Core answer. Because route of administration is set by the size ladder, not by formulation ambition. Aspirin is ~180 Da, a peptide ~2000 Da, insulin ~58 kDa, an antibody ~150 kDa. Small molecules are usually oral; biologics are degraded if given orally and must be injected or infused.

核心答案。 因為給藥途徑是由分子大小的階梯決定，而非由配方的企圖決定。Aspirin 約 180 Da，一個肽約 2000 Da，insulin 約 58 kDa，一個抗體約 150 kDa。小分子通常可口服；生物製品若口服會被分解，因此必須注射或輸注。

Mechanism — why it follows. Peptides have two structural problems named in the lecture: rapid protease degradation, and poor oral bioavailability because of stomach acid. A protein is larger again, so the same barriers apply with a bigger molecule. The consequence is a chain, not a single inconvenience: parenteral administration → a sterile injectable product → heat- and light-sensitive material with a microbial contamination risk → cold chain → administration by a trained person or a device → a clinic or infusion visit for anything that cannot be given subcutaneously.

機制 —— 為何如此推得。 肽在課堂中被點出兩個結構問題：快速的蛋白酶降解，以及因胃酸而導致的口服生物利用度偏低。蛋白的分子更大，因此同樣的屏障會以更大的分子規模出現。其後果是一條連鎖，而非單一的不便：腸胃外給藥 → 無菌注射產品 → 對熱與光敏感、具微生物污染風險的物料 → cold chain → 由受訓人員或裝置給藥 → 對於任何無法皮下給藥者，需前往診所或接受輸注。

Where it stops being true. This is a rule about size and chemistry, not a law of biology. The lecture's whole peptide programme exists because cyclisation, D-/unnatural amino acids and lipidation make peptides survive the gut — and lipidation works partly by increasing serum-protein binding, which prolongs circulation. So "biologic" does not strictly mean "must be injected"; it means a molecule that has not been engineered around its own degradation.

邊界 —— 何時不再成立。 這是一條關於大小與化學的規則，而非生物學定律。課堂中整個肽計劃之所以存在，是因為環化、D-／非天然胺基酸與脂化讓肽得以在腸道中存活——而脂化的作用部分來自增加血清蛋白的結合，從而延長循環時間。因此「生物製品」嚴格來說並不表示「必須注射」；它指的是一個尚未被工程化改造以繞過自身降解的分子。

Core answer. Because it converted biologics from an extraction industry into an information industry. Before it, the product was defined by a tissue you had to collect; after it, the product is defined by a gene you can write down.

核心答案。 因為它把生物製品從一個萃取產業轉變為一個資訊產業。在此之前，產品由你必須採集的組織所定義；在此之後，產品由一個你可以寫下的基因所定義。

Mechanism — why it follows. Insulin was discovered in 1921–22 and reached its first patient (Leonard Thompson, 11 Jan 1922) only after Collip purified the extract; thereafter it was animal-sourced, notably from pig pancreas. That supply is bounded by how many pancreases exist, and an animal protein is foreign to a human. Recombinant DNA broke both constraints: 1972 Berg recombined DNA from two different viruses; 1973 Cohen & Boyer put recombinant DNA into bacteria, creating the first GMOs and showing bacteria could express a foreign gene; 1977 Genentech produced somatostatin in engineered E. coli as proof of concept; 1978 the human insulin gene was cloned into E. coli; 1982 the FDA approved Humulin, the first approved recombinant therapeutic.

機制 —— 為何如此推得。 Insulin 於 1921–22 年被發現，且要等到 Collip 純化該萃取物之後，才首次用於患者身上（Leonard Thompson，1922 年 1 月 11 日）；此後它來自動物來源，尤其是豬胰臟。該供應量受制於存在多少個胰臟，而動物蛋白對人體而言是外來的。重組 DNA 突破了這兩項限制：1972 年 Berg 重組了來自兩種不同病毒的 DNA；1973 年 Cohen 與 Boyer 將重組 DNA 置入細菌，創造出首批 GMOs，並顯示細菌可以表現外來基因；1977 年 Genentech 在經工程改造的 E. coli 中生產 somatostatin 作為概念驗證；1978 年人類 insulin gene 被選殖進 E. coli；1982 年 FDA 核准 Humulin，即首項獲核准的重組治療藥物。

Where it stops being true. Keep the two milestones apart: 1977 is the first recombinant protein, 1982 is the first approved recombinant drug. And the platform is not universal — bacterial hosts cannot perform human glycosylation, so any glycoprotein product pushes you into mammalian cell culture, and each new host is a new process to solve.

邊界 —— 何時不再成立。 要把兩個里程碑分開：1977 年是第一個重組蛋白，1982 年是第一個獲核准的重組藥物。而該平台並非通用——細菌宿主無法進行人類的糖基化，因此任何 glycoprotein 產品都會把你推向哺乳動物細胞培養，而每一個新宿主都是一個待解決的新製程。

Core answer. Because a biologic cannot be fully defined by its structure, so approval cannot attach to the molecule alone — it attaches to the process that made it. Small molecules are characterised precisely and fully defined; biologics are complex and batch-variable, which is what the phrase "the process is the product" means.

核心答案。 因為一項生物製品無法單憑其結構被完整定義，因此核准不能只附著於分子本身——它附著於製造該分子的製程。小分子可被精確鑑定並完整定義；生物製品則複雜且批次間存在變異，這正是「製程即產品」這句話的意思。

Mechanism — why it follows. If the product is a process, then two products with identical amino-acid sequences made in different host cells need not be the same product, because the glycan differs by host. Because the glycan governs circulation time, solubility, folding and receptor binding, a process difference shows up as a clinical difference, not just an assay difference. Regulatory similarity therefore has to be demonstrated, not asserted from sequence.

機制 —— 為何如此推得。 若產品即是一套製程，那麼在相同胺基酸序列下於不同宿主細胞製造的兩項產品，未必是同一項產品，因為 glycan 會隨宿主而不同。由於 glycan 主導循環時間、溶解度、摺疊與受體結合，製程差異會以臨床差異的形式顯現，而不只是檢測上的差異。因此，法規上的相似性必須被證明，而不能僅從序列逕行宣稱。

Where it stops being true. It is a spectrum, not a binary. The more tightly a molecule can be defined, the more it behaves like a small molecule: an aptamer, for instance, is chemically synthesised with no cell culture at all, which makes it cheaper and more reproducible. And for a protein whose activity depends on sequence alone, the glycan may matter less than the framework implies.

邊界 —— 何時不再成立。 這是一道光譜，而非二元劃分。一個分子愈能被嚴謹定義，其行為就愈像小分子：舉例而言，一個 aptamer 是以化學方式合成、完全無需細胞培養，這使它更便宜且更可重現。而對於活性僅取決於序列的蛋白而言，glycan 的影響可能比這個框架所暗示的更小。

Core answer. Because the chemistry addresses the exact property that made peptides commercially unattractive — stability — and the remaining problem is now a formulatory one rather than a fundamental one.

核心答案。 因為這些化學所處理的，正是使肽在商業上缺乏吸引力的那項特性——穩定性——而剩下的問題如今是配方層面的問題，而非根本性的問題。

Mechanism — why it follows. Peptides have two core problems: rapid protease degradation and poor oral bioavailability. Cyclisation attacks both: a rigid ring resists enzymes (longer half-life), improves binding affinity and specificity by fixing the conformation, improves membrane crossing, reduces immunogenicity, and gives a more favourable PK profile so dosing can be less frequent. Two routes exist — disulfide side-chain-to-side-chain (oxytocin) and head-to-tail amide. Ziconotide shows disulfide constraint alone counts: three disulfides and a pseudocyclic structure, from sea-snail venom. Add D-/unnatural residues (Gramicidin) and lipidation, and you have three levers on one problem.

機制 —— 為何如此推得。 肽有兩個核心問題：快速的蛋白酶降解，以及偏低的口服生物利用度。環化同時針對兩者：剛性的環能抵抗酶（更長的半衰期）、藉由固定構形提升結合親和力與專一性、改善穿膜能力、降低免疫原性，並帶來更有利的 PK 特性，使給藥頻率得以降低。存在兩條路徑——雙硫鍵的側鏈對側鏈（oxytocin）與頭對尾的醯胺鍵。Ziconotide 顯示僅靠雙硫鍵的拘束就能奏效：三個雙硫鍵與一個擬環結構，來自海螺毒液。再加上 D-／非天然殘基（Gramicidin）與脂化，你就有三個作用於同一問題的槓桿。

Where it stops being true. Stability is not permeability. The lecture makes a pointed observation: humans do not produce cyclic peptides — nature's cyclic peptides come from plants and bacteria. That means a cyclotide scaffold is foreign, so an immunogenicity question remains even if cyclisation reduces it, and it means the permeability problem is separate from the protease problem. A peptide can be perfectly stable and still not be absorbed.

邊界 —— 何時不再成立。 穩定性不等於通透性。課堂提出一個尖銳的觀察：人類不產生 cyclic peptides——自然界的 cyclic peptides 來自植物與細菌。這意味著一個 cyclotide 骨架是外來的，因此即使環化降低了免疫原性，相關疑問依然存在；也意味著通透性問題與蛋白酶問題是彼此獨立的。一個肽可以完全穩定，卻仍然不被吸收。

Core answer. Size. A camelid single-domain antibody is ~12–15 kDa against ~150 kDa for an IgG, so it penetrates tissue better and can reach hidden or densely packed epitopes — and because it has no Fc, it also loses everything the Fc does.

核心答案。 大小。一個 camelid 單域抗體約 12–15 kDa，相對於一個 IgG 的約 150 kDa，因此它能更好地穿透組織，並能到達隱藏或排列緊密的 epitopes——而由於它沒有 Fc，它也同時失去了 Fc 所負責的一切功能。

Mechanism — why it follows. Take the advantages first: small size means better tissue penetration and access to epitopes a full antibody cannot reach; high stability against extreme pH and temperature; and ease of production in bacterial and yeast systems, which is a cost advantage. Now the bill: rapid clearance through the kidneys gives a short half-life; no Fc means no Fc receptor engagement, so no ADCC, no CDC, no phagocytosis, and no FcRn recycling to extend circulation; and camelid sequence is foreign to humans.

機制 —— 為何如此推得。 先看優勢：小尺寸意味著更好的組織穿透，以及能接觸到完整抗體無法到達的 epitopes；對極端 pH 與溫度具高穩定性；以及易於在細菌與酵母系統中生產，這是一項成本優勢。接著是帳單：經由腎臟的快速清除造成半衰期短；沒有 Fc 就沒有 Fc receptor 的接合，因此沒有 ADCC、沒有 CDC、沒有吞噬作用，也沒有 FcRn 回收以延長循環；而 camelid 序列對人類而言是外來的。

Where it stops being true. The lecture lists both "low immunogenicity" and "immunogenicity concerns" for the same molecule, and both can be honest: a simplified structure is less likely to provoke a response, while a camelid framework may be recognised as foreign. Which dominates is a framework-engineering question, not a property of the format — so treat the claim as a design intent to be tested, not a given. Likewise, losing effector function only costs you something if your mechanism needed it.

邊界 —— 何時不再成立。 課堂對同一個分子同時列出「低免疫原性」與「免疫原性疑慮」，而兩者都可以是誠實的：一個簡化的結構較不易引發反應，而一個 camelid 骨架則可能被辨識為外來物。何者佔主導是一個骨架工程的問題，而非該格式本身的屬性——因此應把這個 claim 視為待驗證的設計意圖，而非既定事實。同樣地，失去 effector function 只有在你的機制需要它時，才會造成損失。

Core answer. Because the Fc is one module performing two unrelated jobs, and each job is read by a different receptor with its own structural requirements. You are not choosing between two settings of one dial; you are trading one interface against another on the same surface.

核心答案。 因為 Fc 是一個執行兩項互不相關任務的模組，而每一項任務由不同的受體讀取，各有其自身的結構要求。你不是在同一個旋鈕的兩種設定之間選擇；你是在同一個表面上，以一個介面換取另一個介面。

Mechanism — why it follows. The Fc binds Fc receptors on macrophages, neutrophils and NK cells, producing phagocytosis and ADCC (NK cells release perforin and granzymes, inducing apoptosis in the target); it binds C1q, producing complement-mediated lysis; and it binds the neonatal Fc receptor, FcRn, which protects the antibody from degradation and gives both the long half-life and placental transfer. Fc glycosylation — including fucose content, the standard glycoengineering lever for raising ADCC via Fcγ receptor affinity — sits on the same region that carries FcRn binding. Strengthen one and you risk perturbing the other, and the glycan pattern is set by the host cell line, so it is a CMC problem as well as a design one.

機制 —— 為何如此推得。 Fc 會與 macrophages、neutrophils 與 NK cells 上的 Fc receptors 結合，產生吞噬作用與 ADCC（NK cells 釋放 perforin 與 granzymes，誘導標的細胞凋亡）；它會與 C1q 結合，產生補體介導的裂解；它也會與新生兒 Fc receptor，即 FcRn 結合，後者保護抗體免於降解，並同時賦予長半衰期與胎盤轉運。Fc 的糖基化——包括 fucose 含量，即以 Fcγ 受體親和力提升 ADCC 的標準糖工程槓桿——位於與 FcRn 結合相同的區域之上。強化其中一項，你就有擾動另一項的風險，而 glycan 的模式由宿主細胞株決定，因此這既是設計問題，也是 CMC 問題。

Where it stops being true. The conflict dissolves when neither function is the mechanism of action. A mAb that blocks a receptor or neutralises a cytokine does not need ADCC at all — and may be better without it, because the same cytolytic activity that kills a tumour cell will also deplete healthy cells carrying the same antigen in a patient dosed for years. In that case silencing effector function is not a compromise, it is the correct design, and the only optimisation left is PK.

邊界 —— 何時不再成立。 當兩項功能都不是作用機制時，這個衝突便消解。一個阻斷受體或中和 cytokine 的 mAb 完全不需要 ADCC——而且沒有它可能更好，因為在一個需用藥多年的患者身上，同一種殺死腫瘤細胞的細胞裂解活性，也會耗竭帶有相同抗原的健康細胞。在這種情況下，沉默 effector function 不是妥協，而是正確的設計，而唯一剩下的優化就是 PK。

Core answer. Because a bispecific built from four chains needs two different heavy chains and two different light chains to pair correctly inside one cell, and biology provides no natural pairing mechanism.

核心答案。 因為一個由四條鏈構成的 bispecific，需要在同一個細胞內讓兩條不同的重鏈與兩條不同的輕鏈正確配對，而生物學並未提供天然的配對機制。

Mechanism — why it follows. Co-express the chains and they assemble at random, giving a mixture of mispaired species alongside the intended molecule. Each wrong species is a product-related impurity that has to be detected, quantified and removed — which converts a clever binding concept into a release-specification problem. The standard fix, knob-into-hole mutations, is engineering that solves chain pairing rather than anything pharmacological. Compare a normal mAb, where manufacturability is standard: that gap is the bottleneck.

機制 —— 為何如此推得。 共同表現這些鏈時，它們會隨機組裝，產生一批錯配物種，與目標分子混在一起。每一種錯誤物種都是產品相關的不純物，必須被檢測、定量與移除——這把一個巧妙的結合概念轉變成一個放行規格的問題。標準的解決方式，即 knob-into-hole 突變，是一種解決鏈配對的工程手段，而非任何藥理學上的作為。對比一個正常的 mAb，其可製造性是標準配備：這個落差就是瓶頸。

Where it stops being true. The argument is specific to IgG-like bispecifics assembled from four chains. Non-IgG formats such as scFv constructs avoid the Fc, and with it the effector functions and the long half-life, and are themselves hard to manufacture. And a chemical route can sidestep pairing entirely: the pBsAb attaches an EGFR-binding peptide to an existing antibody through a bifunctional linker, cutting conjugation from 2–3 days to 6 hours without denaturing the protein. Two arms, one expression system — no chain-pairing statistics at all.

邊界 —— 何時不再成立。 這項論證是針對由四條鏈組裝而成的 IgG 樣 bispecifics。非 IgG 格式如 scFv 構建體避開了 Fc，連帶也避開了 effector functions 與長半衰期，而它們本身同樣難以製造。而一條化學路徑可以完全繞開配對問題：pBsAb 透過一個雙功能 linker，將一個結合 EGFR 的肽接附到既有的抗體上，把接合時間從 2–3 天縮短到 6 小時，且不使蛋白變性。兩個臂，一個表現系統——完全沒有鏈配對的統計問題。

Core answer. Because each format asks a different thing of the target cell. Blocking signalling, delivering a cytotoxic payload, and recruiting an immune effector cell are three separate biological events, and the tumour has to be capable of the one you chose.

核心答案。 因為每一種格式對標的細胞有不同的要求。阻斷訊息傳遞、遞送一個細胞毒性 payload，以及募集一個免疫 effector cell，是三件彼此獨立的生物學事件，而腫瘤必須有能力完成你所選的那一件。

Mechanism — why it follows. A naked IgG works by inhibitory mechanisms (receptor blocking, ligand neutralisation, receptor downregulation) or by effector mechanisms (ADCC, phagocytosis, CDC). An ADC requires the antigen to be internalised: bind → endocytosis → linker cleavage → cytotoxic action → selective killing → apoptosis. Notice where the linker sits in that sequence — after internalisation. A bispecific recruits a second cell type: the pBsAb engages EGFR with one arm and the SIRP-α/CD47 axis with the other, blocking the "don't eat me" signal so macrophages phagocytose the tumour cell. The two arms carry deliberately different binding energies (−8.3 kcal/mol and −6.7 kcal/mol) because one is the address and one is the function.

機制 —— 為何如此推得。 Naked IgG 透過抑制性機制（受體阻斷、配體中和、受體下調）或透過 effector 機制（ADCC、吞噬作用、CDC）發揮作用。一個 ADC 需要抗原被內化：結合 → 胞吞作用 → linker 裂解 → 細胞毒性作用 → 選擇性殺傷 → 凋亡。注意 linker 在這序列中的位置——在內化之後。一個 bispecific 募集第二種細胞類型：pBsAb 以一個臂接合 EGFR，以另一個臂接合 SIRP-α/CD47 軸，阻斷「不要吃我」的訊號，使 macrophages 吞噬腫瘤細胞。兩個臂刻意帶有不同的結合能（−8.3 kcal/mol 與 −6.7 kcal/mol），因為一個是地址，一個是功能。

Where it stops being true. The format fails on a condition you cannot engineer around. If the antigen is not internalised, the ADC's linker never meets its trigger and the payload stays outside the cell, however potent it is — you are left with a circulating conjugate releasing a toxin that is by definition too dangerous to give free. If the tumour does not display the second target, the bispecific's second arm is inert and you have paid for complexity that does nothing.

邊界 —— 何時不再成立。 格式會在一項你無法以工程繞過的條件上失效。若抗原不被內化，ADC 的 linker 永遠遇不上它的觸發條件，payload 便留在細胞外，無論它多麼有效——你只剩下一個循環中的接合物，持續釋放一種按定義過於危險而不可能單獨給藥的毒素。若腫瘤不呈現第二個標的，bispecific 的第二個臂便是惰性的，而你為一個毫無作用的複雜度付出了代價。

Core answer. Because the sequence is cheap and fast to change, while getting a large, charged, nuclease-labile polymer into the right cell is neither.

核心答案。 因為序列的更改既便宜又快速，而要把一個巨大、帶電、易受 nuclease 作用的聚合物送入正確的細胞，則兩者皆非。

Mechanism — why it follows. DNA can be built two ways: solid-phase chemical synthesis, where ammonium hydroxide cleaves the chain off the resin and strips the protecting groups, or enzymatically by PCR. Contrast L2, where proteins must be expressed by living cells. That asymmetry is the whole commercial argument for mRNA: no complex protein process to build, the same manufacturing line with a different sequence, no mammalian culture to scale up. But the molecule still has to survive nucleases, cross a membrane it cannot cross, reach the correct tissue and avoid rapid clearance. Every one of those is a delivery problem, which is why the lecture spends its second half on liposomes, micelles, mesoporous particles, exosomes, red blood cells and cell-penetrating peptides.

機制 —— 為何如此推得。 DNA 可以用兩種方式建構：固相化學合成，即以氫氧化銨將鏈從樹脂上切下並剝除保護基；或以 enzymatic 方式經 PCR 建構。對比 L2，在那裡蛋白必須由活細胞表現。這項不對稱正是 mRNA 的整個商業論據：無需建立複雜的蛋白製程、同一條生產線只需更換序列、無需放大哺乳動物培養。但該分子仍必須存活於 nucleases 之中、跨越它無法跨越的膜、到達正確的組織並避免快速清除。上述每一項都是遞送問題，這就是為何課堂把後半段用在 liposomes、micelles、中孔粒子、exosomes、紅血球與細胞穿透肽上。

Where it stops being true. The claim is about scarcity, not importance. Where delivery to the target organ is already solved, the vehicle becomes reusable and the sequence becomes the variable again — exactly what makes Lumasiran's silencing of HAO1 in liver a platform rather than a one-off. And a superb vehicle carrying the wrong sequence helps nobody: both inputs are necessary, but only one is the current rate-limiting step.

邊界 —— 何時不再成立。 這個 claim 談的是稀缺性，而非重要性。當遞送至標的器官的問題已被解決，載體便變得可重複使用，序列則再次成為變數——這正是使 Lumasiran 在肝臟中沉默 HAO1 成為一個平台、而非一次性成果的原因。而一個承載錯誤序列的卓越載體幫不了任何人：兩項投入都是必要的，但只有一項是當前限速的步驟。

Core answer. Because a polysaccharide alone does not reliably activate T cells, so it produces a weak response with poor memory — and the people who need protection most, adults over 65 and the immunocompromised, are exactly those who cannot compensate for that weakness.

