\), a new cohort receives \(D_{k+1}\).

Accelerated escalation — Design A2

One patient per dose level in the low-dose region, reverting to 3-patient cohorts once activity or toxicity appears. Everything else matches the conventional design: the same step-up regimen, the same between-cohort review, no within-patient escalation.

This exists as a design rather than as a sensitivity analysis because it is the honest comparator. A conventional design requiring 3 patients at every extremely low dose is a straw man, published TCE experience includes accelerated early cohorts with small cohort sizes across very large dose ranges, and scout-and-backfill should beat a reasonably efficient conventional design rather than an intentionally inefficient one. Making it a design also means the micro-cohort saving is reported on its own instead of arriving inside the scouting result.

Simon’s design 3 with option A is the specific form to use: accelerated single-patient cohorts, double dose steps, no intrapatient escalation, reverting to the standard design at the first first-course DLT or the second first-course grade 2 event.

Two parameters:

  • \(n_{\rm low}\) — patients per cohort in the low-dose region. 1 here, 3 in the conventional design.
  • The rule that ends the accelerated phase. Simon’s is a toxicity event. A pharmacodynamic version, ending the accelerated phase at the first sign of B-cell movement, is worth running as a variant, since it is the same signal scout-and-backfill stops on and it separates the trigger from the escalation mechanism.

Catch-up escalation — Design B

Between-patient escalation and opening of new dose levels are identical to the conventional design.

However, patients enrolled at earlier low target doses can subsequently increase their own target dose once the higher dose has been cleared by the trial.

Example. A patient enrolled in the 1 mg cohort receives the step-up dose and then the 1 mg target:

\[ 0.3\rightarrow1 \]

and under the conventional design would continue at 1 mg. Under catch-up the same patient moves up as each higher level is cleared:

\[ 0.3\rightarrow1\rightarrow3\rightarrow10. \]

Each arrow is one weekly administration. The higher levels are opened by the trial’s between-patient escalation, not by this patient, so catch-up changes which patients sit at a dose level and not when that level opens.

Expected result:

  • Little or no improvement in \(T_{3,\rm eval}\).
  • Substantial reduction in patient-time at clearly subactive doses.

This distinguishes patient benefit from program acceleration.

Scout-and-backfill — Design C

This is the main experimental design and the one the project exists to evaluate.

One or more scout patients undergo prospective dose escalation.

After initial priming:

\[ D_0\rightarrow D_1\rightarrow D_2\rightarrow D_3\rightarrow... \]

with escalation every 7 days. Hold the step interval at one week for v1. It matches the safety gate \(W_{\rm decision}\) in Section 15 and the PD turnaround in Section 11, so nothing else in the engine has to move to support it. Shorter intervals are a later question, not a v1 parameter.

Continue escalation when \(B\ge5/\mu L\) and the protocol-defined safety gate permits another dose.

Stop escalation when \(B<5/\mu L\).

Then nominate the corresponding target-dose neighborhood and immediately open conventional backfill.

No arbitrary maximum number of IPDE steps

There should be no cap such as “maximum three escalations.”

If the patient remains pharmacologically underdosed and meets safety criteria, continuing escalation is precisely the hypothesis being tested.

However, distinguish:

unbounded number of escalation steps

from

literally unlimited dose.

A real FIH protocol will always have a maximum permitted dose/exposure based on nonclinical safety, manufacturing, protocol amendment, etc. The simulation needs a technical upper dose/horizon as well so an inactive virtual molecule cannot escalate forever.

That upper bound should be treated as administrative censoring / protocol feasibility, not as the scientific stopping rule.

Published escalation schedules provide conceptual precedent: some cohorts and patients traverse many successive doses during Cycle 1 rather than being limited to an arbitrary number of within-patient increases. Verify a specific citation before relying on this point.

A published within-patient escalation ladder — MCLA-117

The MCLA-117 (tepoditamab) first-in-human study in AML escalated within patients from a 25 µg first dose to a 400 mg target dose, a factor of 16,000, with cohort 1 taking seven doses in 22 days. A long within-patient ladder is precedented and does not have to be argued for here.

Figure 1: MCLA-117 Cycle 1 dose escalation. Each line is one cohort’s within-patient ladder; the dose axis is logarithmic and spans the 25 µg first dose of cohort 1 to the 400 mg target dose of cohort 12.

Three features of the scheme bear on this design, and the MCLA-117 notes carry the dose tables and the sourcing caveat.

  • The number of within-patient steps was not capped, which is what the subsection above assumes.
  • The first dose moved slowly between cohorts, by 4-fold at most and mostly 3-fold or less, while the target dose rose 600-fold over the same sequence. The constraint sits on the first dose in a drug-naive patient, not on the ladder above it. The next subsection takes this up.
  • From cohort 9 the step-up prefix is fixed at 5, 15 and 25 mg and only the target changes, which is the structure Section 7 assumes for the backfill regimen.

The step interval is where the precedent and this design differ. Cohort 1 escalates at 2 to 3-day intervals; scout-and-backfill holds one week, so the simulated ladder is slower than the precedent and any acceleration it reports is conservative on that axis.

The scout’s first dose is constrained, and a second scout re-runs the ladder

Scout-and-backfill as specified lets a scout begin at \(D_{\rm start}\) and climb without limit. Two constraints break that, and they interact.

The first dose in a drug-naive patient escalates slowly between cohorts. It is governed by the same nonclinical safety argument that sets the FIH starting dose, and it moves only by a factor a safety committee will accept: 3-fold typically and 4-fold at most in the scheme above. A second scout therefore cannot begin where the first scout stopped.

One scout is one patient. If the scout depletes at \(D^*\), the nomination in Section 7 rests on a single observation of a single patient’s B-cell trajectory. A programme that will open a backfill cohort on that basis may reasonably want a second scout to reproduce it.

Together these cost calendar time only if the scouts run one after the other. If a second scout must start near the bottom and does so after the first has finished, the second ladder repeats most of the first and confirmation takes nearly as long as the original discovery. Running the scouts concurrently removes that cost, and the next subsection sets out how.

Parameterize this explicitly:

  • \(F_{\rm start}\) — the maximum permitted between-cohort multiplier on the first dose in a drug-naive patient. Base 3; sensitivities 2 and 10, where 10 stands for effectively unconstrained and reproduces the current specification.
  • \(n_{\rm scout}\) — the number of scout patients. Base 1; sensitivities 2 and 3.
  • Scout entry policy, one of: a fixed low first dose for every scout; the highest dose already cleared in a previous scout; or the previous scout’s first dose times \(F_{\rm start}\).

