利用 Claude 赋能的抗体设计开发食物过敏 IgE 抑制剂

通过 Claude 赋能的抗体设计打造同类最佳的食物过敏 IgE 干扰剂

Ezra Erives ∗, Emilio Ferrara ∗, Sean Michael, Nikhil Mallipeddi, Bowen Jing ‡

∗这些作者贡献相当 ‡通讯联系人：bowen@divergencelabs.com

摘要

全球估计有 2.2 亿至 2.5 亿人受食物过敏困扰，严重反应的持续风险对他们的生活质量造成了沉重负担。过敏反应由肥大细胞和嗜碱性粒细胞上结合在其高亲和力受体 Fc εRI 上的过敏原特异性 IgE 交联触发，导致脱颗粒、炎症，在严重情况下会导致过敏性休克。奥马珠单抗（Xolair）是 FDA 批准的唯一用于食物过敏的生物制剂，它能中和循环中的 IgE；然而，它不能有效地从 Fc εRI 上清除 IgE，IgE 在该受体上可以持续存在数周，甚至少量的残留量也能维持敏感性。在此，我们描述了一项借助 Claude 进行的自主计算设计行动，旨在开发一种能够将结合的 IgE 从 Fc εRI 上物理移除的颠覆性抑制剂抗体。在我们研究团队的指导下，Claude 寻求开发一种双表位分子，它将一个能将 IgE 从 Fc εRI 上置换出来的干扰臂与一个能将分子固定在受体结合 IgE 上的非竞争性锚定臂结合在一起。在 91 种经实验表征的干扰剂设计中，有 20 种通过竞争性生物膜层干涉技术显示出对受体结合 IgE 的活性，并且在剂量滴定测定中，三种变体在治疗相关的浓度（25–400 nM）下比奥马珠单抗清除了更多的 IgE。锚定臂的从头设计以 3.4–235 nM 的表观亲和力结合不同表位的 IgE。这些亚组分为进一步表征以开发同类最佳的 IgE 干扰剂提供了起点。

1 引言

食物过敏是一种特应性疾病，其特征是对原本无害的环境或食物抗原产生过量的反应性 IgE。过敏原诱导结合在肥大细胞和嗜碱性粒细胞高亲和力受体 Fc εRI 上的 IgE 发生交联，释放出预先形成和新合成的炎症介质，从而产生从轻度荨麻疹到严重、致命的过敏性休克等典型症状（图 2b）[1]。在美国，该病影响高达 8% 的儿童和约 10% 的成人，总计约 3300 万人，其中约 40% 的患病儿童对一种以上的食物产生反应 [2–4]。直到最近，食物过敏患者的唯一选择仍是严格避免接触过敏原，以及在意外接触时使用可自我注射的肾上腺素 [5]。食物过敏的基本途径是过敏原特异性 IgE 与肥大细胞和嗜碱性粒细胞的高亲和力受体 Fc εRI 的相互作用，这是人类免疫学中最强的蛋白质-受体相互作用之一。IgE 与 Fc εRI α 结合的亲和力接近 10 −10 M，解离速度极慢，无论是否存在过敏原，肥大细胞和嗜碱性粒细胞都保持完全致敏状态 [6–8]，这由 B 细胞持续产生 IgE 提供支持 [9–11]。虽然游离 IgE 的血清半衰期约为两天，但受体结合的 IgE 在人类皮肤中可持续数周至数月 [9, 12–14]。此外，肥大细胞和嗜碱性粒细胞对 IgE 交联极其敏感。作为参考，一个嗜碱性粒细胞带有约 10 5 个 Fc εRI，然而只需几千个过敏原特异性 IgE 分子（占受体库的百分之一或二）即可产生半最大反应，只需几百个就足以引发分泌 [15–17]。因此，食物过敏原 1a 治疗论文与机制

干扰臂打开 C ε3 锚定臂结合受体结合的 IgE IgE 离开，无交联设计目标：干扰臂锚定臂 b Claude 策划的设计战役目标分解与策略文献 · 机制 · 建模工具 · 阶段计划干扰臂重新设计设计、批评、过滤 11,273 个设计 → 91 个被提名从头锚定设计设计、批评、过滤 23,600 个设计 → 57 个被提名 Claude 执行、Claude 引导、人类参与 c动力学表征竞争性 BLI（干扰臂） IgE 结合 IgE + 干扰剂缓冲液 omalizumab 干扰剂 91 个 VHH 中有 20 个去除了 IgE，速度比 omalizumab 快 3 倍。标准 BLI（锚定臂）响应时间 57 个从头锚定中有 3 个结合了 IgE，KD= 3.4、32.5 和 235 nM。 d 走向治疗性双表位锚定干扰剂双表位 NHP 交叉反应性人源化可开发性细胞功能与致过敏性

· · ·

体内

图 1. 双表位破坏性 IgE 抑制剂设计战役概述。(a) 我们试图设计一种双表位 VHH，其锚定臂结合受体结合的 IgE，其干扰臂迫使 IgE 进入无 Fc εRI α 竞争能力的État（状态），从而使其离开受体而不发生交联（来自图 2d）。(b)

设计活动，简化自图 3。Claude 在人类的输入下分解了目标并运行了两个设计阶段。(c) 生物膜层干涉技术表征：用于干扰臂的竞争性 BLI 和用于锚定位的标准 BLI。(d) 通过几何分析对两个臂进行明智配对，将为进一步的体外和体内表征提供起点，以开发治疗性双表位抗体。反应性是 B 细胞持续产生过敏原特异性 IgE 驱动的完全致敏且高度反应性的肥大细胞和嗜碱性粒细胞基线状态的结果。在过去的十年中，食物过敏的管理和治疗发生了革命性的变化。最值得注意的是，2024 年人源化抗 IgE 抗体奥马珠单抗（Xolair）成为首个获得 FDA 批准的用于减少对多种食物过敏反应的药物 [5, 18, 19]。在 OUtMATCH 试验中，16 至 20 周的奥马珠单抗治疗使得 67% 的多食物过敏患者能够耐受 600 毫克的花生蛋白，而安慰剂组为 7% [18]。这是食物过敏治疗的一大进步，但距离完全攻克该疾病还有很大差距。只有 42% 的受试患者完成了两种食物全部 6,044 毫克累积激发试验，三种食物全部完成的比例为 24%。此外，在经历完整的 16 至 20 周治疗后，14% 的患者甚至无法摄入 30 毫克的花生 [18]。治疗方案也很繁琐。由于奥马珠单抗以化学计量方式中和 IgE，剂量由体重和总 IgE 决定。患者必须每两周或四周通过注射服用 75 至 600 毫克的药物，对于体重和/或总 IgE 超出获批说明书范围的患者，尚无确定的给药方案，这排除了很大一部分人群。此外，由于奥马珠单抗不会改变 IgE 的产生，停药后游离 IgE 会恢复，因此患者需要无限期接受治疗 [5, 19]。因此，奥马珠单抗可防范意外暴露，但仍建议患者继续避开过敏原并随身携带肾上腺素 [5, 19]。尽管如此，奥马珠单抗仍代表了食物过敏治疗的突破，并证实了 IgE 是该疾病的核心介质。其局限性源于 IgE:Fc εRI α 相互作用的生物学特性以及 20 世纪 90 年代为奥马珠单抗设定的开发目标（图 2）[20, 21]。2a IgE Fc 几何构型与构象异质性闭合 C ε3：游离、弯曲的 IgE-Fc

Fc εRI 位点关闭；CD23 位点开放

Cε3 Cε3Cε4 Cε4Cε2 对

(后面)

37.1 Å 开放 C ε3：受体结合

Fc εRI α D2 跨越两端的 C ε3

Cε2 对 Cε3 Cε3Cε4 Cε4D1 D2 Fc εRI α

41.6 Å b Fc εRI 结合的 IgE 与交联

肥大细胞 / 嗜碱性粒细胞

D2 D1 D2 D1 致敏、静息

IgE 在 Fc εRI 上停留数周

无过敏原，无信号

D2 D1 D2 D1 过敏原过敏原交联两个 IgE

受体簇集；细胞在几分钟内脱颗粒

激活脱颗粒 c 奥马珠单抗：抑制与破坏中和游离 IgE

C ε3 上的表位与 Fc εRI 位点重叠

奥马珠单抗

KD ≈ 0.2 nM 游离 IgE

无法加载未受损的受体结合 IgE

仅在高浓度下破坏结合的 IgE

部分破坏复合物上的奥马珠单抗衍生 scFv

Cε3Cε4 Cε4D2 D1 Fc εRI α 被杠杆撬动

D2 旋转 14°，移动 5 Å

奥马珠单抗衍生 scFv

在打开的亚位点中第二个拷贝撬动受体

d 双表位破坏机制

Cε3Cε3Cε4Cε4Cε2

Fc εRI α 离开干扰臂

打开 C ε3 亚位点

锚定位

非竞争性表位

在 IgE-Fc 上

(与 a 相同的相机) 臂位点是占位符

1 个锚定臂结合受体结合的 IgE；2 个干扰臂进入开放的 C ε3 位点；3 个 IgE 随 VHH 离开；无交联

在细胞上促进解离

Cε2Cε3 Cε3Cε4 Cε4D1 D2 Cε2Cε3 Cε3Cε4 Cε4D1 D2 Cε2Cε3 Cε3Cε4 Cε4D1 D2 scFv scFv Cε2Cε3 Cε3Cε4 Cε4D1 D2 scFv scFv 子位点 1 子位点 2