核心答案。 因為單獨一個 polysaccharide 無法可靠地活化 T 細胞，因此產生微弱且記憶不良的反應——而最需要保護的人，即 65 歲以上的成人與免疫受損者，正是那些無法補償這項弱點的人。

Mechanism — why it follows. PPSV23 covers 23 serotypes of Streptococcus pneumoniae. The problem is not the antigen choice but the immune route it takes. Protein vaccines run the full chain — APC takes up antigen, displays fragments on MHC, migrates to the lymph node, activates naive T cells, helper T cells release cytokines that stimulate B cells, B cells become plasma cells and memory B cells, and antibody plus cytotoxic T cells give durable protection. A bare polysaccharide skips the T-cell step. The fix is to conjugate it to a carrier protein, an inactivated bacterial toxin, which recruits T-cell help and converts the response into the T-dependent one. The same principle sits behind Hib and pneumococcal conjugate vaccines.

機制 —— 為何如此推得。 PPSV23 涵蓋 Streptococcus pneumoniae 的 23 種血清型。問題不在抗原的選擇，而在它所走的免疫路徑。蛋白疫苗走完完整鏈條——APC 攝取抗原、在 MHC 上呈現片段、遷移至淋巴結、活化 naive T 細胞，helper T 細胞釋放 cytokines 刺激 B 細胞，B 細胞成為漿細胞與記憶 B 細胞，而抗體加上細胞毒性 T 細胞提供持久保護。一個裸露的 polysaccharide 跳過了 T 細胞這一步。解決方式是將其接合到一個載體蛋白，即一種去活化的細菌毒素，藉此募集 T 細胞的協助，並將反應轉換為 T 依賴型。同一原理也支撐著 Hib 與肺炎鏈球菌接合疫苗。

Where it stops being true. The fix improves the response, it does not abolish the host problem: T-cell help still has to come from the patient, so an immunocompromised patient may respond poorly to a conjugate too — a clinical judgement, not a rule. And the carrier adds its own manufacturing burden, a controlled conjugation chemistry plus a second biological component, so the trade is not free.

邊界 —— 何時不再成立。 此解決方式改善反應，但並未消除宿主方面的問題：T 細胞的協助仍必須來自患者本身，因此免疫受損的患者對一個接合疫苗也可能反應不佳——這是臨床判斷，而非通則。此外，載體本身帶來製造負擔，即一套受控的接合化學，加上第二項生物學成分，因此這項取捨並非沒有代價。

Core answer. Because the glycan is determined by the host cell line and the process, not by the gene. You cannot reproduce a glycoprotein from its amino-acid sequence alone, so choosing the host is choosing the product.

核心答案。 因為 glycan 由宿主細胞株與製程決定，而非由基因決定。你無法僅憑胺基酸序列重現一個 glycoprotein，因此選擇宿主就是選擇產品。

Mechanism — why it follows. Glycosylation is not decoration; it does four load-bearing jobs, each with a failure mode. It protects the protein from degradation, so a wrong glycan means rapid clearance and more frequent dosing. It enhances solubility in blood, so a wrong glycan means aggregation and injection-site problems. It is required for correct folding, so a wrong glycan means misfolding and loss of function. It is essential for receptor binding, so a wrong glycan means the drug simply does not work. In cetuximab the consequences are visible on four axes: glycosylation affects binding efficiency to EGFR, Fc-region glycosylation affects ADCC, proper glycosylation lowers the chance of the antibody being seen as foreign, and the glycan influences distribution and concentration at the tumour site.

機制 —— 為何如此推得。 糖基化不是裝飾；它承擔四項支撐性的任務，每一項都有其失效模式。它保護蛋白免於降解，因此錯誤的 glycan 意味著快速清除與更頻繁的給藥。它增進在血液中的溶解度，因此錯誤的 glycan 意味著聚集與注射部位的問題。它是正確摺疊所必需的，因此錯誤的 glycan 意味著錯誤摺疊與功能喪失。它對受體結合至關重要，因此錯誤的 glycan 意味著藥物根本無法發揮作用。在 cetuximab 中，其後果在四個軸向上可見：糖基化影響與 EGFR 的結合效率，Fc 區的糖基化影響 ADCC，適當的糖基化降低抗體被視為外來物的機會，而 glycan 影響在腫瘤部位的分布與濃度。

Where it stops being true. The moat is protein-dependent. Not every biologic is glycosylated, and the tighter a product is defined — a chemically synthesised peptide or aptamer — the less the argument applies. Nor is more glycan better: the requirement is a defined pattern, not abundance, as the heterodimeric glycoprotein hCG (237 amino acids) illustrates.

邊界 —— 何時不再成立。 這道護城河取決於蛋白。並非每一種生物製品都被糖基化，而一個產品被定義得愈嚴謹——例如一個化學合成的肽或 aptamer——這項論證的適用性就愈低。也不是 glycan 愈多愈好：所需的是被明確定義的模式，而非豐度，異二聚體 glycoprotein hCG（237 個胺基酸）正說明了這一點。

Core answer. Because the payload is chosen for potency, and the linker decides whether that potency is released inside a tumour cell or somewhere else. The antibody supplies the address; the linker supplies the schedule — and the schedule is what determines how much drug you can safely give.

核心答案。 因為 payload 是依效力挑選的，而 linker 決定那股效力是在腫瘤細胞內、還是在其他地方被釋放。抗體提供地址；linker 提供時程——而時程決定了你能安全給予多少藥物。

Mechanism — why it follows. The ADC has three components with three distinct roles: the antibody provides targeting specificity, the cytotoxic payload is a toxin too powerful to be administered systemically on its own, and the linker is designed to release the drug once the ADC reaches its target. Run the six steps: binding → internalisation by endocytosis → linker cleavage → cytotoxic action (for example interference with DNA replication) → selective killing → cell death. The critical feature is the ordering. Release is gated by an intracellular event that only occurs after uptake, so the selectivity is a property of the linker's trigger, not of the antibody's affinity alone. If release happens prematurely in circulation, a toxin that was never safe to dose freely is now dosed freely.

機制 —— 為何如此推得。 這個 ADC 有三個組件，各有三種不同的角色：抗體提供標靶專一性；細胞毒性 payload 是一種效力過強、無法單獨全身給藥的毒素；而 linker 的設計是在 ADC 抵達其標的後釋放藥物。走完這六個步驟：結合 → 經胞吞作用內化 → linker 裂解 → 細胞毒性作用（例如干擾 DNA 複製）→ 選擇性殺傷 → 細胞死亡。關鍵特徵在於其排序。釋放由一個只有在攝取之後才會發生的細胞內事件所把關，因此選擇性來自 linker 觸發條件的屬性，而非僅來自抗體的親和力。若釋放在循環中過早發生，一種從來都不安全、不能自由給藥的毒素，如今卻被自由地給藥。

Where it stops being true. The lecture poses the framework question directly — is an ADC targeted delivery? Is it controlled release? — and the answer is both. So "the linker is the moat" is really a statement that antibody and linker are two halves of one gated system, each useless alone: an antibody without a linker-payload delivers nothing, and a linker without an internalising antibody releases a toxin indiscriminately. The hard boundary is antigen internalisation — no endocytosis, no cleavage, no selectivity.

邊界 —— 何時不再成立。 課堂直接提出這個框架性問題——一個 ADC 是標靶遞送嗎？它是控制釋放嗎？——而答案是兩者皆是。因此「linker 是護城河」實際上是在說，抗體與 linker 是一個受控閘門系統的兩半，各自單獨存在時皆無用：一個沒有 linker-payload 的抗體什麼也遞送不了，而一個沒有內化抗體的 linker 則無差別地釋放毒素。硬的邊界是抗原的內化——沒有胞吞作用，就沒有裂解，也沒有選擇性。

Core answer. Because cost per base is the wrong denominator. The right one is cost per correct answer, and some answers are only reachable with reads long enough to span the region in question.

核心答案。 因為每鹼基成本是錯誤的分母。正確的分母是每個正確答案的成本，而有些答案只有用長到足以跨越該段區域的 reads 才能取得。

Mechanism — why it follows. Read length runs from about 1000 bp (Sanger) down to 55–700 bp across second-generation platforms, then back up to 3000 and as much as 15,000 (PacBio) and around 9545 (Nanopore). Repeats are the decisive case, and the clean example is HTT, where the same locus with the same mechanism gives a normal phenotype below 27 CAG repeats and disease at 40 or more — the phenotype is decided purely by repeat count. A short read cannot span the repeat, so the assembler collapses it into a length it guesses, and the number you report is wrong. Chromosome-level variation is the same argument: aneuploidy, deletion, insertion, duplication, inversion, translocation and extrachromosomal DNA are mapping and assembly problems, not substitution problems. A wrong assembly therefore does not produce a slightly worse answer, it produces a confident wrong one — and redoing the programme is the most expensive possible outcome.

機制 —— 為何如此推得。 Read 長度從約 1000 bp（Sanger）一路降到第二代平台的 55–700 bp，然後再回升到 3000，乃至高達 15,000（PacBio）與約 9545（Nanopore）。重複序列是決定性的案例，而最乾淨的例子是 HTT：同一個基因座、同一套機制，在少於 27 個 CAG 重複時表現出正常表現型，在 40 個或以上時則致病——表現型完全由重複次數決定。一個 short read 無法跨越該重複，因此組裝器會把它壓縮成一個它猜測的長度，而你所報告的數字就是錯的。染色體層級的變異是同樣的論證：非整倍體、缺失、插入、重複、倒位、易位與染色體外 DNA 都是定位與組裝的問題，而非鹼基替換的問題。因此，錯誤的組裝不會產生一個稍差的答案，而是產生一個自信的錯誤答案——而重做整個計劃是最昂貴的結果。

Where it stops being true. Long reads are not better per base, and the error column proves it: PacBio and Nanopore both generate insertion–deletion errors, while Illumina/Solexa and AB SOLiD generate substitution errors, and Sanger and Nanopore make all three types. So for a substitution question — a SNP, an exome, a tumour re-sequenced for novel point mutations — short reads are cheaper and more accurate, and depth does the work. The lecture's own answer is that a real project may combine multiple techniques.

邊界 —— 何時不再成立。 Long reads 並非在每鹼基上更好，誤差那一欄證明了這點：PacBio 與 Nanopore 都產生插入–缺失錯誤，而 Illumina/Solexa 與 AB SOLiD 產生鹼基替換錯誤，Sanger 與 Nanopore 則會產生全部三種類型的錯誤。因此對於一個替換類問題——一個 SNP、一個 exome、一個為尋找新點突變而重新測序的腫瘤——short reads 更便宜且更準確，深度就足以完成工作。課堂本身的答案是，一個真實的計劃可能結合多種技術。

Core answer. Because one rearrangement defines the patient and the target at the same time. The biomarker and the mechanism of action are the same observation, which is not true for most point mutations.

核心答案。 因為一次重排同時界定了患者與標的。biomarker 與作用機制是同一個觀察，而這對大多數 point mutations 而言並不成立。

Mechanism — why it follows. A translocation produces a chimeric protein absent from normal cells, so the therapeutic window is built into the biology rather than the dosing. The tumour then depends on that fusion product — the gain-of-function situation represented by KRAS as an oncogene, an active protein with a druggable activity, in contrast to TP53, whose loss of function means you would have to restore something rather than inhibit it. Imatinib (Gleevec) is the illustration: the Philadelphia chromosome was observed in CML in 1960, BCR-ABL1 was defined molecularly, a molecule was designed against the fusion product, and a fatal leukaemia became a managed condition. Detection is well-defined too, by sequencing (INTEGRATE, STAR-fusion, ARRIBA, CICERO) or non-sequencing methods (PCR with Sanger sequencing, hybridisation), with OncoKB as the actionable-gene reference.

機制 —— 為何如此推得。 一個 translocation 產生一個正常細胞中不存在的嵌合蛋白，因此 therapeutic window 是內建於生物學之中，而非內建於給藥之中。腫瘤接著依賴那個 fusion 產物——這種功能增益的情況以 KRAS 作為 oncogene 為代表，即一個具有可成藥活性的活性蛋白，與 TP53 形成對比，後者的功能喪失意味著你必須恢復某樣東西，而非抑制它。Imatinib（Gleevec）就是例證：Philadelphia chromosome 於 1960 年在 CML 中被觀察到，BCR-ABL1 在分子層面被界定，一個分子被針對該 fusion 產物設計出來，而一種致命的白血病成為一種可管理的狀況。檢測方式也定義明確，可經由 sequencing（INTEGRATE、STAR-fusion、ARRIBA、CICERO）或非 sequencing 方法（PCR 搭配 Sanger sequencing、雜交），並以 OncoKB 作為可作用基因的參考。

Where it stops being true. Fusions are not uniformly actionable, and frequency varies by tissue: haematological, soft tissue and epithelial, the last running rare to common. A fusion can be subclonal, so a negative result is not proof of absence. And because detection is a method question — an exon-restricted DNA panel can miss a rearrangement with an intronic breakpoint — test and drug must be co-designed.

邊界 —— 何時不再成立。 Fusions 並非一律可作用，且其頻率隨組織而異：血液、軟組織與上皮，最後一類從罕見到常見。一個 fusion 可以是次選殖性的，因此一個陰性結果不是不存在的證明。而由於檢測是一個方法學問題——一個僅限於 exon 的 DNA panel 可能漏掉一個帶有 intron 斷點的 rearrangement——檢測與藥物必須共同設計。

Core answer. Because each captures a different fraction of a genome whose interpretable part is very small, so the choice sets what you can answer, what you cannot, and what the result is worth to a payer.

核心答案。 因為每一種都捕捉一個基因組的不同部分，而其中可解讀的部分極為微小，因此這項選擇決定了你能回答什麼、不能回答什麼，以及該結果對 payer 而言值多少。

Mechanism — why it follows. The genome is ~3 Gb, there are ~20,000 protein-coding genes, and the coding genome is ~50 Mb — roughly 1.6% of the sequence. WGS captures everything including introns and intergenic space; WES captures the exons; a targeted panel captures defined regions isolated by target enrichment. The asymmetry that drives the decision is that coding variation is largely predictable — missense changes, and indels that are often frameshifting — whereas most non-coding mutations' functions remain unknown, with TERT promoter as the named exception. So WGS buys a great deal of sequence you cannot yet interpret, while a panel does more than reduce cost: it concentrates reads on defined regions so low-frequency somatic variants become visible, which is the same logic behind tumour mutation burden, and it is the format on which a diagnostic product can be built.

機制 —— 為何如此推得。 基因組約 3 Gb，約有 20,000 個 protein-coding genes，而編碼基因組約 50 Mb——大約是序列的 1.6%。WGS 捕捉一切，包括 introns 與基因間區；WES 捕捉 exons；一個標靶 panel 捕捉以標靶富集分離出來的既定區域。驅動這項決定的不對稱在於，編碼變異在很大程度上是可預測的——誤義改變，以及常常造成移碼的 indels——而大多數非編碼突變的功能仍屬未知，並以 TERT promoter 作為點名的例外。因此 WGS 買下大量你尚無法解讀的序列，而一個 panel 所做的不只是降低成本：它把 reads 集中到既定區域，使低頻的體細胞變異得以顯現，這是腫瘤突變負荷背後的同一套邏輯，而它也正是可以據以建立一項診斷產品的格式。

Where it stops being true. Panel blind spots are structured, not random, so the failure is predictable and therefore avoidable. A fusion needs intronic coverage and will be missed by an exon-only design. CNV is a read-depth computation — log2 ratio = normalised read depth in the case sample over the control — which is why it works on any platform, but it detects large changes only; a point mutation is not a copy-number event. Down syndrome shows the positive side: trisomy 21 is a copy-number change, the answer to the lecture's own opening question, and the read-depth measurement that handles it is the same one behind non-invasive prenatal testing, built on the cell-free DNA work of Lo et al. No single design dominates.

邊界 —— 何時不再成立。 Panel 的盲點是有結構的，而非隨機的，因此其失效是可預測的，因而也是可避免的。一個 fusion 需要 intron 的覆蓋，會被僅限 exon 的設計漏掉。CNV 是一項讀取深度的計算——log2 ratio = 病例樣本中經標準化的讀取深度除以對照組——這就是為何它在任何平台上都可行，但它只能偵測大型的改變；一個 point mutation 不是一個 copy-number 事件。Down syndrome 展示了正面的一面：trisomy 21 是一項 copy-number 改變，正是課堂自身開場問題的答案，而處理它的讀取深度測量，也正是非侵入性產前檢測背後的同一種測量，建立於 Lo 等人游離 DNA 的研究之上。沒有任何單一設計佔盡優勢。

Core answer. Because the product is engineered for one administration while the payer's accounting runs annually, and patients move between payers — so the entity that pays the cost is often not the entity that captures the benefit.

核心答案。 因為產品是為單次給藥而設計的，而 payer 的會計卻按年度運行，且患者會在不同 payers 之間流動——因此支付成本的實體，往往不是取得效益的實體。

Mechanism — why it follows. The price list is not a set of outliers; it is a cluster: LUXTURNA $0.85M (RPE65), ZOLGENSMA $2.1M (SMN1), CASGEVY $2.2M (BCL11A), ZYNTEGLO $2.8M (HBB), ROCTAVIAN $2.9M (F8), SKYSONA $3.0M (ABCD1), LYFGENIA $3.1M (HBB), ELEVIDYS $3.2M (DMD), BEQVEZ and HEMGENIX $3.5M each (both F9, haemophilia B), LENMELDY $4.25M (ARSA). Roughly half sit near $3M. A therapy dosed once but booked as a one-period cost against a budget cycle creates a mismatch no discount can fix, because the mismatch is temporal rather than financial.

機制 —— 為何如此推得。 這份價目表不是一組離群值，而是一個群集：LUXTURNA $0.85M（RPE65）、ZOLGENSMA $2.1M（SMN1）、CASGEVY $2.2M（BCL11A）、ZYNTEGLO $2.8M（HBB）、ROCTAVIAN $2.9M（F8）、SKYSONA $3.0M（ABCD1）、LYFGENIA $3.1M（HBB）、ELEVIDYS $3.2M（DMD）、BEQVEZ 與 HEMGENIX 各 $3.5M（兩者皆為 F9，haemophilia B）、LENMELDY $4.25M（ARSA）。大約一半落在 $3M 附近。一項只給藥一次的治療，卻在一段預算週期中被記為單一期間的成本，造成一種沒有任何折扣能修補的錯配，因為這個錯配是時間性的，而非財務性的。

Where it stops being true. It does not invalidate the model; it constrains the indications. A one-time $3M is defensible where the comparator is a lifetime of expensive chronic treatment — precisely the haemophilia situation, with F9 appearing twice at about $3.5M and F8 at $2.9M as alternatives to factor replacement — or where the disease is fatal, rare and has no alternative. It is not defensible at high prevalence: multiplying a $3M price by a common disease is fiscally and politically impossible, which rules out taking a gene therapy into a large population at this price point. Note too that two products against the same gene at the same price means competition is possible even in an ultra-orphan market.

邊界 —— 何時不再成立。 它並沒有使該模式失效；它限制的是適應症。當對照組是一輩子的昂貴慢性治療時，一次性的 $3M 是站得住腳的——這正是 haemophilia 的情況，F9 出現兩次約 $3.5M，F8 為 $2.9M，作為凝血因子補充的替代方案——或者當該疾病是致命的、罕見的且沒有替代方案時則亦然。在高盛行率下則站不住腳：把 $3M 的價格乘上一種常見疾病，在財政上與政治上都不可能，這排除了在這個價位上把一項基因治療帶入大型族群。另請注意，兩項針對同一個基因的產品以相同價格存在，意味著即使在超孤兒藥市場中，競爭也是可能的。

Core answer. Check first-pass metabolism — specifically the extent of hepatic (and gut-wall) extraction. If extraction is high, no excipient change will rescue the molecule, and the programme stops being a formulation problem and becomes a route problem.

核心答案。 先檢查 first-pass metabolism ——具體而言是肝臟（以及腸壁）extraction 的程度。若 extraction 很高，任何賦形劑的更動都救不了這個分子，該開發計畫便不再是一個配方問題，而變成一個給藥途徑問題。

Mechanism — why it follows. An oral dose is absorbed into the portal circulation and passes the liver before it reaches systemic circulation. Phase I enzymes oxidise, reduce or hydrolyse the molecule (classically adding an –OH); Phase II conjugation then makes it water-soluble enough to excrete. The lecture's own figure is that the first-pass effect exceeds 90% for some drugs — so under a tenth of the swallowed dose is left to act. Potency measured in a dish never sees this gate, which is exactly why in-vitro potency and in-vivo oral activity can diverge completely.

機制 —— 為何如此推得。 一個口服劑量被吸收進入肝門循環，並在抵達全身循環之前先經過肝臟。Phase I 酶將分子氧化、還原或水解（典型作法是加上一個 –OH）；隨後 Phase II conjugation 使其水溶性足以被排泄。課程本身給出的數字是，某些藥物的 first-pass effect 超過 90% ——因此吞下的劑量中剩下不到十分之一可供作用。在培養皿中量測的效力從未經過這道關卡，這正是為何 in-vitro 效力與 in-vivo 口服活性可以完全分歧。

Where it stops being true. Two limits. First, the liver is the major site of metabolism but not the only one — gut, lungs, skin and kidneys also biotransform drugs, so hepatic extraction alone can misattribute the loss. Second, bypassing the liver is a property of the route, not the molecule: sublingual, rectal, inhalation, IV, IM and SC all skip or reduce first-pass, so a compound that is un-developable orally may be perfectly developable another way.

