Dose Optimization After Project Optimus (2026): Early-Phase Oncology Designs and Data Capture

TL;DR

Takeaway: In the analysed registry cohort, randomized allocation became more common among trials flagged as same-drug multi-dose, while the flagged multi-dose share varied little across the defined periods. Among industry trials that register two or more experimental arms of the same drug at different doses, the randomized share rose from 34.0% (starts 2018 to 2022) to 50.5% (draft-guidance period) and 53.2% (after the August 2024 final guidance), while the multi-dose share of all trials stayed near one in ten. The tolerability and patient-reported measures the guidance asks sponsors to compare remain rare as registered outcomes: dose-modification outcomes in 6.3% of trials and PRO-CTCAE by name in 0.5%. That naming gap prompts a protocol review; actual completeness needs review of those studies’ specifications and records.

Three computed facts frame the argument, all from a cohort of 6,046 industry-sponsored oncology phase 1, 1/2 and 2 drug trials started between 1 January 2018 and 25 July 2026 [1]. First, same-drug multi-dose designs are 11.3% of the cohort (682 trials) and moved only from 11.9% to 9.4% to 11.2% across the three guidance periods, whereas randomized allocation within them rose from 34.0% to 50.5% to 53.2%. Second, a dose-modification or toxicity-discontinuation outcome is registered by 6.3% of trials, relative dose intensity by 1.1% and PRO-CTCAE by 0.5% (29 trials); even in randomized same-drug multi-dose trials started after the final guidance, 64.3% register a pharmacokinetic outcome but 4.8% register PRO-CTCAE. Third, explicit dose-optimization language or a named dose-optimization part grew from 1.3% of pre-draft starts to 3.6% and then 7.5%, and from 2.3% to 11.4% in phase 1/2 records.

The FDA guidance describes the randomized parallel dose-response trial as "a recommended trial design", not a requirement [2]. It also names the measures it expects sponsors to compare across dosages: duration of exposure, the proportion receiving all planned doses, interruptions, reductions and discontinuations for adverse reactions, time to first modification, and patient-reported symptomatic toxicity. Those measures are derived variables. Reliable analyses depend on appropriately specified planned/administered doses, modification reasons and dates, PRO items and PK sample times across the relevant cohorts, which is a database, randomization and eCOA design decision made before first patient in.

What the August 2024 guidance asks, quoted with its conditions

Takeaway: The guidance recommends comparing multiple dosages, names a randomized parallel dose-response trial as a recommended design, sizes that trial for "sufficient assessment" rather than superiority, and lists the tolerability measures to compare. Every one of those sentences uses "should" or "may be considered", and the document itself defines "should" as recommended, not required.

In August 2024 the FDA Oncology Center of Excellence, the Center for Drug Evaluation and Research and the Center for Biologics Evaluation and Research issued the final guidance Optimizing the Dosage of Human Prescription Drugs and Biological Products for the Treatment of Oncologic Diseases [2]. It is the written output of Project Optimus, the OCE initiative whose stated purpose is to "reform the dose optimization and dose selection paradigm in oncology drug development" and whose goals include strategies "including randomized evaluations of a range of doses" [3]. The draft was published for comment in the Federal Register on 18 January 2023 (88 FR 2932, docket FDA-2022-D-2827) [4].

The document's status is stated in its standard clause: "the word should in FDA guidance means that something is suggested or recommended, but not required" [2]. Everything below is a recommendation with a stated condition, and the article quotes it that way.

The problem the guidance addresses is historical. Dose-finding trials for oncology drugs "have historically been designed to determine the maximum tolerated dose (MTD)", a paradigm "developed for cytotoxic chemotherapies" [2]. The FDA-ASCO workshop page of September 2023 puts the consequence plainly: targeted therapies may be taken "for long periods of time, increasing the importance of tolerability", and the 2022 workshop "concluded that oncology needs a paradigm shift to optimize dosage selection for new drugs" [5]. FDA reviewers who examined initial oncology approvals from 2019 to 2021 described dose optimization through randomized dose evaluation as "not routinely conducted" [6], and the 2021 perspective that announced the program framed the issue as a drug-dosing conundrum in which less can be more [7].

The guidance's background section names the data an MTD-driven trial tends to miss: such trials "may not adequately consider other data, such as low-grade symptomatic toxicities (i.e., grade 1-2), dosage modifications, drug activity, pharmacokinetics, pharmacodynamics, and dose- and exposure-response relationships" [2]. That list is, in effect, the data-capture agenda for the rest of this article.

Section III.B carries the design recommendation. "Multiple dosages should be compared in a trial(s) that is designed to assess antitumor activity, safety, and tolerability to support the proposed recommended dosage(s) listed in a marketing application" [2]. "A recommended trial design to compare multiple dosages is a randomized, parallel dose-response trial", because "Randomization (rather than enrolling patients to non-randomized dosage cohorts) promotes comparability of patients receiving each dosage, minimizing bias in estimation of dose- and exposure-response relationships" [2]. Blinding is conditional: "Blinding patients and investigators to dosage arm assignment may be considered as there could be bias that higher dosages are associated with greater activity" [2].