子位点 1 打开并重新闭合，抗体结合暴露的位点，受体释放

完整复合物

两个子位点均被固定

瞬时打开

位点 1 暴露，位点 2 固定

处于打开位点的抗体

第二个拷贝撬开位点 2

以非结合状态捕获的 IgE

受体游离，结合瞬时暴露界面的抗体可捕获 IgE 并阻止其重新结合

橙色：子位点 1 和 2 的受体接触残基；状态 2–4 来自捕获的中间体，状态 1 来自天然复合物

Cε2 Cε3 Cε4 IgE Fab（V 轻链，C 中）Fc εRI α 奥马珠单抗（omalizumab，V 轻链，C 深色）锚定臂干扰臂过敏原膜受体子位点 1 受体子位点 2 图 2. IgE 生物学与解离问题。(a) IgE-Fc 构象。游离的 IgE-Fc（PDB 2WQR）呈弯曲状，C ε2 对折回且 C ε3 结构域闭合；受体结合的 IgE-Fc（PDB 1F6A）具有开放的 C ε3 结构域，由 Fc εRI α D2 跨接。标出了 C ε3 中心到中心的距离；右侧的 C ε2 对来自 PDB 7SHT（虚线）。(b) 过敏原诱导的效应细胞上 IgE:Fc εRI 聚集激活细胞并释放颗粒介质。(c) 奥马珠单抗结合 C ε3 上的游离 IgE 并阻断受体加载，但其表位在受体结合的 IgE 上被覆盖。右侧为部分解离的复合物（PDB 7SHT）。(d) 双表位干扰剂概念。锚定臂（蓝色）结合非重叠表位并将分子固定在受体结合的 IgE 上；干扰臂（green）在 C ε3 位点打开时进入并将其受体撬开。臂的位置具有示意性（E3_53 和 E2_79 位点，PDB 7MXI）。图注续下一页。

3图 2. IgE 生物学与解离问题（续）。(e) IgE:Fc εRI α 的促进解离

由奥马珠单抗（omalizumab）及其衍生的抗体变体组成的复合物。IgE:Fc εRI α 复合物在其两个 C ε3 子位点之一（橙色；子位点 1 浅色，子位点 2 深色）短暂打开，而另一个子位点保持结合；结合暴露位点的抗体可阻止重新结合，第二个子位点释放，随后 IgE 在非结合状态下与抗体保持结合离开。状态 1，天然复合物（PDB 1F6A，来自 7SHT 的 C ε2 对）；状态 2–4，捕获的解离中间体（PDB 7SHT），在打开的子位点 1 中结合有源自奥马珠单抗的 scFv，第二个拷贝在子位点 2 处撬动 Fc εRI α（14 ◦，5 Å），图中未显示抗体 (2) 和受体 (4)。结构叠加在 C ε4 对上，单一视角；结构域绘制为重原子轮廓。IgE 存在两种形式：循环中的游离态以及结合在细胞表面 Fc εRI 上的结合态，在治疗相关剂量下，奥马珠单抗仅中和游离态（图 2c）。由于 IgE 通过结合态的聚集触发反应，因此中和游离态 IgE 使细胞解除武装的速度仅与其现有复合物的周转速度一样快：嗜碱性粒细胞在大约十天内失去大部分表面 IgE，而皮肤中 Fc εRI 阳性肥大细胞在一周后没有变化，仅在十天后减少 [22–24]。因此，对食物过敏患者最具潜力的药物不仅能中和游离 IgE，还能清除已结合在效应细胞上的 IgE（图 2）。

2 中和与破坏性 IgE 抑制剂

奥马珠单抗通过与 IgE C ε3 链上的两个对称相关表位结合来中和游离 IgE，这种结合方式不仅直接与 Fc εRI α 的结合构象重叠，还可防止被认为能稳定 IgE:Fc εRI α 复合物的 C ε2 结构域折回构象（图 2a）[20, 21, 25, 26]。正因如此，人们最初认为奥马珠单抗一旦与 Fc εRI α 结合，就无法再结合 IgE。然而，在奥马珠单抗最初开发后的后续研究发现，它实际上可以破坏 IgE:Fc εRI α 复合物，但仅在远高于患者体内达到的浓度时才能实现。在这种情况下，Fc εRI α 的短暂部分解离使奥马珠单抗能够占据其在 C ε3 上部分重叠的表位，从而阻止其重新结合 Fc εRI α

(图 2e) [25]。事实上，研究表明，在微摩尔浓度下，奥马珠单抗可加速预形成复合物的解离，并在数小时内从嗜碱性粒细胞上清除 IgE。对于结合态 IgE 的破坏，奥马珠单抗的 50% 破坏能力（DC50）约为 200 μM（SPR），而游离 IgE 抑制的 KD ≈ 0.2 nM [25, 27]。

正构破坏性抑制剂在这项工作的同时，人们发现了锚蛋白重复区段蛋白 (DARPin) E2_79，发现在细胞测定中它比奥马珠单抗能更有效地抑制介质释放 [28, 29]。2012 年，开发出了一种用于分类 IgE 抑制剂的荧光测定法，证明 E2_79 的破坏作用不是作为 Fc εRI α 表位经典竞争剂，而是通过仅 400 Å 3 的模拟空间重叠来实现，相比之下奥马珠单抗的模拟值则为 1200 Å 3，这表明大的重叠可能并非必需，甚至可能对促进破坏起反作用 [8, 25, 30]。随后，研究表明 E2_79 能够在几分钟内从过敏供体嗜碱性粒细胞中去除表面 IgE，消除 IgE 依赖性激活，并保护携带人 Fc εRI α 的小鼠免受被动皮肤过敏反应 [7]。然而，显示出促进解离的 DARPin 作为药物起点十分困难，受到免疫原性、可开发性以及该支架没有获批先例的限制 [27, 31]。基于这项工作，DARPin 背后的学术团队转而对奥马珠单抗本身进行工程改造以确立破坏效力 [32, 33]。他们进化出了能够在数百纳摩尔浓度下去除人源化肥大细胞表面的变体，与部分解离复合物结合的变体的冷冻电镜 (cryo-EM) 结构证实了空间促进解离是破坏机制 [32]。这些变体的肘部柔性后代在单次给药后 36 小时内就能使小鼠脱敏，而奥马珠单抗则与溶剂对照相当 [33]。Excellergy 是一家基于此项工作创立的生物技术公司，目前已进入 1 期临床试验 [33, 34]。

变构破坏性抑制剂与此同时，生物技术公司 Ablynx 通过无偏倚的美洲驼免疫计划，分离出了针对 IgE 的单域抗体 (VHH)，并发现其中有几个不仅能阻断 IgE 与 Fc εRI 的结合，还能从细胞上置换出受体结合的 IgE [35–37]。其中一种 VHH 即 sdAb026 的机制后来得到了解析：它结合 Cε3–C ε4 连接处的 CD23 (Fc εRII) 位点，并锁定无法结合 Fc εRI α 的关闭构象，代表了从远端表位而不是在界面处起作用的第二条破坏途径 [38]。在细胞测定中，sdAb026 能在几分钟内从嗜碱性粒细胞中剥离 IgE。

双位点抗原结合体（Biparatopics）正构和变构破坏性谱系共同确立了驱动 IgE:Fc εRI α 复合物解离的两种不同机制：空间位阻促进解离机制（其中结合物插入受体界面的瞬时暴露子位点并竞争性驱逐受体），以及变构构象捕集机制（其中结合物在远端 CD23 (Fc εRII) 位点将 C ε3 锁定在无法维持受体的闭合构象中）[8, 38]。在这两种情况下，无论采用何种模态，这类单价破坏剂在细胞上都需要达到数百纳摩尔至微摩尔的高浓度才能介导解离。为了显著提高解离效价，研究人员通过将单价破坏臂与针对非竞争性表位的第二个锚定臂相结合，构建了双位点构建体。将 E2_79 与识别受体结合态 IgE 远端 C ε3–C ε4 表面的 DARPin 进行融合，使效率提高了约 100 倍 [27, 39]。此外，knob-in-hole IgG1 Fc 融合体 KIH_E07_79 将全血嗜碱性粒细胞上的半数最大解离浓度降至 2 nM，约为 E2_79 效价的 500 倍，并能在体内 9 小时内清除嗜碱性粒细胞和组织肥大细胞上的 IgE [30, 39, 40]。双位点 VHH 通过将 sdAb026 与非竞争性锚定 VHH 配对，也表现出显著的解离效价提升 [41]。由于对 IgE:Fc εRI α 复合物产生亲合力效应，双位点构建体表现出更高的解离能力。非竞争性臂结合受体结合态 IgE，并将该分子锚定在 IgE:Fc εRI α 复合物上。这反过来使破坏臂保持在较高的局部浓度，随时准备在复合物瞬时呈现其识别状态时发挥作用（图 2d）[27, 39, 41]。

类过敏原性（Anaphylactogenicity）同一种双位点机制通过对单个受体结合态 IgE 分子的亲合力顺式结合（cis-binding）实现了显著更高的解离效价，但也带来了反式结合（trans-binding）两个不同受体结合态 IgE 分子的风险，这可能导致交联并引发过敏反应 [32]。作为一个临床实例，omalizumab 变体 HAE1 的开发在 2 期试验中因两名参与者出现不良事件而中止。尽管事件的性质尚未公开，但细胞实验表明 HAE1 可以激活来自某些供体的嗜碱性粒细胞，这归因于其在完整的 IgE:Fc εRI α 复合物上停留时间更长 [32]。双位点药物灵活的结合几何构型要求必须给予高度重视，以避免这种不希望发生的交联。虽然初步的实验验证表明，例如 KIH_E07_79 是无激活性的 [39, 40]，但随后在其他供体中进行的测试报告了高度的类过敏原活性 [41]。进一步的表征强有力地表明，臂的几何构型、连接部分（串联或 Fc 融合）与取向，以及各臂本身的亲和力和动力学特性，是决定类过敏原性的关键因素 [39-41]。因此，真正同类最佳（best-in-class）的破坏性 IgE 抑制剂必须以足够强的强度结合受体结合态 IgE 以破坏该复合物，同时在物理上又不能桥连同一细胞上的两个 IgE 分子。

3 旨在开发同类最佳 IgE 抑制剂的计算设计工作

有效且安全地阻断结合态 IgE 需要对分子几何构型进行精确控制。传统的通过展示或免疫发现抗体的方法对结合表位或位姿的控制力很小，使得像 IgE 阻断这样复杂的治疗机制很难作为靶点。另一方面，生物分子生成模型可以在制造任何分子之前指定或预测几何构型，从而实现合理的优化或选择。我们假设，一项利用现代生物分子生成模型、由 Claude [42] 指导、并参考先前文献以及合理的表位和位姿选择的计算设计活动，可以加速通往同类最佳 IgE 抑制剂的道路：该抑制剂能够以高亲和力抑制游离 IgE，以治疗相关剂量的效力阻断过敏效应细胞上的 IgE:Fc εRI，并且在这样做时不会交联和激活效应细胞。我们着手设计一种具有两个功能性 VHH 臂的双paratopic（双表位特异性）抗体，其中一个臂可阻断 IgE:Fc εRI α

