FIH Starting Dose: The NOAEL/HNSTD to MABEL/PK-PD Framework
FIH 起始劑量：從 NOAEL/HNSTD 安全底線到 MABEL 與 PK/PD 框架
English
The decision about what dose to give the first human patient in an oncology trial is simultaneously one of the most consequential and least standardized decisions in clinical pharmacology. Unlike later dose optimization decisions — which have accumulating human data to draw on — the FIH starting dose must be chosen almost entirely from nonclinical evidence, mechanistic reasoning, and modeling assumptions. Getting it wrong in the direction of too high exposes the first cohort to potentially irreversible toxicity. Getting it wrong in the direction of too low condemns a succession of patients to prolonged treatment at doses with no meaningful pharmacological activity, extending the trial and straining the ethical bargain that advanced cancer patients accept when they enroll in a phase 1 study.
The regulatory vocabulary for starting dose selection spans several frameworks that were developed at different times and for different drug classes. Understanding which framework applies to which drug is the foundational clinical literacy that any oncologist evaluating a FIH protocol should possess.
NOAEL → HED → MRSD: The healthy volunteer language. The FDA’s 2005 guidance on maximum safe starting doses established NOAEL (no observed adverse effect level) as the anchor point: the highest dose in animal studies that produces no statistically or biologically significant increase in adverse effects. The NOAEL is converted to a human equivalent dose (HED) using body surface area or weight scaling, then divided by a safety factor (typically 10) to arrive at the maximum recommended starting dose (MRSD). This framework was designed for first-in-human studies in healthy adult volunteers, where no disease benefit exists to offset safety risk. Its conservatism is appropriate in that context. It is less appropriate as the sole starting dose logic for a drug being tested in patients with advanced cancer who have limited treatment alternatives.
STD10 / HNSTD: The oncology patient language. ICH S9, the nonclinical guidance specifically for anticancer pharmaceuticals, acknowledges that patients with advanced malignancies accept a higher risk-benefit ratio than healthy volunteers. For small-molecule cytotoxic drugs tested in rodent species, the conventional anchor is the STD10 — the dose that causes severe toxicity in 10% of animals — divided by 10 to give the human starting dose. When the non-rodent species is more predictive (as with many biologics or drugs with primate-specific targets), HNSTD (highest non-severely toxic dose) divided by 6 may be used instead. These are still backwards-looking safety anchors: they characterize what the animal can tolerate rather than what pharmacological activity the human body is predicted to encounter.
MABEL / PAD: The mechanism-first language. For drugs where the primary human risk is not classical toxicity from cellular death or organ damage, but rather from pharmacological action itself, the NOAEL and HNSTD frameworks can be dangerously misleading. The clearest examples are immunostimulatory agents: T-cell engagers, CD3 bispecific antibodies, cytokine agonists, and checkpoint costimulatory antibodies. For these agents, a monkey may show no toxicity at a given dose because the monkey’s immune system responds differently to the mechanism — but human T-cells or cytokine networks may be dramatically more sensitive. MABEL (minimal anticipated biological effect level) addresses this by anchoring the starting dose to the lowest concentration expected to produce detectable biological activity in the most sensitive human-relevant system, typically an in vitro or ex vivo assay. PAD (pharmacologically active dose) is a related concept — the minimum dose producing target pharmacological activity in the relevant animal model — and is often used alongside MABEL to bracket the expected activity range.
The FDA’s analysis of CD3 bispecific constructs (published in 2017) demonstrated that FIH doses derived from NOAEL-based or HNSTD-based approaches were often dangerously close to or above the clinically tolerable dose when the drug was given naively to a first patient, because the animal toxicology did not capture the first-dose cytokine release dynamics. The analysis found that doses corresponding to 10–30% pharmacological activity (as measured by receptor occupancy or T-cell activation assays) were safer starting points. This was a mechanistic correction, not a toxicological one.
PK/PD modeling: Integrating the frameworks. Modern FIH starting dose selection for complex biologics increasingly uses population PK/PD modeling to synthesize the above evidence into a probabilistic prediction of what dose will produce what concentration, what receptor occupancy, and what biological effect in the first human cohort. This is not a replacement for nonclinical safety data — it is a way of making the reasoning transparent, quantifying uncertainty, and defining what early clinical PK/PD measurements can confirm or refute about the starting dose hypothesis. The SNX281 STING agonist case (discussed in starting-dose-beyond-mabel) exemplifies this: the starting dose was derived from a population concentration-response model fitted to ex vivo whole-blood cytokine induction data, converted to a predicted human PK profile, and then subjected to a safety factor. When first-cohort PK was measured, it agreed with predictions within a factor of two — meaning the model was a testable hypothesis, not a black box.
