FDA-AACR 2025 Dose Optimization Trilogy: Start, Explore, Register

FDA-AACR 2025 劑量最佳化三部曲：起始、探索、註冊

English

In 2025, the FDA and the American Association for Cancer Research (AACR) co-published a three-article series in Clinical Cancer Research that is, collectively, the most complete educational framework for understanding modern oncology dose optimization available in the peer-reviewed literature. The three papers are not redundant — they are sequential, covering different stages of the drug development timeline: how to choose a FIH (first-in-human) starting dose, how to run early dose exploration so that multiple candidate doses accumulate comparable evidence, and how to justify the dose chosen for a registration trial. Reading them as a unit reveals something important: dose optimization is not a single decision made at the end of phase 1. It is a continuous, evidence-accumulating process that begins at the very first protocol and should not conclude until the registrational dose is backed by a totality of evidence.

The first paper (Zhu et al., PMID 41036712) tackles FIH starting dose selection. Its central argument is that the FIH dose is not just a safety exercise — it is the opening move in an evidence generation strategy. A good FIH protocol asks: what is the minimal dose that avoids unacceptable first-dose harm while still providing a biologically meaningful signal? This requires integrating nonclinical toxicology (NOAEL, HNSTD), pharmacological activity estimates (MABEL, PAD), PK/PD modeling, and — critically — a forward-looking escalation design that will actually be able to learn something from each dose level. The paper emphasizes that the starting dose, the escalation rules, and the decision criteria should all be pre-specified and connected to the same biological hypotheses. It is not enough to declare “3+3 escalation with a starting dose of X mg” without explaining what observable outcome at each level would change the next decision.

The second paper (Okusanya et al., PMC 12614637) moves the conversation to early-phase dose exploration, which is where most of the real dose-optimization work happens. The paper introduces the concept of “fit-for-purpose trial elements” — design components chosen specifically because they will answer a needed dose-optimization question. These include integral biomarkers (built into the primary endpoint decision), integrated biomarkers (analyzed alongside clinical endpoints), and exploratory biomarkers (hypothesis-generating). More concretely, the paper discusses backfill cohorts — the practice of re-enrolling patients at dose levels that have been declared safe but where evidence is sparse, specifically to build a richer evidence base at that dose. It also discusses randomized dose-expansion cohorts, where patients are deliberately assigned to one of several candidate doses rather than defaulting to the highest tolerated one. The clinical utility index (CUI) is proposed as a multi-attribute scoring tool that can combine response probability and toxicity probability into a single framework for comparing doses. All of these tools share a common goal: to arrive at the marketing application with data that can actually answer whether the proposed dose is better than a lower alternative.

The third paper (Shord et al., PMID 41036557) deals with registration trial dose justification — a stage that clinicians sometimes overlook, assuming that by the time a drug reaches a pivotal trial its dose is settled science. This paper argues the opposite: many registration trials inherit their doses from phase 1 escalation studies by the path of least resistance (highest tolerated dose = RP2D = phase 3 dose), and this chain often has gaps. Exposure-response analysis, model-informed approaches, seamless trial designs that allow dose adjustment based on accumulating data, and CUI can all help bridge these gaps — but they have to be planned in advance, not retrofitted after the fact. The paper is candid about the statistical challenges: adaptive and seamless designs that use the same data for dose selection and efficacy estimation require careful pre-specification of the analysis rules, or the type I error and interpretation of the pivotal result become ambiguous.

Alongside the trilogy, a 2025 Clinical Pharmacology & Therapeutics paper by Zhu et al. (PMID 40248986) proposed a molecular-class framework that is an excellent companion reading. The framework divides oncology drugs into four classes based on their mode of action, and argues that each class has a different primary dose-optimization question. For small-molecule targeted therapies and ADCs, the question is whether higher exposure is actually producing better target inhibition or tumor control — or just more off-target toxicity. For large-molecule antagonists (like checkpoint inhibitors in their blocking function), the question is receptor occupancy and the clinical consequences of incomplete vs. complete blockade. For cancer immunotherapy agonists (T-cell engagers, checkpoint agonists), the question is whether you can step up to a dose where the immune activation is therapeutically relevant without triggering systemic toxicity. For molecules with limited single-agent activity, the question is almost philosophical: how do you dose-optimize something that doesn’t have clear efficacy signals to optimize toward?