The sizing sentence is the one most often misquoted. "The trial should be sized to allow for sufficient assessment of safety and antitumor activity for each dosage", and the guidance adds that powering it "to demonstrate statistical superiority of a dosage or statistical non-inferiority among the dosages using Type I error rates which would be used in registrational trials" is unnecessary [2]. The guidance also allows the comparison to be built into the escalation study: "It may be useful to evaluate additional dose-level cohorts or add more patients to existing dose-level cohorts (i.e., backfill cohorts) in the dose-finding trial for dosages which are being considered for further development" [2]. An adaptive design that stops enrollment to one or more dosage arms after an interim assessment "could be considered", and multiple dosages "may also be compared prior to a registration trial(s) or as part of a registration trial(s) by adding an additional dosage arm(s)", in which case the design "should provide strong control of Type I error" [2].

Section III.C is where the operational shift sits. When selecting dosages, "safety and tolerability should be compared across the multiple dosages, including: duration of exposure; proportion of patients who are able to receive all planned doses; percentage of patients that require dosage interruptions, dose reductions, and drug discontinuations for adverse reactions; time to the first dosage modification; length of the dosage interruptions(s); percentage of patients with serious adverse reactions (including fatal adverse reactions), and percentage of patients with certain toxicities of interest for the product" [2]. Stopping rules for excessive toxicity "should be pre-specified", and the protocol "should clearly state what actions will be taken" if modification, serious adverse reaction or toxicity percentages "are too high for one or more of the dosages" [2]. Persistent grade 1 to 2 symptomatic reactions "may significantly affect a patient's ability to remain on the drug for extended periods", and "Inclusion of PROs should be considered to enhance the assessment of tolerability in dose-finding trials, as well as subsequent trials" [2].

Pharmacokinetics are treated as continuous infrastructure rather than a phase 1 deliverable. "A PK sampling and analysis plan should be included in each protocol", the plan for all clinical trials "should be sufficient to support population PK" and "dose- and exposure-response analyses", and population PK analyses "should be initiated early and updated as additional data become available" [2].

Two further sentences change how a sponsor should read the rest. On formulation: "Perceived difficulty in manufacturing multiple dose strengths is an insufficient rationale for not comparing multiple dosages in clinical trials" [2]. On the consequence of a poorly justified dosage: "An approach where a dosage is chosen for a trial without adequate justification or consideration of all relevant data may not be acceptable, because FDA may determine that patients are exposed to unreasonable and significant risk, or there is insufficient information to determine risk, or the design of the trial is deficient to meet its stated objectives and may place a protocol on clinical hold" [2].

Scope matters. The guidance excludes "selection of the starting dosage for first-in-human trials" and "dosage optimization for radiopharmaceuticals, cellular and gene therapy products, oncolytics, microbiota, or cancer vaccines", and it leaves pediatric drug development outside its specific scope [2]. For a new combination or indication where sufficient relevant data to support the proposed dosage are lacking, "additional dose-finding should be conducted" [2].

What changed in registered designs: randomization within multi-dose designs

Takeaway: In 6,046 industry oncology early-phase drug trials started from 2018 to July 2026, the share testing two or more doses of the same drug stayed near 11%, but the randomized share of those trials rose from 34.0% before the draft to 53.2% after the final guidance. Period assignment by start date is descriptive; the guidance is one candidate explanation among several.

The cohort comes from ClinicalTrials.gov through the AACT relational copy (registry export 25 July 2026, AACT snapshot 1 August 2026) [1]. Of 596,685 registered studies, 455,356 are interventional, 136,726 carry a phase of early phase 1, phase 1, phase 1/2 or phase 2, 59,436 started between 1 January 2018 and 25 July 2026, 26,251 have an industry lead sponsor, 6,160 map to the MeSH condition tree for neoplasms, and 6,046 include at least one drug or biological intervention. By phase the cohort holds 152 early phase 1, 2,678 phase 1, 1,542 phase 1/2 and 1,674 phase 2 trials [1].

Three periods are defined by registered start date: before the draft (1 January 2018 to 31 December 2022, 3,558 trials), the draft period (1 January 2023 to 31 July 2024, 1,080 trials) and after the final guidance (1 August 2024 to 25 July 2026, 1,408 trials) [1][4]. A trial is same-drug multi-dose when at least two experimental arms carry different dose signatures (a numeric dose and unit, a dose-level identifier such as "dose level 2" or "DL2", or a dose label such as "Dose A", "dose regimen 1", "RP2D-1" or "low dose") and those arms share at least one registered intervention record; crossover designs are excluded. A trial is randomized same-drug multi-dose when it also registers randomized allocation [1].

The first result is the stable one. Same-drug multi-dose designs are 682 of 6,046 trials (11.3%): 423 of 3,558 (11.9%) before the draft, 101 of 1,080 (9.4%) in the draft period and 158 of 1,408 (11.2%) after the final guidance [1]. The flagged proportions are similar across these pooled periods; the scan cannot measure every dose comparison or isolate a 2018-to-present practice change.

The second result is the one that does. Within same-drug multi-dose trials, randomized allocation rose from 144 of 423 (34.0%) before the draft to 51 of 101 (50.5%) in the draft period and 84 of 158 (53.2%) after the final guidance [1]. By start year the share climbs from 23.1% in 2018 to 41.4% in 2021, 50.0% in 2023, 52.7% in 2024 and 54.3% in 2025, with 50.0% in the partial 2026 year [1]. Randomized same-drug multi-dose trials as a share of the whole cohort moved from 4.0% to 4.7% to 6.0%.