相互作用，另一个臂以几何相容的表位将分子锚定在受体结合的 IgE 上。在 Claude 协助审查现有文献后，我们选择进行 (1) sdAb026 的局部重新设计，5文献回顾与综合

Decompose biparatopic objective Characterize disruptive mechanisms orthosteric disruptors allosteric disruptors Select computationally amenable scope ( )Select modeling artifacts 36 40 44 48 4GT7 5NQW 1F6A 2Y7Q Critic A Cε3 centroid distance (Å) 36 40 44 48 4GT7 5NQW 1F6A 2Y7Q Critic B Cε3 centroid distance (Å) 0.6 0.7 0.8 0.9 wild type D33A R44A E108A D110A Y112A Critic A Confidence score 0.80 0.84 0.88 0.92 wild type D33A R44A E108A D110A Y112A Critic C Confidence score

Objective decomposition and strategy

To disruptor redesign To anchor design Characterization of interface geometry Clarify optimization scopes Cε3 interface, C ε4 interface, and CDR2 scopes Methodological extensions Redesign arms Tempered redesign (TR) Granular interface geometry loss design step temperature

S I DTS TNIVLDENTSASVDT

TR 7→ Cε3interface TR 7→ Cε4interface IF 7→ Cε3interface IF 7→ Cε4interface IF 7→ CDR2 Redesign 7→CDR2 IF = Inverse Fold Enumeration of design scopes Cε2Cε3–C ε4 Cε2 design Cε3–C ε4 design Scope-specific design ablations Ablation-refined Cε3–C ε4 design Ablation-refined Cε2 design Cε3–C ε4hit expansion Score and nominate for competitive BLI n10 0.886 n16 0.905 n47 0.908 Score and nominate for BLI n81 0.893 n83 0.896 n64 0.912

Disruptor redesign Anchor design

Phase execution Objective decomposition & strategy Near-autonomous Claude execution Claude-driven with human interaction

图 3. Claude 引导的设计任务分解。目标分解与策略、干扰体重新设计和从头锚定设计作为三个阶段列；蓝色方框表示近乎自主的 Claude 阶段执行，绿色边框表示有人机交互的 Claude 驱动的目标分解。一个具有观察到的干扰活性和已知 IgE 结合结构的现有 VHH [38, 43]；以获得我们的干扰臂；以及 (2) 从头设计一个锚定臂，以多样化和控制锚定臂的几何形状。我们最终鉴定了三种干扰体变体，其活性与 sdAb026 相当，且在治疗相关浓度下比奥马珠单抗更有效. 我们还设计并表征了对 IgE 具有个位数纳摩尔亲和力的从头锚定体。这些结果代表了开发双特异性干扰性 IgE 抑制剂等复杂抗体疗法的众多关键步骤中的第一步，并为持续的治疗优化提供了关键的模块化子组件。

3.1 Claude 引导的设计任务分解与编排

设计过程在很大程度上由与 Claude 的交互推动。首先，在我们的研究团队配合下，Claude 统筹了高级设计规划，并将既定目标分解为不同的工作阶段。其次，在我们的研究团队的全程支持下，它对这些工作阶段进行了近乎自主的执行，正如 Claude 最近针对针对不同靶点的 de novo 结合体设计所做的那样 [42]。该设计活动总共包含 22 个不同的自主执行阶段，其中 15 个主要用于计算设计。我们在图 3 中提供了工作流的高级摘要。为了启动设计活动，Claude 对学术界和临床前研究中具有代表性的破坏性方法进行了广泛的审查，并将其归纳为两个显著的机制类别：首先，正构破坏剂（orthosteric disruptors），包括延伸 DARPin E2_79 的单价和双表位谱系 [8, 27]、经过亲和力进化和破坏选择的奥马珠单抗变体 [32, 33, 39]，甚至在足够高浓度下的奥马珠单抗本身 [25, 26]，它们的作用原理是利用与 Fc εRI α 之间微小的、短暂释放的空间重叠来占据结合位点并随后驱动解离。其次，变构破坏剂（allosteric disruptors）充当 CD23 功能同源物 [44]，对 6IgE Fc C ε3–C ε4 的非受体结合态闭合构象表现出选择性亲和力，其中的例子包括 sdAb026 [38]。随后，Claude 分析了每一类方法进行计算重新设计以及最终纳入双表位构建体的可行性，确定了对促进破坏至关重要的每一类的关键功能和结构考量，而所采用的计算模型必须准确重现这些考量。为此，Claude 识别出实验衍生且公开可用的数据，作为直接且定量评估特定功能建模性能的手段。例如，Claude 对多个与设计相关的已解析结构进行了结构预测，包括由 7 号克隆奥马珠单抗变体诱导的正构破坏中间状态（PDB 7SHT）[32, 45]，以及 sdAb026 的破坏性结合模式（PDB 5NQW）[38, 43]，观察到在识别 C ε3–C ε4 的开放/闭合状态方面两类结构均表现出高度一致的准确性，但对变构破坏剂表现出更高的置信度。这些观察结果、VHH sdAb026 中经验证的破坏剂的存在，以及用于指导重新设计的已解析结构的存在，最终促成了模仿变构破坏机制而非正构破坏机制的决策。 1

最后，为了更好地为下游优化中协同折叠模型（co-folding models）的选择提供参考，Claude 识别出了经实验验证的 sdAb026 功能丧失突变体 [38]，并以此为基础建立了一个校准机制，借此挑选出两个不同的协同折叠模型，其置信度最准确地还原了野生型相对于公开突变体的破坏性排序。Claude 进一步明确了所需双表位的功能规范，将其细分为两个臂各自的设计以及必需的连接部分。结合关键的结构或功能见解，它将设计活动分解为后续阶段。这一设计规划阶段的整体流程总结在图 3 左侧。

3.2 破坏剂重新设计

为了启动干扰素重新设计，Claude 首先分析了 PDB 5NQW [38, 43] 中解析出的 sdAb026 与 IgE C ε3–C ε4 结合模式的界面几何结构，识别出了三个可用于开展靶向、结构局部化重新设计的范围。这些范围包括：由 CDR3 环的一部分和一组骨架区域介导的 C ε4、受体远端界面（Cε4 界面范围）；由 CDR1 和 CDR3 介导的 C ε3、受体近端界面（Cε3 界面范围）；以及由远离界面的 CDR2 环组成的第三个范围（CDR2 范围）。第四个范围由 CDR1 和 CDR3 的残基组成，它们不作为接触参与界面，但可以推断其起到支持、稳定界面的作用，该范围最初由 Claude 提出，并贯穿了最初的反向折叠和退火重新设计轮次；然而，为了将资源从结合界面分散到远端的 CDR2 范围，我们选择在后续的攻关阶段中放弃它。关于范围选择的更多细节可以在附录 A.1 中找到，图 3 中包含了一个简要的视觉概览。随后，Claude 选择将随后的干扰素重新设计工作拆分为六个不同的子阶段，每个子阶段都包含重新设计范围和重新设计范式的选择。对于 C ε3 界面和 C ε4 界面范围，Claude 选择进行靶向诱变，以获得具有更优干扰效果的双突变体、三突变体和四突变体。为此，采用了两种主要方法：随机采样子范围的反向折叠 [46]，以及一种无需事先指定突变范围的、定制的基于梯度的诱变方法——退火重新设计。Claude 实现了这一关键方法学扩展的实时概念化、实现、后续验证和部署，并在更广泛的范围内针对攻关项目的特定上下文精心定制了设计工作流（包括开放的和内部的）。最后，针对 CDR2 范围同时应用了反向折叠和基于梯度的重新设计。在该轮结束时，Claude 使用 3.1 节中指定的评分模型重新折叠了所有设计，按 10-seed 平均结合对 ipTM 进行排序，并从四个 C ε3 界面和 C ε4 界面设计范围中各提名了 13 种不同的设计，连同来自 CDR2 范围的另外 40 种设计一起，用于后续的实验验证。有关轮次划分的更详细概览可在图 3（中间列）中找到，结果摘要报告在图 4 中。有关上述各个阶段的更详细论述，请读者参阅 3.2 节；关于退火重新设计的更多细节，请参阅附录 C。

1在 7SHT 和 5NQW 上的相对模型性能可能部分是其分别于 2021 年 10 月和 2018 年 1 月存放（与模型训练截止时间相关）的人为产物。

70.85 0.86 0.87 0.88 0.89 0.90

Mean critic score (10 seeds) Cε3 × CDR1,3 inverse folding Cε3 × CDR1,3 tempered redesign Cε4 × CDR3 inverse folding Cε4 × CDR3 tempered redesign CDR2 inverse folding CDR2 redesign sdAb026 Disruptor optimization 2345678 Substitutions vs sdAb026 0.80 0.82 0.84 0.86 0.88 0.90 0.92 Mean critic score (10 seeds) Cε2 · B1 consensus epitope Cε2 · epitope scan Cε2 · unconditioned Cε3–Cε4 · epitope scan Cε3–Cε4 · unconditioned Anchor design Cε3 × CDR1,3 · inverse folding mean critic score 0.876 n10 Binders in green Curated subset of nominated anchors anchor n83 with disruptor n10 mean critic scores 0.880 / 0.876 n83 + n10 (G 4S) 1KIH linker proxy anchor n83 in tandem with disruptor n10 mean critic scores 0.880 / 0.876 n83 + n10 (G 4S) 3tandem linker proxy abcdefg

Figure 4. Design campaign results. (a) disruptor redesign (a-c): every scored design per arm (gray) and the nominated designs (colored) by (a) mean critic score over ten seeds and by (b) substitution count against reference sdAb026 (b); dashed line, sdAb026; (c) right, nominee n10 on the C ε3–C ε4 dimer. An-chor design (d-g): (d) the same strip per design round; (e) a curated subset of nominated anchor variants on both C ε2 and C ε3–C ε4; (f) a (G4S)2 KIH linker proxy applied to disruptor variant n10 and anchor n83, and (g) a (G3S)3 tandem linker proxy applied to the same combination.

