AAI Internal Wiki

ctDNA-Informed Dose Finding: From Exploratory Appendix to Decision Framework

閱讀時間約 13 分鐘

ctDNA-Informed Dose Finding: From Exploratory Appendix to Decision Framework

ctDNA 導向劑量探索：從探索性附錄到決策框架

English

One of the structural weaknesses of traditional phase 1 oncology trial design is the temporal mismatch between what can be measured quickly and what matters clinically. Dose-limiting toxicities (DLTs) are observable within 21–28 days. A radiographic response — the conventional measure of drug activity — requires 6–12 weeks of treatment before imaging is meaningful. This gap has historically made efficacy invisible during dose escalation, forcing escalation committees to rely almost entirely on toxicity signals when deciding whether a given dose level is worth expanding or continuing toward. The practical consequence is that dose escalation has often been a one-dimensional exercise, with the biological plausibility of a dose being an afterthought rather than a co-equal input.

ctDNA offers a potential bridge across this temporal gap. Because ctDNA in plasma reflects the sum behavior of tumor cells shed into circulation, its kinetics can change within days to weeks of starting treatment — far earlier than a CT scan can detect volumetric change. The 2025 paper by Chen, Mozgunov, Baird, and Jaki in Statistical Methods in Medical Research proposed a formal biomarker-informed Bayesian dose-finding design that integrates repeated ctDNA measurements alongside toxicity and activity outcomes. In simulation studies, incorporating ctDNA trajectory into the escalation model improved target dose identification, reduced the time patients spent at doses with low biological activity, and shortened overall trial duration. The key conceptual move is treating ctDNA not as a secondary correlative analysis but as a prospectively defined input into the dose-finding algorithm itself.

What this requires in practice is more demanding than simply adding a ctDNA sample to the collection schedule. The assay must have defined analytic performance characteristics: limit of detection, limit of quantification, and — critically — the handling of samples where ctDNA is undetectable at baseline (which in some tumor types affects 30–50% of patients). The molecular response threshold must be pre-specified: is a 30% decline meaningful? A 50% decline? Complete clearance? The sampling schedule must be aligned with escalation decision cycles rather than convenient clinical visit windows. And the relationship between ctDNA change and downstream clinical endpoints — progression-free survival, overall response rate, overall survival — must be characterized, at least in the relevant disease setting, before ctDNA can be promoted from a noisy early signal to a decision-informing readout.

The 2025 JECCR perspective review proposed a pragmatic solution: phase 1 trials do not necessarily need deep mutational profiling for ctDNA to serve dose-finding purposes. Low-pass whole-genome sequencing, fragmentomics, and aneuploidy-based assays can track tumor fraction at lower cost and with faster turnaround than comprehensive genomic profiling. If the goal is to ask “did this dose level change tumor biology?” rather than “which resistance mutation emerged?”, a tumor fraction readout may be sufficient for early cohort decisions — and fast enough to matter.

The GIM21 ctDNA-RECIST proof-of-concept study (Journal of Experimental & Clinical Cancer Research 2026) introduced an important caveat that any clinician teaching ctDNA-informed decisions must confront: ctDNA waves. In 50 patients with HER2-positive metastatic breast cancer receiving trastuzumab emtansine (T-DM1), the investigators mapped 466 ctDNA time points against 113 RECIST-1.1 imaging assessments. They found that ctDNA responses were often deeper than radiographic responses — but also that serial ctDNA measurements showed oscillating increases and decreases that made any single ctDNA time point an unreliable basis for stopping treatment. Applying a naive rule of “stop at first ctDNA increase” would have prematurely terminated treatment in patients who went on to achieve durable responses. The paper therefore proposed a personalized ctDNA-guided algorithm using multiple consecutive measurements rather than single-point triggers — a finding that generalizes directly to FIH trials: if a ctDNA-informed escalation or stopping rule is going to be written into a protocol, it must account for biological noise and require confirmatory measurements.