邊界 —— 何時不再成立。 有兩項限制。第一，肝臟是代謝的主要場所，但不是唯一的場所——腸道、肺、皮膚與腎臟同樣會對藥物進行 biotransformation，因此單看 hepatic extraction 可能把損失錯誤歸因。第二，繞過肝臟是給藥途徑的性質，而非分子的性質：sublingual、rectal、吸入、IV、IM 與 SC 全都跳過或減少 first-pass，因此一個無法以口服開發的化合物，換一條路徑可能完全可以開發。

Core answer. Because only the unbound fraction is pharmacologically active. Blood volume is not the issue; the free concentration is. Displacement does not create drug — it converts a large inert reservoir into active drug, and it does so multiplicatively rather than additively.

核心答案。 因為只有未結合的部分具藥理活性。問題不在血量，而在游離濃度。置換並不製造藥物——它把一個龐大而惰性的儲庫轉化為活性藥物，而且是以乘法而非加法的方式進行。

Mechanism — why it follows. The card is precise: only the unbound/free fraction is active and available for distribution, metabolism and excretion, while high protein binding acts as a reservoir limiting immediate action. Bound drug behaves as a macromolecule and cannot even pass the glomerulus. Run the arithmetic on 97% binding: 3% is free. Displace a third of the bound drug and you add roughly 2 percentage points of free drug — against a 3% baseline, close to a doubling of active drug. For a drug whose entire identity is titrated anticoagulation, that is the difference between protection and haemorrhage. The same transporter-sharing logic recurs at excretion, where aspirin and methotrexate compete for tubular secretion.

機制 —— 為何如此推得。 教材說得精確：只有未結合／游離的部分具活性，且可供分佈、代謝與排泄，而高蛋白結合則扮演儲庫、限制即時作用。與蛋白結合的藥物行為如同大分子，甚至無法通過腎絲球。以 97% 結合率來算這筆數：3% 是游離的。置換其中三分之一的結合藥物，就大約增加 2 個百分點的游離藥物——相對於 3% 的基準，接近活性藥物翻倍。對於一個身分完全建立在調整式抗凝的藥物而言，那就是保護與出血之間的差別。同樣的共享轉運體邏輯在排泄環節重現，aspirin 與 methotrexate 會在腎小管分泌上競爭。

Where it stops being true. Displacement only becomes a clinical event when the displaced drug is highly bound and has a narrow margin between therapeutic and toxic exposure; displace a drug with a wide margin and you get a measurable pharmacokinetic change with no visible consequence. The mechanism also assumes the remaining free drug is not cleared so fast that the displaced fraction is removed before it can act.

邊界 —— 何時不再成立。 置換只有在被置換的藥物為高蛋白結合且治療與毒性暴露之間的邊際狹窄時，才成為臨床事件；置換一個邊際寬闊的藥物，你得到的是一個可量測的 pharmacokinetic 變化，卻沒有可見的後果。此機制也假設剩餘的游離藥物不會被清除得太快，以致被置換出的部分在能作用之前就已被移除。

Core answer. Competitive antagonism is surmountable — a parallel shift right at the same maximum; non-competitive/allosteric antagonism shifts right and lowers the maximum. The ceiling exists because the two drugs remove receptor availability differently: one competes for it, the other takes it out of play.

核心答案。 Competitive antagonism 是可克服的——在相同最大值下平行右移；non-competitive／allosteric antagonism 則右移且降低最大值。天花板之所以存在，是因為這兩類藥物以不同方式移除受體的可用性：一者與之競爭，另一者使它退出戰場。

Mechanism — why it follows. A competitive antagonist occupies the same site reversibly, so raising agonist concentration out-competes it at equilibrium — the plateau is unchanged, and the Schild relationship quantifies how much agonist is needed per unit of antagonist. A non-competitive antagonist removes receptors from the recruitable pool, so past a point no amount of agonist restores the response. Physiological antagonism is the same distinction in another dress: two agonists pulling on different receptors in opposite directions, leaving each pathway's ceiling intact.

機制 —— 為何如此推得。 Competitive antagonist 以可逆方式佔據同一個位點，因此提高 agonist 濃度便能在平衡時勝過它——高原期不變，而 Schild relationship 量化了每單位 antagonist 需要多少 agonist。Non-competitive antagonist 則把受體自可召募的池中移除，因此超過某一點之後，再多的 agonist 都無法恢復反應。Physiological antagonism 是同一組區分的另一種裝扮：兩個 agonist 拉動不同受體、方向相反，使每一條路徑的天花板保持完整。

Where it stops being true. The flat top of a competitively-antagonised curve is not solely the antagonist's doing — it depends on spare receptors. The card states that a maximal tissue response can be elicited without full occupancy, and that spare receptors reduce the apparent effect of inhibitors. That reserve is the buffer a competitive antagonist must chew through before the maximum falls. In a tissue with little reserve, a competitive antagonist will depress the maximum too, and the clean distinction degrades into a matter of degree.

邊界 —— 何時不再成立。 受 competitive 拮抗的曲線之所以頂部平坦，並不只是 antagonist 的功勞——它取決於備用受體。教材指出，無需完全佔滿即可引出組織的最大反應，而備用受體會減低抑制劑的表面效應。那份儲備就是 competitive antagonist 必須先啃穿、最大值才會下降的緩衝。在儲備很少的組織中，competitive antagonist 同樣會壓低最大值，而那條乾淨的區分便退化為程度之別。

Core answer. Neither is better — they answer two different questions on one axis. The full agonist buys ceiling; the partial agonist buys a ceiling. Which one is the drug depends on whether the clinical risk is under-treatment or over-treatment.

核心答案。 兩者都沒有更好——它們在同一條軸線上回答兩個不同的問題。Full agonist 買到的是天花板；partial agonist 買到的是一個天花板。哪一個才是藥，取決於臨床風險是治療不足還是治療過度。

Mechanism — why it follows. Affinity is the probability of occupying a receptor at any instant; intrinsic efficacy is the ability to activate it once bound. These are set by different molecular properties, so occupancy and response are not interchangeable. A full agonist with spare receptors can saturate the tissue response while three-quarters of the receptors are still empty; a partial agonist, however high its affinity, cannot exceed its own intrinsic activity no matter how much of the pool it occupies. The card puts numbers on it: 5–10% occupancy for the full agonist, 100% for a submaximal partial response.

機制 —— 為何如此推得。 親和力是在任一瞬間佔據某個受體的機率；intrinsic efficacy 則是結合之後將其活化的能力。兩者由不同的分子性質決定，因此佔據率與反應不可互換。帶有備用受體的 full agonist 可以在四分之三的受體仍然空著時就使組織反應飽和；而 partial agonist 無論親和力多高、佔據了池中多大比例，都無法超越自身的 intrinsic activity。教材把數字放上去：full agonist 為 5–10% 佔據率，次最大 partial 反應則需 100%。

Where it stops being true. The comparison assumes the same receptor and the same tissue reserve, because potency is only meaningful between agonists sharing a receptor. Receptor populations are not fixed either — they can be upregulated, downregulated or sequestered — so the occupancy-to-response mapping that holds on day one may not hold after chronic dosing. Where baseline endogenous tone is already high, a partial agonist's ceiling can fall below what the disease needs, converting a safety advantage into a plain efficacy failure.

邊界 —— 何時不再成立。 這項比較假設相同的受體與相同的組織儲備，因為 potency 只有在共用同一受體的 agonist 之間才有意義。受體族群也並非固定——它們可以被 upregulated、downregulated 或 sequestered ——因此第一天成立的佔據率對反應的對應關係，在長期給藥之後可能不再成立。當基線的內源性張力已經很高時，partial agonist 的天花板可能低於疾病所需，把一項安全優勢轉變成單純的效力失敗。

Core answer. Because β₁ and β₂ sit on different tissues and do different jobs. Sparing β₁ is a benefit when you want airways; it is a defect when the heart is the organ you need to treat.

核心答案。 因為 β1 與 β2 位於不同組織，做不同的工作。當你要的是呼吸道時，避開 β1 是優點；當心臟才是你需要治療的器官時，它就是缺陷。

Mechanism — why it follows. β₂ receptors sit on bronchioles and vessels and mediate bronchodilation and vasodilation; β₁ sits on the heart and drives rate and contractility. Both are Gs-coupled and both raise cAMP — the difference is purely anatomical. Salbutamol exploits that geography, and the card supplies the proof of what happens without selectivity: isoprenaline acts on all β receptors, giving bronchodilation but also β₁-driven increases in automaticity and heart rate with arrhythmias common, and it was superseded in asthma by specific β₂ agonists with fewer cardiac effects. Flip the patient and the polarity reverses: in heart failure the intervention is β₁ blockade (metoprolol, carvedilol), so a molecule that avoids the heart does nothing useful there.

機制 —— 為何如此推得。 β2 受體位於細支氣管與血管上，媒介支氣管擴張與血管擴張；β1 則位於心臟，驅動心率與收縮力。兩者都與 Gs 偶聯，也都提高 cAMP ——差別純粹是解剖位置。Salbutamol 利用的正是這份地理分布，而教材提供了沒有選擇性時會發生什麼的證據：isoprenaline 作用於所有 β 受體，帶來支氣管擴張，同時也帶來 β1 驅動的自律性與心率上升，心律不整常見，因而在氣喘治療中被具較少心臟效應的特定 β2 agonist 取代。把病人翻轉過來，極性也反轉：在心臟衰竭中，介入正是 β1 阻斷（metoprolol、carvedilol），因此一個避開心臟的分子在那裡毫無用處。

Where it stops being true. Selectivity is relative, not absolute, and the preference reverses again in obstructive airways: nonselective β blockade increases airway and peripheral vascular resistance, producing bronchospasm and worsening peripheral vascular disease through β₂ blockade. The heart failure benefit is also not a simple receptor effect — it appears only after roughly three months of low-dose, slow titration, so acute haemodynamic reasoning would not have predicted it.

邊界 —— 何時不再成立。 選擇性是相對的，而非絕對的，而在阻塞性呼吸道中偏好再度反轉：非選擇性 β 阻斷增加呼吸道與周邊血管阻力，透過 β2 阻斷產生支氣管痙攣並使周邊血管疾病惡化。心臟衰竭的效益也不是單純的受體效應——它只在低劑量、緩慢滴定約三個月之後才出現，因此急性血液動力學的推理不會預測到它。

Core answer. Because one of the receptors you blocked was a brake on the system you were trying to quiet. α₂ is presynaptic and restrains norepinephrine release; blocking it releases more transmitter, which partially cancels the α₁-mediated vasodilation you wanted.

核心答案。 因為你阻斷的其中一個受體，正是你想讓它安靜下來的那個系統的煞車。α2 位於突觸前，抑制 norepinephrine 釋放；阻斷它會釋出更多傳遞物質，部分抵消你想要的 α1 媒介血管擴張。

Mechanism — why it follows. The card places α₂ presynaptically as negative feedback on norepinephrine release, with three subtypes (α_2A, α_2B, α_2C) all on sympathetic terminals. Norepinephrine is cleared mainly by reuptake, so synaptic transmitter concentration is a balance of release and removal — and α₂ sets the release side. Block it and you raise norepinephrine at every terminal, competing with your own α₁ blockade at the vascular smooth muscle. The card states the consequence directly: α₂ blockade reduces the force of the vasodilation. A selective α₁ blocker such as prazosin (also doxazosin, terazosin) dilates without touching the brake.

機制 —— 為何如此推得。 教材把 α2 置於突觸前，作為 norepinephrine 釋放的負回饋，並有三種 subtype（α2A、α2B、α2C），全都位於交感神經末梢。Norepinephrine 主要經由再回收清除，因此突觸傳遞物質的濃度是釋放與移除之間的平衡——而 α2 設定的是釋放那一側。阻斷它，你就提高每一個末梢的 norepinephrine，在血管平滑肌上與你自己的 α1 阻斷競爭。教材直接陳述後果：α2 阻斷減弱血管擴張的力量。選擇性 α1 blocker 如 prazosin（另有 doxazosin、terazosin）則能在不觸碰煞車的情況下擴張。

Where it stops being true. The mechanism also predicts the inverse — that occupying α₂ rather than blocking it lowers sympathetic outflow, which is why clonidine is an agonist used as an antihypertensive. Direction of effect follows occupation, not receptor name. The reasoning also assumes an intact sympathetic reflex arc: where sympathetic outflow is already minimal there is no brake to release.

邊界 —— 何時不再成立。 此機制也預測了反向情形——佔據 α2 而非阻斷它會降低交感輸出，這就是為何 clonidine 是以 agonist 身分用作降壓藥。效應的方向跟隨佔據與否，而非受體的名稱。此推理也假設交感反射弧完整：在交感輸出本已極少之處，沒有煞車可供釋放。

Core answer. Because heart failure is a syndrome, not a contractility deficit. The two drugs pull opposite levers in the same equation: one buys pump force, the other buys relief from chronic sympathetic drive and better diastolic filling. Their time courses differ, which is the whole point.

核心答案。 因為心臟衰竭是一種症候群，而非單純的收縮力不足。這兩種藥物在同一條方程式中拉動相反的拉桿：一者買到泵浦的力量，另一者買到從慢性交感驅動中解脫，以及更好的舒張期填充。它們的時程不同，而這正是重點。

Mechanism — why it follows. Digoxin, from foxglove and described by Withering in 1775, inhibits Na⁺/K⁺-ATPase, so intracellular Na⁺ rises, less Ca²⁺ is extruded via the Na⁺/Ca²⁺ exchanger, cytosolic Ca²⁺ increases, and excitation–contraction coupling strengthens — direct, immediate contractility. β-blockers do the reverse acutely: lower contractility, heart rate, blood pressure and oxygen demand, while prolonging diastolic filling for better perfusion. The card's key observation is what happens next — they paradoxically improve heart failure, but only after about three months, starting low and titrating slowly. The acute direction of a haemodynamic effect and its long-term outcome effect are not the same thing.

機制 —— 為何如此推得。 Digoxin 來自毛地黃，由 Withering 於 1775 年描述，它抑制 Na+/K+-ATPase，使細胞內 Na+ 上升，較少 Ca2+ 經 Na+/Ca2+ exchanger 排出，細胞質 Ca2+ 因此增加，excitation–contraction coupling 隨之加強——直接而即時的收縮力。β-blockers 在急性期則反向而行：降低收縮力、心率、血壓與氧需求，同時延長舒張期填充以改善灌流。教材的關鍵觀察是接下來發生的事——它們悖論地改善心臟衰竭，但只在約三個月之後，且須從低劑量開始緩慢滴定。血液動力學效應的急性方向與其長期結局效應，並非同一回事。

Where it stops being true. The card frames heart failure management as five patient-centred aims — quality of life, independence, exercise capacity, survival, reduced resource use — and notes that no single drug hits all five, which is why polypharmacy (diuretic + ACEI + β-blocker + digoxin) is deliberate design rather than failed focus. The reasoning also depends on titration; a β-blocker started at target dose, or judged on a short-term contractility reading, would be read as harmful.

邊界 —— 何時不再成立。 教材把心臟衰竭處置框定為五項以病人為中心的目標——生活品質、獨立性、運動能力、存活、減少資源使用——並指出沒有任何單一藥物能命中全部五項，這就是為何 polypharmacy（diuretic + ACEI + β-blocker + digoxin）是刻意的設計，而非失去焦點。此推理也取決於滴定；一個以目標劑量起始、或以短期收縮力讀數來評判的 β-blocker，都會被讀成有害。

Core answer. Because the RAAS is not a blood-pressure system that happens to touch potassium — it is a volume-and-electrolyte system whose vasoconstrictor arm also raises pressure. Block the axis and you reduce both jobs at once; the potassium rise is the same mechanism seen from the other side.

核心答案。 因為 RAAS 不是一個恰好會碰到鉀的血壓系統——它是一個容積與電解質系統，其血管收縮那一臂同時也提高血壓。阻斷這條軸線，你一次減弱了兩項工作；鉀的上升就是從另一側看見的同一個機制。

Mechanism — why it follows. Renin from the kidney cleaves angiotensinogen to angiotensin I; ACE in the lungs converts it to angiotensin II, which produces vasoconstriction and blood-volume expansion. ACE inhibitors (lisinopril, captopril, enalapril, perindopril, ramipril) cut angiotensin II and give vasodilation plus volume reduction — the therapeutic effect. But aldosterone is a downstream effector of the same axis, and aldosterone at the distal tubule drives sodium and water reabsorption while governing potassium handling. That is why the potassium-sparing agents in the same lecture — spironolactone and amiloride — work by antagonising aldosterone there: they produce the same potassium retention ACE inhibition produces.

機制 —— 為何如此推得。 腎臟分泌的 renin 將 angiotensinogen 切割成 angiotensin I；肺部的 ACE 再將它轉為 angiotensin II，產生血管收縮與血容量擴張。ACE inhibitors（lisinopril、captopril、enalapril、perindopril、ramipril）切斷 angiotensin II，帶來血管擴張加上容積減少——這就是治療效應。但 aldosterone 是同一條軸線的下游效應物，而遠曲小管的 aldosterone 一方面驅動鈉與水的再吸收，一方面掌管鉀的處理。這就是為何同一堂課中的保鉀藥物—— spironolactone 與 amiloride ——是靠在該處拮抗 aldosterone 來作用：它們產生與 ACE 抑制相同的鉀滯留。

Where it stops being true. The card's caution sharpens the boundary: loop and thiazide diuretics cause hypokalaemia and are commonly co-prescribed with RAAS blockers, so the two directions of potassium movement collide in one patient and monitoring is not optional. The risk is also concentrated in the populations the card singles out — the elderly have decreased kidney size, renal blood flow, functional nephrons and tubular secretion, hence lower GFR — so the same dose is safer in a young patient with intact renal function.

邊界 —— 何時不再成立。 教材的告誡使界線更為銳利：loop 與 thiazide diuretics 造成 hypokalaemia，且常與 RAAS blockers 共同處方，因此鉀移動的兩個方向在同一位病人身上相撞，監測並非可選項。風險也集中在教材特別指出的族群——老年人腎臟體積、腎血流、功能性腎元與腎小管分泌均下降，因而 GFR 較低——所以同一劑量在腎功能完好的年輕病人身上較為安全。

Core answer. Because the property that makes a benzodiazepine desirable — immediate enhancement of inhibitory tone — is the very property that drives adaptation on repeated exposure. Dependence is not a manufacturing impurity; it is the pharmacological premise, delayed.

核心答案。 因為使 benzodiazepine 令人嚮往的那項性質——即時增強抑制性張力——正是驅動反覆曝露後產生適應的同一項性質。成癮不是製程雜質；它是被延遲顯現的藥理前提。

Mechanism — why it follows. Benzodiazepines potentiate GABA-A, the brain's inhibitory brake. The card splits the subtypes meaningfully: α₁ carries sedation, amnesia and anticonvulsant effects, α₂ carries anxiety relief — one molecule, two receptor jobs, so anxiolysis arrives bundled with sedation. Sustained enhancement of an inhibitory signal pushes the system to adapt, and the card separates the two mechanisms precisely: homologous desensitisation leaves receptors present but unresponsive to the same ligand, whereas down-regulation is an actual decrease in receptor number appearing over hours to days of prolonged exposure. Either way the same dose buys less effect, and removing the drug unmasks the inhibition no longer being supplied. SSRIs, SNRIs and tricyclics act more slowly but are preferred long-term precisely because they rarely cause dependence.

機制 —— 為何如此推得。 Benzodiazepines 增強 GABA-A，即大腦的抑制性煞車。教材有意義地拆分其 subtype：α1 承載鎮靜、失憶與抗痙攣效應，α2 承載焦慮緩解——一個分子、兩份受體工作，因此抗焦慮與鎮靜是綁在一起到來的。抑制性訊號的持續增強會推動系統去適應，而教材精確區分兩種機制：homologous desensitisation 使受體仍然存在，卻對同一配體不再反應；而 down-regulation 則是受體數目真正減少，在長時間曝露數小時至數日之後出現。兩種情況下，同一劑量都買到更少的效果，而停藥則暴露出不再被供應的抑制。SSRIs、SNRIs 與 tricyclics 作用較慢，但長期而言更受偏好，正因為它們很少造成成癮。

Where it stops being true. This is not an argument that the fast drug is wrong. The card's contrast is speed-plus-dependence versus slow-plus-safer, so speed is a genuine advantage wherever the need is acute and time-limited — procedural sedation, status epilepticus, acute agitation. The reasoning breaks when used to justify chronic therapy, not acute therapy.

邊界 —— 何時不再成立。 這不是在主張快效藥是錯的。教材對比的是快速加上成癮，對上緩慢加上較安全，因此在需求屬急性且有限時，速度是真正的優勢——程序性鎮靜、status epilepticus、急性躁動。此推理是在被用來正當化長期治療時才失效，而非急性治療。

Core answer. Carbidopa is a peripheral dopa-decarboxylase inhibitor — it protects the dose of levodopa from being converted to dopamine outside the brain, so more of it arrives where the disease actually is. Its value is pharmacokinetic, not pharmacodynamic.

核心答案。 Carbidopa 是一個周邊 dopa-decarboxylase inhibitor ——它保護 levodopa 的劑量不被在大腦之外轉化為 dopamine，使更多藥物抵達疾病真正所在之處。它的價值是 pharmacokinetic，而非 pharmacodynamic。

Mechanism — why it follows. Parkinson's is dopamine deficiency in the basal ganglia, and the card frames treatment as restoring an ACh/dopamine balance from two ends: anticholinergics (benztropine, trihexyphenidyl, procyclidine) suppress excess acetylcholine, and dopaminergic agents raise deficient dopamine. Levodopa is the dopamine precursor, but dopa-decarboxylase is present peripherally too, and dopamine generated out there cannot cross the blood-brain barrier. Every molecule converted peripherally is a wasted dose. Carbidopa blocks that peripheral conversion, concentrating the productive step inside the CNS.