3.3 从头锚定设计

锚定臂是针对在受体结合的 IgE 上保持暴露的两个表面进行从头设计的（图 2d）。第一个是 C ε2 结构域对，这是在双特异性格式中改善 sdAb026 的 VHH 锚定物的表位类别 [41]。第二个是 C ε3–C ε4 表面，该表面被 E3_53 识别，E3_53 是第一个双特异性 DARPin 的锚定物 [27, 39]。Claude 首先使用所选的共折叠评论器对已发表的 C ε2 结合 VHH [41] 进行了重新折叠，以检查模型是否将它们放置在一致的表位上，并使用共识足迹作为第一个目标。锚定物是通过结构预测器从随机起点进行反向传播，跨所有三个 CDR，在十二个临床 VHH 框架上设计的（表 2）。在 C ε2 上进行了三个阶段的设计：一个针对共识表位，一个针对整个结构域随机抽取的表位补丁，以及一个无表位条件限制的设计（附录 B.2）。在 C ε3–C ε4 上，进行了两个阶段：第一阶段使用与 C ε2 类似随机抽取的表位补丁，并带有惩罚与 Fc εRI α 接触的额外规避损失

以及 sdAb026 足迹，随后进行了一次完全没有表位条件限制的后续运行。两轮运行之后，都跟进了基于反向折叠的迭代求精方案 [46–49]。总共在 1,310 个 H100 小时内对外评分了 23,600 个独特的设计，其中 57 个被提名，42 个针对 C ε2，15 个针对 C ε3–C ε4（图 4，表 3）。请读者参考附录 B 以了解更多技术讨论，参考图 3（右列）以全面了解该轮设计，并参考图 4 了解关于后续实验验证的平板提名的设计结果。83.4 双特异性组装

在设计活动的最后阶段，我们让 Claude 利用第 3.2 节和第 3.3 节中表征设计方案的破坏效力和结合亲和力的计算结果，结合对姿态和结合几何构型的模型预测，来组装出有前景的双表位抗体。正如第 2 节所讨论的，连接部分的选型在先例破坏剂中各不相同，并且在一个更大的设计空间内，该空间既包括单链串联连接子（例如双表位 DARPin bi53_79 [27] 中发现的 (G 4S) 2 连接子），也包括更庞大的“杵臼结构”（KIH）、IgG 融合构建体 [39, 41]。为了进一步缩小设计范围，我们要求 Claude 将其搜索范围限制为仅包含串联的 (G 4S) k 连接子，其中

k ∈ { 3, 4, 5, 6, 7, 8}。尽管搜索空间有限，但 IgE Fc 的二聚体对称性，加上 C ε2 的不稳定性质，使得对假定的双表位抗体可能的顺式结合模式进行几何表征仍然是一项艰巨的任务。长而柔性的连接子对于靶向 IgE 的双表位抗体来说是一个严重的隐患。锚定臂结合可提高局部有效破坏臂浓度，从而增强分子的破坏效力，这种预期的双表位效应依赖于较短而非较长的连接子。连接子长度的作用，以及更广泛的连接子和手臂几何构型的作用，已被广泛认为在特定构建体的过敏反应隐患中起着主导作用，其重要性超越了亲和力和动力学 [39, 41]。事实上，更长的连接子和更具柔性的几何构型能够适应不同受体结合分子的反式结合，进而导致稳定的桥联、激活以及最终的过敏反应。在此背景下，Claude 着手将任务分为两部分：首先，Claude 通过识别 PDB 中存在这些链的所有已存结构（2Y7Q、7SHT、8YWA），对结合了 Fc εRI α 的潜在 IgE Fc C ε2–C ε4 构象进行了系统枚举。对于每个假定的锚定臂和破坏臂对，枚举了所有结构以及锚定臂和破坏臂的对称相关结合位点所有组合，从而得出总共 12 种粗略的候选顺式结合几何构型。其次，Claude 推导出一个算法，对于每个固定的几何构型和 k ∈ { 3, 4, 5, 6, 7, 8} 的选择，该算法尝试构建 (G 4S) k 串联连接子残基的阿尔法碳排布，以便在成对的互补 VHH 末端之间产生畅通的路径。对于后者，我们需要进行少量的干预，以确保排除病理性连接子几何构型（即穿过 IgE C ε3–C ε4 开放构象的“甜甜圈孔”），并妥善考虑某些空间位阻冲突。然后，组合程序会提名能够构建无阻碍连接子的最短 k，以及相应的构建方案。图 4（右下角）给出了其中一种连接子的构建方案。

4 结果

We nominated designs from both the disruptor and anchor design campaigns for experimental validation and assessed their functional activity via bio-layer interferometry (BLI). For the disruptor arm, we per-formed a multi-phase, competitive BLI experiment to measure the extent to which the designed antibody in solution can displace IgE bound to an immobilized soluble Fc εRI α ectodomain, mimicking the physio-logical context of Fc εRI on mast cells and basophils. For the anchor arm, we performed standard kinetic characterization of free IgE binding to immobilized VHH.

4.1 Designed disruptor variants displace IgE from bound Fc εRI α

We first nominated 91 disruptor designs for experimental screening, alongside the parent sdAb026 and single alanine mutants of its C ε4-facing CDR3 previously reported to remove functional activity [38], at a single concentration. Twenty designs removed IgE faster than every negative control. We classified these 20 sequences as positive hits for disruptive activity. Six hits appeared to improve upon the activity of sdAb026. Hits originated from every design scope and method, albeit at different rates (Figure 5D). Redesign of CDR2 was by far the most productive route, accounting for most of the hits and most of the designs that appeared to exceed the disruptive activity of the parent. However, activity did not depend on staying close to the parent sequence, and functional activity was retained for sequences differing by as many as eight substitutions (Figure 5B), demonstrating the ability of biomolecular models to navigate 9Fc εRI loads onto the probe

1 IgE binds Fc εRI

2 VHH binds IgE

3 IgE is displaced

Time BLI response

IgE loaded IgE displaced 1234

0 1 2 3 4 5 6 7 8

Mutations vs sdAb026

55 60 65 70 75 80 85 90 95

IgE remaining after 120 s (%)

V8 V4 V1 V2 V7 V3 V6 V5

sdAb026

E108A D110A Y112A

active (n = 20) inactive (n = 71)

1 2 3

kon (10 5 M−1 s−1)

510 15 20 25 30 35 40 45

IgE lost after 120 s (%)

V1 V2 V3 V4 V5 V6 V7 V8

sdAb026 Pearson r = -0.15 (p = 0.16, n = 91) 2 3 4 5 6 7 8

Substitutions vs sdAb026

Cε3 × CDR1,3 inverse folding

Cε3 × CDR1,3 tempered redesign

Cε4 × CDR3 inverse folding

Cε4 × CDR3 tempered redesign

CDR2 inverse folding

CDR2 redesign

Not disruptive (71) Disruptive, below sdAb026 (14) Exceeds sdAb026 (6)

020 40

Designs

best worst

27 28 30 31 33 51 52 53 54 55 56 57 58 99 100 101 102 103 104 105 107 108 109 110 111

20 40

IgE lost (%)

sdAb026 E108A CDR1 CDR2 CDR3

sdAb026 Omalizumab V1 V2 V3 V4 V5 V6 V7 V8 400 nM 76.6 81.1 77.0 81.8 83.0 75.9 76.4 73.6 80.0 74.4 200 nM 72.2 82.8 74.5 82.1 82.3 78.3 80.9 78.9 86.5 78.4 100 nM 70.7 79.3 71.2 80.7 80.5 79.0 80.9 81.6 87.2 75.6 50 nM 71.5 80.1 73.6 79.8 80.0 78.6 82.5 83.2 88.3 74.6 25 nM 69.6 76.6 70.1 78.7 76.7 75.2 79.8 81.1 85.5 74.3

0 60 180 360 600 0.5 0.6 0.7 0.8 0.9 1.0

Fraction of IgE bound

400 nM

sdAb026 Omalizumab Designed variants (8) Y112A

0 60 180 360 600 0.5 0.6 0.7 0.8 0.9 1.0

Fraction of IgE bound

200 nM

0 60 180 360 600 0.5 0.6 0.7 0.8 0.9 1.0

100 nM

0 60 180 360 600 0.5 0.6 0.7 0.8 0.9 1.0

50 nM

0 60 180 360 600 0.5 0.6 0.7 0.8 0.9 1.0

25 nM

A BC D E

IgE bound at 600 s (%), puri fi ed VHH

F GH I J K

加入干扰剂后的时间 (s) 图 5. 干扰臂设计的实验表征。说明见下一页。 10 图 5. 干扰臂设计的实验表征。(A) 竞争性 BLI 测定的原理图和理想化传感器图。IgE 的置换量通过峰值 BLI 响应相对于总 IgE 负载的下降程度来量化。(B–E) 通过 120 秒解离阶段对 91 种设计进行的初步筛选。(B) 解离阶段结束时剩余的 IgE 与来自 sdAb026 的取代数量的关系。20 种变体（蓝色）的表现出的表观 IgE 解离速率超过了所有阴性对照；推进至纯化阶段的八种变体标记为 V1–V8。(C) 干扰效应与标准结合动力学测定中的结合速率无关。(D) 按重新设计范围与方法以及取代数量分类的设计。(E) 在每个 CDR 位置发生突变的设计数量（顶部）以及其中的最佳和最差 IgE 置换量（底部）。(F–K) 纯化变体在 600 秒解离以及五个临床相关浓度下与奥马珠单抗（omalizumab）一同进行了测定。(F) 在 400 至 25 nM 下结合的 IgE 百分比；绿色单元格比奥马珠单抗移除了更多的 IgE。(G–K) 400、200、100、50 和 25 nM 下的时间进程；箭头标出了干扰剂的加入。序列功能图景零样本。在比亲本更具活性的设计中，整个 CDR2 以及 CDR1 中的两个位置出现了取代（图 5E），而面向 C ε4 的 CDR3 中的大部分位置在活性设计中从未出现过，这与该环丙氨酸突变体中活性的丧失一致。随后，我们纯化了八种活性设计，连同 sdAb026 和无活性的丙氨酸突变体 Y112A，在一系列临床相关浓度下将它们彼此进行比较，并与奥马珠单抗进行比较（图 5F–K）。为此，我们执行了延长的干扰性 BLI 测定，将干扰阶段延长至 10 分钟。所有变体在所有浓度下均保持了大于 Y112A 对照的干扰活性。令人鼓舞的是，在测试的每个浓度下，三种变体 V1、V4 和 V8 比奥马珠单抗移除了更多的 IgE。然而，这些变体似乎并未能在亲本 sdAb026 的基础上有所改进，其中最佳变体 V1 在大多数浓度下提供相似的响应。为了确认干扰活性不仅仅是结合的函数，我们对游离 IgE 与固定化 VHH 的结合进行了单点 BLI 测定。几乎所有设计都保留了结合活性，许多设计在测量极限处具有非常慢的解离常数（附录 E）。表观结合速率 kon 并不能预测干扰活性（图 5C）。此外，丙氨酸突变体阴性对照与 sdAb026 结合 IgE 的紧密程度几乎相同，但在干扰性 BLI 筛选中移除的 IgE 要少得多（图 5B）。因此，C ε4 界面处的单个取代（即阴性对照）可以在几乎不改变亲和力的同时消除干扰作用。