The shift is concentrated by phase. In phase 2, randomized same-drug multi-dose designs went from 78 of 1,027 pre-draft starts (7.6%) to 26 of 264 (9.8%) and 49 of 383 (12.8%) after the final guidance [1]. In phase 1/2 the randomized multi-dose share moved from 4.2% to 4.6% to 5.6%, while explicit dose-optimization language or a named dose-optimization part in the record went from 2.3% to 4.6% to 11.4%. In phase 1 the randomized multi-dose share stayed near 2% (1.9%, 2.2%, 2.2%) while dose-optimization language grew from 1.2% to 3.7% to 7.8% [1]. Backfill language appears in 0.3% of pre-draft records and 2.2% of post-final records. The literal phrase "Project Optimus" appears in one record.

Read together, the two phase 1 numbers describe a design that is invisible to the arm table. A seamless protocol may register several dose comparisons within a single group. Conversely, different dose signatures or a shared combination-treatment intervention can generate false-positive same-drug flags. The two indicators can therefore diverge. Without protocol-level validation, neither is a proven lower bound or a verified dose-comparison census.

These associations coincide with the guidance timeline; causation remains unproven. Candidate explanations include the growth of antibody-drug conjugate and bispecific pipelines whose regimens invite dose and schedule comparison, the concentration of randomized dose-ranging among larger sponsors with pharmacometrics capacity, and the spread of seamless phase 1/2 protocols that move the comparison into the early-phase record. A sensitivity check using a stricter numeric-dose-only definition shows the same direction, with the randomized share moving from 35.4% to 53.1% to 55.1% across the three periods [1].

How many dose arms, how often blinded and how large

Takeaway: Among the 279 randomized same-drug multi-dose trials, two dose arms is the modal design (165 trials), 225 are open-label, and median registered enrollment is 102 overall and 120 for post-final starts. These are benchmarks of current practice, not thresholds.

The 279 randomized same-drug multi-dose trials give the operational baseline a design team asks for [1]. Two distinct dose arms appear in 165 trials, three in 54, four in 22 and five or more in 38. The counts come from dose quantities, dose-level identifiers and dose labels in experimental-arm text, so a study with unlabeled arms can be under-counted, and a trial with five or more signatures is often a multi-part protocol rather than a five-arm comparison.

Masking is the exception. Of the 279 trials, 225 register no masking, 25 quadruple masking, 12 single, 10 double and 7 triple [1]. The guidance's conditional sentence on blinding, "may be considered as there could be bias that higher dosages are associated with greater activity" [2], describes a design few sponsors have chosen; independent response review can be considered where assessment bias is relevant, with the relative use of that control requiring a separate analysis (the companion BICR analysis covers what registry wording shows about central review and what it leaves open).

Registered enrollment is modest in this cohort and higher in the post-final period. Median enrollment across the 279 trials is 102: 92 for pre-draft starts, 78 for draft-period starts and 120 for post-final starts [1]. By phase, 153 of the 279 are phase 2, 71 phase 1/2, 54 phase 1 and 1 early phase 1. Those medians are consistent with the guidance's sizing sentence, which asks for "sufficient assessment of safety and antitumor activity for each dosage" rather than a powered superiority test [2]. Per-arm allocation ratios and estimation targets require review beyond this extraction, so the medians describe practice and not a recommended size.

Two public randomized dose comparisons and the data they consumed

Takeaway: The sotorasib postmarketing requirement and DESTINY-Lung02 are two selected public randomized dose comparisons. One was open-label and post-approval, the other blinded and used in a registration programme; both depended on randomization, per-arm exposure, PK and dose-modification data that had to exist before the comparison could be read.

The sotorasib case shows what happens when the comparison comes after approval. The May 2021 accelerated-approval letter for NDA 214665 lists postmarketing requirement 4071-2: "Conduct a multicenter, randomized clinical trial to further characterize serious adverse events, including gastro-intestinal toxicity and compare the safety and efficacy of sotorasib 960 mg daily versus a lower daily dose in patients with locally advanced or metastatic, KRAS G12C mutated, non-small cell lung cancer who have received at least one prior systemic therapy" [8]. The sponsor's timetable in the letter sets trial completion for October 2022 and final report submission for February 2023 [8].

The comparison was published in 2024 [9]. In a randomized, open-label phase 2 study (NCT03600883), adults with previously treated KRAS G12C-mutated advanced NSCLC received sotorasib 960 mg (n = 104) or 240 mg (n = 105) once daily. Objective response rate was 32.7% at 960 mg and 24.8% at 240 mg; median progression-free survival was 5.4 and 5.6 months; median overall survival at a median follow-up of 17.5 months was 13.0 and 11.7 months; AUC and Cmax were 1.3-fold numerically higher at 960 mg; and the authors describe the study as "not powered for formal statistical hypothesis testing" [9]. The abstract's closing sentence is the data-capture point: treatment-emergent adverse events "were manageable with label-directed dose modifications" [9]. Which dose belongs on the label is a regulatory judgment outside the scope of this article. What the record shows is that reading a fourfold nominal dose difference against a 1.3-fold exposure difference required randomized allocation, PK sampling in both arms and dose-modification data that could be compared by arm.

DESTINY-Lung02 shows the comparison done before registration and blinded. Goto and colleagues report a blinded, multicenter phase 2 study in which 152 patients with previously treated HER2-mutant metastatic NSCLC were randomly assigned 2:1 to trastuzumab deruxtecan 5.4 or 6.4 mg/kg every three weeks, with confirmed objective response rate by blinded independent central review as the primary end point [10]. Confirmed ORR was 49.0% with 5.4 mg/kg and 56.0% with 6.4 mg/kg; grade 3 or higher drug-related treatment-emergent adverse events occurred in 38.6% and 58.0% of patients; adjudicated drug-related interstitial lung disease occurred in 12.9% and 28.0%; median treatment duration was 7.7 and 8.3 months; and the authors conclude that the safety profile favored 5.4 mg/kg [10].