機制 —— 為何如此推得。 Parkinson 是基底核中 dopamine 的缺乏，教材把治療框定為從兩端恢復 ACh／dopamine 的平衡：anticholinergics（benztropine、trihexyphenidyl、procyclidine）抑制過多的 acetylcholine，而 dopaminergic agents 提高不足的 dopamine。Levodopa 是 dopamine 的前體，但 dopa-decarboxylase 也存在於周邊，而在那裡生成的 dopamine 無法穿越 blood-brain barrier。每一個在周邊被轉化的分子都是一劑浪費。Carbidopa 阻斷那道周邊轉化，把有產出的步驟集中到中樞神經系統之內。

Where it stops being true. Carbidopa only addresses what happens before the drug reaches the brain. It does nothing about dopamine already in the CNS being broken down — that is a different lever, the Selegiline arm of MAO-B inhibition. So carbidopa is a third, orthogonal mechanism layered on a two-arm strategy, and its benefit depends on peripheral decarboxylation being a material share of the loss.

邊界 —— 何時不再成立。 Carbidopa 只處理藥物抵達大腦之前所發生的事。它對中樞神經系統中已被分解的 dopamine 毫無作為——那是另一根拉桿，即 MAO-B inhibition 的 Selegiline 那一臂。因此 carbidopa 是疊加在雙臂策略之上的第三個、正交的機制，而它的效益取決於周邊 decarboxylation 在損失中占有實質的比例。

Core answer. Because asthma is reversible bronchial hyperresponsiveness driven by inflammation, while COPD is structural damage that is not fully reversible. The two respond to different mechanisms, and only one of them offers a reversibly obstructed airway to restore.

核心答案。 因為氣喘是由發炎驅動、可逆的支氣管過度反應，而 COPD 是不完全可逆的結構性損傷。兩者回應不同的機制，而只有其中一者提供了可逆的氣道阻塞可供恢復。

Mechanism — why it follows. Asthma is chronic inflammation of the bronchioles that makes airways hypersensitive to allergens and irritants, with smooth muscle tightening and mucus overproduction. The arm that matches that mechanism is anti-inflammatory: inhaled corticosteroids inhibit phospholipase A₂, reducing arachidonic acid release; leukotriene modifiers act on the same mediator class either at the receptor (montelukast) or by blocking 5-lipoxygenase (zileuton). COPD is chronic bronchitis plus emphysema, caused by smoking, pollution and fumes — not curable. Its levers differ: anticholinergics (ipratropium, tiotropium) lean COPD more than asthma, and roflumilast, a PDE₄ inhibitor that raises cAMP, is used in severe COPD to reduce exacerbations — and is explicitly not a bronchodilator nor for acute bronchospasm.

機制 —— 為何如此推得。 氣喘是細支氣管的慢性發炎，使氣道對過敏原與刺激物過度敏感，並伴隨平滑肌收緊與黏液過度分泌。對應此機制的是一臂抗發炎作用：inhaled corticosteroids 抑制 phospholipase A2，減少 arachidonic acid 釋放；leukotriene modifiers 則作用於同一類媒介物，或作用於受體（montelukast），或阻斷 5-lipoxygenase（zileuton）。COPD 是慢性支氣管炎加上 emphysema，由吸菸、污染與煙霧造成——無法根治。它的拉桿不同：anticholinergics（ipratropium、tiotropium）更偏向 COPD 而非氣喘，而 roflumilast 這種提高 cAMP 的 PDE4 inhibitor 用於嚴重 COPD 以減少急性惡化——且明確不是 bronchodilator，也不用於急性支氣管痙攣。

Where it stops being true. "Reversible" is a tendency, not a law. A COPD patient still receives bronchodilators, and a long-standing asthmatic can develop fixed obstruction; the label does not change what the remaining mechanism has room to do. The card's framing of inhaled corticosteroids also holds in both diseases — they prevent attacks rather than relieve acute bronchospasm — so the anti-inflammatory and bronchodilator arms occupy different clinical occasions even when co-prescribed.

邊界 —— 何時不再成立。 「可逆」是一種傾向，而非定律。COPD 病人仍然接受 bronchodilators，而長期氣喘患者也可能發展出固定的阻塞；仿單並不會改變剩餘機制還有多少空間可以發揮。教材對 inhaled corticosteroids 的框定在兩種疾病中同樣成立——它們預防發作，而非緩解急性支氣管痙攣——因此抗發炎那一臂與 bronchodilator 那一臂即使被共同處方，仍佔據不同的臨床時機。

Core answer. It is a safety and therapeutic-window advantage produced purely by changing the route, and that is what manufactures the efficacy story. The molecule is unchanged; only its exposure distribution is.

核心答案。 那是一項純粹由改變給藥途徑所產生的安全性與 therapeutic window 優勢，而正是它製造出了效力故事。分子沒有改變，改變的只是它的曝露分布。

Mechanism — why it follows. Inhalation delivers medicine directly to the lungs, acting on lung tissue while minimising systemic side effects. An inhaled corticosteroid inhibits phospholipase A₂ locally, reducing arachidonic acid release and so airway inflammation, and because the drug is deposited where it acts, a smaller dose produces the same local effect with fewer systemic consequences than oral dosing. That is a therapeutic-index improvement with no chemical novelty at all. Note what the route does to the clinical occasion: inhaled corticosteroids prevent attacks rather than relieve acute bronchospasm, whereas short-acting β₂ agonists give quick relief as needed and long-acting ones provide regular control as an adjunct when ICS alone fails.

機制 —— 為何如此推得。 吸入將藥物直接遞送到肺部，作用於肺組織，同時把全身性副作用降至最低。一個 inhaled corticosteroid 在局部抑制 phospholipase A2，減少 arachidonic acid 釋放，從而減少氣道發炎；而由於藥物沉積在它作用之處，較小的劑量即可產生相同的局部效果，且全身性後果少於口服給藥。那是以完全沒有任何化學新穎性換得的 therapeutic index 改善。注意這條途徑對臨床時機做了什麼：inhaled corticosteroids 預防發作，而非緩解急性支氣管痙攣；相對地，短效 β2 agonists 視需要提供快速緩解，長效者則在 ICS 單獨使用失效時作為輔助提供規律控制。

Where it stops being true. The dose advantage is a property of the route, not of the corticosteroid — give the same molecule orally and it disappears. It is also conditional on delivery: the card specifies inhalant products as finely powdered or liquid drugs carried by special devices such as low-pressure aerosol containers, with inhaled anti-infectives given as dry powder or nebuliser solution. The "smaller dose" outcome depends on a device actually depositing drug in the airway.

邊界 —— 何時不再成立。 這項劑量優勢是途徑的性質，而非 corticosteroid 的性質——把同一個分子改為口服，它就消失。它也取決於遞送：教材把吸入產品定義為由特殊裝置攜帶的細粉末或液體藥物，例如低壓氣霧容器，而吸入型抗感染藥則以乾粉或 nebuliser solution 給藥。「更小劑量」的結果取決於裝置是否真的把藥物沉積在氣道中。

Core answer. A cytotoxic usually fails on therapeutic window — you cannot give enough drug; a biologic usually fails on the target — the biology matters less than the model said, or the molecule does not reach or stay where it acts. Efficacy and toxicity move together in the first case and are largely decoupled in the second.

核心答案。 Cytotoxic 通常在 therapeutic window 上失敗——你無法給到足夠的藥；biologic 則通常在標靶上失敗——生物學的重要性不如模型所說，或分子無法抵達、無法停留在它作用之處。在第一種情況下，效力與毒性同向移動；在第二種情況下，兩者大致脫鉤。

Mechanism — why it follows. Chemotherapy interferes with the cell cycle and kills all rapidly dividing cells, not just cancer cells — hence damage to skin, hair follicles, mouth and digestive tract, and bone marrow, giving the lecture's side-effect list from fatigue and hair loss to lower blood counts with anaemia and neutropenia. Neutropenia is the dose-limiting toxicity driving infection risk and treatment delays, which is why blood counts are monitored between cycles. The failure mode is dose-capping, and scheduling and infusion timing matter as much as the milligrams. Biologics are the opposite shape of problem: made from substances produced by living cells via recombinant DNA, they are large complex protein molecules broken down by digestion — hence injection or infusion only, with heat-sensitivity and contamination risk built into the supply chain. Their side effects are mostly injection-site reactions, so the failure that matters is target or delivery failure.

機制 —— 為何如此推得。 化療干擾細胞週期，殺死所有快速分裂的細胞，而不只是癌細胞——因此造成皮膚、毛囊、口腔與消化道以及骨髓的損傷，產生課程所列的副作用清單，從疲倦與掉髮，到血球數下降伴隨 anaemia 與 neutropenia。Neutropenia 是劑量限制性毒性，驅動感染風險與治療延遲，這就是為何必須在療程之間監測血球數。其失敗模式是劑量封頂，而排程與輸注時機的重要性不亞於毫克數。Biologics 是形狀相反的問題：以重組 DNA 由活細胞所產生的物質製成，它們是大型複雜的蛋白質分子，會被消化分解——因此只能注射或輸注，且熱敏感性與污染風險已內建於供應鏈之中。它們的副作用主要是注射部位反應，因此真正要緊的失敗是標靶失敗或遞送失敗。

Where it stops being true. The split is not clean. A biologic can still produce mechanism-derived harm through the immune pathway it modulates, and format alone confers nothing — the card's example, Herceptin targeting HER2-positive breast cancer, shows a validated target is what makes the model pay. Conversely a cytotoxic can fail outside its intended axis: bleomycin can awaken dormant breast cancer cells in vitro, with resulting metastatic relapse.

邊界 —— 何時不再成立。 這條分界並不乾淨。Biologic 仍可透過它所調節的免疫路徑產生由機制推導出的傷害，而單憑形式本身不能賦予任何東西——教材的例子 Herceptin 靶向 HER2-positive 乳癌，顯示出經過驗證的標靶才是讓該模式划算的原因。反過來，cytotoxic 也可能在預定軸線之外失敗：bleomycin 在 in vitro 中能喚醒休眠的乳癌細胞，導致轉移性復發。

Core answer. It exploits mechanistic independence. Resistance is the tumour's adaptation to a specific mechanism, so a second agent acting differently restores selective pressure. If the second agent shares the mechanism, or the tumour has already adapted to it, the addition adds nothing.

核心答案。 它利用的是機制上的獨立性。Resistance 是腫瘤對某一特定機制產生的適應，因此第二個作用方式不同的藥物能恢復選擇性壓力。若第二個藥物共享同一機制，或腫瘤已經對它產生適應，那麼加入它就等於沒有加入。

Mechanism — why it follows. The card states the logic plainly: using more than one drug type increases effectiveness because different types target cancer cells differently, and it reduces the likelihood of resistance. The defining feature of cancer cells here is that they become virtually immortal and multiply out of control, while healthy cells die by apoptosis when no longer useful and typically repair themselves after chemotherapy. The differential exploited is that repair capacity, not a magic bullet — and the tumour's ability to adapt around one mechanism is exactly what a second, independent mechanism removes. The card's economics follow: resistance drives combination and sequential regimens and next-line products, each new mechanism becoming a market when prior lines fail. Topoisomerase inhibitors and antimetabolites, for instance, attack DNA handling at entirely different points, which is what makes them combinable.

機制 —— 為何如此推得。 教材把邏輯說得明白：使用多於一種藥物類型能提高效力，因為不同類型以不同方式靶向癌細胞，同時降低產生 resistance 的可能性。此處癌細胞的定義性特徵是它們變得幾乎不死、失控增殖，而健康細胞在不被需要時經由 apoptosis 死亡，且通常在化療之後能自行修復。被利用的差異就是那份修復能力，而不是什麼神奇子彈——而腫瘤繞過某一機制去適應的能力，正是第二個獨立機制所移除的。教材的經濟學隨之而來：resistance 驅動組合療程與序貫療程以及後線產品，每一個新機制在先前的線別失敗時就成為一個市場。例如 topoisomerase inhibitors 與 antimetabolites 在完全不同的環節攻擊 DNA 的處理，這正是它們可以組合的原因。

Where it stops being true. Two boundaries. The card draws a hard line between acquired resistance — the tumour adapts so the same agent loses efficacy — and plain treatment failure from a wrong diagnosis; only the former is addressable by adding a mechanism, and treating the second as the first delays correct therapy. Second, combination logic works only on the resistance axis. It does not reduce toxicity: both cytotoxics damage the same rapidly dividing tissues, so the dose-limiting marrow suppression and the rest of the side-effect profile remain. Addition buys independence of mechanism, not of harm.

邊界 —— 何時不再成立。 有兩項界線。教材在 acquired resistance ——腫瘤產生適應，使同一藥物失去效力——與因診斷錯誤而造成的單純治療失敗之間劃下明確界線；只有前者可以藉由加入一個機制來處理，而把後者當作前者會延誤正確的治療。第二，組合邏輯只在 resistance 這條軸線上運作。它不會降低毒性：兩個 cytotoxics 損害的是同一批快速分裂的組織，因此劑量限制性的骨髓抑制與其餘副作用輪廓依然存在。加入所買到的是機制上的獨立，而非傷害上的獨立。

Core answer. Because they push the same axis in opposite directions, and the two directions fail differently: one fails loudly as toxicity, the other silently as loss of efficacy.

核心答案。 因為它們把同一條軸線推向相反的方向，而這兩個方向以不同方式失敗：一者以毒性大聲地失敗，另一者以失去效力無聲地失敗。

Mechanism — why it follows. Grapefruit inhibits CYP450 — CYP3A4 in the card's version — so atorvastatin, lovastatin and simvastatin are not cleared as fast, levels rise, and toxicity follows; the documented statin adverse reaction is muscle degeneration, conspicuous enough that a patient reports it. St John's wort does the reverse: a potent liver-enzyme inducer, it lowers drug levels, and the worked example is digoxin, where reduced levels mean reduced efficacy and worsening congestive heart failure. Same enzyme, inverse direction — and the induction event is easy to miss because nothing abnormal appears; the drug simply stops working while the disease deteriorates. These are deterministic, not random idiosyncrasies: predictable exposure changes, which is why they sit in a table of food-herb interactions rather than a list of unexplained ADRs.

機制 —— 為何如此推得。 葡萄柚抑制 CYP450 ——在教材的版本中是 CYP3A4 ——因此 atorvastatin、lovastatin 與 simvastatin 的清除不那麼快，濃度上升，毒性隨之而來；有紀錄的 statin 不良反應是肌肉退化，顯眼到病人會主動報告。St John's wort 則反向而行：它是強效的肝酶誘導劑，會降低藥物濃度，而書中演練的例子是 digoxin，其濃度降低意味著效力降低與鬱血性心臟衰竭惡化。同一個酶，相反的方向——而誘導事件很容易被漏掉，因為沒有出現任何異常；藥物就只是停止作用，而疾病持續惡化。這些是確定性的，而非隨機的特異反應：可預測的曝露變化，這就是為何它們被放在食物—草本交互作用的表格中，而非一份無法解釋的 ADR 清單。

Where it stops being true. The interaction is patient-behaviour dependent, so it cannot be engineered out of the product. And the herbal half sits largely outside the regulated system: TCM botanicals, dietary supplements and vitamins do not need FDA approval before sale and are sold over the counter without a prescription. There is no label to harmonise, no mandated interaction study, no post-market obligation on the herb — so the approved product carries a burden its counterpart does not. The boundary is regulatory asymmetry, not pharmacological: the enzyme does not care which side of the approval system the inhibitor came from.

邊界 —— 何時不再成立。 這項交互作用取決於病人行為，因此無法從產品端以工程手段移除。而草本的那一半大致位於受規管體系之外：中藥材、膳食補充劑與維他命在販售前不需要 FDA 核准，且不憑處方即可在櫃檯購得。沒有仿單可以協調，沒有強制性的交互作用研究，對該草本也沒有上市後義務——因此獲准上市的產品背負著對手不必背負的負擔。這條界線是法規上的不對稱，而非藥理上的：酶並不在乎抑制劑來自核准體系的哪一側。

Core answer. It tells you the trial can only resolve the common end of the scale. Frequency grading runs from "very common" at 1 in 10 through "rare" at 1 in 1,000 to 1 in 10,000, and a few thousand patients is not enough denominator to characterise anything below roughly 1 in 1,000.

核心答案。 它告訴你，試驗只能解析這個量尺上偏常見的那一端。頻率分級從 1 in 10 的「非常常見」，經 1 in 1,000 的「罕見」，一路到 1 in 10,000，而數千名病人不足以作為分母去刻畫大約低於 1 in 1,000 的任何情況。

Mechanism — why it follows. Run the card's own perspective against the card's own trial size. If an ADR affects 1 in 10,000 people, then 9,999 of 10,000 are not expected to have it — so in a few thousand patients you may observe zero cases even when the risk is real. The lecture's trial blind spot is usually described qualitatively (small populations, exclusion of special populations, stringent criteria, unstudied genetic/gender/environmental factors, incomplete interaction data), but it is also quantitative, and that arithmetic is the quantitative half. Hence the card's position that minimal safety information is known during development and that continuous observation under real-world use is required.

機制 —— 為何如此推得。 把教材自己的觀點套用到教材自己的試驗規模上。若某項 ADR 影響 1 in 10,000 的人，那麼 10,000 人中有 9,999 人預期不會有它——因此在數千名病人中，即使風險真實存在，你也可能觀察到零個案例。課程對試驗盲點的敘述通常是定性的（族群小、排除特殊族群、條件嚴苛、遺傳／性別／環境因素未經研究、交互作用資料不完整），但它同時也是定量的，而那筆算術就是定量的一半。因此教材的立場是：開發期間所知的安全性資訊極少，且必須在真實世界使用下持續觀察。

Where it stops being true. The distinction is between failing to detect and demonstrating absence. Not seeing an event in a few thousand patients does not establish that the rate is below 1 in a few thousand; it establishes that the study had no power to answer the question. Two limits compound it: the fixed exclusion of children, pregnant and lactating women and the elderly removes the subgroups most sensitive to ADRs, and even upstream animal work is skewed toward males, limiting how well preclinical data generalise.

邊界 —— 何時不再成立。 此處的區分在於未能偵測與證明其不存在。在數千名病人中沒有看到某事件，並不能確立發生率低於每數千名病人 1 例；它確立的是該研究沒有足夠檢定力回答這個問題。有兩項限制使之疊加：固定排除兒童、懷孕與哺乳婦女以及老年人，移除了對 ADR 最敏感的次族群；甚至更上游的動物研究也偏向雄性，限制了臨床前資料的可外推程度。

Core answer. Because Reduction is achieved through better statistics and design rather than by accepting weaker evidence, so the ethical choice and the information-efficient choice are the same choice. And because animal work is approval-gated, every animal used without a design justification is simultaneously an ethical exposure and a regulatory delay.

核心答案。 因為 Reduction 是透過更好的統計與設計來達成，而非靠接受更弱的證據，因此倫理的選擇與資訊效率的選擇是同一個選擇。也因為動物工作是須經核准才能進行的，每一隻在缺乏設計理由下被使用的動物，同時是一項倫理曝險與一項法規延誤。

Mechanism — why it follows. Read the three definitions together. Replacement asks for non-animal models at every opportunity appropriate for the science; Refinement asks for procedures chosen to minimise pain or distress; Reduction asks for proper statistical analyses, appropriate timelines and comparison groups so the fewest animals are used — and that wording is decisive, because the lever it names is design quality, not sample-size shaving. The institutional pressure is explicit: animal research is permitted only under specific circumstances with strict regulation on humane care, approval must be obtained before work begins, the position is codified in Hong Kong law through the Department of Health of the HKSAR, animals are used only where no other technique can deliver the answer, and new treatments are studied in animals first for evidence of benefit and risk before any human use.

機制 —— 為何如此推得。 把三個定義放在一起讀。Replacement 要求凡科學上適宜的每一個機會都採用非動物模型；Refinement 要求選擇能將疼痛或不適降至最低的程序；Reduction 要求妥善的統計分析、合適的時程與對照組，使所使用的動物最少——而那段措辭是決定性的，因為它指名的槓桿是設計品質，而非削減樣本數。制度性的壓力是明白的：動物研究只在特定情況下才被允許，並對人道照護有嚴格規範；工作開始之前必須取得核准；此立場透過香港特別行政區衛生署載入香港法律；動物只在沒有任何其他技術能提供答案時才被使用；而新療法在任何用人之前，先在動物身上研究以取得效益與風險的證據。

Where it stops being true. The framework does not mean animals are broadly replaceable. The card's position is that non-animal alternatives — human research, computer models, cell cultures and other in vitro work, statistical modelling, epidemiology — already cover many projects, but that some element remains essential with no alternative means of obtaining the information. The reasoning also assumes the alternative model is appropriate; the lecture's caution about AI shows the trap, since a system trained on skewed data gives conclusions that do not transfer.

邊界 —— 何時不再成立。 這套框架並不表示動物可以被廣泛取代。教材的立場是：非動物替代方案——人體研究、電腦模型、細胞培養與其他 in vitro 工作、統計模型、流行病學——已經涵蓋許多計畫，但仍有一些環節是不可或缺的，且沒有其他取得資訊的替代方式。此推理也假設替代模型是適宜的；課程對 AI 的告誡顯示了這個陷阱，因為以偏斜資料訓練出來的系統，會給出無法轉移的結論。

Core answer. Because gene similarity is a claim about conserved biology, not about identical exposure, metabolism or physiology. The model is deliberately narrow — that narrowness is why it gives clean causal answers, and also why it cannot represent the whole market.