The FDA workshop on ctDNA as an early drug development endpoint (August 2024) identified three questions that any sponsor using ctDNA in a phase 1 decision framework must answer. First: what constitutes a pre-specified molecular response — what magnitude of change, over what time frame, measured by what assay, is clinically meaningful? Second: how will low or undetectable baseline ctDNA be handled — can patients without detectable ctDNA participate, and if so, are they evaluable for the ctDNA-informed decision endpoint? Third: what is the bridging evidence that links ctDNA change to outcomes that regulators and patients care about?

The ASCO 2025 clinical update (BJC Reports) reinforced a distinction that is easy to blur in early enthusiasm: prognostic value is not the same as predictive utility, and predictive utility is not the same as clinical utility. ctDNA correlates with survival in many settings — that is prognostic value. Whether a ctDNA-informed treatment decision improves outcomes relative to the counterfactual is a different and harder question. SERENA-6, which randomized patients to switch endocrine therapy when ctDNA detected an ESR1 mutation before imaging progression in HR-positive/HER2-negative metastatic breast cancer, demonstrated improved PFS — suggesting that ctDNA-guided intervention can have clinical utility in the right context. But the DYNAMIC-III trial in resected stage III colon cancer showed that ctDNA-informed treatment escalation did not improve recurrence-free survival. The lesson for FIH: ctDNA entering a dose-finding decision must have a plausible mechanistic case for why ctDNA change at that dose level, in that tumor type, under those pharmacological conditions, should be treated as a reliable early signal — not just because ctDNA is a sophisticated technology.

In summary, the transition from ctDNA as exploratory appendix to ctDNA as dose-finding input requires four things: a pre-specified molecular response definition, a sampling schedule aligned with cohort decisions, an assay capable of reliable turnaround, and a conceptual framework linking ctDNA change to the biological activity question the trial is designed to answer. Without all four, ctDNA remains decorative — adding scientific texture to a paper without adding decision quality to the trial.

中文

傳統第一期腫瘤臨床試驗設計存在一個結構性弱點：可快速量測的指標與臨床真正重要的指標之間存在時間差。劑量限制毒性（DLT）在 21 至 28 天內可觀察到。而影像腫瘤反應——傳統的藥物活性衡量標準——需要 6 至 12 週的治療才能有意義地評估。這個時間差讓藥物有效性在劑量升量過程中幾乎隱形，迫使升量委員會在決定某個劑量層級是否值得擴增或繼續時，幾乎完全依賴毒性訊號。實際結果是：劑量升量往往成為一維度練習，劑量的生物學合理性被當成事後考量，而非與毒性並列的核心輸入。

ctDNA 提供了跨越這個時間差的橋梁。由於血漿中的 ctDNA 反映腫瘤細胞釋放到循環中的總和行為，其動態可在開始治療後數天至數週內變化——遠早於電腦斷層可偵測到體積變化。Chen、Mozgunov、Baird 與 Jaki 2025 年在 Statistical Methods in Medical Research 提出的生物標記導向貝氏劑量探索設計，正式將重複 ctDNA 測量與毒性及活性結果整合。模擬研究顯示，將 ctDNA 動態軌跡納入升量模型可改善目標劑量的識別、減少病人停留在生物活性很低劑量的時間，並縮短整體試驗時間。這個概念的關鍵轉移，是把 ctDNA 從次要的相關性分析，提升為預先定義的升量演算法輸入。

這在實踐中的要求比簡單地在採樣計劃中加入 ctDNA 採血更高。分析方法必須有明確的性能特性：偵測下限、定量下限，以及——關鍵地——對基線 ctDNA 不可偵測樣本的處理方式（在某些癌種中這影響 30–50% 的病人）。分子反應門檻必須預先規定：下降 30% 有意義嗎？50%？完全清除？採血時程必須與升量決策週期對齊，而不只是配合方便的門診就診時間。ctDNA 變化與下游臨床終點——無惡化存活期、整體反應率、整體存活期——之間的關係，至少在相關疾病情境中必須有所描述，ctDNA 才能從雜訊早期訊號晉升為影響決策的讀出值。