The two studies consumed the same categories of data even though their designs differed. Both needed randomized allocation recorded separately from the dose dispensed. DESTINY-Lung02 needed dose-level blinding, a central review of response and an adjudication process for the toxicity of interest [10]. The sotorasib comparison needed PK sampling that could resolve exposure across arms and dose-modification records that could be summarized by arm [9]. Neither result could have been assembled from an adverse-event table alone, The publications support the importance of those data categories, but do not document every implementation choice or when each database field was built.

The tolerability measures the guidance names are rarely registered

Takeaway: The guidance asks sponsors to compare exposure, dose modifications and patient-reported symptoms across dosages, yet in the cohort a dose-modification outcome is registered by 6.3% of trials, relative dose intensity by 1.1% and PRO-CTCAE by 0.5%. Pharmacokinetics, by contrast, is registered by 58.6%. Registered outcomes are a naming surface, but the asymmetry is a reason to check the specification, not evidence that the corresponding records were missing.

Conventional early-phase outcomes dominate the registered rows. Across the 6,046 trials, a dose-limiting toxicity outcome is registered by 47.2%, a pharmacokinetic outcome by 58.6% (and as a primary outcome by 11.3%), and a pharmacodynamic outcome by 23.2% [1]. The tolerability measures in Section III.C are marginal by comparison. A dose-modification, interruption, reduction or toxicity-discontinuation outcome appears in 379 trials (6.3%), rising from 5.3% before the draft to 7.2% and 8.0% [1]. Relative dose intensity is a registered outcome in 64 trials (1.1%). Any patient-reported outcome appears in 472 trials (7.8%), and that share drifted down across periods, from 8.4% to 7.4% to 6.5%. PRO-CTCAE is named in 29 trials (0.5%): 0.3% before the draft, 0.3% in the draft period and 1.1% after the final guidance, reaching 3.1% (12 of 383) in phase 2 post-final starts [1]. An exposure-response outcome is named in 28 trials (0.5%).

The pattern holds inside the trials that are actually comparing doses. In the 84 randomized same-drug multi-dose trials started after the final guidance, 64.3% register a PK outcome, 15.5% a dose-modification outcome, 11.9% any PRO, 4.8% PRO-CTCAE by name (4 trials) and 2.4% relative dose intensity (2 trials) [1]. Among the 106 post-final trials with an explicit dose-optimization signal, 85.8% register PK, 21.7% register dose modification and 1.9% register any PRO [1].

One boundary should be stated once and clearly. An outcome row on ClinicalTrials.gov is a name, not a data specification. A sponsor can collect dose modifications on every cycle and never register them as an outcome, so the percentages above measure registration and only registration. The asymmetry shows which terms are detectable in registered outcome fields. Protocols and analysis plans supply the fuller record of comparative endpoints and requirements. The companion analysis of COA registration language found the same naming-surface limit for PRO and ClinRO tokens across the whole registry.

The EU comparator reinforces the same measures. The EMA guideline on the clinical evaluation of anticancer medicinal products, Revision 6, adopted by the CHMP on 18 November 2023, says that dose-finding for molecularly targeted agents "should not only focus on safety endpoints, but also on determining an optimal biologically active dose" [11]. On timing it observes that "in phase I trials of MTAs, more than half of the patients present with their first grade 3-4 toxicity after cycle 1", and that "Lower grade toxicity over longer periods of time that affect tolerability and the possibility of maintaining the intended dose intensity may need to be addressed in the DLT and MTD definitions" [11]. Its tolerability section states that "Outcomes such as dose adjustments and discontinuation rate often provide important information on tolerability" and that patient-reported outcomes "may be complementary tools for assessing the tolerability" [11]. The EMA page also lists a concept paper for Revision 7 with a consultation period from 30 April to 31 July 2025 [11].

Patient-reported symptomatic toxicity has a standard instrument. The National Cancer Institute's PRO-CTCAE item library holds "124 items representing 78 symptomatic toxicities drawn from the CTCAE"; items evaluate "frequency, severity, interference, amount, presence/absence", each toxicity "is assessed by 1-3 attributes", the recall period "is the past 7 days", responses "are scored from 0 to 4 (or 0/1 for absent/present)", and the instrument is "publicly available for all to use in their clinical research" [12]. The NCI overview cautions against assuming one standard approach for every longitudinal analysis [12]. Published composite grading algorithms do exist, including a tested method that combines attributes for an individual symptomatic adverse event; that is different from a universally mandated overall toxicity score. Preserve item responses and pre-specify the chosen scoring and longitudinal method [13].

FDA's core-PRO guidance, finalized in October 2024, supplies the selection and scheduling frame. It recommends collecting and separately analyzing five core PROs: disease-related symptoms, symptomatic adverse events, an overall side effect impact summary measure, physical function and role function [14]. For symptomatic adverse events it recommends "selecting a concise set of the most important symptomatic AEs that are expected to occur from an item library", names PRO-CTCAE as "an example of one acceptable item library", and states that PRO data on symptomatic AEs "are intended to complement, not replace, safety data" [14]. On frequency it recommends a baseline assessment, that "Assessment frequency should be higher within the first few treatment cycles", that the schedule "should take into account the administration schedule of the treatment(s) under study", and it illustrates a strategy "that assesses PROs more frequently in the first 8 weeks of treatment" [14]. The August 2024 dosage guidance points to this document for instrument selection and assessment frequency [2]. A published example of the schedule in practice collected PRO-CTCAE at baseline, days 1 and 15 of each cycle and off treatment, and summarized each symptomatic adverse event by maximum post-baseline score and a 12-week area under the curve [15].