核心答案。 因為基因相似性是一項關於生物學保守性的主張，而非關於曝露、代謝或生理完全相同的保證。這個模型是刻意狹窄的——那份狹窄正是它能給出乾淨因果答案的原因，也是它無法代表整個市場的原因。

Mechanism — why it follows. The case for animal models is control: short life-cycles allow lifespan and multi-generation study, the environment can be fixed, and known genetics let the researcher isolate one variable. That strength is also the limitation. First, it is explicitly a model: an organ's activity may differ even where basic chemistry, cell structure and organisation are conserved. Second, metabolism varies with species, sex, health and disease state — the L1 lecture notes that albumin, liver function, genetics and age all shape it — so the exposure producing an effect in the animal is not the exposure that will occur in the patient. Third, and most concretely documented, is sex: most researchers prefer male mice and rats, yet male and female responses differ physiologically, paracetamol metabolism in adult females is around 60% of adult males, and female ADR incidents are roughly twice as high. A male-animal package is systematically less informative about half the eventual users.

機制 —— 為何如此推得。 支持動物模型的理由是控制：短的生命週期使壽命與多世代研究成為可能，環境可以被固定，而已知的遺傳背景讓研究者得以隔離單一變數。這項長處同時也是限制。第一，它明確地是一個模型：即使基本化學、細胞結構與組織方式在演化上保守，器官的活動仍可能不同。第二，代謝隨物種、性別、健康與疾病狀態而變化—— L1 課程指出 albumin、肝功能、遺傳與年齡都會形塑它——因此在動物身上產生效果的曝露，並不是病人身上將會發生的曝露。第三，也是紀錄得最具體的，是性別：大多數研究者偏好雄性小鼠與大鼠，然而雄性與雌性的反應在生理上不同，成年雌性對 paracetamol 的代謝約為成年雄性的 60%，而女性的 ADR 事件大約高出兩倍。一份以雄性動物為主的資料包，對最終使用者中的一半而言，系統性地較不具資訊量。

Where it stops being true. This is not an argument against animal data — they are the mandatory evidence step before any consideration of human use, and decades of mouse and rat data underpin new and repurposed drug development. It is an argument that no single-species, single-sex package can be the whole safety argument, which is why pharmacovigilance is continuous across the lifecycle. There is also an inherited-bias limit on the obvious fix: AI alternatives trained on existing animal-testing results inherit whatever skew those data hold, making data bias a validation problem rather than a technical footnote.

邊界 —— 何時不再成立。 這不是在反對動物資料——它們是在任何考慮人體使用之前的強制性證據步驟，而數十年來小鼠與大鼠的資料支撐著新藥與老藥新用的開發。這是在主張：沒有任何單一物種、單一性別的資料包可以構成完整的安全論證，這就是為何 pharmacovigilance 必須貫穿整個生命週期而不間斷。對那個顯而易見的補救方式，還有一項繼承性偏誤的限制：以既有動物試驗結果訓練的 AI 替代方案，會繼承那些資料所帶有的任何偏斜，使資料偏誤成為一個驗證問題，而非技術上的註腳。

Core answer. The patent gives the right to exclude others; it does not give the owner the right to practise the invention. Those are two different questions and only the first is answered by your own grant.

核心答案。 專利所賦予的是排除他人實施的權利；它並不賦予權利人實施該發明本身的權利。這是兩個彼此完全不同的問題，而你自己所獲得的授權，只回答了其中的第一個問題。

Mechanism — why it follows. A granted claim is a negative right: it lets you stop others from making, using or selling the claimed subject matter within that jurisdiction. Whether you may make and sell depends on whether anyone else holds an earlier, broader claim your product falls inside. The L1 grid frames the owner's position as "make, use, sell", but the same card is explicit: a patent gives the right to exclude others — not the right to practise — and rights are territorial.

機制 —— 為何如此推得。 一項已獲授予的權利要求屬於消極權利：它讓你能夠在該司法管轄區之內，阻止他人製造、使用或銷售該項所主張的標的。至於你本人是否可以製造及銷售，則取決於是否有其他人持有更早、範圍更寬廣的權利要求，而你的產品落入該權利要求的範圍之內。L1 表格把權利人的地位表述為「製造、使用、銷售」，但同一張卡片本身已清楚說明：專利所賦予的是排除他人的權利，而非實施的權利；而且權利具有地域性。

Where it stops being true. The two statements coincide only where no third party holds a right covering your product in that jurisdiction. In biologics the blocking claim is usually upstream — composition-of-matter on the target, a platform claim on the format, a manufacturing claim. And a US grant says nothing about selling in China, because the rights do not travel.

邊界 —— 何時不再成立。 只有在沒有第三方於該司法管轄區之內持有一項涵蓋你產品之權利的情況下，這兩項陳述才會互相吻合。在生物製品領域，構成阻礙的權利要求通常位於上游，例如針對標的的組合物權利要求、針對格式的平台權利要求，以及製造權利要求。而美國的授權對於在中國銷售一事毫無意義，因為這些權利並不會隨着地域而轉移。

Core answer. None, on the facts as given. Copyright protects the expression of an idea, never the idea itself. The SOP text is protected; the scientific concept behind it is not, and implementing a concept is not copying an expression.

核心答案。 就上述所陳述的事實而言，答案是不存在任何侵權。版權所保護的對象，是一個思想的表達方式，而絕不是該思想本身。SOP 的文字受到保護；但隱藏在其背後的科學概念則不受保護，而落實一個概念並不等同於複製一種表達。

Mechanism — why it follows. Copyright attaches to a work fixed in a tangible medium and gives the owner the exclusive right to copy, distribute, adapt, display and perform it. That is a right over the document. The lecture's own examples put SOPs, manuals, booklets, writings and software under copyright, while secret "recipes" and processes sit under trade secret — which is precisely because the process itself is not the copyrightable thing.

機制 —— 為何如此推得。 版權附着於一件已固定在有形媒介之上的作品，並賦予權利人複製、分發、改編、展示及表演該作品的專屬權利。由此可見，這是一項以文件本身為對象的權利，而不是一項以方法為對象的權利。課程教材本身所舉的例子，把 SOP、手冊、小冊子、文字作品及軟件歸入版權的範圍之下，而保密的「配方」與流程則歸屬於 trade secret 之下——作出這種歸類的理由，恰恰是流程本身並不是那件可以受版權保護的東西。

Where it stops being true. Two limits. If the competitor reproduces the SOP's text, tables or figures, that is infringement — the right does bite on the wording. Second, the source's rule is that copyright covers the expression and the exam line is drawn there: a biotech SOP counts, "but the scientific concept behind it does not."

邊界 —— 何時不再成立。 這裏有兩項限制。第一，如果該競爭對手複製了 SOP 的文字、表格或圖示，那就構成侵權——換言之，這項權利確實會咬住文字的表述方式。第二，資料來源所立的規則是：版權涵蓋的是表達，而考試的分界線正劃在此處——生物科技的 SOP 算在涵蓋範圍之內，「但其背後的科學概念則不算」。

Core answer. Because the rights have different terms, different requirements and different failure modes, and each covers a different thing. The portfolio is what remains standing when any one right lapses.

核心答案。 因為這些權利各自有不同的期限、不同的要求與不同的失效模式，而且每一項權利所涵蓋的對象也不相同。當其中任何一項權利失效的時候，仍然站得住腳的，正是這個權利組合本身。

Mechanism — why it follows. A patent runs 20 years from filing and requires full public disclosure — that disclosure is the price of the bargain. A trademark lasts indefinitely while it is in use and stays distinctive. A trade secret lasts indefinitely while secrecy is maintained through active measures such as NDAs. Copyright runs for life plus 70 years and requires only fixation. So the molecule and method sit in the patent, the brand and source-identity sit in the trademark, the manufacturing know-how sits in the secret that was never published, and the SOPs sit in copyright.

機制 —— 為何如此推得。 專利自提出申請之日起計為期 20 年，並且要求申請人作出完整的公開披露——而這項公開披露，就是整筆交易所要付出的代價。商標只要仍然在使用之中、並且保持其顯著性，便可以無限期地一直存續下去。Trade secret 只要透過 NDA 等積極措施維持其保密狀態，同樣可以無限期地存續下去。版權則存續至作者身故後 70 年，而且只要求作品已被固定即可。因此，就最終的分配而言，分子與方法落在專利之中，品牌與來源識別落在商標之中，製造方面的 know-how 落在那項從未公開的機密之中，而各項 SOP 則落在版權之中。

Where it stops being true. The bundle only bites to the extent each right is enforceable against the competitor's chosen route. A generic challenger attacking the molecule does not need your SOP, and a trade secret gives the owner the exclusive right to use it only "provided it is not discovered by others" — it confers no exclusion against independent discovery.

邊界 —— 何時不再成立。 這個捆綁起來的組合，只有在每一項權利都能夠針對競爭對手所選擇的路徑加以執行時，才會真正發揮作用。一名攻擊該分子的學名藥挑戰者，並不需要用到你的 SOP 這類文件；而 trade secret 只在「未被他人發現」的前提之下，才賦予權利人專屬使用該機密的權利——它並不賦予任何對抗獨立發現的排他權。

Core answer. The invented word wins on protection, not on communication. The spectrum runs generic (no protection) → descriptive (protectable only with secondary meaning) → suggestive → arbitrary → fanciful (strongest), and the source's worked example of the downside is "Aspirin", which is generic.

核心答案。 自創的詞是在保護這個層面上勝出，而不是在傳達訊息這一方面勝出。換言之，問題並不在於名稱能否說明藥物的作用，而在於它能否在市場上取得保護。這條光譜依序排列為 generic（完全不受保護）→ descriptive（只有在取得 secondary meaning 之後，才可受保護）→ suggestive → arbitrary → fanciful（效力最強的一類）；而資料來源用以說明這種做法有哪些弊端的現成實例，就是「Aspirin」這個詞，它本身已經歸屬於 generic。

Mechanism — why it follows. A trademark protects the mental association in the consumer's mind between a product and its source, and it must be inherently distinctive, connected with the goods, and used in commerce. A mark that is the common term for the category cannot perform that function at all, which is why generic terms get nothing. A descriptive name starts weak and must earn protection by acquiring secondary meaning; a coined word has no prior meaning tied to the goods, so it is inherently distinctive and cannot slide toward generic by being descriptive.

機制 —— 為何如此推得。 商標所保護的，是消費者心中對某項產品與其來源之間所建立的心理聯想；而商標本身必須具備固有的顯著性、與該商品之間有所關聯，並且確實被使用於商業活動之中。一個屬於該類別通用名稱的標誌，根本無法發揮上述這項功能，這正是 generic 詞彙一無所獲的原因所在。一個描述性名稱在起步時較為薄弱，必須透過取得 secondary meaning 才能贏得保護；至於一個自創的詞，由於與該商品之間並無任何既有的意義關聯，因此本身即具備顯著性，也不會因為具描述性而逐步滑向 generic 的境地。

Where it stops being true. Strongest is not automatically best. The infringement standard is likelihood of confusion, not identical copying, so a very distinctive mark is easier to police — but a fanciful name carries no information to a prescriber and needs marketing spend to acquire meaning. And protection is not permanent: a mark can still go generic if it becomes merely descriptive in the market's usage, which is exactly how a strong position is lost.

邊界 —— 何時不再成立。 最強並不自動等同於最好。侵權的判斷標準是產生混淆的可能性，而不是完全相同的複製，因此一個高度顯著的商標在維護與執法上較為容易；然而一個 fanciful 名稱對處方者而言並不帶有任何可供辨識的資訊，必須投入市場推廣的費用才能逐步取得意義。而且保護並非永久不變：一個商標如果在市場的用法之中變得僅具描述性，它仍然可能變成 generic，而這正是強勢地位喪失的典型方式。

Core answer. The trade is disclosure for enforceable exclusion versus secrecy for indefinite but fragile protection. Neither dominates; what decides it is whether the process is visible in the product.

核心答案。 這是一項取捨：以披露換取可以執行的排他權，對比以保密換取無限期但脆弱的保護。兩者之間並無絕對的優劣之分；真正作出決定的，是這項製程到底能否從產品之中看出來。

Mechanism — why it follows. A patent is a bargain with the public: you share the invention in full disclosure in exchange for 20 years of exclusivity — and that exclusivity runs against everyone, including an independent inventor who got there the same way. A trade secret publishes nothing and lasts forever provided active measures (NDAs) are taken, but the owner's exclusive right to use it holds only "provided it is not discovered by others", so it gives no remedy against independent discovery or reverse engineering.

機制 —— 為何如此推得。 專利是與公眾之間所訂立的一項交易，其內容如下：你以完整披露的方式向公眾分享該發明，藉此換取為期 20 年的排他權——而這項排他權可以對抗所有人，包括一名在完全不知情之下、以相同方式獨立完成該發明的發明人。Trade secret 則不公開任何內容，只要採取積極的保密措施（例如簽訂 NDA）便可以一直永久存續下去；但權利人專屬使用該機密的權利，只在「未被他人發現」的前提之下才成立，因此它對於他人獨立發現該機密，或對該產品進行反向工程，並不提供任何救濟。

Where it stops being true. The deciding fact is detectability. If a competitor can analyse the product and reconstruct the process, secrecy buys you nothing and you have given away the patent for free. If the process is not ascertainable from the product — an internal cell line, a fermentation recipe — secrecy can outlast the patent term. The limit is that the whole asset is the secrecy itself, so a single leak converts a perpetual right into nothing. Note that design right is not a route here: it protects ornamental, non-functional appearance.

邊界 —— 何時不再成立。 真正具決定性的事實是可偵測性。如果一名競爭對手能夠分析你的產品，並據此重建該製程，那麼保密對你毫無助益可言，而你已經白白放棄了專利。如果該製程無法從產品之中查明——例如一株內部的細胞株、一份發酵配方——保密便可能比專利期限存續得更久。其限制在於：整個資產就是這份保密本身，因此單單一次洩漏，便足以把一項本來屬於永久的權利化為烏有。須注意的是，design right 在此並不是一條切實可行的路徑：它所保護的，是裝飾性、非功能性的外觀。

Core answer. Novelty and obviousness compare the claim to the prior art; enablement and written description are about whether the specification ever taught and possessed the invention. The first two are claim-comparison problems; the third is a disclosure problem.

核心答案。 Novelty 與顯而易見性，都是把權利要求拿來與 prior art 互相比較；而 enablement 與 written description 所關乎的，則是說明書本身到底有沒有教示過這項發明、並且擁有過這項發明。前兩者屬於權利要求比對上的問題；第三者則是披露上的問題，而它所涉及的是文件本身。

Mechanism — why it follows. Novelty asks whether the claimed subject matter was already disclosed — and the trap is that your own public disclosure destroys it, with no take-back even if a post is deleted after three days. Obviousness asks whether the differences from the prior art would have been obvious before the effective filing date to a person of ordinary skill, anchored in 35 U.S.C. §103, framed by the Graham v. John Deere three-step inquiry and applied with the teaching-suggestion-motivation plus common sense approach of KSR v. Teleflex (2007). Enablement requires teaching how to make and use the invention without undue experimentation, and written description requires showing you possess it — in biotech, via biological deposits and sequence listings.

機制 —— 為何如此推得。 Novelty 所追問的是：該項所主張的標的是否已經被披露過——而這裏的陷阱在於，你自己所作的公開披露，就足以把它破壞掉，即使那則貼文在三天之後被刪除，也無法把它收回，而且沒有任何補救的機會。顯而易見性所追問的是：該發明與 prior art 之間的差異，在有效申請日之前，對於該所屬領域中具有一般技術的人來說，是否屬於顯而易見——其依據是 35 U.S.C. §103，並由 Graham v. John Deere 所確立的三步檢驗架構所框定，再配合 KSR v. Teleflex（2007）所採用的 teaching-suggestion-motivation 加上常識的取徑予以適用。Enablement 所要求的，是教示他人如何製造及使用該發明，而無須進行過度的實驗；而 written description 則要求顯示你確實擁有該發明——在生物科技領域之中，這一點通常要透過生物材料的寄存以及序列表來達成，才能夠成立。這兩項要求所檢驗的，都是申請人究竟在文件之中教示了什麼。

Where it stops being true. The routes out of obviousness are evidentiary and live inside the record: unexpected effect, synergy, teaching away, commercial success, long-felt but unsolved need, failure of others — and the source notes that because biology and chemistry are unpredictable arts, a surprising technical effect is especially probative there. What you cannot do is reason your way past a disclosure gap: enablement and possession are properties of the document as filed. The source does not spell out refiling mechanics, so the safe reading is that a scope defect is met within the family while a possession defect is about what the original application taught.

邊界 —— 何時不再成立。 走出顯而易見性的各種途徑，全都屬於證據性質，並且存在於紀錄之內：意外的效果、協同作用、teaching away、商業上的成功、長期存在而未能解決的需求、他人的失敗——而資料來源亦指出，正因為生物學與化學屬於不可預測的技藝，一項令人意外的技術效果在該領域之中尤其具有證明力。你無法做到的，是單憑推理就繞過一項披露上的缺口：enablement 與擁有，是申請當時該份文件本身的固有性質。資料來源並沒有說明重新申請的具體機制，因此較為穩妥的一種解讀是：範圍上的缺陷可以在同一專利族之內補救；而擁有上的缺陷，所關乎的則是原始申請當時究竟教示了什麼。

Core answer. Two different questions with two different answers. The new-use claim may well be patentable, because a novel indication is a different claimed invention from the compound; and WIV could still not commercialise without Gilead's permission.

核心答案。 這是兩個不同的問題，並且有兩個不同的答案。該項新用途權利要求很可能具備可專利性，因為一項新的適應症與該化合物本身，屬於不同的所主張發明；而 WIV 在未取得 Gilead 許可的情況之下，仍然無法把它商業化。

Mechanism — why it follows. The source records Gilead filing numerous applications from 2010 onward, including the composition of matter and use in treating coronavirus infections (SARS, MERS), with the specification covering coronaviridae; WIV filed on 21 January 2020 for COVID-19 use. A patent is a right to exclude, not permission to act — so a grant to WIV would let it stop others practising that claimed use, while making or selling Remdesivir itself still falls inside Gilead's earlier compound claim.

機制 —— 為何如此推得。 資料來源記載，Gilead 自 2010 年起已提交了大量申請，其中包括針對組合物的申請，以及針對用於治療冠狀病毒感染（SARS、MERS）的用途的申請，而其說明書的涵蓋範圍，更及於 coronaviridae；另一方面，WIV 則在 2020 年 1 月 21 日就 COVID-19 的用途提出了申請。專利是一項排他權，而不是一項行動的許可——因此，授予 WIV 的一項專利，會讓它能夠阻止他人實施該項所主張的用途；但這項專利並不會讓 WIV 取得製造或銷售 Remdesivir 本身的自由，因為該行為仍然落入 Gilead 較早取得的化合物權利要求範圍之內。

Where it stops being true. Everything turns on claim scope and jurisdiction. The FTO question is whether the earlier claim actually covers your space, so if Gilead's Chinese position did not reach the relevant use or territory the analysis changes. And the source presents this as the lecture's worked example of the patentability/FTO gap, not as an adjudicated outcome.

邊界 —— 何時不再成立。 一切取決於權利要求的範圍以及司法管轄區。FTO 的問題在於：較早的那項權利要求是否實際涵蓋了你的空間；因此，如果 Gilead 在中國的權利範圍並沒有觸及相關的用途或相關的地域，整個分析就會隨之改變。而資料來源是把這宗案例，當作課程中說明 patentability 與 FTO 之間落差的實例來呈現，而不是把它當作一項已經裁決的結果。

Core answer. Eligibility turns on whether the claimed molecule exists in nature. cDNA is built by lab technicians from mRNA with the non-coding introns removed, yielding a molecule distinct from any in the body, so it is not naturally occurring. Isolated gDNA — though structurally changed, lacking the covalent bonds linking it to the rest of the gene — was not functionally different enough, and a claim to it would "improperly tie up the future use and study of these genes."

核心答案。 適格性所取決的，是所主張的那個分子到底是否存在於自然界之中。cDNA 是由實驗室技術人員從 mRNA 構建而成，其中非編碼的內含子已被移除掉，因而產生一個與體內任何分子都不相同的分子，所以它本身就並不是天然存在之物。分離出來的 gDNA 儘管在結構上已經改變、缺少了把它與基因其餘部分連結起來的共價鍵，但在功能上的差異並不足夠，而對它主張權利將會「不當地把這些基因的未來使用與研究圈佔起來」。

Mechanism — why it follows. Laws of nature, natural phenomena and abstract ideas sit outside patentable subject matter. A naturally occurring DNA segment is a product of nature and "not patent eligible merely because it has been isolated." The escape is a markedly different article — different in structure, function or properties — or a claim integrated into a practical application, such as treating a disease or limiting to a tangible physical product.

機制 —— 為何如此推得。 自然法則、自然現象以及抽象概念，都位於可專利標的的範圍之外。一段天然存在的 DNA 片段，是自然的產物，「不會僅僅因為它被分離出來就具備專利適格性」。可行的出路，在於一項顯著不同的製品——在結構、功能或性質上有所不同——或者一項已被整合於實際應用之中的權利要求，例如治療某種疾病，或者把權利要求限定為一項有形的實體產品。換言之，單靠把它分離出來，並不足以取得適格性。

Where it stops being true. Two boundaries. The source notes a synthesised product identical to its natural counterpart is eligible only if markedly different — so synthesising is not by itself enough. And the drafting lesson is jurisdiction-bounded: the same isolated gene with a specified practical application is described as patentable in the EU and, per the lecture, in China, while post-Myriad it is rejected in the US.

邊界 —— 何時不再成立。 這裏有兩項邊界。資料來源指出，一項與其天然對應物完全相同的合成產品，只有在顯著不同的情況之下才具備適格性——因此單憑合成本身並不足夠。而草擬上的教訓，同樣是受司法管轄區所限制的：同一個分離出來的基因，若具備一項指定的實際應用，在歐盟便會被描述為可以取得專利，而按照課程所述，在中國亦然；然而在 Myriad 案之後的美國，它則會被拒絕授予專利。

Core answer. Validity and design-around live in the same claim language, so every limitation you add buys validity at the cost of scope. There is no scope that is stable independent of the prior art you can actually see.