4.2 表位多样化的锚臂命中化合物

我们随后通过 BLI 提名并实验测试了 57 种锚定臂设计，其中 VHH 固定在探针上。3 种变体表现出清晰解析的动力学拟合，命中率约为 5%（图 6A）。这些命中结果来自三个表位特异性设计模块中的两个：C ε2 侧面，位于变体 B1、E1 和 F4 共识结合位点的远端 [41]；以及 C ε3–C ε4 表面。有点令人惊讶的是，没有一种针对变体 B1、E1 和 F4 共识表位的设计表现出结合。印证了近期的文献 [42]，评论者评分并不能完全预测哪些设计会结合；锚定设计 1 是 19 个针对远端 C ε2 表位被提名者中排名最低的，而锚定设计 2 和 3 分别是 15 个针对 C ε3–C ε4 表位被提名者中第二好的。将结合速率对解离速率作图（图 6C）显示，这三个命中产物以相似的速率结合，因此它们从 3.4 到 235 nM 的表观亲和力范围几乎完全由其解离速率产生。然而，我们提醒说，这些亲和力测量可能受到亲合力的混淆；由于 IgE 呈现每个 Fc 结构域两次，与相邻固定 VHH 的结合事件可能不是独立的。不过，奇怪的是，锚定设计 1 和 3 似乎占据了两个 IgE 单体交汇处一个近乎对称的单一结合位点（图 6A）。因此，亲合力的实验表征将提供一种不同寻常的机制，用于直接探测预测结合姿态的准确性。利用预测的结合姿态，当后者结合到相同的 IgE 二聚体时，我们探讨了将锚与原始干扰剂 sdAb026 连接所需的连接体长度（图 6D）。因为在此 11 anchor Cε2Cε3Cε4

090 180 270 360 时间 (s) 0246 响应 (nm) KD3.4 nM 传感器图 1:1 拟合 090 180 270 360 时间 (s) 0246KD32.5 nM 090 180 270 360 时间 (s) 0246KD235 nM Cε2近端 0/23 Cε2远端 2/19 Cε3–Cε41/15 0123456 峰值响应 (nm) 31 2 10 410 510 6 表观 kon (M −1s−1) 10 −5 10 −4 10 −3 10 −2 10 −1 koff (s −1) 低于测定的解离速率下限 1 nM 10 nM 100 nM 1 μM 123 50 100 末端间距 (Å) anchor –sdAb026 sdAb026 –anchor anchor –sdAb026 sdAb026 –anchor anchor –sdAb026 sdAb026 –anchor 123 345678连接体 (G 4S) k,k=

A 锚定设计 1

Cε2 表位（预测）

锚定设计 2

Cε3–Cε4 表位（预测）

锚定设计 3

Cε2 表位（预测）

B C D

sdAb026 90° 较近副本较远副本

图 6. 从头设计的锚定臂的实验表征。(A) 三种锚定设计结合了 IgE，显示了在 IgE-Fc 上的预测姿态以及带有单指数拟合（虚线）的 BLI 传感器图（黑色）。(B) 所有 57 种从头锚定设计按设计模块的峰值结合响应；低于虚线（0.2 nm）则未解析出结合。(C) 三个命中产物的表观结合和解离速率，带有相等 KD 的线。(D) 锚的预测姿态表明了 (G 4S) k 的可能长度

可以将 sdAb026 组合成串联 VHH 双表位构建体的连接体，具有不同的方向。距离范围源自两种 C ε2 构象。由于存在两个相同的 sdAb026 结合位点和两种连接顺序，这两个臂有四种可能的方向。对于所有三个锚定位点，至少有一种方向允许连接体长度在 3–5 个 (G 4S) 重复序列的分布范围内，这是治疗性抗体的典型范围。锚定位点 2 和 3 距离 sdAb026 结合位点更近，因此似乎比另一种情况更兼容某种串联 VHH 方向。由于这些距离具有一定的一致性（取决于 C ε2 的构象），我们没有限定单一的连接体，而是构建了一个文库（使用我们的干扰变体代替 sdAb026），该文库涵盖了每种锚定位点的两个方向以及一系列连接体长度（第 3.4 节），以便进行进一步表征。12 5 讨论与未来工作

在本报告中，我们证明了计算设计能够生成双表位抗体的两个功能臂，这是迈向同类最佳 IgE 破坏性抑制剂的关键第一步。我们的并行设计活动产生了：(a) 一个能够将 IgE 从 Fc εRI 上移除的破坏臂，以及 (b) 一个在非竞争性表位结合 IgE 的锚定臂。通过 Claude 代理式编排生物分子生成模型 [42]，两者都在几周内获得，而传统筛选工作流程则需要数月时间。从此前报道具有破坏活性的单结构域抗体 sdAb026 [38] 开始，我们获得了 20 个具有可测量破坏活性的变体。这些变体与亲本相比有多达 8 个氨基酸取代，并且代表了一个功能多样化的家族，彼此之间的残差多达 10 个（在 31 个 CDR 残基中）。这些发现与使用生物分子基础模型设计远离原始起点功能序列的类似结果相呼应 [50–52]。我们将继续探索代理编排的计算设计活动 [42]，以便进一步探索已知的序列空间，在保持功能的同时，引导实现其他感兴趣的属性。虽然我们在受控剂量下没有测出任何功能活性优于 sdAb026 的变体，但现有功能序列的多样化为独立优化功能效力之外的属性（如可开发性、双表位兼容性和交叉反应性）提供了多个起点。与过去进行此类多样化所需的定制文库规模筛选方案相比，这显著降低了时间和成本 [33, 33]。结合我们的锚定设计，我们期望双表位构建（作为串联 VHH 或 VHH-Fc）能够显著提高效力和破坏功效。事实上，据报道，亲本破坏剂 sdAb026 只有在与锚定 VHH 结合时，才会在治疗相关的浓度下表现出破坏作用 [35, 37, 41]。同样，正构破坏剂 DARPin E2_79 在转换为双表位格式后，效力也得到了极大改善 [27, 39]。有鉴于此，我们很受鼓舞地看到，即使我们的设计以单价构建体表达，其活性也优于奥马珠单抗。这里追求的破坏剂设计方法的一个关键局限性是依赖于具有所需功能活性的分子物种的现有序列（以及事实上对接的结构）。未来的工作将探讨 Claude 是否可以在没有参考序列的情况下运用推理策略来设计功能所需序列；或者更雄心勃勃的是，当所需功能的结构规范（在本例中是通过使 IgE Fc 倾向于闭合构象来诱导破坏）完全不可用时。由 Claude 在此生成的子组件将为开发候选双表位奠定良好基础，以便进一步表征，从而打造同类最佳的 IgE 破坏性抑制剂。在对游离和结合 IgE 抑制进行生物物理验证后，候选药物将在嗜碱性粒细胞和肥大细胞测定中进行测试，以评估其在不同供体间的破坏、激活和防过敏原激发保护作用，随后在动物体内进行进一步的先导优化以及功能、药代动力学和药代动力学研究。一个关键风险是避免通过受体交联产生过敏原性；我们预计，利用 Claude 进行双表位格式的几何推理和分析将在下一阶段的发现中非常有用。放眼 IgE 之外，这项工作中开发的战略和方法为进一步开展合理设计的抗体运动奠定了强大基础，这些抗体能够实现更复杂的生物学功能，为人类疾病的同类最佳药物打开了大门。

致谢

我们感谢 Adu Subramanian、Nathan Mallipeddi 和 Dan Levin 对我们论文早期阶段的反馈；感谢 Nathan Robertson、Pascaline Mary、David Holmes、Assel Ismoldayeva、Sam Sledzieski、Jack Kenney 以及整个 Divergence Labs 充满决心的技术爱好者团队在实验指导和项目支持方面给予的帮助；特别感谢 Shahram Salek、Leonard Presta 和 Joao Vieira 为该项目提供的宝贵专业知识与战略指导，同时也感谢整个 Divergence Labs 研发委员会和董事会给予的支持。 13 参考文献

[1] Scott H. Sicherer and Hugh A. Sampson. Food allergy: A review and update on epidemiology, patho-genesis, diagnosis, prevention, and management. Journal of Allergy and Clinical Immunology , 141 (1):41–58, 2018. doi: 10.1016/j.jaci.2017.11.003. [2] Ruchi S. Gupta, Elizabeth E. Springston, Manoj R. Warrier, et al. The prevalence, severity, and distribution of childhood food allergy in the United States. Pediatrics , 128(1):e9–e17, 2011. doi: 10.1542/peds.2011-0204. [3] Ruchi S. Gupta, Christopher M. Warren, Bridget M. Smith, et al. Prevalence and severity of food aller-gies among US adults. JAMA Network Open , 2(1):e185630, 2019. doi: 10.1001/jamanetworkopen. 2018.5630. [4] Christopher M. Warren, Ozge Nur Aktas, Lorenzo J. Manalo, Tami R. Bartell, and Ruchi S. Gupta. The epidemiology of multifood allergy in the United States: A population-based study. Annals of Allergy, Asthma & Immunology , 130(5):637–648.e5, 2023. doi: 10.1016/j.anai.2022.12.031. [5] U.S. Food and Drug Administration, Center for Drug Evaluation and Research. NDA/BLA multi-disciplinary review and evaluation: sBLA 103976/S-5245, Xolair (omalizumab). FDA review, BLA 103976, supplement 5245, 2024. [6] Scott C. Garman, Beth A. Wurzburg, Svetlana S. Tarchevskaya, Jean-Pierre Kinet, and Theodore S. Jardetzky. Structure of the Fc fragment of human IgE bound to its high-affinity receptor Fc εRI α.