Relative dose intensity is the derived measure that ties these together. It is a ratio of delivered to planned dose intensity, and the systematic review and meta-analysis by Nielson and colleagues, covering advanced solid-tumor chemotherapy, is one review of its association with survival [16]. That evidence is observational and concerns chemotherapy regimens, so it supplies no threshold for a targeted agent. Its relevance here is narrower: a measure defined over planned and delivered dose across time resists transcription on a single form; it has to be computed from dosing records that carry planned dose, administered dose, dates and reasons.

The literature supplies a separate measure of publication attention, which cannot be ranked against registry adoption. Europe PMC records whose title or abstract matches "dose optimization" plus a cancer term rose from 26 in 2018 to 50 in 2022, 64 in 2023, 92 in 2024, 139 in 2025 and 177 in 2026 to date; records matching "Project Optimus" first appear in 2022 (3) and reach 31 in 2024, 19 in 2025 and 23 in 2026; "PRO-CTCAE" records rose from 13 in 2018 to 54 in 2023 and 56 in 2025; "relative dose intensity" with a cancer term rose from 41 in 2018 to 67 in 2025 and 76 in 2026 [17]. These are query counts, not evidence of practice, and 2026 is a partial year.

Randomized expansion, backfill and seamless phase 1/2: what each demands

Takeaway: The randomized parallel dose-response design is thirty years old in ICH E4; expansion-cohort and seamless protocols are possible settings for its use, without a claim that those approaches are newly invented. Each placement changes when randomization, cohort identifiers, PK sampling and estimand definitions have to be in place.

ICH E4, adopted in March 1994, describes the design the FDA guidance now recommends: "Randomization to several fixed dose groups (the randomized parallel dose-response study) is simple in concept and is a design that has had extensive use and considerable success" [18]. E4 also describes an ethical constraint on placebo-controlled comparisons; the historical reasons for oncology’s design choices need separate evidence. Parallel dose-response designs with placebo "would not be acceptable in the study of some conditions, such as life-threatening infections or potentially curable tumors, at least if there were effective treatments known" [18]. The 2024 guidance says it "should be considered along with" E4 [2], which places targeted oncology back inside the general dose-response framework while leaving the ethical caveat where E4 put it.

The first placement is the expansion cohort. FDA's March 2022 guidance on expansion cohorts describes first-in-human designs that use "multiple, concurrently accruing subject cohorts, where individual cohorts assess different aspects of the safety, pharmacokinetics, and antitumor activity of the drug product" [19]. A randomized dose-comparison cohort is one such cohort, and it inherits the same protocol, database and monitoring structure as the escalation part it follows.

The second placement is backfill. The dosage guidance's sentence on backfill cohorts allows a sponsor to add patients at dose levels already cleared while escalation continues, "to allow for further assessment of safety and activity prior to initiating a trial to compare multiple dosages" [2]. In the non-randomized backfill design considered here, patients enroll without randomization; backfill is not universally non-randomized, and and its patients enter at different calendar times from the escalation patients at the same level. A 2026 simulation study by Cheng and Li compares exactly these two options, "the two-stage randomized expansion design, in which dose escalation is completed at the first stage and then patients are randomized across selected dose levels for further evaluation, and the one-stage backfill design, which allows enrollment at previously explored doses while escalation is ongoing", across rule-based, utility-based and model-based dose-selection methods and with and without imbalance factors [20]. The abstract reports no winner, and this article infers none. The design lesson is that the two options place different demands on the data: backfill requires cohort and dose-level identifiers that stay stable across escalation and backfill so the pooled dose-level data can be read as one series, while randomized expansion requires an allocation system that is live before the expansion opens.

The third placement is the seamless phase 1/2 protocol, where the registry's language signal is rising fastest (2.3% to 11.4% of phase 1/2 records) [1]. Gao and colleagues describe the analytical demand of that setting: dose optimization in oncology is multidimensional, must weigh "long-term tolerability beyond dose-limiting toxicities", and is best approached "through the lens of Totality of Evidence and with the mindset of model-informed drug development", building "a quantitative knowledge base integrating diverse sources of data" [21]. A knowledge base of that kind needs consistent definitions and a documented common analysis model across escalation, backfill and randomized cohorts; source forms need not be identical if mappings preserve meaning. If the parts use different dosing forms, different PK schedules or different PRO items, the pooled exposure-response analysis the guidance asks for requires reconciliation before it can be assembled.

The estimand is the fourth design decision, and it is a data decision as much as a statistical one. ICH E9(R1) defines intercurrent events as "events occurring after treatment initiation that affect either the interpretation or the existence of the measurements associated with the clinical question of interest" and gives "discontinuation of treatment due to toxicity" as an example of a granular intercurrent event [22]. In a two-arm dose comparison, discontinuations, reductions and interruptions are exactly the events that differ between arms, so the strategy for handling them has to be chosen before the data are unblinded, and the reason for every dose change has to be recorded in a way the strategy can use.

The data-capture specification from cohort one

Takeaway: The guidance's comparison measures and the registry gaps translate into a short list of raw fields and derivations. The matrix below maps each design feature to what has to exist from cohort one; the specification table names the fields, the derivation and the current registry signal. Frequencies and thresholds stay protocol decisions.