核心答案。 有效性與迴避設計，兩者都存在於同一套權利要求的文字之中，因此你每加入一項限定，都是在以範圍為代價來換取有效性。在這之中，沒有任何一個範圍，能夠獨立於你實際看得見的 prior art 而保持穩定。

Mechanism — why it follows. Infringement requires all claim elements to be present, literally or equivalently — A+B+C infringes, A+B does not — so narrowing the claim simultaneously narrows the set of products that infringe. Meanwhile the transition clause "comprising" is open-ended, so adding extra elements to an accused product does not escape infringement. Adding a limitation therefore has two effects at once: it moves the claim away from the prior art and away from real products.

機制 —— 為何如此推得。 侵權要求所有權利要求要素都必須存在，無論是在字面上存在，或是在等同的意義上存在——A+B+C 構成侵權，A+B 則不構成——因此收窄權利要求，同時也就收窄了那一組會被判定為構成侵權的產品。另一方面，轉接詞「comprising」本身屬於開放式用語，因此為被控產品加入額外要素，並不能藉此擺脫侵權的認定。由此可見，加入一項限定會同時產生兩種效果：它既使權利要求遠離 prior art，也使它同時遠離市場上真實存在的產品。

Where it stops being true. The loop only runs as far as your search does. Prior art 3 defeats the narrowed claim because you had not found it, so the analysis rests on a prior-art search that may be incomplete — and the source's own framing is that a too-broad claim is easy to invalidate while a too-narrow claim is easy to design around and fails to cover many infringers. The target is a definite claim that clearly defines the invention and is supported by the description and written description.

邊界 —— 何時不再成立。 這個循環只會在你檢索所及的範圍之內運行。Prior art 3 之所以能夠擊敗收窄之後的權利要求，是因為你當初並沒有找到它，因此整個分析建立在一項可能並不完整的 prior art 檢索之上——而資料來源本身的框架是這樣的：過於寬廣的權利要求容易被宣告無效，而過於狹窄的權利要求則容易被迴避設計，並且無法涵蓋眾多的侵權者。真正的目標，是一項明確的權利要求，既能清楚界定該發明，並且同時獲得說明書與 written description 的支持。

Core answer. It means eligibility is jurisdictional, not a property of the invention. A claim type that carries the whole asset in one market carries nothing in another, so the plan has to change by changing claim type, not by abandoning markets.

核心答案。 這意味著適格性是取決於司法管轄區的，而不是該發明本身的屬性。一種在某個市場能夠承載整個資產的權利要求類型，在另一個市場可能什麼都承載不了；因此計劃必須透過改變權利要求的類型來調整，而不是透過放棄市場來調整。

Mechanism — why it follows. The source's eligibility table lists a method of treating a disease with the note "US OK; CN and EP NOT OK." At the EPO, methods practised on the human body are not patentable — the rationale is to protect the physician's freedom to treat — while substances and devices used in those methods are patentable. In China, in-vitro diagnostic methods are patentable but methods on the living human body generally are not.

機制 —— 為何如此推得。 資料來源的適格性表格所列出的其中一項，是「一種治療疾病的方法」，並在其後附註「US OK；CN 及 EP NOT OK」。在 EPO，任何對人體實施的方法都不具備可專利性——其理據在於保護醫生的治療自由——但用於實施這些方法之中的物質與器械，則仍然可以取得專利。在中國，體外診斷方法可以取得專利，但針對活體人體所實施的方法則一般不可以取得專利。換言之，這條界線所劃分的，是方法與產品，而不是發明本身。

Where it stops being true. The fix is claim-type selection: draft composition-of-matter, or substance-for-use claims, so the same scientific asset is protected where the method claim is not. The compound per se remains protectable, so "method claims are unavailable" does not mean "no protection". And the split runs the other way for other subject matter: an isolated gene with a specified practical application is described as patentable in the EU and China yet rejected in the US post-Myriad.

邊界 —— 何時不再成立。 補救之道在於權利要求類型的選擇：草擬組合物權利要求，或者草擬物質用途權利要求，使同一項科學資產在方法權利要求無法獲得保護的地方仍然得到保護。化合物本身則依然可以受到保護，因此「方法權利要求不可用」並不意味著「沒有任何保護」。而就其他標的而言，這種分野的方向恰好相反：一個具備指定實際應用的分離基因，在歐盟與中國被描述為可以取得專利，但在 Myriad 案之後的美國則被拒絕。

Core answer. It settled the invention-versus-discovery question in the industry's favour: a human-made life form is a product of human ingenuity, not a product of nature. The Court held the oil-degrading bacterium had "markedly different characteristics from any found in nature" and the potential for significant utility, and read Congress as intending patentable subject matter to include "anything under the sun that is made by man."

核心答案。 它把「發明」與「發現」之間長久以來的爭議，一次性地在有利於產業的一方確定下來：人造的生命形式屬於人類智慧的產物，而不是自然的產物。法院認定，該株能夠降解石油的細菌本身「具備與自然界中所能找到的任何細菌顯著不同的特徵」，並且具有產生顯著實用性的相當潛力；法院亦把國會的意圖解讀為：可專利標的應當包括「太陽之下由人所製造的任何事物」在內。

Mechanism — why it follows. The examiner had rejected the claim on the ground that living things are not patentable, and the bacterium had no natural equivalent. Once engineered organisms became eligible subject matter, the industry's core assets — cell lines, transgenic animals, later cDNA — became claimable. Every later biotech-patent position is a logical extension of this one.

機制 —— 為何如此推得。 審查員當初是以「生物不可取得專利」為理由，駁回該項權利要求的，而該株細菌在自然界之中，並沒有任何對應物。一旦經過工程改造的生物體成為適格的標的，這個產業的核心資產——例如細胞株、轉基因動物，以及其後的 cDNA——便都可以主張權利。其後每一項生物科技專利的立場，基本上都可以說是這一項裁決的邏輯延伸。

Where it stops being true. The decision is not "life is patentable." The same opinion reaffirmed that a new mineral discovered in the earth or a new plant discovered in the wild is not patentable subject matter, and the isolation line later excluded Myriad's gDNA. The operative distinction is human intervention, not the living status of the material. The 5–4 split is also the measure of how close the alternative path came.

邊界 —— 何時不再成立。 這項裁決所說的，並不是「生命可以取得專利」這回事。同一份意見書同時重申，在地底發現的一種新礦物，或在野外發現的一種新植物，都不屬於可專利標的；而其後的分離界線，更把 Myriad 的 gDNA 排除在外。真正起作用的區分，在於其間是否有人為介入，而不是該材料本身是否具有生命。而 5 比 4 這樣的分裂本身，同時也衡量了另一條路徑曾經有多麼接近成真。

Core answer. Because patentable subject matter is defined by each jurisdiction's own law and each office's own test — and in Europe, ethics is expressly part of examination rather than a policy footnote.

核心答案。 因為可專利標的，是由每一個司法管轄區自身的法律，以及每一個專利局自身所採用的檢驗標準所界定的——而在歐洲，倫理明確地屬於審查過程的一部分，而不是政策上的一則附註而已。

Mechanism — why it follows. The USPTO granted a claim to a "transgenic non-human mammal" as a logical extension of Chakrabarty. The EPO initially rejected it and then granted after a "utilitarian balancing test" weighing animal suffering against potential medical benefit to humanity — the first explicit ethical calculation in patent examination. Canada rejected the mouse itself on the ground that a "higher life form" is not a "manufacture" or "composition of matter", and that this was a social and ethical question for Parliament rather than the courts.

機制 —— 為何如此推得。 USPTO 授予了一項「轉基因非人類哺乳動物」的權利要求，作為 Chakrabarty 案的一項邏輯延伸。EPO 最初駁回了該項申請，其後在一項「功利主義的平衡檢驗」之下——權衡動物所承受的痛苦與對人類的潛在醫療效益——予以授予，這也是專利審查的歷史上首次出現的一項明確倫理計算。加拿大則駁回了針對該小鼠本身的權利要求，其理由是所謂「較高生命形式」並非「製品」或「組合物」，並且認為這是一個應當交由國會、而非交由法院處理的社會與倫理問題，因此不應由法院代為作出答案。

Where it stops being true. The Canadian outcome did not reject the invention in full — the process of making the mouse remained patentable. So "you cannot patent this here" is usually a statement about the claim type, not about the underlying invention. And the case must not be over-generalised into "the West patents animals": three Western jurisdictions with similar legal traditions reached three different answers on the same invention.

邊界 —— 何時不再成立。 加拿大的結果並沒有全盤否定該發明——製造該小鼠的製程本身，仍然可以取得專利。因此，「你在此地不能為這個申請專利」通常是一項關於權利要求類型的陳述，而不是一項關於底層發明究竟是什麼的陳述。而且這宗案例不應被過度概括為「西方國家一律為動物授予專利」：三個具有相似法律傳統的西方司法管轄區，竟然對同一項發明得出了三個互不相同的答案。

Core answer. It is a designed feature. TRIPS sets a baseline and deliberately grants wide leeway; by writing the morality exception into the treaty it lets each member impose stricter ethical rules — which is exactly why the US, EU and China diverge despite all being WTO members.

核心答案。 它是一項經過刻意設計的特徵。TRIPS 設定了一條底線，並且刻意給予寬闊的迴旋空間；把道德例外寫入條約之中，便讓每一個成員得以在其境內施加更嚴格的倫理規則——這正是美國、歐盟與中國儘管同為 WTO 成員，卻仍然會出現分歧的原因所在。

Mechanism — why it follows. Article 27.1 requires patents for "any inventions, whether products or processes, in all fields of technology" if new, involving an inventive step and capable of industrial application. Articles 27.2 and 27.3 permit exclusions for ordre public or morality (including human, animal or plant life or health), for diagnostic, therapeutic and surgical methods, and for plants and animals other than micro-organisms and essentially biological processes for producing plants or animals. Note what 27.3 leaves inside the requirement: micro-organisms are not excluded, so most of biotech stays in. The morality exception then lands in national law — EPC Article 53(a) excludes commercial exploitation contrary to ordre public or morality, and Article 5 of the Chinese Patent Law prohibits inventions contrary to the laws of the state or social morality.

機制 —— 為何如此推得。 第 27.1 條要求，成員必須對「任何發明，無論是產品或者是方法，在所有技術領域」授予專利，只要該發明屬於新穎、具備 inventive step，並且能夠作工業上的應用。第 27.2 及 27.3 條則允許成員在特定情況之下作出排除，總共涉及三類客體：第一類是 ordre public 或道德，其中包括了人類、動物或植物的生命或健康在內；第二類是診斷、治療及外科手術方法；第三類則是微生物以外的植物與動物，以及那些主要為生物學的植物或動物生產方法。須特別注意第 27.3 條究竟把什麼留在該項要求之內：微生物並沒有被排除在外，因此生物科技的大部分領域，實際上仍然留在這項要求所涵蓋的範圍之中。而道德這項例外，其後必須落在各國國內法之中，才得以真正落實——EPC 第 53(a) 條排除任何違反 ordre public 或道德的商業利用，而中國專利法第 5 條則禁止任何違反國家法律或社會公德的發明在境內取得專利。

Where it stops being true — the two positions. For harmonisation: a single global rule would let a company plan one claim set, and its absence means a claim valid in the US can be excluded in Europe on grounds that are not technical at all. Against it: the leeway is what lets a member's own moral convictions govern its own market, which is what the drafters wrote in on purpose; and the exclusion sits in the same space as the human body and embryo exclusions, which is where the CRISPR, germline and AI disputes will land.

邊界 —— 兩種立場。 支持協調統一的一方認為：單一的全球規則會讓一家公司只需規劃一套權利要求即可，而缺乏這套規則，便意味着一項在美國有效的權利要求，可能基於一些完全非技術性的理由而在歐洲遭到排除。反對的一方則認為：正是這份迴旋空間，讓每一個成員自身的道德信念得以管治其自身的市場，而這正是起草者當初刻意把它寫入條約的內容；而且這項排除與人體及胚胎的排除位於同一個空間，亦正是 CRISPR、生殖系與人工智能這幾場爭議，日後將會落腳的地方。

Core answer. Because ownership of excised tissue and consent to its use are two different legal questions, and only the second one survived.

核心答案。 因為已切除組織的歸屬問題，與對它加以使用時的同意問題，本來就是兩個彼此不同的法律問題，而在這宗案例之中，只有第二個問題得以成立。

Mechanism — why it follows. On property, the court held against Moore: once tissue is removed the patient no longer holds a property right, on the reasoning that a "litigation lottery" would hinder socially important research. Separately and independently, the physicians breached their fiduciary duty and failed to obtain proper informed consent by not disclosing their research and financial interests. The wrong was not the use of the tissue; it was the concealment.

機制 —— 為何如此推得。 在財產這一邊，法院作出了對 Moore 相當不利的裁定：一旦組織被切除之後，病人便不再持有任何財產權，而其所持的推理是，「訴訟抽獎」將會妨礙那些對社會具有重要意義的研究。另外，並且是獨立地，該等醫生亦違反了他們對病人所負的受信義務，並且因為沒有披露自身的研究與財務利益，而未能取得一份適當的 informed consent。不當之處並不在於使用該組織這件事情本身，而是在於對此事加以隱瞞。

Where it stops being true. The holding is narrow. It denies a conversion claim in excised cells — it does not license research use without consent. The consent duty survives intact and independently, which is why the lecture states that informed consent is non-negotiable and that clinical-trial and tissue-use consent must be watertight and transparent about commercial interests.

邊界 —— 何時不再成立。 這項裁定的範圍其實是相當狹窄的。它所否定的，只是就已切除細胞所提出的一項 conversion 主張——但它並不表示可以在未經同意的情況之下把研究使用合法化。同意的義務則完整而獨立地存續下來，這正是課程內容指出 informed consent 不容妥協，以及臨床試驗與組織使用的同意必須滴水不漏、並且對商業利益保持透明的原因。

Core answer. No. In Hong Kong, IP protection operates on a first-to-file basis: the first person to file with the Intellectual Property Department is the rightful owner. It does not matter who had the idea first — it matters who filed the paperwork first. And even inventorship is not the same thing as ownership.

核心答案。 不對。在香港，知識產權保護是以先申請的原則運作的：最先向知識產權署提出申請的那個人，就是該項知識產權的合法擁有人。誰先有這個想法並不重要，真正重要的是誰先把文件提交上去。換言之，起決定作用的是提交文件的先後，而不是構思的先後。而且，即使某人具備 inventorship 的身分，那也並不等同於擁有權。

Mechanism — why it follows. Two separate questions. First, inventorship: the inventor is the "actual deviser", which means conceiving the core idea, using initiative to go beyond a brief to reach the inventive concept, or identifying the significance of a result that led to the invention. Routine work, offering advice, supervising a project without devising the concept, testing the invention and follow-up work do not on their own make someone an inventor. Second, ownership: the general principle is that an employer owns IP created by employees in the course of their employment, and universities and institutes normally own IP created by staff or students in the course of employment or that makes significant use of institutional resources, funds, time, facilities or equipment.

機制 —— 為何如此推得。 這是兩個彼此獨立的問題，必須把它們分開來處理。第一是 inventorship：所謂發明人，是指「實際的構思者」，意思是指構思出核心的想法，或者主動運用判斷力以超越既定的任務範圍，從而達到該發明構思，或者辨識出一項導致該發明出現的結果，究竟具有什麼樣的重要性。例行性的工作、單純提供意見、監督一項項目而沒有構思出該概念、對該發明進行測試，以及後續的跟進工作，本身都不會使某人因此而成為一名發明人。第二是擁有權：一般的原則是，僱主擁有其僱員在受僱期間所創造的知識產權；而大學及研究院所通常擁有由教職員或學生在受僱或就學期間所創造出來的知識產權，或者擁有那些顯著地使用了機構資源、資金、時間、設施或設備的知識產權。

Where it stops being true. The source is candid that Hong Kong law does not clearly state who owns IP an employee creates during employment — employer ownership is stated as a general principle, not a crisp statutory rule. So the institutional policy is what actually governs, and a student who has not read it does not know their position.

邊界 —— 何時不再成立。 資料來源坦承，香港法律並沒有清楚訂明，僱員在受僱期間所創造的知識產權究竟由誰擁有——僱主擁有權是以一般原則的方式陳述的，而不是一項明確的成文規則。因此，真正起管治作用的，是所屬機構本身所訂立的政策；而一名沒有讀過該政策、也不清楚政策內容的學生，實際上並不知道自己身處什麼位置。

Core answer. It is a research-integrity failure, not primarily a copyright one — and the integrity frame is what makes it actionable, because a fabricated reference has no right-holder to misappropriate.

核心答案。 這是研究誠信上的一種失誤，而不是主要屬於版權的問題——而誠信這個框架，正是使它可被追究的原因所在，因為一條捏造出來的參考文獻，並沒有任何權利人可以供它侵害。

Mechanism — why it follows. Plagiarism is a failure of attribution against someone's expression; copyright protects that expression. A hallucinated citation appropriates nothing, because the source does not exist: the wrong is against the honesty and trust the scholarly record rests on. And the source is explicit that hallucination is structural, not rare — it arises from insufficient training data, incorrect model assumptions and biased training data, it is very difficult to get rid of, and the practical consequence is that references in GenAI text often contain errors or simply do not exist. Fabrication has also become cheap, which shifts the integrity question from "did you copy?" to "is this real?" — and generating figures, images or interview transcripts and presenting them as collected data is fabrication whatever tool produced them, even when the model was tuned sentence by sentence until the output looked like a real micrograph.

機制 —— 為何如此推得。 Plagiarism 是對他人表達的一種標註上的失誤；而版權所保護的，正是該種表達本身。一條由模型幻覺所產生出來的引用，並沒有侵佔任何東西，因為那個來源根本並不存在：它所違背的，正是學術記錄所賴以成立的誠實與信任。而資料來源明確指出，幻覺是結構性的問題，而不是罕見的例外——它源於訓練數據不足、模型假設不正確，以及訓練數據帶有偏見；它極難被徹底消除，而其實際後果就是：GenAI 文本之中的參考文獻往往含有錯誤，又或者根本並不存在。捏造這件事本身也已變得廉價，這把誠信問題從「你有沒有抄襲？」轉向「這是不是真的？」——而生成圖表、圖像或訪談逐字稿，並把它們當作所收集到的真實數據來呈現，無論是由什麼工具產生出來，都屬於捏造；即使模型曾被逐句調校，直至其輸出看起來像一張真實的顯微照片，情況亦是如此。換言之，到了這個時候，問題的關鍵已經不再只是有沒有標明出處。

Where it stops being true. The two frames overlap rather than replace each other. GenAI output may contain the same content as web pages and websites owned by others, so the student who publishes it can also be the one infringing. And both frames depend on the same remedy logic: an integrity process acts on the record and on standing — the source's cases show retraction, revocation of a doctorate, dismissal and demotion, and NSFC grant bans of 3 to 7 years with recovery of allocated funds — while a copyright dispute needs a right-holder and can simply be settled or never brought.

邊界 —— 何時不再成立。 這兩個框架是互相重疊的，而不是彼此取代對方。GenAI 的輸出可能含有與他人所擁有的網頁及網站相同的內容，因此把它發表出來的那名學生，本身也可能正是構成侵權的那一方。而這兩個框架都取決於相同的補救邏輯：其中誠信程序所作用的對象，是紀錄與身分——資料來源的案例所顯示的後果，包括撤稿、撤銷博士學位、被解僱與降職，以及 NSFC 資助禁令 3 至 7 年並追回已撥款項——而版權爭議則需要有權利人的存在，並且可以透過私下和解了結，或者根本不被提起。因此，兩者所處理的對象並不相同。

The trade-off. The narrow trial maximises the probability of a clean win and produces a financeable asset; the all-comers trial produces the label and the addressable market, and only it generates the data a payer reimburses on. One optimises for the next financing; the other optimises for launch.

取捨所在。 窄範圍試驗將乾淨勝出的機率最大化，並產出一個可融資的資產；all-comers 試驗則產出 label 與可及的市場，而且只有它能產生 payer 願意據以作出 reimbursement 的數據。前者為下一輪融資而優化，後者為上市而優化。

What decides it. What the company needs next, not what the biology says. If the next milestone is a partner or a priced round, a positive signal in a defined population is a licensable asset and the rest can be the partner's problem. If the next milestone is launch, all-comers data is the product, and a biomarker-only label permanently caps the market. The decision is a financing decision wearing a scientific costume.

真正的決定因素。 取決於公司下一步需要甚麼，而不是生物學說甚麼。若下一個里程碑是找到合作方或完成一輪定價融資，那麼在明確界定的人群中出現正面信號，就是一個可授權的資產，其餘的可以留給合作方去操心。若下一個里程碑是上市，all-comers 數據本身就是產品，而僅限 biomarker 的 label 會永久封住市場上限。這個決定，是一個披著科學外衣的融資決定。

Where it stops being true. Prevalence of the biomarker. If it marks a large share of patients, the narrow trial is not actually narrow and you may get both answers at once; if it marks a small share, all-comers dilutes the treatment effect toward the null and an underpowered miss becomes likely.

邊界 —— 何時不再成立。 biomarker 的患病率。若它標記了相當大比例的病人，窄範圍試驗其實並不窄，你甚至可能同時得到兩個答案；若它只標記一小部分，all-comers 會把治療效果稀釋至趨向 null，令一個檢定力不足的失敗變得更可能。

What would change my mind. A partner willing to fund the biomarker work alongside the broad trial, which dissolves the either/or entirely.

甚麼會改變我的立場。 若有合作方願意在廣泛試驗之外同時資助 biomarker 相關工作，這個二選一就會徹底消失。

The trade-off. Speed and cost versus acceptability. Single-region enrolment reaches a readout sooner and cheaper. Multi-region enrolment costs time and money and buys a dataset the later regulator will accept without a bridging study.