Together, these papers make the case that modern oncology FIH trials should be designed not as dose-finding exercises but as dose-learning systems. Each phase of development — starting dose, escalation, expansion, registration — should generate data that feeds back into and updates the dose hypothesis. The physician’s role in this system is not passive — to receive the RP2D announced at the end of phase 1 and accept it. The physician’s role is to ask, at every stage: has this development program generated enough multi-attribute evidence to justify this dose for the patients who will actually receive it in clinical practice?

中文

2025 年，FDA 和美國癌症研究協會（AACR）在《Clinical Cancer Research》聯合發表了三篇系列文章。這套文章合在一起，是現有同行評審文獻中，理解現代腫瘤藥物劑量最佳化最完整的教育框架。三篇文章並非重複——它們是循序漸進的，分別涵蓋藥物開發時間線的不同階段：如何選擇 FIH 起始劑量、如何進行早期劑量探索以積累多個候選劑量的可比較證據，以及如何為註冊試驗選擇的劑量提供理由。將三者作為一個整體閱讀，會揭示一個重要事實：劑量最佳化不是在第一期結束時做出的單一決策，而是一個持續積累證據的過程，從第一份方案開始，直到最終由全面證據支持的註冊劑量才告一段落。

第一篇（Zhu 等人，PMID 41036712）處理 FIH 起始劑量選擇。核心論點是：FIH 起始劑量不只是安全性練習——它是證據生成策略的第一步。一份好的 FIH 方案要問：在避免不可接受的第一劑傷害的同時，什麼是能提供有意義生物學訊號的最低劑量？這需要整合非臨床毒理學（NOAEL、HNSTD）、藥理活性估計（MABEL、PAD）、PK/PD 建模，以及最關鍵的——一個真正能從每個劑量層級學到東西的前瞻性遞增設計。文章強調，起始劑量、遞增規則和決策標準都應事先規定，並與同一套生物學假說相連。只宣告「以 X mg 起始劑量進行 3+3 遞增」而不解釋每個層級的什麼可觀察結果會改變下一個決策，是不夠的。

第二篇（Okusanya 等人，PMC 12614637）將討論移向早期劑量探索——大部分真正的劑量最佳化工作發生在這裡。文章引入「符合目的的試驗元素」概念——特別選擇以回答所需劑量最佳化問題的設計組件。這包括整合性生物標記（納入主要終點決策）、整合型生物標記（與臨床終點並行分析）和探索性生物標記（假說生成）。更具體地說，文章討論了回填 cohort（backfill）——在已被宣告安全但證據稀少的劑量層級重新納入病人，專門用來在該劑量建立更豐富的證據基礎。它還討論了隨機劑量擴增 cohort，病人被刻意分配到幾個候選劑量之一，而不是自動採用最高耐受劑量。臨床效益指數（CUI）被提出作為一種多屬性評分工具，可將反應概率和毒性概率結合成比較劑量的單一框架。所有這些工具共享一個共同目標：在提交上市申請時，帶有能真正回答「所提議劑量是否優於較低替代劑量」的資料。