The matrix is derived from the guidance's named measures, the core-PRO guidance and the registry findings. It states requirements to specify, not claims that any system meets them.

Four rows carry most of the work. Randomized dose arms need allocation recorded separately from dispensing, with stratification factors and timestamps, and where dose-level blinding is used, roles that keep dose level from the people whose assessments could be biased; the companion RTSM analysis covers what registry flags tell a team about that configuration and what they leave open. Backfill needs cohort and dose-level identifiers that stay constant between escalation and backfill, and compatible case report forms or documented mappings that preserve cohort-specific differences. The tolerability comparison needs planned and administered dose per cycle, modification type, dates and a reason code with an adverse-event link where the change is caused by an adverse event, so that time to first modification, length of interruption, proportion receiving all planned doses and relative dose intensity are derived rather than transcribed. The PRO row needs a baseline assessment and a denser schedule in the first cycles, with item selection recorded against the expected mechanism-based toxicities [14].

The specification table turns those rows into fields.

Three implementation points follow from the table. First, the derived measures the guidance lists are only as good as the dosing record. A dosing form that records only total cycles completed or cumulative dose yields neither time to first modification nor length of interruption; the form has to carry planned dose, administered dose, start and stop dates and a modification reason for every cycle, with edit checks that request a reason when planned and administered dose differ, while allowing explained rounding, administrative or other non-AE changes. Second, PRO-CTCAE has item-level measures and published per-event composite grading methods [12][13], so the analysis plan must specify which summaries will be used (for example, maximum post-baseline score or an area under the curve over a defined window, as in the published schedule example [15]), and the eCOA build has to time-stamp each administration against the dosing calendar. Third, PK sampling forms have to record actual sample times relative to actual dosing, with food status for oral agents, because the guidance's expectation that population PK be "initiated early and updated as additional data become available" [2] presumes that early-cohort samples are analyzable alongside later ones.

None of this requires a particular vendor or product. It requires that the first-in-human protocol, the database specification, the randomization plan and the eCOA schedule be written together, before the first cohort opens, with the comparison the sponsor expects to run in mind.

Frequently asked questions

Does FDA require sponsors to compare two doses?

There is no universal two-dose mandate in this nonbinding guidance. Product-specific requirements, such as the sotorasib postmarketing requirement described above, can require a comparison [8]. The August 2024 guidance says multiple dosages "should be compared" and calls the randomized parallel dose-response trial "a recommended trial design", and it defines "should" as "suggested or recommended, but not required" [2]. It sets no number of doses. In the registry cohort, two dose arms is the modal design among randomized same-drug multi-dose trials (165 of 279), with three arms in 54 [1]. The guidance does pair its recommendations with a stated consequence: a dosage chosen "without adequate justification or consideration of all relevant data may not be acceptable" and FDA "may place a protocol on clinical hold" [2].

How many early-phase oncology trials randomize a dose comparison now?

In the 6,046-trial cohort, 279 trials (4.6%) register a randomized comparison of two or more doses of the same drug. Within trials that compare doses of the same drug at all, the randomized share rose from 34.0% (2018 to 2022 starts) to 50.5% (draft period) and 53.2% (after the final guidance); in phase 2 the randomized multi-dose share rose from 7.6% to 12.8% [1]. The multi-dose share of all trials stayed near one in ten.

Is PRO-CTCAE required in dose-optimization trials?

No. The dosage guidance says "Inclusion of PROs should be considered to enhance the assessment of tolerability in dose-finding trials" [2], and the core-PRO guidance names PRO-CTCAE as "an example of one acceptable item library" for symptomatic adverse events, to be used as "a concise set of the most important symptomatic AEs that are expected to occur" [14]. In the cohort, PRO-CTCAE is named in 0.5% of trials and in 4.8% of randomized same-drug multi-dose trials started after the final guidance [1].

What is relative dose intensity, and why does it need derivation-ready fields?

Relative dose intensity is the ratio of delivered to planned dose intensity over a defined period; the evidence linking it to survival comes from observational chemotherapy studies and supplies no threshold for a targeted agent [16]. Because it is computed across cycles from planned dose, administered dose, dates and modification reasons, it has to be derived rather than entered on a single form. In the cohort it is a registered outcome in 1.1% of trials [1].

Does the guidance apply to combinations, cell therapies or pediatric programs?

For a new combination or indication where sufficient relevant data to support the proposed dosage are lacking, the guidance says "additional dose-finding should be conducted" [2]. It excludes first-in-human starting doses, radiopharmaceuticals, cellular and gene therapy products, oncolytics, microbiota and cancer vaccines, and it leaves pediatric development outside its specific scope [2]. The FDA-ASCO 2023 workshop noted that combinations "have additive or potentially synergistic effects but also greater toxicity" than monotherapies [5].

Methodology and limitations

Takeaway: One computed registry lane, one computed literature lane and primary FDA, NCI, ICH and EMA documents, all fetched first-hand. Registry findings are descriptive associations bounded by what submitted records can show.

Registry cohort. ClinicalTrials.gov via AACT, registry export 25 July 2026 and AACT snapshot 1 August 2026 [1]. Filters, in order: interventional study type; phase early phase 1, phase 1, phase 1/2 or phase 2; start date 1 January 2018 to 25 July 2026; industry lead sponsor; MeSH "neoplasms" in the registry's condition mapping (direct term or ancestor); at least one drug or biological intervention. Funnel: 596,685; 455,356; 136,726; 59,436; 26,251; 6,160; 6,046.