A practical framework for oncology clinicians evaluating FIH starting dose rationale is to ask five questions in sequence. First: what is the primary risk mechanism for the first human patient? Classical cytotoxicity or organ toxicity (use NOAEL/HNSTD); immune activation or cytokine release (use MABEL/PAD); on-target normal tissue engagement (use a combination, noting which normal tissues express the target). Second: is the animal model pharmacologically relevant to the human target? If the target is human-specific and not expressed in the test species, the animal NOAEL may be nearly meaningless as a safety anchor. Third: what is the human-relevant assay used to derive MABEL, and how was variability across donors handled? Population-level uncertainty matters; a MABEL based on the average of five donor samples is less robust than one based on a population model that accounts for inter-individual variability. Fourth: does the escalation design allow the trial to learn quickly from early PK/PD data? A starting dose that is biologically plausible but still conservative should be paired with an escalation plan that can reach a pharmacologically active range within a reasonable number of cohorts, so that patients enrolled early in the trial have some probabilistic chance of receiving a dose with meaningful activity. Fifth: are the monitoring and stopping rules commensurate with the mechanism-specific risk? A MABEL-based start for a T-cell engager should be accompanied by sentinel dosing, hospitalization for the first dose, cytokine monitoring, step-up protocols, and pre-defined thresholds for pausing escalation — not just standard DLT criteria.
The 2024 FDA dose optimization guidance was explicit that it does not address FIH starting doses — that guidance pertains to dose optimization for drugs with existing clinical data seeking registration. This is an important clarification for clinicians who may conflate Project Optimus (which is about finding the right dose in the post-FIH development continuum) with the starting dose question (which is about not harming the first patient). They are connected — the starting dose must make the subsequent escalation informative — but they are not the same question.
中文
決定腫瘤試驗中第一位人類受試者應接受什麼劑量，是臨床藥理學中後果最重要也最缺乏標準化的決策之一。與後續的劑量最佳化決策不同——那些決策可以依賴累積的人體資料——FIH 起始劑量必須幾乎完全從非臨床證據、機轉推理與建模假設中選出。在過高方向出錯，可能讓第一個世代暴露於潛在不可逆的毒性。在過低方向出錯，則讓一連串病人長期接受沒有有意義藥理活性的劑量治療，延長試驗並損害晚期癌症病人在參加第一期研究時接受的倫理交換。
起始劑量選擇的監管詞彙涵蓋多個框架，這些框架在不同時期為不同藥物類別開發。了解哪個框架適用於哪種藥物，是任何評估 FIH 試驗方案的腫瘤科醫師應具備的基礎臨床素養。
NOAEL → HED → MRSD：健康志願者的語言。 FDA 2005 年最高安全起始劑量指引以 NOAEL（未觀察到不良反應劑量）作為錨點：動物研究中未產生統計或生物學顯著不良效應增加的最高劑量。NOAEL 透過體表面積或體重換算轉換為人體等效劑量（HED），再除以安全係數（通常為 10）得到最高建議起始劑量（MRSD）。這個框架是為健康成人志願者的首次人體研究設計，在那個情境中不存在疾病利益來抵消安全風險，其保守性是合適的。對於在治療選擇有限的晚期癌症病人中測試的藥物，它作為唯一的起始劑量邏輯就不太適合了。