Rolling scout entry

There is no reason for a second scout to wait for the first to finish. Once the leading scout has taken a dose and cleared it, that rung is cleared for anyone, so a following scout can start and climb behind. Scouts enter on a rolling basis, each one one or more rungs behind the one ahead, and the ladder carries several patients at once.

Two things constrain how tight the stagger can be, and they are what “as soon as it clears” has to mean in a protocol.

  1. A follower’s next dose waits on the leader’s result, not on the leader’s dose. The leader is dosed at rung \(k\), and the PD result and safety review for that rung arrive \(T_{\rm PD,ready}\) plus the review lag later (Section 11). A follower must not be dosed at rung \(k\) before that arrives, or the trial has put two patients at an uncleared dose and gained nothing over dosing them together. The minimum stagger is therefore the turnaround, not the escalation interval.
  2. Sentinel rules constrain the first dose in each patient. A protocol that staggers first doses between patients applies to scouts too, and that gap is usually longer than the PD turnaround.

Two rules complete it. A follower never enters a rung the leader has not cleared. And when the leading scout triggers the stop, every following scout climbs to that dose and stops there, so each becomes an observation at the nominated dose, which the declaration in Section 4 needs anyway.

What rolling entry costs is patients inside the latency window. Section 15 already establishes that a scout escalates on acute safety before delayed toxicity from earlier rungs is observable. One scout carries that risk alone. Three scouts climbing a rung apart carry it concurrently, so a delayed signal that appears at rung \(k\) arrives when three patients have been dosed past it rather than one. That is the real price of the concurrency and it is not visible in \(\Delta T\). Count it: report the number of patients dosed above a rung at the moment that rung’s delayed-toxicity information becomes available, and carry it alongside the burden rows in Section 19.

Parameterize the stagger as \(\Delta_{\rm scout}\), the gap between successive scout entries. Base one rung; sensitivities two rungs and the serial case, where the next scout starts only after the previous one stops. The serial case is the comparator that says what the concurrency bought.

What real protocols do here is the open question, and it is the same question the GB261 entry on the references page asks: what ended the accelerated phase, and whether the patient who triggered it stopped alone or everyone stopped. Answer it from a protocol before fixing the rule.

One patient stops the ladder; three declare the dose

The first scout to reach \(B<5/\mu L\) ends the accelerated phase. Not two of two, not a confirmed pattern. Act on the first value below the threshold and confirm afterwards for the record; a repeat sample before acting only delays the stop.

The asymmetry against the three patients required to declare the dose in Section 4 is deliberate, and someone will challenge it, so state the reason. The two errors point in opposite directions. One patient’s depletion is a noisy observation, and an atypical subject can cross the threshold at a dose the population would not, but that error ends the accelerated phase early and costs only speed. Requiring two or three would keep scouts climbing past a dose already shown to be active in someone, which is the error that cannot be taken back. Declaring \(D_{\rm true}\) on one patient would be the opposite mistake, an overclaim, which is why the declaration keeps its three. One to stop, three to declare.

Simon accepts the same asymmetry: the accelerated phase there ends on the first qualifying event in any patient rather than on a confirmed pattern. Reusing an escalation-slowing trigger that a protocol already carries also makes the ethics discussion a question of precedent rather than a fresh argument.

The confirmatory-sample question in Section 13 is a different rule and keeps its answer. That one governs \(\theta_{\rm move}\), the threshold that slows the ladder, where a false positive costs acceleration across the remaining rungs. This one governs the stop, where a false positive costs one rung.

Where two scouts do deplete at different dose levels, the nomination rule in Section 7 assumes one \(D^*\). Nominating the lower is the conservative choice and costs the non-confirmation cycles of Section 7 when it is too low.

7. Candidate backfill regimen after IPDE

This needs to be defined explicitly.

Preferred base implementation:

If scout depletion first occurs following target \(D^*\), nominate the base-case conventional regimen \(0.20D^*\rightarrow D^*\) on approximately D1/D4, matching the step-up structure of the conventional arm so that the two are comparable. Under the two-step sensitivity, nominate \(0.04D^*\rightarrow0.20D^*\rightarrow D^*\) on D1/D4/D8 instead.

Then enroll 3 fresh backfill patients on that regimen.

This is preferable to claiming that the scout’s complete escalation sequence is itself the candidate regimen.

Sensitivity analysis:

Use the immediately preceding IPDE dose as the step-up dose of the backfill regimen, or the preceding two under the two-step sensitivity.

For example, if scout history was:

\[ 1\rightarrow3\rightarrow10\rightarrow30 \]

and 30 mg produces depletion, test \(3\rightarrow10\rightarrow30\) as the conventional backfill regimen.

Both approaches are reasonable. Start with the fixed-ratio approach because it makes comparisons cleaner.

Expect the nomination to sit low

The scout’s stopping signal arises from the cumulative exposure and the cumulative B-cell trajectory of the whole ladder, not from \(D^*\) in isolation. Drug from earlier doses is still present when the signal fires: the retained PK gives an effective half-life of roughly 8 days (\(\ln 2 \times V_{ss}/CL\)) against a 7-day escalation interval. B-cell recovery (\(t_{1/2}\approx30\) days) is slower still, so partial depletion carries forward from step to step. The nominated \(D^*\) can therefore sit one or more grid steps below the dose whose conventional regimen depletes a fresh patient. The PD-guided escalation rule in Section 13 is the main lever against this, since a scout that decelerates once the marker moves stops nearer to the dose responsible for the depletion. That is the dynamics working as specified, and it means backfill non-confirmation is an expected event the protocol must handle, not an edge case.

Backfill confirmation and non-confirmation rules

  • Confirmation. The backfill regimen confirms when its patients reach \(B<5/\mu L\) at the prespecified PD assessment time. In the deterministic phases this is all-or-none in the reference patient; an \(x/3\) rule (e.g. at least 2 of 3) is deferred to Phase 4 with IIV.
  • Non-confirmation. Escalate the candidate target dose one grid step and enroll a new 3-patient conventional cohort at that level. Repeat until a cohort confirms or the administrative dose cap is reached. After a failed backfill the IPDE arm is therefore running conventional escalation that starts at \(D^*\) instead of \(D_{\rm start}\), and the calendar time this consumes stays inside \(T_{3,\rm eval}\).
  • Patients in a failed backfill cohort may step up to the next candidate dose under the catch-up rules for their own benefit, but they no longer count toward \(T_{3,\rm eval}\).

8. PK model

Use published teclistamab PK as the reference exposure model.

Published teclistamab PopPK uses:

  • SC first-order absorption.
  • 2-compartment disposition.
  • Parallel time-independent and time-dependent clearance.