取捨所在。 速度與成本對比可接受程度。單一地區入組能更快、更便宜地取得 readout。多地區入組則要付出時間與金錢，換來一個日後的監管機構願意接受、無須再進行 bridging study 的數據集。

What decides it. Whether the launch-market regulator actually accepts foreign data alone for this indication. If it does, split enrolment is pure cost. If it does not, a single-region trial buys a readout you cannot file with — the worst outcome, because you paid for data and still owe a local study. Secondary: standard of care and drug metabolism differ by population, so a multi-region trial is also the only way to observe the effect size you will actually get at launch.

真正的決定因素。 取決於上市市場的監管機構是否真的單憑境外數據就接受這個適應症。若接受，分散入組就是純粹的成本。若不接受，單一地區的試驗只買到一個你無法據以申報的 readout——這是最差的結果，因為你付了錢取得數據，卻仍欠一項本地研究。其次是：standard of care 與藥物代謝因人而異，所以多地區試驗也是唯一能觀察到你在上市時實際會得到多大 effect size 的方法。

Where it stops being true. In a rare disease, multi-region is the only way to enrol at all, so geography is forced and the question becomes which regions, not how many. And where an accelerated pathway exists, the bridging requirement may be deferred rather than removed.

邊界 —— 何時不再成立。 在罕見疾病中，多地區是唯一能入組的途徑，因此地域是被迫的，問題就變成選哪些地區，而不是選幾個。而在設有 accelerated approval 途徑的地方，bridging 要求可能只是被延後，而非被取消。

What would change my mind. A reliance or recognition route that formally accepts the foreign dataset — a procedural fact, not a scientific one, and one that moves.

甚麼會改變我的立場。 存在一條 reliance 或 recognition 途徑，正式接受境外數據集——這是程序上的事實，而非科學上的事實，而且會隨政策變動。

The trade-off. Amending is fast and cheap but inherits a statistical problem: a hypothesis generated from the same data used to test it is not independent evidence, and every reviewer knows it. A separate prospective study is clean but costs a trial and a delay. You are trading time against the credibility of the claim.

取捨所在。 修訂快速且便宜，卻繼承了一個統計問題：由同一批用於檢定的數據所產生的假說，並非獨立證據，而每一位評審都心知肚明。另做一項前瞻性研究則乾淨俐落，但代價是一整項試驗與一段延誤。你是在用時間換取這個主張的可信度。

The mechanism. Retrospective subgroup cuts multiply comparisons, and with enough subgroups something will look significant by chance. Pre-specification is what converts a correlation into evidence. Where the subgroup is also biologically motivated and was predicted before the cut, the penalty is smaller but never zero.

機制。 回溯性的 subgroup 切分會令比較次數倍增，只要有足夠多的 subgroup，總會有某一個因偶然看起來顯著。只有事前指定（pre-specification）才能把相關性轉化為證據。若該 subgroup 同時有生物學依據、並且在切分之前就已被預測，代價會較小，但永遠不會是零。

What decides it. Whether the subgroup hypothesis predates the data. If a mechanism predicted it, an amendment with a pre-specified analysis and an adjusted alpha is defensible. If the cut was exploratory, only a new study produces a claim anyone will act on.

真正的決定因素。 取決於這個 subgroup 假說是否早於數據而存在。若有機制預測了它，那麼以一項事前指定的分析與經過校正的 alpha 所作的修訂是站得住腳的。若這個切分只是探索性的，就只有一項新的研究才能產生任何人願意據以行動的主張。

What would change my mind. High biomarker prevalence plus a large effect size, which makes a confirmatory study small and quick enough that the delay is trivial.

甚麼會改變我的立場。 高 biomarker 患病率，加上一個很大的 effect size——這會令確認性研究變得小而快，以致那段延誤微不足道。

The trade-off. Reading it as target risk stops you and preserves capital. Reading it as molecule risk lets you continue and possibly be the one who gets it right. The first is cheap and often correct; the second is the only way any field advances.

取捨所在。 把它解讀為 target 風險，會令你停下來並保住資本。把它解讀為分子風險，則讓你繼續前進，並有可能成為把它做對的那一個。前者便宜，而且往往是對的；後者是任何領域得以進步的唯一途徑。

The mechanism. A Phase III failure carries two kinds of information: target validity, and execution — dose, exposure, population, endpoint, formulation. If their exposure at the failed dose was below yours, or their population broader than your selection, the failure says little about the target. If they achieved good target engagement at an adequate dose and still missed on a hard endpoint, the target is the problem, and a structural difference in your molecule will not rescue it.

機制。 一次 Phase III 失敗承載兩種資訊：target 的有效性，以及執行——劑量、exposure、人群、終點、配方。若他們在失敗劑量下的 exposure 低於你，或他們的人群比你篩選的更廣泛，那麼這次失敗對 target 的說明就很有限。若他們在足夠的劑量下達到了良好的 target engagement，卻仍在一個硬終點上落空，那麼問題就出在 target，而你分子上的結構差異救不了它。

What decides it. Whether the competitor published target-engagement data. A failure with confirmed engagement at an adequate dose is a target signal. A failure without it is uninformative and should not be allowed to kill your programme by default.

真正的決定因素。 取決於競爭對手是否公布了 target-engagement 數據。在足夠劑量下已確認 engagement 的失敗，是一個關於 target 的信號。沒有這項數據的失敗則不具資訊價值，不應被容許默認地扼殺你的項目。

What would change my mind. Your own exposure-response data showing you reach comparable engagement with a materially better therapeutic index — then the molecule difference is the thesis and it is worth the trial.

甚麼會改變我的立場。 你自己的 exposure-response 數據顯示，你以實質上更好的 therapeutic index 達到相若的 engagement——那麼分子上的差異就是你的核心論點，值得為此進行試驗。

The trade-off. Mechanism buys understanding and reduces the chance that the next trial is wrong for a reason you cannot see. Sample size buys statistical confidence and reduces the chance of an underpowered miss. One improves the probability the hypothesis is right; the other improves your ability to detect it. They are not substitutes.

取捨所在。 機制換來理解，並降低下一次試驗因一個你看不見的原因而出錯的機會。樣本量換來統計信心，並降低檢定力不足而落空的機會。前者改善假說為真的機率；後者改善你偵測到它的能力。兩者不能互相替代。

The mechanism. An underpowered trial answers nothing, because a null is uninterpretable — you cannot separate "no effect" from "not enough patients". But a well-powered trial on the wrong dose or the wrong population answers the wrong question confidently. Mechanism data is what tells you which dose and which population a larger sample should test.

機制。 一項檢定力不足的試驗甚麼也回答不了，因為 null 結果無法解讀——你無法把「沒有療效」與「病人不夠多」分開。但一項檢定力充足、卻用錯劑量或錯人群的試驗，會自信地回答了錯誤的問題。機制數據才能告訴你，更大的樣本應該去檢定哪個劑量、哪個人群。

What decides it. The phase. Before dose and population are fixed, mechanism is the higher-yield spend because it defines the design. Once the design is frozen and the only question is whether the effect is real, sample size is. Where it stops being true: if the mechanistic question is not decision-relevant. Engagement markers that will not change a single protocol choice are expensive curiosity, not de-risking.

真正的決定因素。 取決於期別。在劑量與人群確定之前，機制是回報更高的支出，因為它界定了設計。一旦設計凍結、唯一的問題只是療效是否真實時，就是樣本量。何時不再成立：若那個機制問題與決策無關。那些不會改變任何一項 protocol 選擇的 engagement 標記，只是昂貴的好奇心，而不是去風險。

What would change my mind. A regulatory or payer requirement that forces the comparison, which removes the choice.

甚麼會改變我的立場。 一項監管機構或 payer 的要求，強制進行這項比較，也就取消了這個選擇。

The trade-off. As a substitute, RWE is cheap and occasionally accepted — confirmatory use for a well-characterised mechanism, or where randomisation is unfeasible. As a complement, it characterises the population a trial excluded. As a lever, it is not evidence about efficacy at all; it is evidence about use, and its value is in the conversation rather than the file. Choosing the wrong role means designing the wrong study.

取捨所在。 作為替代品，RWE 便宜，而且偶爾會被接受——用於已有充分刻劃的機制的確認性用途，或在無法隨機化的情況下。作為補充，它刻劃了試驗所排除的人群。作為槓桿，它根本不是關於療效的證據；它是關於使用情況的證據，其價值在於對話之中，而非在申報檔案之中。選錯了角色，就等於設計了錯誤的研究。

The mechanism. RWE's central weakness is confounding by indication — the patients who receive a drug differ systematically from those who do not, often because of the reason they received it. Randomisation is the only thing that reliably breaks that link. RWE is therefore strong where the comparator is not a judgement call (within-patient before-and-after, or a natural experiment) and weak wherever it is.

機制。 RWE 的核心弱點是適應症混雜（confounding by indication）——接受某藥物的病人，與沒有接受的病人，存在系統性差異，而原因往往就是他們接受該藥的理由本身。隨機化是唯一能可靠地切斷這條連結的東西。因此，RWE 在對照組的選擇不涉及判斷時很強（病人自身的前後對比，或自然實驗），而在涉及判斷的地方則很弱。

What decides it. Whether the claim is causal and confirmatory, or descriptive and supportive. Adherence, safety in the real population, and outcomes in groups trials excluded are where RWE is genuinely strongest.

真正的決定因素。 取決於這個主張是因果性且確認性的，還是描述性且支援性的。服藥依從性、真實人群中的安全性，以及試驗所排除群體的結果，正是 RWE 真正最強的地方。

What would change my mind. A regulator formally accepting a specific RWE design for a specific claim. That is jurisdiction-specific and moves with policy, so the answer becomes procedural rather than scientific.

甚麼會改變我的立場。 監管機構正式接受某一特定的 RWE 設計，用於某一特定的主張。這因司法管轄區而異，並隨政策變動，所以答案就變成程序性的，而非科學性的。

The trade-off. You want to price against the broadest label you can defend and launch where the price is highest, but those two optima rarely coincide. Pricing against the broad label while launching in the narrow market overstates your addressable population; pricing against the narrow one while launching in the broad market anchors you low.

取捨所在。 你想以你能守住的最寬 label 作為定價依據，並在價格最高的地方上市，但這兩個最優點很少重合。以寬 label 定價卻在窄市場上市，會高估你可及的人群；以窄 label 定價卻在寬市場上市，則會把你錨定在低位。

The mechanism. Price and label are not independent. A payer prices a drug against the alternatives available to the specific patients on the label, so a narrow label means a narrower comparator set, often a higher price per patient, and a smaller volume. The label is the denominator of the entire value argument.

機制。 價格與 label 並非互相獨立。payer 是相對於 label 上那些特定病人所能獲得的替代方案來為藥物定價，所以窄 label 意味著更窄的對照組集合、往往更高的每名病人價格，以及更小的銷售量。label 是整個價值論證的分母。

What decides it. Which market you actually intend to earn in. If the higher-price market is the launch market, price against that label and treat the other as an expansion opportunity. If prices are similar, the tie-break is which decision is more defensible to revisit later.

真正的決定因素。 取決於你實際上打算在哪一個市場賺錢。若高價市場就是上市市場，就以該 label 定價，並把另一個市場視為擴張機會。若價格相近，決勝的關鍵就是哪一個決定在日後重新檢視時更站得住腳。

What would change my mind. A post-approval commitment that harmonises the two labels, which collapses the trade-off at the cost of one more study.

甚麼會改變我的立場。 一項獲批後的承諾，把兩個 label 統一，這會令取捨消失，代價是多做一項研究。

The trade-off. Running it produces directly comparative evidence, which is the strongest possible input to an HTA and is frequently the difference between reimbursement and refusal. Refusing preserves capital and the launch timeline and bets that an indirect comparison will hold. Negotiating redefines the question — which subgroup, which endpoint, at which price.

取捨所在。 做這項試驗會產生直接比較的證據，那是對 HTA 而言最強的輸入，而且往往是 reimbursement 與拒絕之間的分野。拒絕則保住資本與上市時程，賭的是間接比較能夠成立。談判則是重新界定問題——哪個 subgroup、哪個終點、以甚麼價格。

The mechanism. HTA bodies do not evaluate absolute efficacy; they evaluate incremental benefit against the comparator actually used in that country. Without a head-to-head, the comparison is indirect, which adds uncertainty — and uncertainty gets priced, usually as a discount or a restricted population. So the trial is purchasing the removal of a discount.

機制。 HTA 機構評估的並不是絕對療效；它們評估的是相對於該國實際使用的對照方案所帶來的增量效益。沒有 head-to-head，比較就是間接的，這會增加不確定性——而不確定性會被定價，通常以折扣或限制人群的形式出現。所以這項試驗買的，是取消一項折扣。

What decides it. How large the price premium you are defending is. If the premium is small, the trial is not worth its cost. If it is large, the trial is cheap relative to the revenue it protects. Reimbursement risk is an economic question, so it is answered with arithmetic, not with science.

真正的決定因素。 取決於你要守住的價格溢價有多大。若溢價很小，這項試驗不值得它的成本。若溢價很大，相對於它所保護的收入，試驗就很便宜。reimbursement 風險是一個經濟問題，所以要用算術來回答，而不是用科學。

What would change my mind. A payer willing to accept a well-conducted indirect comparison together with a managed-entry agreement capping their downside — then the trial can safely be deferred past launch.

甚麼會改變我的立場。 一名 payer 願意接受一項執行良好的間接比較，並配以一項為其下行風險設限的 managed-entry 協議——那麼這項試驗就可以安全地延後到上市之後。

The trade-off. On the direct case — revenue minus cost — the answer is usually no. The indirect case is what you are really buying: a reference approval other regulators rely on, a quality and manufacturing system inspected to that standard, or continued access as a condition of a partnership. You are buying an option, not a market.

取捨所在。 就直接的帳目——收入減成本——而言，答案通常是「不」。你真正買到的是間接的帳目：一個其他監管機構會援引的參考批准、一套按該標準接受查驗的品質與生產體系，或作為合作條件而維持的市場準入。你買的是一個選擇權，而不是一個市場。

The mechanism. Regulatory decisions are used as inputs by other regulators, tenders and payers through reliance and recognition pathways. An approval from a stringent authority functions as a certificate. An approval from a small market with a long queue is often just a cost, unless that market's decision is itself referenced elsewhere or unless someone else is paying for the filing.

機制。 透過 reliance 與 recognition 途徑，監管決定會被其他監管機構、招標方與 payer 當作輸入使用。來自嚴格監管機構的批准，起到的是一紙證書的作用。來自一個排隊很長的小市場的批准，往往只是一項成本——除非該市場的決定本身會被其他地方援引，或除非申報的費用是由別人支付。

What decides it. Whether the approval actually functions as an input to a regulator you care about, and who bears the cost. If a partner or a government funds the local filing, the cost is theirs and the option is yours — which is the only version of this that is cleanly positive.

真正的決定因素。 取決於這項批准是否真的對你在意的某個監管機構起到輸入的作用，以及由誰承擔成本。若由合作方或政府出資進行本地申報，成本是他們的，而選擇權是你的——這是唯一一種乾淨地屬於正回報的情況。

What would change my mind. A change in a reliance pathway, or a procurement rule that accepts only locally registered product. Either one converts a cost centre into a gate.

甚麼會改變我的立場。 reliance 途徑出現變化，或出現一條只接受本地註冊產品的採購規則。兩者之中任何一項，都會把一個成本中心變成一道關卡。

The trade-off. Publishing adds to the public record and stops others repeating the work, and it costs you a defensible position — a disclosed failure tells competitors which target or scaffold is dead, and can be read by investors as evidence of a weak pipeline. Silence protects both, at the expense of everyone else.

取捨所在。 公布會為公共紀錄增添內容，並阻止其他人重複這項工作，但代價是失去一個可守的位置——被披露的失敗會告訴競爭對手哪個 target 或骨架已經死掉，也可能被投資者解讀為 pipeline 薄弱的證據。保持沉默則能同時保住這兩者，代價由其他所有人承擔。

The mechanism. The information asymmetry is real. A negative result is a free research subsidy to whoever reads it: it narrows the global search space at your expense. But silence is also a systemic cost, because the literature becomes biased toward positives and the same failed hypotheses get retried across companies for decades. You are choosing between a private gain and a public good.

機制。 這種資訊不對稱是真實存在的。一個陰性結果，對任何讀到它的人來說都是一筆免費的研究補貼：它縮小了全球的搜尋空間，而代價由你承擔。但沉默同樣是一種系統性成本，因為文獻會偏向正面結果，同一批失敗的假說會在數十年間於不同公司之間被反覆重試。你是在私人利益與公共財之間作選擇。

What decides it. Whether the result is about your molecule or about the mechanism. A molecule-specific failure is a legitimate trade secret. A mechanism-level finding is the kind the field cannot accumulate without, and withholding it moves from competitive caution toward a public-good problem.

真正的決定因素。 取決於這個結果是關於你的分子，還是關於機制。針對特定分子的失敗，是正當的 trade secret。機制層面的發現，則是這個領域缺少了就無法累積的那一類知識，把它扣下不發，就從競爭上的謹慎，變成了公共財的問題。

What would change my mind. A registry, or a disclosure requirement at funding or listing level. Once disclosure is expected, the calculus flips and the cost becomes compliance rather than competitive disadvantage.

甚麼會改變我的立場。 出現一個登記庫，或在融資或上市層面出現披露要求。一旦披露成為預期，整個計算就會翻轉，成本就變成合規成本，而不再是競爭劣勢。

The trade-off. Treat it as a pricing problem and you take the discount. Treat it as an evidence problem and you spend on a study whose result you cannot guarantee. The first trades margin for access; the second preserves price and risks another negative outcome. Getting the diagnosis wrong wastes the expensive option.

取捨所在。 把它當成定價問題處理，你就接受折扣。把它當成證據問題處理，你就把錢花在一項結果無法保證的研究上。前者以利潤換取市場準入；後者保住了價格，卻冒上又一個陰性結果的風險。診斷錯了，就浪費了那個昂貴的選項。

The mechanism. A refusal is usually a statement about uncertainty, not about price — the payer cannot resolve something. It might be the comparator set, the durability of effect, the generalisability of the trial population, or a budget impact too large at any offered price. Each of those has a different remedy, and a discount only addresses one of them.

機制。 一次拒絕，通常是關於不確定性的表述，而不是關於價格——payer 有某件事無法釐清。那可能是對照組的集合、療效的持久性、試驗人群的可外推性，或是無論開甚麼價格都過大的預算衝擊。這每一項都有不同的補救方法，而折扣只能處理其中一項。

What decides it. The payer's stated reason, read literally and not charitably. If the objection is uncertainty about durability, no discount fixes it. If the objection is budget impact at any price, the answer may be a managed-entry agreement rather than another trial.

真正的決定因素。 取決於 payer 所陳述的理由，要照字面讀，而不要往善意處讀。若反對的理由是對持久性的不確定，任何折扣都解決不了。若反對的理由是無論甚麼價格都有預算衝擊，答案可能是 managed-entry 協議，而不是另一項試驗。

What would change my mind. An outcomes-based agreement shifting risk to the manufacturer and making the payer's exposure conditional. That can convert a refusal into coverage without new efficacy data — but only where the outcome is measurable and attributable.

甚麼會改變我的立場。 一份以結果為基礎的協議，把風險轉移給製造商，並令 payer 的風險敞口變成有條件的。這可以在沒有新療效數據的情況下，把一次拒絕轉化為納入 reimbursement 的範圍——但只在結果可量度且可歸因的情況下才行。

The trade-off. It is a question about who carries the binary and who captures the optionality. Licensing early converts an uncertain large number into a certain smaller one and hands Phase III risk to the partner. Waiting preserves the upside but converts your whole equity story into the outcome of one trial.

取捨所在。 這是一個關於由誰承擔這個二元結果、由誰捕捉這份選擇權的問題。提早授權，會把一個不確定的大數字轉換成一個確定的小數字，並把 Phase III 風險交到合作方手上。等待則保住上行空間，卻把你整個股權故事都押在一項試驗的結果之上。

The mechanism. A licence price is roughly risk-adjusted future cash flow, discounted at the buyer's required return. Waiting does more than raise the price: it changes who is pricing the asset. A partner with a portfolio can absorb a Phase III miss; a single-asset company cannot. The discount you accept is the market price of not being able to absorb that miss — not a failure of negotiation.

機制。 授權價格大致就是經風險調整的未來現金流，以買方要求的回報率折現。等待不只是把價格抬高：它改變了由誰來為這個資產定價。擁有組合的合作方能夠承受一次 Phase III 的落空；只有單一資產的公司則不能。你所接受的折讓，就是你承受不了那次落空的市場價格——而不是談判失敗。

What decides it. Whether a miss would end the company or merely delay it. If a miss is survivable, waiting is rational because you are selling from strength. If it is not, licensing at Phase II is insurance and should be described internally that way.

真正的決定因素。 取決於一次落空會終結這家公司，還是僅僅令它延後。若這次落空是可承受的，等待就是理性的，因為你是在強勢中出售。若承受不了，那麼在 Phase II 授權就是保險，而且在內部就應該這樣描述它。

What would change my mind. A co-development or NewCo structure in which you retain economics while sharing cost and risk. That preserves optionality without carrying the whole binary, and it changes the arithmetic more than any price negotiation does.

甚麼會改變我的立場。 一種共同開發或 NewCo 結構，讓你在分擔成本與風險的同時保留經濟權益。這既保住了選擇權，又不必獨自承擔整個二元結果，而且它對帳目的改變，比任何價格談判都大。

The trade-off. A secret is indefinite and requires no disclosure, but it is only as strong as your ability to keep it contained and to detect misappropriation. A patent is time-limited and public, but it binds anyone including independent inventors. Neither dominates; they fail differently.