STD10 / HNSTD：腫瘤病人的語言。 專門針對抗癌藥物的非臨床指引 ICH S9 承認，晚期惡性腫瘤病人接受的風險效益比高於健康志願者。對在囓齒類動物中測試的小分子細胞毒性藥物，傳統錨點是 STD10——造成 10% 動物嚴重毒性的劑量——除以 10 得到人體起始劑量。當非囓齒類動物物種更具預測性時（如許多生物製劑或具有靈長類特異性靶點的藥物），可改用 HNSTD（最高非嚴重毒性劑量）除以 6。這些仍然是往後看的安全錨點：它們描述動物能耐受什麼，而非預測人體將遇到什麼藥理活性。
MABEL / PAD：機轉優先的語言。 對於主要人體風險不是來自細胞死亡或器官損傷的古典毒性，而是來自藥理作用本身的藥物，NOAEL 和 HNSTD 框架可能產生危險的誤導。最清楚的例子是免疫刺激劑：T 細胞接合器、CD3 雙特異性抗體、細胞激素促效劑，以及免疫檢查點共刺激抗體。對這些藥物，猴子在某個劑量下可能沒有毒性，因為猴子的免疫系統對這種機轉的反應方式不同——但人類 T 細胞或細胞激素網絡可能對此敏感得多。MABEL（預期最低生物效應劑量）通過將起始劑量錨定到預期在最敏感的人體相關系統（通常是體外或離體分析）中產生可偵測生物活性的最低濃度來解決這個問題。PAD（藥理活性劑量）是相關概念——在相關動物模型中產生靶點藥理活性的最低劑量——常與 MABEL 一起使用以劃定預期活性範圍。
FDA 對 CD3 雙特異性構建體的分析（2017 年發表）顯示，當藥物樸素地給予第一位病人時，從 NOAEL 或 HNSTD 方法推導的 FIH 劑量，往往危險地接近或高於臨床可耐受劑量，因為動物毒理學未能捕捉首劑細胞激素釋放動態。分析發現，對應 10–30% 藥理活性（以受體佔有率或 T 細胞活化分析測量）的劑量是更安全的起始點。這是機轉修正，而非毒理學修正。
PK/PD 建模：整合框架。 現代複雜生物製劑的 FIH 起始劑量選擇越來越多使用族群 PK/PD 建模，將上述證據合成為第一個人體世代中什麼劑量將產生什麼濃度、什麼受體佔有率、以及什麼生物效應的概率預測。這不是取代非臨床安全資料——而是使推理透明化、量化不確定性，並定義早期臨床 PK/PD 測量可確認或反駁起始劑量假說的內容。SNX281 STING 促效劑案例（在 starting-dose-beyond-mabel 中討論）就是典範：起始劑量來自用體外全血細胞激素誘導資料建立的族群濃度—反應模型，轉換為預測的人體 PK 曲線，再套用安全係數。當第一個世代的 PK 被測量時，它與預測結果在兩倍以內吻合——意味著模型是可檢驗的假說，而非黑箱。
評估 FIH 起始劑量合理性的腫瘤科醫師實用框架，是依序問五個問題。第一：第一位人類病人的主要風險機轉是什麼？ 古典細胞毒性或器官毒性（使用 NOAEL/HNSTD）；免疫活化或細胞激素釋放（使用 MABEL/PAD）；靶點正常組織接合（使用組合，注意哪些正常組織表現靶點）。第二：動物模型對人類靶點是否具有藥理相關性？ 若靶點是人類特異性且在測試物種中不表現，動物 NOAEL 作為安全錨點可能幾乎沒有意義。第三：用於推導 MABEL 的人體相關分析方法是什麼，以及如何處理供體間的變異性？ 族群不確定性很重要；基於五個供體平均值的 MABEL，比基於考慮個體差異的族群模型的 MABEL 可靠性更低。第四：升量設計是否允許試驗從早期 PK/PD 資料中快速學習？ 生物學上合理但仍保守的起始劑量，應配合能在合理世代數內達到藥理活性範圍的升量計劃，使試驗早期入組的病人有一定概率接受有意義活性的劑量。第五：監測和停試規則是否與機轉特異性風險相稱？ T 細胞接合器的 MABEL 起始應配合哨兵給藥、首劑住院、細胞激素監測、逐步升量方案，以及暫停升量的預先定義門檻——而不只是標準 DLT 標準。
2024 年 FDA 劑量最佳化指引明確指出，該指引不涉及 FIH 起始劑量——那份指引針對的是具有現有臨床資料尋求申請核准的藥物的劑量最佳化。這是臨床醫師可能混淆 Project Optimus（關於在 FIH 後開發連續體中找到正確劑量）與起始劑量問題（關於不傷害第一位病人）的重要澄清。它們是相連的——起始劑量必須使後續升量具有資訊性——但它們不是同一個問題。
Key Concepts | 核心概念
| 術語 | 定義 | 最適用情境 |
|---|---|---|
| NOAEL | 動物研究中未觀察到不良效應的最高劑量 | 健康志願者、傳統小分子 |
| HED | 用體表面積換算的人體等效劑量 | NOAEL 的轉換步驟 |
| MRSD | HED 除以安全係數後的最高建議起始劑量 | FDA 2005 指引框架 |
| STD10 | 造成 10% 動物嚴重毒性的劑量 | 囓齒類動物的抗癌藥物 |
| HNSTD | 最高非嚴重毒性劑量（非囓齒類） | 靈長類相關生物製劑 |
| MABEL | 預期最低生物效應劑量（人體相關分析） | 免疫活化藥、T 細胞接合器 |
| PAD | 動物模型中最低藥理活性劑量 | 與 MABEL 並用以定義活性範圍 |
| Population PK/PD model | 整合不確定性的概率預測框架 | 複雜生物製劑、缺乏動物相關性時 |
Related Pages | 相關頁面
- starting-dose-beyond-mabel — Moving beyond MABEL for immune-activating drugs and bispecifics
- immune-engager-fih-rp2d-evidence-chain — How MABEL-based starts connect to RP2D without MTD
- ctdna-as-translational-endpoint — Biological evidence that complements starting dose rationale
- sy-5007-ret-inhibitor-case-study — A modern FIH applying these principles to a RET-targeted kinase inhibitor