Published typical estimates:

Published teclistamab population-PK typical estimates
Parameter Value
\(CL_1\) 0.449 L/day
\(V_c\) 4.13 L
\(Q\) 0.0390 L/day
\(V_p\) 1.34 L
\(k_a\) 0.133 day\(^{-1}\)
\(F\) 0.718
time-dependent \(CL_2\) 0.547 L/day
\(k_{\rm DES}\) 0.0292 day\(^{-1}\)

For this project:

\[ \boxed{\text{omit time-dependent }CL_2} \]

and retain \(CL=0.449\ {\rm L/day}\).

Rationale: time-varying clearance is real for teclistamab but is not central to the trial-design question and unnecessarily complicates the simulation.

The resulting model is therefore:

A published-teclistamab-structure PK model simplified to time-independent disposition.

Do not claim that it reproduces the complete approved teclistamab PK model.

The retained parameterization provides exposure persistence on approximately the weekly scale, which is the behavior required for this exercise.

Use a 74-kg reference patient initially because that was approximately the reference body weight in the published model. Covariates are unnecessary in v1.

PK IIV can initially be omitted or reduced to CL/Ka variability. The principal design conclusions should not depend on reproducing all teclistamab covariate relationships.

9. B-cell pharmacodynamic model

Use a simple indirect-response / turnover model with drug-stimulated B-cell loss.

One convenient form is:

\[ \frac{dB}{dt} = k_{\rm in} - k_{\rm out}B - k_{\rm kill}(C)B \]

where

\[ k_{\rm kill}(C) = k_{\max} \frac{C^h}{EC_{50}^h+C^h}. \]

At baseline \(k_{\rm in}=k_{\rm out}B_0\).

This is intentionally much simpler than published TCE QSP models. Mechanistic mosunetuzumab models describe T-cell activation and B-cell killing across blood and tissues, demonstrating precedent for using B-cell depletion as a TCE PD output, but that level of complexity is unnecessary here.

Suggested starting values are \(B_0=100\ {\rm cells/\mu L}\) and \(t_{1/2,B}\approx30\ {\rm days}\), so \(k_{\rm out}=\ln 2/30\).

Use approximately

\[ k_{\max}=1\ {\rm day}^{-1} \]

as an initial value: at saturating drug effect, this gives roughly 95% depletion over approximately 3 days absent replenishment.

Use \(h=1-2\).

These are simulation assumptions, not claimed clinical estimates.

10. Define the “true active dose” through calibration

Do not choose an arbitrary EC50 and then discover what dose happens to be active.

Instead, define the desired scenario in terms of a true active target dose.

For each scenario choose \(D_{\rm true,active}\).

Then numerically solve for \(EC_{50}\) such that the standard candidate regimen ending at \(D_{\rm true,active}\), which in the base case is:

\[ 0.20D_{\rm true} \rightarrow D_{\rm true} \]

produces \(B=2.5/\mu L\) at the prespecified PD assessment time in a typical patient, and verify that the same regimen ending one grid step lower leaves \(B>10/\mu L\). Calibrating to exactly \(5/\mu L\) would place the defining scenario on the decision boundary, where solver tolerance and assessment timing can flip the outcome by a full grid step; the margin keeps the grid separation clean while preserving the definition below.

This makes the parameterization transparent and makes prediction error easy to manipulate.

The active dose should therefore mean:

The first target-dose level on the protocol dose grid whose standard step-up regimen produces profound peripheral B-cell depletion in the reference patient.

It does not mean that a single isolated administration of that dose necessarily does so.

11. PD sampling and decision latency

Separate biology from operational availability.

B cells may fall rapidly, but the decision requires:

sample -> assay -> QC -> review -> dose decision.

Use \(T_{\rm PD,ready}=3\ {\rm days}\) as a reasonable initial scenario.

Sensitivity analyses \(2,\ 3,\ 5,\ 7\ {\rm days}\).

If the actual program has a known operational turnaround, replace these with real values.

The next weekly IPDE dose can only occur if the PD result is available.

If \(B<5/\mu L\) before the next scheduled dose, IPDE stops.

12. Prediction-error framework

This should be a major simulation axis.

Separate three quantities:

  • \(D_{\rm start}\) — the FIH starting target dose.
  • \(D_{\rm pred}\) — the translationally predicted biologically active target dose.
  • \(D_{\rm true}\) — the actual active target dose in the simulation.

Define:

\[ R_{\rm predicted-gap} = \frac{D_{\rm pred}}{D_{\rm start}} \]

and

\[ E_{\rm pred} = \frac{D_{\rm true}}{D_{\rm pred}}. \]

Therefore:

\[ \frac{D_{\rm true}}{D_{\rm start}} = R_{\rm predicted-gap} \times E_{\rm pred}. \]

Suggested values:

\[ R_{\rm predicted-gap}=10,\ 100,\ 1000 \]

and

\[ E_{\rm pred}=1/10,\ 1/3,\ 1,\ 3,\ 10,\ 30. \]

That separates:

  • How conservative the starting dose is.
  • How accurately the pharmacology was predicted.

This is particularly useful for comparing novel targets versus another BCMA TCE.

13. Dose escalation factor

Set the step size from the PD signal rather than fixing it. The scout is already measured before every escalation decision (Section 11), and the measurement says which regime the ladder is in. While the B-cell count has not moved, the scout is far below the active dose and a large step costs nothing but crosses more of the gap. Once the count starts to move, the active dose is close and a large step overshoots it.

The rule operates on \(B/B_0\), the most recent available B-cell count as a fraction of that scout’s own baseline:

The PD-guided escalation rule
PD state Condition Next step
No movement \(B \ge \theta_{\rm move} B_0\) \(F_{\rm quiet}\), base 10-fold
Movement \(\theta_{\rm move} B_0 > B \ge 5/\mu L\) \(F_{\rm move}\), base 3-fold
Profound depletion \(B < 5/\mu L\) Stop; nominate (Section 7)

with \(\theta_{\rm move}=0.70\), a 30% decline from baseline, as the base case.

The rule is one-way in practice. B cells do not recover while dosing continues, so the ladder decelerates once and does not re-accelerate, and no oscillation between the two step sizes is possible.

Three parameters, and a comparator:

  • \(\theta_{\rm move}\) — the fraction of baseline at which the marker counts as having moved. Base 0.70; sensitivities 0.50 and 0.85.
  • \(F_{\rm quiet}\) — the step while the marker has not moved. Base 10; sensitivities 3 and 30.
  • \(F_{\rm move}\) — the step after it has. Base 3; sensitivity 2.
  • Fixed multiplier, \(D_{k+1}=FD_k\) with \(F\) = 2, 3 or 4 throughout, as the reference arm. Every claim for the PD-guided rule is a difference against this.