Nature , 406(6793):259–266, 2000. doi: 10.1038/35018500. [7] Mary D. Holdom, Anna M. Davies, Joanne E. Nettleship, et al. Conformational changes in IgE con-tribute to its uniquely slow dissociation rate from receptor Fc εRI. Nature Structural & Molecular Biology , 18(5):571–576, 2011. doi: 10.1038/nsmb.2044. [8] Beomkyu Kim, Alexander Eggel, Svetlana S. Tarchevskaya, Monique Vogel, Heino Prinz, and Theodore S. Jardetzky. Accelerated disassembly of IgE-receptor complexes by a disruptive macro-molecular inhibitor. Nature , 491(7425):613–617, 2012. doi: 10.1038/nature11546. [9] Rodrigo Jiménez-Saiz, Derek K. Chu, Talveer S. Mandur, et al. Lifelong memory responses perpetuate humoral TH2 immunity and anaphylaxis in food allergy. Journal of Allergy and Clinical Immunology ,140(6):1604–1615.e5, 2017. doi: 10.1016/j.jaci.2017.01.018. [10] Miyo Ota, Kenneth B. Hoehn, Weslley Fernandes-Braga, et al. CD23 +IgG1 + memory B cells are poised to switch to pathogenic IgE production in food allergy. Science Translational Medicine , 16 (733):eadi0673, 2024. doi: 10.1126/scitranslmed.adi0673. [11] Joshua F. E. Koenig, Niels Peter H. Knudsen, Allyssa Phelps, et al. Type 2–polarized memory B cells hold allergen-specific IgE memory. Science Translational Medicine , 16(733):eadi0944, 2024. doi: 10.1126/scitranslmed.adi0944. [12] Monica G. Lawrence, Judith A. Woodfolk, Alexander J. Schuyler, Leland C. Stillman, Martin D. Chap-man, and Thomas A.E. Platts-Mills. Half-life of IgE in serum and skin: Consequences for anti-IgE therapy in patients with allergic disease. Journal of Allergy and Clinical Immunology , 139(2):422– 428.e4, 2017. doi: 10.1016/j.jaci.2016.04.056. [13] Roger M. Cass and Burton R. Andersen. The disappearance rate of skin-sensitizing antibody ac-tivity after intradermal administration. Journal of Allergy , 42(1):29–35, 1968. doi: 10.1016/ 0021-8707(68)90129-9. [14] Shuichi Kubo, Toshinori Nakayama, Kunie Matsuoka, Hiromichi Yonekawa, and Hajime Karasuyama. Long term maintenance of IgE-mediated memory in mast cells in the absence of detectable serum IgE. The Journal of Immunology , 170(2):775–780, 2003. doi: 10.4049/jimmunol.170.2.775. [15] Donald W. MacGlashan. Releasability of human basophils: Cellular sensitivity and maximal histamine release are independent variables. Journal of Allergy and Clinical Immunology , 91(2):605–615, 1993. doi: 10.1016/0091-6749(93)90266-i. 14 [16] Donald W. MacGlashan. Basophil activation testing. Journal of Allergy and Clinical Immunology , 132 (4):777–787, 2013. doi: 10.1016/j.jaci.2013.06.038. [17] Donald W. MacGlashan and Sarbjit S. Saini. Omalizumab increases the intrinsic sensitivity of hu-man basophils to IgE-mediated stimulation. Journal of Allergy and Clinical Immunology , 132(4): 906–911.e4, 2013. doi: 10.1016/j.jaci.2013.04.056. [18] Robert A. Wood, Alkis Togias, Scott H. Sicherer, et al. Omalizumab for the treatment of multiple food allergies. New England Journal of Medicine , 390(10):889–899, 2024. doi: 10.1056/NEJMoa2312382. [19] Genentech, Inc. XOLAIR (omalizumab) injection, for subcutaneous use; XOLAIR (omalizumab) for injection, for subcutaneous use [prescribing information]. U.S. prescribing information, BLA 103976, 2026. [20] Tse Wen Chang, Frances M. Davis, Nai-Chau Sun, Cecily R. Y. Sun, Donald W. MacGlashan, and Robert G. Hamilton. Monoclonal antibodies specific for human IgE-producing B cells: A potential therapeutic for IgE-mediated allergic diseases. Nature Biotechnology , 8(2):122–126, 1990. doi: 10. 1038/nbt0290-122. [21] L. G. Presta, S. J. Lahr, R. L. Shields, et al. Humanization of an antibody directed against IgE. The Journal of Immunology , 151(5):2623–2632, 1993. doi: 10.4049/jimmunol.151.5.2623. [22] D. W. MacGlashan, B. S. Bochner, D. C. Adelman, et al. Down-regulation of Fc εRI expression on human basophils during in vivo treatment of atopic patients with anti-IgE antibody. The Journal of Immunology , 158(3):1438–1445, 1997. doi: 10.4049/jimmunol.158.3.1438. [23] John A. Eckman, Patricia M. Sterba, Denise Kelly, et al. Effects of omalizumab on basophil and mast cell responses using an intranasal cat allergen challenge. Journal of Allergy and Clinical Immunology ,125(4):889–895.e7, 2010. doi: 10.1016/j.jaci.2009.09.012. [24] Lisa A. Beck, Gregory V. Marcotte, Donald MacGlashan, Alkis Togias, and Sarbjit Saini. Omalizumab-induced reductions in mast cell Fc εRI expression and function. Journal of Allergy and Clinical Immunol-ogy , 114(3):527–530, 2004. doi: 10.1016/j.jaci.2004.06.032. [25] Luke F. Pennington, Svetlana Tarchevskaya, Daniel Brigger, et al. Structural basis of omalizumab therapy and omalizumab-mediated IgE exchange. Nature Communications , 7(1):11610, 2016. doi: 10.1038/ncomms11610. [26] Anna M. Davies, Elizabeth G. Allan, Anthony H. Keeble, et al. Allosteric mechanism of action of the therapeutic anti-IgE antibody omalizumab. Journal of Biological Chemistry , 292(24):9975–9987, 2017. doi: 10.1074/jbc.M117.776476. [27] Alexander Eggel, Günther Baravalle, Gabriel Hobi, et al. Accelerated dissociation of IgE-Fc εRI com-plexes by disruptive inhibitors actively desensitizes allergic effector cells. Journal of Allergy and Clin-ical Immunology , 133(6):1709–1719.e8, 2014. doi: 10.1016/j.jaci.2014.02.005. [28] Michael J. Baumann, Alexander Eggel, Patrick Amstutz, Beda M. Stadler, and Monique Vogel. DARPins against a functional IgE epitope. Immunology Letters , 133(2):78–84, 2010. doi: 10.1016/ j.imlet.2010.07.005. [29] A. Eggel, P. Buschor, M. J. Baumann, P. Amstutz, B. M. Stadler, and M. Vogel. Inhibition of ongoing allergic reactions using a novel anti-IgE DARPin-Fc fusion protein. Allergy , 66(7):961–968, 2011. doi: 10.1111/j.1398-9995.2011.02546.x. [30] Alexander Eggel and Theodore S. Jardetzky. Structural and functional insights into IgE receptor interactions and disruptive inhibition. Immunological Reviews , 331(1):e70031, 2025. doi: 10.1111/ imr.70031. [31] Michael T. Stumpp, Keith M. Dawson, and H. Kaspar Binz. Beyond antibodies: The DARPin ® drug platform. BioDrugs , 34(4):423–433, 2020. doi: 10.1007/s40259-020-00429-8. 15 [32] Luke F. Pennington, Pascal Gasser, Silke Kleinboelting, et al. Directed evolution of and structural insights into antibody-mediated disruption of a stable receptor-ligand complex. Nature Communica-tions , 12(1):7069, 2021. doi: 10.1038/s41467-021-27397-z. [33] Daniel Brigger, Pascal Guntern, Luke F. Pennington, Robin van Brummelen, Theodore S. Jardetzky, and Alexander Eggel. High-affinity omalizumab variants with optimized disruptive potency prevent anaphylaxis in vivo. Journal of Allergy and Clinical Immunology , 156(5):1290–1302, 2025. doi: 10.1016/j.jaci.2025.05.028. [34] Excellergy Inc. A phase 1, randomized, double-blind, placebo-controlled, single and multiple ascend-ing dose trial of Exl-111 in healthy participants. ClinicalTrials.gov identifier NCT07356713, 2026. [35] Peter Meerts, Sigrid Cornelis, Guy Hermans, Jo Vercammen, Frank Kamiel Delphina Verdonck, and Stephanie Staelens. Immunoglobulin single variable domains directed against IgE. Patent WO2012175740A1, December 2012. [36] M. Rinaldi, T. Denayer, S. Thiolloy, et al. ALX-0962, an anti-IgE Nanobody ® with a dual mode of action. European Respiratory Journal 42(Suppl. 57), abstract 1765, 2013. [37] Peter Meerts, Sigrid Cornelis, Bruno Dombrecht, Jo Vercammen, and Frank Verdonck. Improved polypeptides directed against IgE. Patent WO2014087010A1, June 2014. [38] Frederic Jabs, Melanie Plum, Nick S. Laursen, et al. Trapping IgE in a closed conformation by mim-icking CD23 binding prevents and disrupts Fc εRI interaction. Nature Communications , 9(1):7, 2018. doi: 10.1038/s41467-017-02312-7. [39] Luke F. Pennington, Pascal Gasser, Daniel Brigger, Pascal Guntern, Alexander Eggel, and Theodore S. Jardetzky. Structure-guided design of ultrapotent disruptive IgE inhibitors to rapidly terminate acute allergic reactions. Journal of Allergy and Clinical Immunology , 148(4):1049–1060, 2021. doi: 10. 1016/j.jaci.2021.03.050. [40] Theodore S. Jardetzky, Luke Pennington, Alexander Eggel, and Pascal Gasser. Heteromultimeric pro-teins for reducing IgE-dependent allergic effector cell activation. Patent WO2022061240A1, March 2022. [41] Edzard Spillner, Sivelle Coline, Lars Harder Christensen, et al. Displacers of IgE-Fc εRI. Patent US12692325B2, July 2026. [42] Claude Science and Amir Shanehsazzadeh. Autonomous de novo protein binder design with Claude. Anthropic technical report, August 2026. URL https://www-cdn.anthropic.com/ 30bf50e22a01388bb29bf077ee3f244531594b7a.pdf .[43] Nick S. Laursen, Frederic Jabs, Edzard Spillner, and Gregers R. Andersen. IgE-Fc in complex with single domain antibody 026. Protein Data Bank entry 5NQW, Worldwide Protein Data Bank, 2017. [44] Balvinder Dhaliwal, Daopeng Yuan, Marie O. Y. Pang, et al. Crystal structure of IgE bound to its B-cell receptor CD23 reveals a mechanism of reciprocal allosteric inhibition with high affinity receptor Fc εRI. Proceedings of the National Academy of Sciences of the United States of America , 109(31):12686– 12691, 2012. doi: 10.1073/pnas.1207278109. [45] Luke F. Pennington and Theodore S. Jardetzky. Structure of a partially disrupted IgE high affinity receptor complex bound to an omalizumab variant. Protein Data Bank entry 7SHT, Worldwide Protein Data Bank, 2021. [46] Justas Dauparas, Gyu Rie Lee, Robert Pecoraro, et al. Atomic context-conditioned protein se-quence design using LigandMPNN. Nature Methods , 22(4):717–723, 2025. doi: 10.1038/ s41592-025-02626-1. [47] Benjamin Fry, Kaia Slaw, and Nicholas F. Polizzi. Zero-shot design of drug-binding proteins via neural iterative selection–expansion. Nature , 656(8126):237–249, 2026. doi: 10.1038/ s41586-026-10670-w. 16 [48] Christopher J. Frank, D. Schiwietz, L. Fuss, Sergey Ovchinnikov, and Hendrik Dietz. AlphaFold2 refinement improves designability of large de novo proteins. bioRxiv, 2024. [49] Yehlin Cho, Griffin Rangel, Gaurav Bhardwaj, and Sergey Ovchinnikov. Protein Hunter: exploiting structure hallucination within diffusion for protein design. bioRxiv, 2025. [50] Ali Madani, Ben Krause, Eric R. Greene, et al. Large language models generate functional protein sequences across diverse families. Nature Biotechnology , 41(8):1099–1106, 2023. doi: 10.1038/ s41587-022-01618-2. [51] Thomas Hayes, Roshan Rao, Halil Akin, et al. Simulating 500 million years of evolution with a language model. Science , 387(6736):850–858, 2025. doi: 10.1126/science.ads0018. [52] Jeffrey A. Ruffolo, Stephen Nayfach, Joseph Gallagher, et al. Design of highly functional genome editors by modelling CRISPR–Cas sequences. Nature , 645(8080):518–525, 2025. doi: 10.1038/ s41586-025-09298-z. [53] Sankar Basu and Björn Wallner. DockQ: a quality measure for protein-protein docking models. PLoS ONE , 11(8):e0161879, 2016. doi: 10.1371/journal.pone.0161879. [54] Claudio Mirabello and Björn Wallner. DockQ v2: improved automatic quality measure for protein multimers, nucleic acids, and small molecules. Bioinformatics , 40(10):btae586, 2024. doi: 10.1093/ bioinformatics/btae586. [55] Salvatore Candido, Thomas Hayes, Alexander Derry, et al. Language modeling materializes a world model of protein biology. bioRxiv, 2026. 17 A Technical details of disruptor redesign