Design flags. Same-drug multi-dose: two or more experimental arms whose title or description carries different dose signatures (number plus unit; "dose level N" or "DLn"; title labels "Dose A", "Dose 1", "dose regimen 1", "RP2D-1", "RD-2", "low/high dose"), where the arms share at least one registered intervention record, excluding crossover designs. Randomized same-drug multi-dose adds randomized allocation. Dose-optimization signal: "dose optimization/optimisation" in title, summaries, arms or interventions, or a design group naming dose optimization, randomized dose, dose confirmation or dose comparison. Outcome flags are regular-expression matches on registered outcome measure and description text (PK, PD, exposure-response, dose limiting toxicity, dose modification or discontinuation for toxicity, relative dose intensity, any patient-reported outcome, PRO-CTCAE). Periods are assigned by registered start date: before 1 January 2023; 1 January 2023 to 31 July 2024; 1 August 2024 onward [2][4].

Literature counts. Europe PMC REST API queried on 12 September 2026 for fixed title/abstract expressions by publication year, 2016 to 2026 [17]. Counts measure publication attention, not practice; 2026 is partial.

Limitations. Registry records are sponsor-submitted descriptions; arm text is heterogeneous and dose signatures depend on pattern matching, so arm-based detection can miss comparisons and can also falsely link regimen, cohort or combination-treatment signatures. Randomization at study level does not prove randomization between the flagged doses. Optimization parts registered as one group may appear only in the language flag. Outcome flags detect names, not data collection. Period assignment uses registered start dates, 168 of 447 trials with 2026 starts carry estimated dates, and the post-final window is approximately 24 months against 60 for the pre-draft window. Period labels use month/year boundaries rather than exact issue days; start date does not establish when a protocol or amendment adopted a design. PRO-CTCAE counts are small (29 trials), so their period changes are fragile. Industry lead sponsorship excludes academic-led trials with industry collaborators. No causal claim about the guidance is made. The two clinical examples are single studies reported as published; neither settles the labeled dose, which is a regulatory decision.

Conclusion

Takeaway: Randomized allocation is a small majority within the flagged post-final subset, while several tolerability terms are uncommon in registered outcomes. Use those signals to check protocol-specific dosing fields, PRO schedules, PK sampling and allocation controls; they do not prove missing capture or a universal design default.

Three conclusions carry the article. Randomized allocation is a majority within the flagged same-drug multi-dose subset of this industry oncology early-phase cohort (53.2% of post-final starts, up from 34.0%), while the flagged multi-dose shares remained similar across the pooled periods [1]. The measures the guidance asks sponsors to compare across dosages, from the proportion receiving all planned doses to patient-reported symptomatic toxicity, include derived variables whose names are uncommon in this cohort’s registered outcomes and that resist reconstruction from an adverse-event table [1][2]. And the two public comparisons that have been read against a label, sotorasib and DESTINY-Lung02, were readable because randomization, PK, dose-modification and adjudicated safety data existed by design [9][10].

The Monday-morning action is a review of the first-in-human specification against the two tables in this article: whether the dosing form yields time to first modification and relative dose intensity without transcription, whether the PRO schedule is denser in the first cycles with items chosen for the expected toxicities, whether PK samples carry actual times relative to actual doses, and whether the allocation system for the optimization part will be live before that part opens.

As an implementation example, EClinCloud's product set covers the four systems this specification touches: EDC, which the company describes as pairing AI-driven study build with OCR-assisted source capture, data collection, remote review and standards management, "engineered around ALCOA+ principles" [23]; RTSM, described as "supporting complex randomization designs and parallel protocols while keeping drug supply precisely under control" [24]; eCOA, which "captures trial data across all five COA dimensions" of ePRO, eClinRO, eObsRO, ePerfO and eDiary with multilingual support [25]; and IRC, which "unifies image upload and reader workflows on one standardized, GxP-aligned platform" [26]. That is product scope, not an endpoint or acceptance promise; the specification above is what a sponsor has to write regardless of the platform that carries it.

Sources

1. ClinicalTrials.gov via AACT, EClinCloud analysis of the 25 July 2026 registry export (AACT snapshot 1 August 2026): industry-sponsored oncology early phase 1, phase 1, phase 1/2 and phase 2 drug or biological trials with start dates from 1 January 2018 to 25 July 2026, n = 6,046; design, outcome and language flags as defined in the methodology; accessed September 12, 2026. Public sources: ClinicalTrials.gov and AACT.

2. U.S. Food and Drug Administration (OCE, CDER, CBER), Optimizing the Dosage of Human Prescription Drugs and Biological Products for the Treatment of Oncologic Diseases, Guidance for Industry, final, August 2024, docket FDA-2022-D-2827, accessed September 12, 2026. Quotations from Sections I to III of the PDF.

3. U.S. Food and Drug Administration, Oncology Center of Excellence, Project Optimus, program page, accessed September 12, 2026.

4. Federal Register, Optimizing the Dosage of Human Prescription Drugs and Biological Products for the Treatment of Oncologic Diseases; Draft Guidance for Industry; Availability, 88 FR 2932, 18 January 2023, docket FDA-2022-D-2827, accessed September 12, 2026.

5. U.S. Food and Drug Administration, Second Annual FDA ASCO Workshop on Getting the Dosage Right, September 6-7, 2023, meeting page, accessed September 12, 2026.