What the rule can get wrong

Both failure directions are simulatable and should be reported separately.

A false brake is a measurement that crosses \(\theta_{\rm move}\) through assay and biological variability rather than through drug effect. It drops the ladder into 3-fold steps early, and the scout then crawls through a range it could have crossed quickly, which is a direct charge against \(\Delta T\). A 30% decline in an absolute B-cell count is not obviously outside the day-to-day variability of the measurement, so \(\theta_{\rm move}\) may need a confirmatory second sample rather than a single crossing. This cannot appear before Phase 4, since Phases 1 to 3 have no variability to generate it.

A missed brake is the opposite: the marker has moved but the sample that would have shown it is not yet available, so a 10-fold step lands past the active dose. The nominated \(D^*\) is then too high, and Section 5 scores it as overshoot rather than as lost time.

Report the number of steps spent in each regime. The interval between \(\theta_{\rm move} B_0\) and \(5/\mu L\) spans a 14-fold change in B cells at \(B_0=100\) cells/uL, and how many 3-fold dose steps that interval takes under the B-cell PD model in Section 9 is what decides whether the brake costs calendar time or merely places the nomination better. That is a simulation output, not something to assert here.

What this changes elsewhere

The brake should raise the nominated \(D^*\) relative to the fixed-ladder case, because the scout approaches the depletion threshold in smaller increments and stops nearer to the dose that actually caused it. If so, the undershoot described in Section 7 shrinks and the non-confirmation cycles with it. That is one of the more valuable things this rule could buy, and it is worth reporting separately from the change in \(T_{3,\rm eval}\).

It also separates this design further from the accelerated titration designs above. Their step size changes on toxicity; this one changes on pharmacology, using a marker measured for the purpose.

Precedent for the step sizes

The MCLA-117 scheme in Section 6 gives within-patient increments of 1.5 to 3-fold in its early cohorts: cohort 1 runs 25, 50, 100, 200, 300, 450 and 675 µg, decelerating from 2-fold to 1.5-fold as it climbs. Larger within-patient increments do appear later, with cohorts 10, 11 and 12 stepping from 25 mg on D8 to 120, 240 and 400 mg on D15, a 4.8 to 16-fold increase. Those larger steps go into doses already cleared in earlier cohorts, so they are not blind escalation. A 10-fold blind step in a scout has no precedent in that scheme, and \(F_{\rm quiet}=10\) should be presented as an extrapolation rather than as established practice.

14. Acute safety / CRS

Do not make CRS the main modeling problem.

Published quantitative models exist for TCE CRS, including repeated time-to-event approaches in which CRS hazard depends on longitudinal exposure and a separate inhibitory/tolerance component representing priming.

This validates the assumption that:

Acute TCE toxicity can depend strongly on both current exposure and prior treatment history.

However, importing an epcoritamab CRS model into a teclistamab-PK/B-cell simulation would create a hybrid molecule with questionable quantitative meaning.

Therefore use one of two approaches.

v1 preferred

Treat acute safety as a protocol gate rather than an important simulated endpoint:

  • Next IPDE dose cannot occur until the acute safety assessment is complete.
  • Assume acute clinically important toxicity is largely observable within approximately 48 hours for the base-case operational discussion.
  • No unresolved acute safety event is permitted.
  • Assume the gate is passed in the base efficiency simulations.

Then perform simple sensitivity analyses introducing acute treatment holds.

Optional later extension

Implement a generic exposure-dependent binary/RTTE acute toxicity model calibrated to plausible CRS rates.

Do not spend much time optimizing this before the speed/PD simulation works.

The ladder is itself a step-up schedule

The document elsewhere treats within-patient escalation only as a risk to be gated. For CRS specifically there is an argument in the other direction, and it should be stated rather than left for a reviewer to raise.

An IPDE patient arrives at each dose level having already received every lower level on the ladder. A conventional patient arrives at the same target dose after a single step-up dose in the base case. On the tolerance/priming component of the published TCE CRS models referenced above, where prior exposure reduces the hazard at a given current exposure, the IPDE patient is the better-primed of the two at that dose. The IPDE ladder is a longer step-up schedule, not an unprimed jump.

Two limits on how far this argument goes:

  • It applies to CRS, where priming is an established and quantified mechanism. It does not transfer to ICANS (Section 16), where the exposure-response relationship is not established well enough to claim priming protects anyone.
  • The IPDE patient also reaches higher absolute doses than any conventional patient of the same era in the trial, so better priming at a given dose does not mean lower total risk.

The net effect on CRS is therefore ambiguous, and that is the honest statement. Record it as a reason not to assume the acute-safety comparison runs against IPDE, and revisit it if the optional RTTE extension is built.

15. Critical point about safety-decision windows

This comparison must be fair.

Do not allow IPDE to escalate after 48 hours while making conventional between-patient escalation wait 28 days unless that genuinely represents the protocols being compared.

If conventional escalation also uses a relatively short acute safety window before opening the next dose level, then both designs can outrun delayed toxicity information at the program level.

The IPDE-specific concern is somewhat different:

The same individual accumulates sequential higher exposures before all delayed consequences of earlier exposures may be known.

Therefore specify the two windows separately:

  • \(W_{\rm decision}\) — the minimum information required to permit the next escalation.
  • \(W_{\rm full}\) — the complete safety evaluation required for a patient to count toward the primary endpoint.

IPDE escalation gate

Base case: \(W_{\rm decision}=7\) days. The next intra-patient escalation uses all safety data accrued over the full 7-day interval since the previous dose, plus the PD result; the ~48-hour acute assessment sits inside that window rather than being the whole gate. Weekly within-patient increases are precedented: approved TCE step-up schedules escalate within patients at 2–4-day intervals (teclistamab doses on D1/D4/D8), so a 7-day step is a slower schedule than those products already use.

Sensitivity: \(W_{\rm decision}=14\) days (escalation every other week), representing a more conservative safety-review committee.

Conventional between-cohort review window

Pin this explicitly; it is the largest single driver of the conventional timeline and therefore of the whole comparison.

\(W_{\rm cohort}\) = time from the last patient in a cohort receiving the target dose until the next dose level may open.

  • 3-patient cohorts: base 28 days (a standard DLT-window review); sensitivities 14 and 21 days.
  • 1-patient accelerated cohorts in the low-dose region: base 7 days; sensitivity 14 days. The base case deliberately makes the accelerated comparator fast, so IPDE is tested against an efficient conventional design rather than a straw man.