We provide extended commentary and technical details for all disruptor redesign phases as outlined in Figure 3.

A.1 Characterization of 5NQW interface geometry and clarification of redesign scopes

sdAb026 bound to IgE Fc Cε3–C ε4 (PDB 5NQW) Cε4×CDR3 Cε3×CDR1,3 CDR1,3 free scope CDR2 scope D99 E100 E101 Y102 A103 L104 G105 P106 N107 E108 F109 D110 Y111 G26 F27 T28 F29 G30 N31 Y32 D33 A97 T98 D99 E100 E101 Y102 A103 G26 F27 T28 F29 G30 A97 T98 A103 L104 G105 F109 Y111 I51 D52 T53 G54 G55 D56 I57 T58 IMGT CDR1 IMGT CDR2 IMGT CDR3

Figure 7. Redesign scopes on sdAb026. sdAb026 bound to IgE Fc C ε3–C ε4 (PDB 5NQW; C ε3 and C ε4rendered in white and light-grey, respectively). Left-middle to right: the four redesign scopes with color-coded scope residues highlighted according to their respective CDRs.

Determination of design scope After analyzing the binding mode of sdAb026 in PDB 5NQW, Claude nominated the C ε3 interface, C ε4 interface, and free scopes detailed in Figure 7. As reported in the main text, the free scope was eventually dropped in favor of a fourth, interface-distal CDR2 scope so as to increase the diversity and extent of the explored optimization space. We emphasize that chronologically, full redesign was initially carried out against the three non-CDR2 scopes before the conceptualization of the fourth, complementary CDR2 scope reported in Figure 3. Claude then further identified the first three, interface-proximal scopes as being candidates for targeted mutagenesis (edit distance 2-4) rather than full scope redesign, utilizing both inverse folding (Appendix A.2) and a gradient-based mutagenic design scheme developed for precisely this purpose, as described in Appendix A.3.

A.2 Inverse Folding of Scopes 1-3

Sequence-biased inverse folding. To achieve the desired mutagenic effect with inverse folding, Claude initially considered combining a uniform and calibrated sequence bias with SolMPNN [46]. First, N = 250

unbiased draws using SolMPNN at T = 0 .1 were produced for each scope S, with ski,j denoting the k-th recorded logit of the j-th residue member of the i-th draw, after which Claude sought to calibrate a uniform, scope-specific biasing parameter bS so as to maximize the fraction

N M

N

i=1 M

m=1

1 2 ≤ di,m (bS ) ≤ 3 (1) where

di,m (bS ) = X

j∈S

1 ˜r mi,j ̸ = rj

, ˜r mi,j ∼ softmax k

�ski,j + bS 1[k = rj ]/T

(2) of biased resamples predicted to land between two and three edits, where rj is the parent sdAb026 residue at scope position j, ˜r mi,j is the m-th of M residues resampled at that position from the recorded logits ski,j

of unbiased draw i after adding the bias bS to the parent residue’s logit (softmax over the twenty amino acids k, tempered by T ), di,m (bS ) is the number of substitutions within S carried by that resample, and 18 0.5 0.6 0.7 0.8 0.9

Critic A mean binder-pair ipTM 0.5 0.6 0.7 0.8 0.9 Critic B mean binder-pair ipTM aUnbiased probe, every scope position free (T = 0.1) tether (75 unique / 250 draws) anchor (76 unique / 250 draws) free (6 unique / 250 draws) sdAb026 b FR1 CDR1 FR2 –FR3 CDR3 sdAb026 SGFTFGNYDM…CATDEEYALGPNEFDYscope 1 · tether H®= 0.45 bits SGFTFGNYDM…CATDEEYALGPNEFDYscope 2 · anchor H®= 0.59 bits SGFTFGNYDM…CATDEEYALGPNEFDY scope 3 · free H®= 0.08 bits SGFTFGNYDM…CATDEEYALGPNEFDY 25 26 27 28 29 30 31 32 33 34 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 wild-type residue (sdAb026) substitution redesign scope IMGT CDR 012345678 mean edits per draw (over scope positions) 0.0 0.2 0.4 0.6 0.8 1.0 expected fraction of 250 draws unique cDiversity vs edit count under the parent bias tether anchor free T = 0.1 T = 0.2 T = 0.4 probe, observed (T = 0.1, b = 0) MC at b* (T = 0.1)

Figure 8. Sequence-biased inverse folding of scopes 1–3. (a) Critic A against Critic B mean binder-pair ipTM for the unique sequences among the N = 250 unbiased SolMPNN draws per scope at T = 0 .1, with marker size scales with the number of draws yielding the sequence. (b) Sequence logos of the unbiased draws (top) and of the biased resamples of Equation (1) at the calibrated bS (bottom), over sdAb026 with the three scopes marked; black, the sdAb026 residue; red, a substitution. (c) Expected fraction of unique sequences among 250 biased resamples against their mean edit distance as bS is swept, at T = 0 .1, 0.2 and

0.4; filled markers, the unbiased draws.

1[·] is the indicator function. Claude performed this calibration procedure to obtain such a bS for each of the three interface-proximal scopes. Logo diagrams for the averaged unbiased and biased position-specific probability matrices are shown on the top and bottom, respectively, of Figure 8 (b), emphasizing that bias was primarily applied toward a small number of higher entropy indices, so as to yield an undesirable tension between edit distance and redesign diversity, and so that the realization of diverse sequences requires a stronger bias, yielding designs of both high edit distance and low unbiased model likelihood. Indeed, as modeled in Figure 8 (c), the fraction of unique designs rises only modestly within the edit distance interval of interest (2-4 edits) as temperature is increased from T = 0 .1 to T = 0 .4. Subsequent inverse folding and critic evaluation substantiated such biasing as detrimental to model confidence, suggesting that biasing in this fashion yields sequences which are consistent with neither the inverse folding model itself nor the downstream folding critics.

Inverse Folding with Manually Constructed Scopes In light of poor critic confidence on the part of manually calibrated biased inverse folding, Claude elected instead to manually enumerate all double-, triple-, and a select number of quadruple mutants of sdAb026 until a predetermined budget of N = 750

was obtained, and subsequently inverse fold at T = 0 .4 with SolMPNN. Critic confidence with respect to both nominated critics, edit distance, and DockQ are visualized in Figure 10. In particular, we report both mean critic binder-pair ipTM (that is, the average ipTM between the design and the respective IgE dimer chains), and the sub-interface DockQ mean [53, 54], which we define as the average DockQ of the variant with respect to both C ε3 and C ε4 interfaces.