取捨所在。 秘密沒有期限，也不需要披露，但它的強度僅限於你把它守住、並且偵測到不當使用的能力。專利有時間限制而且是公開的，但它約束任何人，包括獨立的發明者。兩者並無高下之分；它們失敗的方式不同。

The mechanism. The funding difference is about verifiability, not strength. A patent is a registered, auditable asset a diligence team can read and price. A trade secret has to be evidenced by process — access controls, NDAs, documented handling — and its value collapses on disclosure, which in diagnostics or software can happen the moment a competitor reverse-engineers the output. Investors price what they can verify, so a secret-only asset attracts funders who know the field well enough to assess the moat themselves, and repels generalists.

機制。 融資上的差異在於可驗證性，而非強度。專利是一項已註冊、可審計的資產，盡職調查團隊可以閱讀並為它定價。trade secret 則必須以流程作證——存取控制、NDA、有紀錄的處理程序——而它的價值會隨披露而崩塌，在診斷或軟件領域，這種事可能在競爭對手逆向解出結果的那一刻就發生。投資者只為他們能驗證的東西定價，所以一個只有秘密的資產，會吸引那些熟悉這個領域、有能力自行評估護城河的資助者，而排斥一般化的投資者。

What decides it. Whether the asset is observable from the product. If the product reveals the method, secrecy is a fiction. If the method is genuinely unobservable, it can be stronger than a patent would have been — the case that pushed parts of diagnostics toward secrecy after eligibility was narrowed.

真正的決定因素。 取決於這個資產是否能從產品中被觀察到。若產品會洩露方法，保密就是一種虛構。若方法確實無法被觀察，它可能比一項專利本來會有的保護更強——這正是專利適格性被收窄之後，把診斷領域的一部分推向保密的那種情況。

What would change my mind. A shift in the legal test that makes the correlation claimable, which would make patenting the lower-risk route again.

甚麼會改變我的立場。 法律檢驗標準出現轉變，令該相關性關係變得可以主張，這會使申請專利再次成為風險較低的路徑。

The trade-off. Designing around is scientifically slowest but leaves you with clean, unencumbered IP — the most valuable outcome if it works, and sometimes impossible. Licensing is fastest and secures freedom to operate, but converts your asset into a royalty-bearing one and hands the upstream holder a veto over your future deals. Challenging is cheapest if it succeeds and existential if it fails, and it is a public declaration of war.

取捨所在。 繞道設計在科學上最慢，但會為你留下乾淨、無負擔的 IP——若成功，這是最有價值的結果，而有時它並不可行。取得授權最快，並確保 freedom to operate，卻把你的資產變成一個須付 royalty 的資產，並把對你未來交易的否決權交到上游持有者手上。提出挑戰若成功則成本最低，若失敗則是存亡問題，而且它是一份公開的宣戰書。

The mechanism. Patentability and freedom to operate are separate questions: you can hold a perfectly valid patent on your improvement and still be blocked from selling by a broader claim. The question is not whether your own IP is good — it is whether someone else's claim reads on your product. That is why the earlier Remdesivir situation mattered: a valid new-use patent in one country, with no freedom to operate because a broader composition claim already covered the space.

機制。 可專利性與 freedom to operate 是兩個不同的問題：你可以就自己的改良持有一項完全有效的專利，卻仍被一項更寬廣的主張擋住、無法銷售。問題不在於你自己的 IP 好不好——而在於別人的主張是否讀到了你的產品。這正是早前 Remdesivir 的情況之所以重要的原因：在某一國家持有一項有效的新用途專利，卻沒有 freedom to operate，因為一項更寬廣的組合物主張已經涵蓋了這個空間。

What decides it. The strength of the upstream claim and how much of your programme's value sits on that one element. A narrow claim invites a design-around. A broad, well-prosecuted claim on an unavoidable target makes licensing inevitable, and the negotiation becomes about terms rather than strategy.

真正的決定因素。 取決於上游主張的強度，以及你的項目價值有多大一部分落在那單一要素之上。狹窄的主張會招來繞道設計。一項針對無可迴避之 target 的寬廣且審查得當的主張，會令取得授權變得無可避免，而談判就變成關於條款，而不是關於策略。

What would change my mind. Your runway. A challenge takes years and the outcome is binary; if your cash runs out before the decision, licensing is the answer whether or not the claim is weak — and the valuation story should say so plainly.

甚麼會改變我的立場。 取決於你的資金跑道。一場挑戰要花數年，而結果是二元的；若你的現金在裁決之前耗盡，那麼無論那項主張是否薄弱，取得授權就是答案——而 valuation 的故事也應該直白地這樣說。

The trade-off. File where you will sell and you protect revenue. File where you will manufacture and you protect process. File where competitors will manufacture and you protect against the substitute you never see. Those are three different lists, and you only get to pick three entries across them.

取捨所在。 在你會銷售的地方申請，你保護的是收入。在你會生產的地方申請，你保護的是製程。在競爭對手會生產的地方申請，你保護的是防範那個你永遠看不見的替代品。這是三份不同的清單，而你只能在這三者之間挑選三個名額。

The mechanism. Patent rights are territorial: a patent only stops someone making or selling within that jurisdiction. A product manufactured in a country you did not file in and sold into one you did is still an import problem, so manufacturing geography is part of the IP strategy rather than a separate operations decision. Filing broadly in markets you will never enter is expensive and almost purely defensive.

機制。 專利權具地域性：一項專利只能阻止他人在該司法管轄區內製造或銷售。一個在你沒有申請的國家製造、再賣進你有申請的國家的產品，仍然是一個進口問題，所以生產的地理佈局是 IP 策略的一部分，而不是一項獨立的營運決定。在你永遠不會進入的市場廣泛申請，既昂貴，又幾乎純粹是防禦性的。

What decides it. Where the money and the molecules are: the largest market by revenue, the country where the active is made or the process is practised, and the market where the nearest competitor manufactures. Filing coverage is therefore a legible claim about your commercial model — investors and partners read it as a statement of intent.

真正的決定因素。 取決於錢與分子在哪裡：按收入計最大的市場、原料藥製造地或製程實施地，以及最接近的競爭對手進行生產的市場。因此，申請覆蓋範圍是一個關於你商業模式的、可被讀懂的主張——投資者與合作方會把它讀作一份意向聲明。

What would change my mind. A change in manufacturing footprint, or a partner taking a territory. Both move the answer, and neither is a legal question.

甚麼會改變我的立場。 生產佈局出現變化，或合作方取走某個區域。兩者都會改變答案，而兩者都不是法律問題。

The trade-off. A single upfront payment matches the economics of a one-time intervention and the value of avoided lifetime care, but it collides with annual payer budgeting and with how patients change insurers. Spreading payment over years matches both budgets and the durability question, but requires someone to keep paying after the patient has moved on.

取捨所在。 一次性的 upfront 付款，符合一次性介入的經濟特性，也符合所避免之終身照護的價值，但它與 payer 的年度預算編列、以及病人轉換保險公司的現實相衝突。把付款分攤到數年，既配合預算，也回應了持久性的問題，但它要求有人在病人已經離開之後仍繼續付款。

The mechanism. The affordability problem is a timing mismatch, not a price objection: benefit accrues over decades while the budget is annual and the insured population turns over. Annuity and outcomes-based models solve the timing but create a harder problem — attribution. You need a measurable outcome, a defined window, and a counterparty still present when the payment falls due. Models that assume a stable payer and a clean outcome measure work on paper and break on churn.

機制。 可負擔性的問題是一個時間錯配，而不是對價格的異議：效益在數十年間累積，而預算是年度的、投保人群又不斷更替。年金式與以結果為基礎的模式解決了時間問題，卻製造了一個更難的問題——歸因。你需要一個可量度的結果、一個界定的時間窗，以及在付款到期時仍然存在的交易對手。那些假設 payer 穩定、結果指標乾淨的模式，在紙上可行，卻會在人群流失時瓦解。

What decides it. How durable the effect is and how measurable the outcome. A therapy with a verifiable endpoint and a demonstrable multi-year effect can carry an outcomes-based contract. One whose benefit is a slow risk reduction over many years cannot, because neither side can settle the trigger.

真正的決定因素。 取決於療效有多持久，以及結果有多可量度。一項擁有可驗證終點與可證明之數年療效的治療，可以承載一份以結果為基礎的合約。若某項治療的效益是多年間緩慢的風險下降，則不能，因為雙方都無法就觸發條件達成一致。

What would change my mind. A reinsurance or risk-pooling mechanism that lifts the long-tail liability off the individual payer. That is the genuinely missing infrastructure; without it most of the elegant models stay theoretical.

甚麼會改變我的立場。 一套再保險或風險匯聚機制，把長尾負債從單一 payer 身上移走。那才是真正缺失的基礎設施；沒有它，大部分精巧的模式都只停留在理論層面。

The trade-off. You give up operational simplicity and a single counterparty, and you take on the cost and capability of commercialising elsewhere yourself. You are protected from the risk that one partner underperforms in your best market, and from financing the entire launch from your own balance sheet. The split trades concentration for control.

取捨所在。 你放棄了營運上的簡便與單一交易對手，並自行承擔在其他地方進行商業化的成本與能力要求。你受到保護，免於一個合作方在你最好的市場表現不濟的風險，也免於要用自己的資產負債表為整個上市融資。這種拆分是以集中換取控制。

The mechanism. A global licence transfers execution risk but also concentrates it. If the partner deprioritises your asset relative to their own portfolio, you cannot fix that from outside. A territory split keeps the markets you value and hands over the ones you cannot build in — but it multiplies your regulatory, supply and commercial overheads, and splits one evidence package across two launch timelines and two reimbursement conversations.

機制。 一份全球授權轉移了執行風險，卻也把風險集中起來。若合作方相對於自己的組合，把你的資產排到較低優先次序，你從外部無法糾正。按區域拆分，保住了你看重的市場，並把你無法在當地建立的市場交出去——但它會令你的監管、供應與商業開支倍增，並把同一套證據包拆分到兩條上市時程與兩場 reimbursement 對話之中。

What decides it. Whether you have, or can hire, commercial capability in the retained territory. If you cannot discharge that obligation, keeping the territory is a liability dressed as an option, and you would have been better selling it.

真正的決定因素。 取決於你在保留的區域中，是否擁有或能否聘得商業能力。若你無法履行那項義務，保留該區域就是一項扮成選擇權的負債，而你本來把它賣掉會更好。

What would change my mind. A partner offering genuine co-development with shared global rights instead of a geographic carve-out. That keeps everyone aligned on one global plan and removes the incentive to deprioritise anyone.

甚麼會改變我的立場。 合作方提出真正的共同開發、共享全球權益，而不是按地域切分。這會令各方在同一份全球計劃上保持一致，並消除把任何一方排到較低優先次序的誘因。

The trade-off. A broad declaration is cheap to comply with and easy to police, but it is uninformative. A granular standard captures what actually matters and imposes real work on honest authors. The choice is between a rule everyone follows and a rule that genuinely discriminates between acceptable and unacceptable use.

取捨所在。 一項籠統的聲明，遵守起來成本低、也易於監管，但它不具資訊價值。一套細緻的標準能捕捉真正要緊的東西，卻對誠實的作者構成實際的負擔。這個選擇，是在一條人人都會遵守的規則，與一條真正能區分可接受與不可接受使用的規則之間。

The mechanism. The harm is not that a tool was used; it is that a reader cannot tell whether a figure represents an observation or a generation. A binary declaration cannot separate those, because "I used AI" is equally true of a spelling check, a plotting script, and a fabricated micrograph. What matters is provenance and reproducibility: whether the image traces back to raw data, and whether the process is repeatable. A declaration that does not establish provenance is a compliance artefact, not evidence.

機制。 傷害不在於使用了某個工具；而在於讀者無法分辨某一幅圖所代表的是觀察，還是生成。一個二元的聲明無法區分這兩者，因為「我使用了 AI」這句話，對一次拼寫檢查、一段繪圖腳本，以及一張偽造的顯微照片而言，同樣成立。要緊的是來源與可重複性：這幅影像是否可追溯到原始數據，以及這個過程是否可重複。一項未能確立來源的聲明，只是合規文件，而不是證據。

What decides it. Which uses have a plausible honest version. Language editing and plotting assistance do; generating an image presented as acquired data does not. So a workable standard requires tool name, version and the specific step, plus that images in results be traceable to source data — which already excludes generated figures regardless of what is declared.

真正的決定因素。 取決於哪些用途存在一個說得通的誠實版本。語言編輯與繪圖輔助有；產生一幅被當成所取得數據來呈現的影像則沒有。所以一套可行的標準，要求提供工具名稱、版本與具體步驟，並且要求結果中的影像可追溯到原始數據——這一點本身就已排除生成的圖片，無論作了甚麼聲明。

What would change my mind. Detection reliable enough to verify provenance independently. Until that exists the standard has to rest on disclosure plus a paper trail, which is exactly why major journals treat generated images as unacceptable rather than merely declarable.

甚麼會改變我的立場。 要有足夠可靠的偵測，能獨立地驗證來源。在這種技術存在之前，標準只能建立在披露加上書面紀錄之上，這正是主要期刊把生成的圖片視為不可接受、而不只是需要聲明的原因。

The trade-off. The instinct is to fix it quietly and quickly, or to wait until you understand the full extent before saying anything. Both fail in the same direction: they delay the moment others can rely on the correction, and they make the eventual disclosure look managed rather than made. The trade is between the comfort of certainty and the duty to inform early.

取捨所在。 本能反應是悄悄而迅速地把它修好，或者等到你完全弄清楚影響範圍之後才開口。兩者失敗的方向相同：它們都延後了別人可以信賴更正的那一刻，並且令最終的披露看起來像是被操控的，而不是主動作出的。這裡的取捨，是在確定性所帶來的安心，與及早告知的責任之間。

The mechanism. The order matters because the first actions are the ones you cannot undo, and because the information belongs to more people than you. Anyone who has cited, built on, or made a decision from that paper is exposed from the moment you know. So the sequence runs outward from disclosure: your co-authors and corresponding author immediately, then the journal, then the correction or erratum as needed.

機制。 次序要緊，是因為最初的幾個行動是你無法收回的，也因為這項資訊屬於你以外的更多人。任何引用過、據以建構、或據以作出決定的人，從你知道的那一刻起就暴露在風險中。所以次序是由披露向外推進：立即通知你的共同作者與通訊作者，然後是期刊，然後按需要發出更正或勘誤。

What decides it. The nature of the error, not its severity. An honest mistake gets a correction. A fabrication or manipulation is a retraction and is not yours to manage alone — it goes to the institution. Delaying in order to determine which it is, is itself the failure, because you cannot be both the person investigating and the person deciding how much to reveal.

真正的決定因素。 取決於錯誤的性質，而不是它的嚴重程度。一個誠實的過失，得到的是更正。一項偽造或操弄，則是撤稿，而且不是你一個人可以處理的——它要交給機構。為了判定它屬於哪一種而拖延，本身就是那個過失，因為你不可能同時是調查的人，又是決定要揭露多少的人。

What would change my mind. If the error is genuinely typographical and affects no result or interpretation, the corrective action still applies but the urgency and the audience are smaller. The principle does not change; only the scope does.

甚麼會改變我的立場。 若該錯誤真的是排印之誤、且不影響任何結果或解讀，更正的行動仍然適用，只是迫切性與受眾較小。原則不變，改變的只是範圍。

The trade-off. Raising it early protects the record and gives the co-author a chance to explain. Raising it formally damages a colleague and a working relationship and may be premature. The honest position is that these are sequential rather than alternative — the informal route must come first, and it must have a deadline attached.

取捨所在。 及早提出，能保護紀錄，並給共同作者一個解釋的機會。正式提出，則會損害一位同事與一段工作關係，而且可能言之過早。誠實的立場是：這兩者是有先後次序的，而不是二選一——非正式的途徑必須先走，而且必須附上一個期限。

The mechanism. Irreproducibility has several innocuous explanations: a missing step, a different analysis version, a file-naming error, a legitimate processing step left undocumented. It becomes something else when the raw data cannot be produced at all, or when the reported values cannot be derived from them under any reasonable procedure. The distinction is between a documentation problem and a data problem, and only the co-author can close the first.

機制。 無法重現有幾種無害的解釋：漏了一個步驟、分析版本不同、檔案命名出錯、一個正當的處理步驟沒有留下紀錄。但當原始數據根本無法提出，或所報告的數值在任何合理的程序下都無法由這些數據推導出來時，它就變成另一回事了。這其中的分別，是紀錄問題與數據問題之別，而只有共同作者能解決前者。

What decides it. Whether the raw data exist and can be produced. If they do and the gap is in the method, it is an internal fix. If they cannot be produced, or appear altered, that is a research-integrity matter and belongs to the institution and the journal — whether or not you wanted it to.

真正的決定因素。 取決於原始數據是否存在、是否能夠提出。若存在，而缺口在於方法，那就是內部可以修好的事。若無法提出，或看起來曾被改動，那就是一宗研究誠信的事，屬於機構與期刊處理——無論你是否希望如此。

What would change my mind. A co-author who engages fully and supplies a complete, reproducible path. That resolves it internally and the escalation never happens. Refusal or indefinite delay is the signal that it has stopped being internal.

甚麼會改變我的立場。 共同作者全面配合，並提供一條完整、可重現的路徑。這樣就能在內部解決，升級也就不會發生。拒絕配合或無限期拖延，就是它已不再是內部事務的信號。

The trade-off. The data exist, the science is valid, and the patients consented to research — but not to this. Treating the existing consent as sufficient because the research is broadly similar is convenient and frequently wrong, because consent is specific and the commercial dimension is precisely the part people object to.

取捨所在。 數據存在、科學有效，而病人也同意了研究——但不是同意這個。因為研究大致相似，就把現有的同意視為足夠，這樣做很方便，卻往往是錯的，因為同意是有特定範圍的，而商業這個面向，恰恰就是人們所反對的部分。

The mechanism. Consent is a permission tied to a described purpose, and a use outside that description is simply not covered, however beneficial. Genomic data compounds this: it is not fully de-identifiable in the way most data is, and it can implicate relatives who never consented to anything. So the constraint is not merely procedural — the old consent is inadequate because a new class of use, with a new class of beneficiary, was never disclosed to anyone.

機制。 同意是一項與所述目的綁定的許可，而在該描述之外的用途，無論多有益，就是不在涵蓋範圍之內。基因組數據令這一點更複雜：它不像大多數數據那樣可以完全去識別化，而且可能牽連到從未同意過任何事的親屬。所以這項約束不只是程序性的——舊同意之所以不足，是因為一種新的使用類別、配上一種新的受益者類別，從未向任何人披露過。

What decides it. Whether the original consent was broad enough to cover unspecified future research, and whether the governance body with authority over those samples can authorise the new use. If the consent was narrow and that body cannot expand it, the honest answer is that the constraint is real and the options are reconsent or a different dataset.

真正的決定因素。 取決於原本的同意是否寬廣到涵蓋未指明的未來研究，以及對那些樣本擁有權限的治理機構能否授權這項新用途。若同意是狹窄的，而該機構又無法擴大它，誠實的答案就是：這項約束是真實的，選項只有重新取得同意，或改用另一套數據集。

What would change my mind. A governance framework with genuine authority to approve the use, plus a benefit-sharing arrangement addressing the very thing the original consent omitted. That can make the use defensible without fresh consent — but only if the authority is substantive rather than a formality.

甚麼會改變我的立場。 一套擁有真正權限去批准該用途的治理框架，再加上一項利益分享安排，針對的正是原同意所遺漏的那件事。這可以令該用途在沒有新同意的情況下仍然站得住腳——但前提是那項權限是實質的，而不是形式上的。

The trade-off. A regulatory submission is advocacy, and selecting a defensible analysis is legitimate — every sponsor does it, and the alternative is a formless data dump. The difference between choosing the best honest presentation and engineering a misleading one is not located in any single choice. It is located in whether the framing survives scrutiny by someone actively trying to overturn it.

取捨所在。 監管申報本身就是一種倡議，選擇一項站得住腳的分析是正當的——每個申辦方都這樣做，而另一個極端則是毫無條理的數據堆。選擇最誠實的最佳表述，與精心設計一套誤導性的表述，兩者之間的差別，並不在於任何單一的選擇。差別在於：這個表述能否經得起一個積極嘗試把它推翻的人的審視。

The mechanism. The harm is not falsity; it is the removal of the reader's ability to correct for the framing. A relative risk reduction and an absolute one can describe identical data and support different decisions, and the manipulation lies in which is foregrounded and whether the other is available. If the reviewer has the numbers, the framing is a matter of judgement. If they do not, a decision has been made on their behalf.

機制。 傷害並不在於虛假；而在於剝奪了讀者對這種表述作出校正的能力。相對風險下降與絕對風險下降，可以描述同一批數據，卻支持不同的決定，而操弄就在於把哪一個放在前面、以及另一個是否可得。若評審手上有數字，表述就是判斷的問題。若他們沒有，那就是有人替他們作了決定。

What decides it. Whether the full data accompany the chosen framing, and whether the choice would change under a different but equally valid analysis. A pre-specified analysis is defensible; a post-hoc selection presented as pre-specified is not — and the two can be indistinguishable in a summary table.

真正的決定因素。 取決於完整的數據是否伴隨所選的表述，以及在另一項同樣有效的分析之下，這個選擇是否會改變。一項事前指定的分析站得住腳；一項事後挑選卻被呈現為事前指定的分析則不——而這兩者在一張摘要表裡可能無法區分。

What would change my mind. A reviewer or regulator who specifies the analysis in advance and has access to source data. That externalises the judgement and removes the incentive, which is exactly why protocols, statistical analysis plans and data transparency exist.

甚麼會改變我的立場。 由一位評審或監管機構事先指定該分析，並可取得原始數據。這把判斷外部化，並消除了誘因，這正是 protocol、統計分析計劃與數據透明之所以存在的原因。