Full evaluability window

For full evaluability, begin with:

\[ W_{\rm full}=28\ {\rm days} \]

and sensitivity analyses \(W_{\rm full}=14,\ 21,\ 28\ {\rm days}\).

Only one primary clock, \(T_{3,\rm eval}\), is needed. There is no need for separate \(T_{48}\) and \(T_{\rm full}\) endpoints.

16. ICANS: critical limitation, not a fabricated model

This is the main unresolved safety issue.

The literature review did not identify an established quantitative TCE exposure-to-ICANS model analogous to either the Friberg neutropenia model or existing CRS models.

For teclistamab, published clinical experience indicates:

  • ICANS occurs in a minority of patients.
  • Median onset is several days after the most recent dose.
  • Events are not reliably confined to the first 48 hours.
  • First occurrence can happen after later doses.
  • Recurrent ICANS can occur.

A key pharmacometric limitation is that the relationship between TCE exposure and neurological toxicity is not sufficiently established for a credible general-purpose quantitative model.

That matters greatly for QW IPDE.

Do not model ICANS as \(P(ICANS\mid D_k)\) followed by a dose-specific onset clock.

With repeated weekly exposure \(D_1\rightarrow D_2\rightarrow ICANS\) there is generally no credible way to identify whether the event would have occurred without \(D_2\), was triggered by \(D_2\), reflects \(D_1\), or reflects the cumulative immune trajectory.

Likewise, an arbitrary model such as \(h_{\rm ICANS}=f(C_{\max},AUC)\) would appear more rigorous than the available evidence supports.

Recommendation for v1

Do not quantitatively model ICANS.

State explicitly:

IPDE may outrun delayed neurotoxicity such as ICANS, and no sufficiently established quantitative exposure-ICANS model currently exists to reliably simulate that risk.

This is a limitation of the proposed development strategy, not merely of the simulation.

The design mitigates—but does not eliminate—the uncertainty by:

  • Stopping escalation as soon as profound PD occurs.
  • Not seeking an MTD.
  • Transitioning immediately to conventional backfill.
  • Using appropriate ongoing safety review.

17. Neutropenia model

Include this because it can illustrate the general phenomenon of a delayed/cumulative toxicity process.

Use the established Friberg semimechanistic model:

\[ Prol \rightarrow Transit_1 \rightarrow Transit_2 \rightarrow Transit_3 \rightarrow ANC. \]

Equations:

\[ k_{tr}=\frac{4}{MTT} \]

\[ \frac{dProl}{dt} = k_{tr}Prol(1-E_{\rm drug}) \left(\frac{ANC_0}{ANC}\right)^\gamma - k_{tr}Prol \]

\[ \frac{dTransit_1}{dt} = k_{tr}(Prol-Transit_1) \]

with analogous equations for subsequent transit compartments, and:

\[ \frac{dANC}{dt} = k_{tr}(Transit_3-ANC). \]

The classic Friberg model uses a proliferating compartment, three maturation compartments and circulating cells; it is a standard semimechanistic framework for delayed myelosuppression.

Reasonable generic system values include approximately:

\[ MTT\approx116-125\ {\rm h} \]

and:

\[ \gamma\approx0.17. \]

Use a concentration-driven effect \(E_{\rm drug}=Slope\times C\) or an Emax formulation if numerical behavior is preferable.

Important interpretation

The TCE-specific SLOPE is not known.

Furthermore, published teclistamab exposure-response analyses have not established a simple clear exposure relationship for grade >=3 neutropenia.

Therefore:

The Friberg simulation is a generic quantitative stress test of delayed/cumulative toxicity, not a validated teclistamab neutropenia prediction.

Its purpose is to illustrate this principle:

\[ \boxed{ \text{toxicity latency} > \text{escalation interval} } \]

means that clinically normal observations at the next dose decision can coexist with a latent toxicity process already developing.

Do not imply that neutropenia is a surrogate for ICANS.

18. Why stopping at B-cell depletion matters

The proposed design is not:

Escalate until toxicity.

It is:

Escalate until sufficient pharmacology OR toxicity.

Thus:

\[ B\ge5/\mu L \quad\&\quad Safety\ acceptable \]

leads to another dose.

Whereas \(B<5/\mu L\) immediately terminates exploratory escalation.

This is the principal protection against unnecessarily traversing far into the biologically active range.

It also makes the scientific purpose very different from within-patient MTD finding.

19. Operational model

Do not initially create a weighted “operational complexity score.”

Report the components separately.

For every simulated design track:

Patient and operational burden components, reported per simulated design track
Element Metric
Additional clinic visits count
Drug administrations count
Individual dose changes count
Pharmacy preparations count
Real-time dose decisions count
Safety committee reviews count
PD results needed before next dose count
Observation/hospitalization patient-hours
Patients dosed above a rung when its delayed-toxicity information arrives count
Total trial calendar time days

Report every row twice, per patient and across the trial. Serial scouts concentrate a heavy burden on very few patients; conventional escalation spreads a lighter one across three patients per rung. The two totals can point in opposite directions, and which one a reader wants depends on whether they are asking what a participant undergoes or what the trial consumes (Section 1).

Record observation and hospitalization in patient-hours, not patient-days. Counting in days hides the case this design creates most of, a short post-dose observation window repeated at every rung of the ladder. Eight hours after each of six escalations and one overnight admission are different things that round to the same number of patient-days. Aggregate to days for reporting where that reads better.

Two of these rows are not patient burden and should not be read as such. Individual dose changes and real-time dose decisions are counts of occasions on which a dose is calculated, prescribed, prepared and verified for one patient rather than issued once for a whole cohort. They are a proxy for protocol complexity and for the number of opportunities for a dosing error, neither of which the simulation models. Report the counts; do not convert them into a patient-facing cost.

The key quantity is incremental burden relative to what patients would otherwise undergo.

For a QW oncology TCE:

  • Weekly visit already planned.
  • Labs already planned.
  • Observation already planned.
  • Changing the dose may add little patient-facing burden.

For an immune-reset program whose eventual treatment might involve only 1–2 administrations:

  • Serial QW IPDE could add multiple visits and doses.
  • Operational and patient burden can therefore be substantial.

Safety latency is not operational burden

Keep this separate.

There should be two axes:

  1. Incremental operational burden.
  2. Latency/uncertainty of clinically relevant safety information.

A study can be operationally easy but scientifically risky because delayed toxicity information is unavailable.