2025 年 JECCR 前瞻性評論提出了務實解決方案：第一期試驗未必需要深度突變分析，ctDNA 才能服務劑量探索目的。低覆蓋率全基因體定序、片段體學與倍數體異常檢測，可以比全面基因體分析更低成本、更快速地追蹤腫瘤分率。如果目標是問「這個劑量層級是否改變了腫瘤生物學？」而非「出現了哪種抗藥性突變？」，腫瘤分率讀出可能對早期世代決策就已足夠——且快得及時發揮作用。

GIM21 ctDNA-RECIST 概念驗證研究（Journal of Experimental & Clinical Cancer Research 2026）引入了任何教授 ctDNA 導向決策的臨床醫師都必須面對的重要警示：ctDNA 會波動。在 50 位接受 trastuzumab emtansine（T-DM1）治療的 HER2 陽性轉移性乳癌病人中，研究者將 466 個 ctDNA 時間點與 113 個 RECIST 1.1 影像評估進行對應。他們發現 ctDNA 反應常比影像反應更深，但連續 ctDNA 測量也顯示出波動性的上升和下降，使任何單一 ctDNA 時間點都不可靠地成為停止治療的依據。套用「一旦 ctDNA 上升就停藥」的簡單規則，將會過早終止某些最終獲得持久反應的病人的治療。這篇論文因此提出使用多個連續測量點而非單點觸發的個人化 ctDNA 導引演算法——這個發現直接推廣到 FIH 試驗：如果 ctDNA 導向的升量或停藥規則要寫進試驗方案，它必須考慮生物學雜訊並要求確認性測量。

FDA 2024 年 ctDNA 作為早期藥物開發終點的研討會確立了三個問題，任何在第一期決策框架中使用 ctDNA 的廠商都必須回答。第一：什麼構成預先規定的分子反應——什麼幅度的變化、在什麼時間框架內、用什麼分析方法、才算有臨床意義？第二：基線 ctDNA 很低或不可偵測時如何處理？第三：連結 ctDNA 變化與監管機構和病人真正關心的結果的橋接證據是什麼？

總結而言，ctDNA 從探索性附錄過渡到劑量探索輸入需要四件事：預先定義的分子反應定義、與世代決策對齊的採樣時程、能可靠及時回報的分析方法，以及將 ctDNA 變化連接到試驗所要回答的生物活性問題的概念框架。缺少其中任何一項，ctDNA 就只是裝飾性的——為論文增添科學質感，卻沒有為試驗提升決策品質。

Key Concepts | 核心概念

  • Biomarker-informed dose-finding: a trial design where an early biomarker (ctDNA) is pre-specified as a decision input alongside toxicity endpoints, not relegated to post-hoc analysis.
  • Molecular response threshold: a pre-specified magnitude of ctDNA change (e.g., ≥50% decline, clearance) that defines a “responder” at the molecular level. Must be established before the trial begins.
  • ctDNA waving: oscillating ctDNA rises and falls observed on serial measurements that do not necessarily indicate treatment failure; single time points are unreliable stopping triggers.
  • Tumor fraction: the proportion of plasma cell-free DNA that is tumor-derived; low baseline tumor fraction limits assay interpretability.
  • Clinical utility vs. clinical validity: clinical validity = ctDNA change correlates with outcomes; clinical utility = acting on ctDNA change improves outcomes. The former is necessary but not sufficient for the latter.
  • Bayesian adaptive design: a statistical framework that updates dose-decision probabilities as data accumulate, enabling incorporation of both toxicity and ctDNA biomarker signals.

Related Pages | 相關頁面

關係圖譜

反向連結