第三篇（Shord 等人，PMID 41036557）處理註冊試驗劑量合理性——臨床醫師有時忽視的一個階段，假設藥物到達關鍵試驗時其劑量已是定論。這篇文章論點恰恰相反：許多註冊試驗按阻力最小的路徑繼承第一期遞增研究的劑量（最高耐受劑量 = RP2D = 第三期劑量），而這條鏈常有缺口。暴露-反應分析、模型輔助方法、允許基於積累資料調整劑量的無縫試驗設計，以及 CUI，都可以幫助填補這些缺口——但必須事先規劃，而不是事後補救。文章坦誠指出統計挑戰：使用相同資料進行劑量選擇和療效估計的適應性和無縫設計，需要仔細預先規定分析規則，否則關鍵結果的一型錯誤和解釋變得模糊。

在三部曲之外，Zhu 等人 2025 年《Clinical Pharmacology & Therapeutics》的文章（PMID 40248986）提出分子類別框架，是很好的配套閱讀。框架依作用模式將腫瘤藥物分為四類，並論證每類有不同的主要劑量最佳化問題。對小分子標靶藥和 ADC，問題是更高暴露是否真的產生更好的目標抑制或腫瘤控制——還是只是更多脫靶毒性。對大分子拮抗劑（如檢查點抑制劑），問題是受體佔有率和不完全與完全阻斷的臨床後果。對癌症免疫治療促效劑（T 細胞接合器、檢查點促效劑），問題是是否能步進達到免疫活化具有治療意義而不引發全身毒性的劑量。對單藥活性有限的分子，問題幾乎是哲學性的：如何對一個沒有明確療效訊號可優化的藥物進行劑量最佳化？

這些文章合在一起，主張現代腫瘤 FIH 試驗應設計為劑量學習系統，而非劑量尋找練習。開發的每個階段——起始劑量、遞增、擴增、註冊——都應產生能反饋並更新劑量假說的資料。醫師在這個系統中的角色不是被動的——接受第一期結束時宣告的 RP2D 並接受它。醫師的角色是在每個階段問：這個開發計畫是否產生了足夠的多屬性證據，來為將在臨床實踐中實際接受治療的病人證明這個劑量的合理性？

Key Concepts | 核心概念

• 三部曲架構 | Trilogy structure: Article 1 = FIH starting dose (biological hypothesis), Article 2 = early exploration (backfill + randomized expansion + biomarkers), Article 3 = registrational justification (totality of evidence). 第一篇 = FIH 起始劑量（生物學假說），第二篇 = 早期探索（回填 + 隨機擴增 + 生物標記），第三篇 = 註冊合理性（全面證據）。
• Backfill cohorts | 回填 cohort: Re-enrollment at already-cleared dose levels to build richer evidence at intermediate doses — turning sparse mid-range data into defensible dose comparisons. 在已通過安全審核的劑量層級重新納入，在中間劑量建立更豐富的證據——將稀少的中段資料轉化為可辯護的劑量比較。
• Clinical Utility Index | 臨床效益指數: A multi-attribute scoring tool combining response probability and toxicity probability — allows true dose comparison rather than defaulting to the highest tolerated. 結合反應概率和毒性概率的多屬性評分工具——允許真正的劑量比較而非預設採用最高耐受。
• Seamless trial design | 無縫試驗設計: Phase 1/2 or 2/3 designs that carry data from earlier stages into later dose decisions — requires careful pre-specification to avoid statistical inflation. 一/二期或二/三期設計，將早期階段資料帶入後期劑量決策——需要仔細預先規定以避免統計膨脹。
• Molecular-class framework | 分子類別框架: Four drug classes, each with distinct dominant dose-optimization questions — replacing one-size-fits-all MTD logic. 四種藥物類別，各有不同的主導劑量最佳化問題——取代一體適用的 MTD 邏輯。

Related Pages | 相關頁面

• project-optimus-dose-optimization — The regulatory initiative driving this series
• fih-starting-dose-selection — Deep dive into Article 1 concepts (MABEL, PAD)
• benefit-risk-dose-decision-framework — Practical implementation of Article 2 multi-attribute tools
• fda-2024-dose-optimization-guidance — The formal guidance document framing the series
• postmarketing-dose-optimization-risk-signals — What happens when the evidence chain is broken