20. Enrollment and trial-operation parameters

At minimum parameterize \(\lambda_{\rm enroll}\) with scenarios such as:

  • 0.5 patients/week.
  • 1 patient/week.
  • 2 patients/week.
  • Accrual not binding: 5 patients/week, or an unlimited queue.

Include:

  • Screening delay.
  • Patient staggering/sentinel rules.
  • SMC/review lag.
  • PD-result turnaround.
  • Cohort-opening delay.

These can dominate calendar time and therefore should not be omitted.

A simple discrete-event engine is preferable to pretending all patients appear instantaneously.

Which resource is binding

Calendar time is set by whichever is slower, waiting for patients or waiting for the safety review between cohorts. Which one binds changes what every other lever buys, and it differs by archetype.

Where accrual is slow, patients and calendar time move together. A 3-patient cohort takes three enrolment slots to fill and the trial waits, so cutting to 1-patient cohorts saves both.

Where accrual is fast, which is the expectation for the autoimmune and immune-reset archetype in Section 22, that inverts. Cohorts fill quickly and calendar time becomes the number of cohorts multiplied by \(W_{\rm cohort}\). A 1-patient cohort then saves patients at subactive doses and saves little time, because the review window runs regardless of how many patients sat in the cohort.

Two things follow, both predictions to check rather than assumptions to build in.

  • The 1-versus-3 cohort size lever in Section 6 reports on the patient axis and not the time axis once accrual is fast. That is the same shape as the catch-up result.
  • What moves the clock under fast accrual is the number of cohorts, which is set by the dose step size. This is Simon’s finding in Section 6, that time is bought by step size rather than by escalating within a patient, arriving from the other direction. It also suggests where scout-and-backfill earns its keep: not by removing waiting, but because seeing PD in the same patient is what makes a larger step defensible. On that reading the PD-guided rule in Section 13 is the time mechanism and the ladder is what licenses it. The simulation should confirm or refute this, since it is the difference between a design that saves time and one that only saves patients.

Backfill pre-screening

Make this an explicit design lever, because it plausibly matters more than PD turnaround.

After the scout depletes, IPDE must enroll 3 fresh backfill patients. At \(\lambda_{\rm enroll}=0.5\)/week that is roughly 6 weeks of accrual before the \(W_{\rm full}\) clocks even start, while the conventional arm has been enrolling continuously throughout. The confirmation penalty \(T_{3,\rm eval}-T_{\rm BCD}\) is therefore an accrual quantity at low enrollment rates, and PD turnaround (Section 11), measured in days, cannot compete with it.

Simulate two policies:

  • Reactive. Backfill screening begins when the scout depletes. This is the base case.
  • Anticipatory. A pool of \(n\) patients is screened and held ready during the scout escalation, so backfill dosing starts after the review lag alone. Screening is speculative and some pool patients are never dosed, so count them in the operational burden of Section 19.

The comparison bounds how much of the confirmation penalty is recoverable by operations rather than by pharmacology. If anticipatory pre-screening removes most of it, that is a cheaper program change than faster assays and should be reported as such.

21. Important design variants / sensitivity analyses

The highest-value sensitivity analyses are:

  1. Prediction error \(D_{\rm true}/D_{\rm pred}\).
  2. Predicted starting-to-active gap.
  3. PD-guided escalation against a fixed 2x, 3x or 4x multiplier, and the movement threshold \(\theta_{\rm move}\) = 0.50, 0.70, 0.85 (Section 13, dose escalation factor).
  4. One (base) vs two standard step-up doses.
  5. The rule that ends the accelerated phase in Design A2: toxicity, as in Simon, or first B-cell movement.
  6. PD turnaround 2–7 days.
  7. Enrollment rate.
  8. Safety decision window.
  9. Full evaluability window.
  10. One vs multiple simultaneous/staggered IPDE scout patients.
  11. Delayed neutropenia strength/latency.
  12. IPDE escalation gate 7 vs 14 days.
  13. Backfill nomination offset (nominate \(D^*\) vs \(3D^*\)) and the cost of non-confirmation cycles.
  14. Reactive vs anticipatory backfill pre-screening (Section 20).
  15. Cap on the between-cohort increase in the first dose given to a drug-naive patient, \(F_{\rm start}\) = 2, 3 or 10 (Section 6).
  16. One vs two vs three scouts, the scout entry policy, and the scout stagger \(\Delta_{\rm scout}\) = one rung, two rungs or serial (Section 6).

Avoid large factorial designs initially. Start with interpretable scenario sets.

22. Suggested archetypes

Patient population

State the population for each archetype. It decides whether \(B_0=100\) cells/uL and the \(5\) cells/uL threshold are sensible, and the two candidate populations differ enough to change the PD model’s starting conditions.

  • Relapsed/refractory oncology, the teclistamab-like setting. Baseline B cells are disease-affected and prior therapy is common.
  • Autoimmune / immune reset, the setting where the eventual regimen may be 1–2 administrations. Some candidates have had prior B-cell-depleting therapy such as rituximab and can enter with baseline counts already low or still recovering. Accrual is expected to be fast here rather than rate-limiting, which changes what cohort size buys (Section 20).

The second population breaks the design’s central signal where a patient enrolls near or below the threshold, because a scout who starts at \(B<5/\mu L\) has no escalation signal at all and a scout who starts low depletes early and nominates a dose far below \(D_{\rm true}\). Handle it as an eligibility criterion rather than as a modeling problem: require a documented baseline B-cell count within the normal range for scout eligibility, and record how many screened patients that excludes.

Pin the floor as a number, not as “the normal range.” Reference intervals for CD19+ counts vary between laboratories, and a multi-site trial reading a site-specific range gives a stopping trigger whose meaning shifts by site. Choose one value and write it into the protocol.

Whether the floor also applies to the three patients declaring the dose in Section 4 is open. They need it for the same reason the scout does, since a patient entering at 30 cells/uL reaching \(B<1\) is a weaker observation than one entering at 250. Applying it to everyone contributing to the depletion endpoint will cost eligibility in an autoimmune population with prior B-cell-directed therapy, so decide it before it arrives as a screening-failure review. Section 19’s operational counters should include screen failures on this criterion.

For v1, run Archetype 1 as oncology with \(B_0=100\) cells/uL. Introduce a low-baseline population when IIV enters in Phase 4, where a distribution of \(B_0\) is representable.

Archetype 1 — novel B-cell-depleting TCE

This should be the primary analysis.

Assumptions:

  • FIH.
  • Substantial uncertainty in active dose.
  • QW treatment.
  • Strong priming effect.
  • Rapid acute safety information (~48 h).
  • Peripheral B-cell PD available within several days.
  • Profound B-cell depletion possible.
  • Target criterion <5 cells/uL.
  • Delayed ICANS remains major unmodeled uncertainty.