6. Fourie Zirkelbach J, Shah M, Vallejo J, et al., Improving Dose-Optimization Processes Used in Oncology Drug Development to Minimize Toxicity and Maximize Benefit to Patients, Journal of Clinical Oncology 2022;40:3489-3500, doi 10.1200/JCO.22.00371, accessed September 12, 2026.

7. Shah M, Rahman A, Theoret MR, Pazdur R, The Drug-Dosing Conundrum in Oncology: When Less Is More, New England Journal of Medicine 2021;385:1445-1447, doi 10.1056/NEJMp2109826, accessed September 12, 2026.

8. U.S. Food and Drug Administration, Drugs@FDA, NDA 214665 Lumakras (sotorasib) accelerated approval letter, May 2021, postmarketing requirement 4071-2, accessed September 12, 2026.

9. Hochmair MJ, Vermaelen K, Mountzios G, et al., Sotorasib (960 mg or 240 mg) once daily in patients with previously treated KRAS G12C-mutated advanced NSCLC, European Journal of Cancer 2024, doi 10.1016/j.ejca.2024.114204, NCT03600883, accessed September 12, 2026.

10. Goto K, Goto Y, Kubo T, et al., Trastuzumab Deruxtecan in Patients With HER2-Mutant Metastatic Non-Small-Cell Lung Cancer: Primary Results From the Randomized, Phase II DESTINY-Lung02 Trial, Journal of Clinical Oncology 2023;41:4852-4863, doi 10.1200/JCO.23.01361, accessed September 12, 2026.

11. European Medicines Agency, CHMP, Guideline on the clinical evaluation of anticancer medicinal products, EMA/CHMP/205/95 Rev.6, adopted by the CHMP 18 November 2023; landing page with the Revision 7 concept paper consultation dates at EMA, accessed September 12, 2026.

12. National Cancer Institute, Division of Cancer Control and Population Sciences, PRO-CTCAE Overview, item library, attributes, recall period and scoring statements, accessed September 12, 2026.

13. Basch E, Becker C, Rogak LJ, et al. Composite Grading Algorithm for the National Cancer Institute’s Patient-Reported Outcomes version of the Common Terminology Criteria for Adverse Events (PRO-CTCAE), Clinical Trials 2021;18(1):104–114, doi:10.1177/1740774520975120, accessed September 13, 2026. Development and testing of per-symptomatic-event composite grades; distinct from a universal longitudinal or overall toxicity scoring rule.

14. U.S. Food and Drug Administration (OCE, CDER, CBER), Core Patient-Reported Outcomes in Cancer Clinical Trials, Guidance for Industry, final, October 2024, docket FDA-2020-D-2303, accessed September 12, 2026.

15. Madariaga A, Mitchell SA, Pittman T, et al., Patient self-reporting of tolerability using PRO-CTCAE in a randomized double-blind, placebo-controlled phase II trial comparing gemcitabine in combination with adavosertib or placebo, Gynecologic Oncology 2022;167:226-233, doi 10.1016/j.ygyno.2022.08.006, accessed September 12, 2026.

16. Nielson CM, Bylsma LC, Fryzek JP, Saad HA, Crawford J, Relative Dose Intensity of Chemotherapy and Survival in Patients with Advanced Stage Solid Tumor Cancer: A Systematic Review and Meta-Analysis, The Oncologist 2021;26:e1609-e1618, doi 10.1002/onco.13822, accessed September 12, 2026.

17. Europe PMC REST API, EClinCloud analysis of records per publication year (2016 to 2026, partial 2026) for fixed title/abstract queries on dose optimization with a cancer term, Project Optimus, PRO-CTCAE and relative dose intensity with a cancer term, Europe PMC web services, accessed September 12, 2026.

18. International Council for Harmonisation, ICH E4 Dose-Response Information to Support Drug Registration, Step 4, 10 March 1994, accessed September 12, 2026.

19. U.S. Food and Drug Administration, Expansion Cohorts: Use in First-In-Human Clinical Trials to Expedite Development of Oncology Drugs and Biologics, Guidance for Industry, final, March 2022, docket FDA-2018-D-2777, accessed September 12, 2026.

20. Cheng H, Li XN, A simulation study comparing backfill and randomized expansion in oncology dose optimization, Journal of Biopharmaceutical Statistics 2026 (online 19 February 2026), doi 10.1080/10543406.2026.2627391, accessed September 12, 2026.

21. Gao W, Liu J, Shtylla B, Venkatakrishnan K, Yin D, Shah M, Nicholas T, Cao Y, Realizing the promise of Project Optimus: Challenges and emerging opportunities for dose optimization in oncology drug development, CPT: Pharmacometrics and Systems Pharmacology 2024;13:691-709, doi 10.1002/psp4.13079, accessed September 12, 2026.

22. International Council for Harmonisation, ICH E9(R1) Addendum on Estimands and Sensitivity Analysis in Clinical Trials, Step 4, adopted 20 November 2019, accessed September 12, 2026.

23. EClinCloud, EDC, Electronic Data Capture, product page, scope statement only, accessed September 12, 2026.

24. EClinCloud, RTSM, Randomization and Trial Supply Management, product page, scope statement only, accessed September 12, 2026.

25. EClinCloud, eCOA, Electronic Clinical Outcome Assessment, product page, scope statement only, accessed September 12, 2026.

26. EClinCloud, IRC, Independent Imaging Review Solution, product page, scope statement only, accessed September 12, 2026.