This is likely the strongest scientific use case because there is an immediate pharmacologic stopping signal.

Archetype 2 — another BCMA TCE

Use principally as a low-uncertainty comparator.

There is already considerable class knowledge about BCMAxCD3 pharmacology and clinically active exposure ranges.

Therefore \(E_{\rm pred}\) should generally be closer to 1.

Prediction-error scenarios might initially emphasize \(1/3,\ 1,\ 3\).

Expected hypothesis:

If class-informed predictions are approximately correct, IPDE provides much less incremental value.

But deliberately include \(E_{\rm pred}=10\) or perhaps 30.

This tests the counterargument:

“We think we know the active range, but what if this molecule is substantially less potent than predicted?”

If the prediction is badly wrong, IPDE can again become useful.

This makes another BCMA TCE a worthwhile team discussion, but not necessarily because the recommendation will be to use IPDE. It is an excellent test of the framework.

Archetype 3 — solid-tumor TCE

Do this after the B-cell model is functional.

The challenge is that there may be no equivalent rapid binary PD endpoint.

Reuse the trial engine but replace B-cell depletion with a generic pharmacologic activity threshold such as \(C_{\rm trough}>C_{\rm active}\) or a target-engagement/activation biomarker if available.

This case may have:

  • Very low incremental operational burden because repeated therapy is already planned.
  • Larger uncertainty around where biological/clinical activity starts.
  • Weaker feedback for deciding exactly when to stop IPDE.

That contrast is scientifically useful.

Do not add an arbitrary tumor-response model just to make the archetype work.

25. The most important figures

Target four figures.

Figure 1 — Patient trajectories

Show conventional vs IPDE:

Conventional

P1:  low -- low -- low
P2:        3x  -- 3x
P3:              9x
...

IPDE

P1:  low -> 3x -> 9x -> 27x -> B<5
                              |
                              v
                           BACKFILL
                         P2 P3 P4

Color by B-cell status.

This explains the concept immediately.

Figure 2 — Primary result

Plot:

\[ T_{3,\rm eval} \]

against \(\log_{10}(D_{\rm true}/D_{\rm start})\) for conventional, catch-up and prospective IPDE.

Once stochastic elements enter, plot the median with a 10th–90th percentile band rather than a single line.

Figure 3 — Prediction error

Plot:

\[ \Delta T = T_{3,\rm conventional} - T_{3,\rm IPDE} \]

versus \(D_{\rm true}/D_{\rm predicted}\).

This directly answers when class/translational knowledge makes IPDE unnecessary.

Figure 4 — Safety-information latency

Illustrate Friberg ANC trajectories alongside escalating doses.

Draw the conventional arm on the same axes. A panel showing only IPDE trajectories reads as an exhibit against IPDE, which misstates the finding. The safety-decision windows in Section 15 establish that a conventional program escalating on short between-cohort windows also commits to new dose levels before delayed toxicity from earlier levels is observable. The figure should show both arms so the reader can see where the two differ:

  • Conventional: latency outruns the program, and each individual patient sits at one dose level.
  • IPDE: latency outruns the program on the same terms, and additionally the same individual accumulates sequential higher exposures inside the latency window.

The second row is the IPDE-specific risk, and it is visible only against the first.

The message is not:

IPDE causes neutropenia.

It is:

Absence of currently observed toxicity is incomplete reassurance when the biological toxicity process has substantial latency.

26. Main anticipated conclusions to test, not assume

The simulation should be capable of disproving these.

Expected hypotheses are:

  1. Big gap in predicted active dose, big gain from intrapatient escalation. IPDE produces substantial acceleration when the true active dose is many escalation steps above the FIH starting dose.
  2. Good prediction of active dose removes benefit of intrapatient escalation. The advantage decreases sharply when the active dose is accurately predicted.
  3. Catch-up helps patients, but not time to evaluate active dose. Catch-up IPDE primarily benefits participants rather than trial calendar time. This replicates the accelerated-titration literature (the prior art in Section 6) rather than establishing something new, and a result contradicting it is a reason to check the engine first.
  4. Backfill shrinks the gain. The advantage of prospective IPDE persists after requiring 3 conventional backfill patients, but is smaller than the apparent gain in \(T_{\rm BCD}\).
  5. Fast PD turnaround raises the value. Rapid PD turnaround strongly increases the value of IPDE.
  6. Operational burden is small when patients are already scheduled for repeated early administrations.
  7. Escalation can outrun safety information. When clinically relevant toxicity latency exceeds the escalation interval, rapid dose escalation outpaces what has been observed.
  8. ICANS stays unquantified. The inability to model it reliably remains a genuine uncertainty rather than something that should be mathematically assumed away.
  9. Fixed-ratio backfill undershoots. Because the scout’s stopping signal arises on cumulative exposure, the backfill at \(D^*\) under-shoots in some scenarios; the resulting non-confirmation cycles reduce, but do not eliminate, the IPDE advantage.
  10. At slow accrual, enrollment dominates. At low enrollment rates the confirmation penalty \(T_{3,\rm eval}-T_{\rm BCD}\) is dominated by enrolling the 3 backfill patients, not by PD turnaround; if so, the fast PD turnaround hypothesis is demoted and anticipatory pre-screening (Section 20) recovers more time than a faster assay.
  11. Priming does not rescue acute safety. Better priming at a given dose does not make the acute-safety comparison favour IPDE, because IPDE patients also reach higher absolute doses (Section 14, CRS).
  12. A second scout repeats the first ladder, and that only costs time when the scouts run in series. Constraining the first dose in a drug-naive patient to a 3-fold between-cohort increase forces a later scout to start near the bottom. Under serial entry that repeats the first ladder and removes much of the scout advantage; under rolling entry (Section 6) the repeat overlaps the first ladder and costs patients rather than weeks.
  13. Most of the low-dose saving is the micro-cohort, not the ladder. The A-to-A2 contrast accounts for a large share of the patients saved at subactive doses and of any calendar-time saving in the low-dose region, and the A2-to-C contrast is smaller than the A-to-C contrast that bundles them. If so, the case for scout-and-backfill rests on the pharmacodynamic stop and the confirmation it enables rather than on speed through the subtherapeutic range.
  14. Rolling scouts are cheap in time and expensive in exposure. Scouts entering a rung apart reach the nominated dose at close to the calendar time of a single scout, and deliver the observations the declaration needs without a separate backfill wait. The price is that a delayed-toxicity signal at a given rung arrives with \(n_{\rm scout}\) patients dosed above it rather than one, so rolling entry buys time with exposure rather than with visits.
  15. PD-guided beats a fixed ladder. PD-guided escalation reaches profound depletion no later than a fixed 3-fold ladder, because the large steps while the marker is quiet buy back the small steps taken after it moves, and it nominates a \(D^*\) closer to \(D_{\rm true}\), so fewer backfill cohorts fail to confirm.
  16. Cohort size is a patient lever, not a clock lever, when accrual is fast. Cutting low-dose cohorts from three patients to one reduces the number treated at subactive doses without materially changing \(T_{3,\rm eval}\), because the between-cohort review window runs regardless of cohort size. Where accrual is slow the same change moves both.

26b. What result would make IPDE not worth doing

How to read the map, once the scenarios have been run. Section 1 says what comes out; this section says which parts of it point away from IPDE.

The simulation quantifies a benefit against a risk it explicitly cannot quantify (Section 16). No arithmetic resolves that trade, so the useful output is not a recommendation but a statement of where the benefit is large enough for the safety discussion to be worth having. The thresholds below are contours for reading the result. A programme carrying this into a real decision replaces them with its own and fixes them before it looks at the numbers.

On the time axis, with \(\Delta T = T_{3,\rm conventional} - T_{3,\rm IPDE}\):

  • \(\Delta T\) below \(\Delta T_{\min}\): IPDE is not worth the unquantified delayed-toxicity risk in that scenario. Recommend the accelerated conventional design and stop. No safety argument is needed, because there is no benefit to weigh against it.
  • \(\Delta T\) above \(\Delta T_{\min}\): the acceleration is material, and the decision moves to the clinical and regulatory judgement in Section 27, which the simulation informs but does not make.

Report \(\Delta T\) against \(\Delta T_{\min}\) for every scenario, and name the scenarios falling on each side. The expected shape of the answer is that IPDE clears the threshold only when prediction error is large (Section 12), which is the same conclusion as the good prediction of active dose removes benefit of intrapatient escalation hypothesis in Section 26, stated as a decision rather than as a trend.

A result that saves patients but not time

\(\Delta T\) alone cannot register a design that takes the same calendar time and reaches the answer having exposed fewer patients to subactive doses. Catch-up escalation is expected to be that design, so the decision structure needs a second arm rather than scoring it as a failure.

Report the pair \((\Delta T,\ \Delta N_{\rm sub})\) for every scenario, where \(\Delta N_{\rm sub}\) is the reduction in patients treated only at subactive exposures against the conventional comparator. Four cases:

  • Neither material. Recommend the accelerated conventional design and stop.
  • \(\Delta T\) material. The acceleration is real and the decision moves to Section 27.
  • \(\Delta N_{\rm sub}\) material, \(\Delta T\) not. The design is a patient-benefit argument rather than a programme-acceleration one, and it is the weaker case for accepting an unquantified delayed-toxicity risk. The patients spared subactive dosing are the same patients carrying the escalation risk, where a calendar-time saving spreads its benefit to the programme and to later patients while the risk stays with the trial’s participants. State the result in those terms rather than as a saving.
  • Both material. The strongest case available.

Set \(\Delta N_{\rm sub,\min}\) the way \(\Delta T_{\min}\) is set, from programme considerations rather than from the simulation.

The third axis in Section 1, what the trial consumes, stays out of this rule. A cheaper trial is worth having and is reported, but cost is not a reason to accept an unquantified delayed-toxicity risk, and a design that saves visits while saving neither time nor patients does not clear this threshold. Say so explicitly, because the argument will be made.

27. Critical assessment of the strategy

There are three arguments a skeptical clinical team is likely to make, and the project should confront them directly.

Objection 1: “The scout regimen isn’t clinically interpretable.”

Correct. That is why the scout never substitutes for backfill. The primary endpoint explicitly requires 3 patients at a conventional regimen.

Objection 2: “You’re outrunning toxicity.”

Potentially correct. This is probably the strongest objection. CRS is relatively tractable and front-loaded, but ICANS is not sufficiently understood quantitatively. This uncertainty cannot currently be eliminated with modeling. Stopping immediately at strong PD and moving to conventional backfill limits exposure, but does not prove safety.

Objection 3: “We can simply use accelerated conventional cohorts.”

Also correct, and this is why \(n=1\) low-dose conventional cohorts must be simulated. If IPDE provides little additional acceleration against that comparator, the result should say so.

That last comparison is particularly important. Otherwise a team can reasonably dismiss the work as comparing IPDE against an unnecessarily slow Phase 1 design.

27b. Future work

Deferred deliberately, with the reason recorded, so that a later version does not have to rediscover the argument. Nothing here is a v1 dependency.

Model-based dose finding. Three families sit behind the exclusion in Section 6 and they are not the same thing. The continual reassessment method and Bayesian logistic regression models fit a dose-toxicity curve and choose the next dose from it, so they are model-based. BOIN and mTPI decide from the data at the current dose against precomputed boundaries, without fitting a curve, so they are model-assisted. 3+3 is neither.

Two reasons defer all of them. The binding constraint in this project is pharmacology in a range believed to be subtherapeutic rather than toxicity statistics, so a better toxicity allocator does not address it. And these designs are less common in the immunology studies this project is aimed at than in oncology dose finding, so adopting one would add machinery the reader has to be sold on before the simple version has been shown to work.

The second reason is a judgement about current practice rather than a documented one, and it is weakest in exactly this space. A B-cell-depleting TCE in autoimmune disease imports oncology dose-finding practice along with the modality. Treat it as a reason to defer rather than as a reason it will never apply, and check it before repeating it to a review board.

What would reopen the question is the \(F_{\rm start}\) question in Section 6. Guo et al. 2025 places intra-patient escalation inside a CRM with an adaptively updated starting dose, which is that question treated formally rather than as a fixed cap. Work applying Bayesian logistic regression models to the same problem has not been located; when it is, it goes on the references page.

The nearer neighbour is not a toxicity allocator at all. This project’s escalation rule fires on pharmacodynamics, and the Project Optimus material listed under “Related, not yet placed” on the references page is the literature for PK/PD-guided dose finding. If anything on that page changes the escalation rule in Section 13, it is more likely to be that than a CRM.

28. References

Every source this document draws on is on the references page, in reading order, each carrying a marker recording whether it has been checked against its source. Nothing is marked checked yet.

Two things there need verifying before anything from this document goes into final work: the teclistamab PK parameters in Section 8 and the Friberg system values in Section 17, both transcribed rather than verified, and the MCLA-117 dose table in Section 6, transcribed from an image of the poster rather than from the poster.

Back to top