A.3 Tempered redesign of Scopes 1-3

In addition to the inverse folding in Appendix A.2, Claude enabled the rapid development and integration of a second, gradient-based design approach we’ve termed tempered redesign , which together with an improved interface geometry loss , yields designs which achieve both higher critic confidence and deviate less significantly from the intended reference geometry across all three initial design scopes, and as visu-alized in Figure 9. A more extensive technical discussion of both techniques can be found in Appendix C. 19 0.5 0.6 0.7 0.8 0.9

mean binder-pair ipTM (10 seeds) 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Mean Sub-Interface DockQ (10 seeds) scope 1 · C ε3 × CDR1,3 0.5 0.6 0.7 0.8 0.9 mean binder-pair ipTM (10 seeds) 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Mean Sub-Interface DockQ (10 seeds) scope 2 · C ε4 × CDR3 0.5 0.6 0.7 0.8 0.9 mean binder-pair ipTM (10 seeds) 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Mean Sub-Interface DockQ (10 seeds) scope 3 · CDR1,3 free 0.70 0.75 0.80 0.85 0.90 mean binder-pair ipTM (10 seeds) 0.60 0.65 0.70 0.75 0.80 0.85 0.90 0.95 Mean Sub-Interface DockQ (10 seeds) scope 4 · CDR2

inverse folding tempered redesign redesign per-method median sdAb026 Figure 9. Summary of sdAb026 redesign designs. Top-left, top-right, and bottom-left: scatters of mean binder-pair ipTM against mean sub-interface DockQ for double, triple, and quadruple mutants generated by tempered redesign (blue) and inverse folding (grey) against the C ε3 interface, C ε4 interface, and free scopes, respectively. Bottom-right: scatter of mean binder-pair ipTM against mean sub-interface DockQ for the redesigns obtained via inverse folding with SolMPNN (grey) and gradient-based redesign (red) against the CDR2 scope.

A.4 CDR2 Redesign

For the CDR2 scope, and due to its distance from both C ε3 and C ε4 interfaces, Claude near-autonomously executed full-scope redesign using a combination of both inverse folding with SolMPNN and gradient-based redesign. Toward this end, SolMPNN (v_48_020) was used to generate 750 redesigns at each rung of the temperature ladder T ∈ { 0.1, 0.3, 0.5, 0.7, 1.0}, obtaining unique-sequence fractions of 18.9%, 61.6%, 87.6%, 97.9% and 100% respectively (750 draws per rung), of which 1,289 unique designs over the ladder 20 0.4 0.5 0.6 0.7 0.8 0.9

inverse folding Critic B mean binder-pair ipTM n = 3,082

scope 1 · tether (C ε3 × CDR1,3)

n = 2,572

scope 2 · anchor (C ε4 × CDR3)

n = 2,067

scope 3 · free (CDR1,3 non-contact)

0.4 0.5 0.6 0.7 0.8 0.9

Critic A mean binder-pair ipTM

0.4 0.5 0.6 0.7 0.8 0.9 tempered redesign Critic B mean binder-pair ipTM n = 538 0.4 0.5 0.6 0.7 0.8 0.9

Critic A mean binder-pair ipTM

n = 567 0.4 0.5 0.6 0.7 0.8 0.9

Critic A mean binder-pair ipTM

n = 520 0.4 0.5 0.6 0.7 0.8 0.9 inverse folding mean binder-pair ipTM (10 seeds) n = 3,082 n = 2,572 n = 2,067 234

edit distance to sdAb026

0.4 0.5 0.6 0.7 0.8 0.9 tempered redesign mean binder-pair ipTM (10 seeds) n = 538 234

edit distance to sdAb026

n = 567 234

edit distance to sdAb026

n = 520 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Mean Sub-Interface DockQ (10 seeds)

Figure 10. Inverse folding and tempered redesign of scopes 1–3. Rows: inverse folding (every unique sequence among the enumerated double, triple and sampled quadruple mutants of sdAb026) and tempered redesign (the production pool at two to four substitutions); columns: the three scopes. Top: Critic A against Critic B mean binder-pair ipTM per sequence, colored by mean sub-interface DockQ to PDB 5NQW over ten seeds; star, sdAb026. Bottom: edit distance to sdAb026 against mean binder-pair ipTM over ten seeds, same coloring; bars, per-edit medians; dashed line, sdAb026. 21 Phase Method H100 h Settings Designs produced / nominated

Inverse folding 7→Cε3 ×CDR1,3 SolubleMPNN on the sdAb026–Fc complex (PDB 5NQW), subset mode 80 v_48_020; every pair and triple and 300 quadruples of the 15 scope positions, 5 draws each; T= 0.4; parent residue excluded; 5 folds per critic 3 237 / 13 Inverse folding 7→Cε4 ×CDR3 as above 60 as above over the 13 scope positions 2 572 / 13 Inverse folding 7→ CDR1,3 free as above 50 as above over the 12 scope positions 2 101 / 0 Tempered redesign 7→ Cε3×CDR1,3 gradient-based redesign with the interface specification and the tempered schedule 265 18 interface groups (one per sdAb026 contact residue and Fc chain, min-k1, 8–10 Å) at total weight 0.5; template 0.5; pLM 0.05; lr 0.1; Sapex = 0.60, N= 120; 876 trajectories; 5 folds per critic 641 / 13 Tempered redesign 7→ Cε4×CDR3 as above 265 as above with S apex = 0.70, N= 120; 874 trajectories 681 / 13 Tempered redesign 7→ CDR1,3 free as above 265 as above with S apex = 0.60, N= 120; 874 trajectories 629 / 0 Inverse folding 7→CDR2 SolubleMPNN on the complex, all CDR2 positions free 25 v_48_020; parent residue excluded at all 8 positions; 1 289 designs over a T= 0.1–1.0 temperature ladder; 5 folds per critic 1 289 / 19 Redesign 7→CDR2 gradient-based redesign from a random CDR2 start, standard cosine schedule 130 same losses and interface groups as the tempered rows; 150 steps, T1.0 →0.01; 750 trajectories; 5 folds per critic 695 / 20 Total 1 140 H100 h; 11 845 unique designs scored under both critics; 91 nominated for BLI.

Table 1. Disruptor optimization: rounds of computation.

were scored under both critics (Table 1). Gradient-based redesign was similarly run with identical interface geometry loss settings to those developed and used in Appendix A.3 to obtain 750 trajectories, yielding 695 distinct designs (93%). The results of both inverse folding and gradient-based redesign are displayed in Figure 9, indicating that direct redesign trades off exact recapitulation of interface geometry as measured by mean sub-interface DockQ for an increase in mean critic confidence in the form of a thicker tail.

A.5 Conclusion and nomination

Claude concluded the round by nominating a total of 91 distinct designs by selecting the top 13 from each of four total C ε3 and C ε4 scope-method combinations, as well as 20-each from each of the two design approaches employed against the CDR2 scope. One of the CDR2 scope variants redesigned using inverse folding was subsequently dropped to accommodate an additional control, yielding the reported total of 91 designs. An overview of the ipTM of the nominated designs and their edit distances to sdAb026 is provided in Figure 4. Additional design details are provided in Table 1.

B Technical details of de novo anchor design

B.1 Computational Prediction of Literature-Described C ε2 Binders

In Figure 11 (a), we list reported affinities and kinetic quantities for immunization-derived, non-humanized nanobodies confirmed to bind IgE Fc [41]. We additionally pursue refolding against both IgE Fc C ε2 and Cε2–C ε4 dimers using one of the two nominated campaign critic models and report intra-variant epitope Jaccard and ipTM results in Figure 11 (b, c). In particular, we obtain a shared, predicted consensus epitope for variants B1, E1, and F4, but fail to reach consensus for variants B3 and D2. Folds against C ε2–C ε4 and Cε2-only dimers are provided in Figure 11 (d).

B.2 Computational design against C ε2 Scope

Scope construction. We consider a global C ε2 scope constructed from PDB 2Y7Q as the set of solvent-accessible residues (per-residue SASA ≥ 15 2 via PyMOL) whose alpha carbon has an unobstructed line of sight of length at least 20 Å to an alpha carbon on C ε3–C ε4. We then further restrict to obtain a local and limited B1-consensus scope as the intersection of this initially constructed C ε2 scope with the set of residues present in the epitope footprint in at least 75% of B1 cofold predictions from Appendix B.1. Both scopes are visualized in Figure 12. We next describe a total of three Claude-driven de novo design phases against these two scopes. 22 sdAb KD koff (s −1)B1 65 nM 4.9 × 10 −2

B3 0.28 nM 2.5×10 −4 D2 81 nM 2.0×10 −2 E1 n.r. n.r. F4 n.r. n.r.

(a) B1 B3 D2 E1 F4 0.0 0.2 0.4 0.6 0.8 1.0 seed-to-seed footprint Jaccard

within C ε2-dimer folds within C ε2–Cε4-dimer folds pooled, both

(b) B1 B3 D2 E1 F4 0.0 0.2 0.4 0.6 0.8 1.0 mean binder-pair ipTM (5 seeds) arm + C ε2 dimer arm + C ε2–Cε4 dimer (c)

(d)

Figure 11. Literature-described C ε2 anchors: reported kinetics and their refolding on the C ε2 dimer and on the C ε2–C ε4 dimer. (a) Reported BLI affinity and dissociation rate on IgE-Fc for five experimentally characterized sdAbs from Spillner et al. [41] (n.r.: not reported.) (b) Seed-to-seed consistency of the 4 Å footprint as the pairwise Jaccard within the five C ε2-dimer folds, within the five C ε2–C ε4-dimer folds, and pooled over all ten (mean ± SEM over pairs; Critic A). (c) Mean binder-pair ipTM ± SEM over the five Critic A seeds on each target. (d) All ten refolds of each sdAb (Critic A and Critic B, five seeds each) superimposed on one reference surface of the target, with sdAb chains rendered in blue.

Design against B1 consensus scope For each of the twelve distinct frameworks in Table 2, Claude ran 102 design trajectories of epitope-conditioned redesign of all three CDRs, for a total of 1224 total design trajectories. Each such design was then subsequently refolded using five distinct seeds and across both campaign critic models, with results reported in Figure 12. We then iteratively refine the top-40 variants identified by 10-seed mean ipTM through repeated inverse folding and folding, similar to the procedures described by Fry et al. [47], Frank et al. [48], Cho et al. [49]. Results are reported in Figure 12, including both 10-seed mean critic confidence and epitope Jaccard to the target epitope.

Epitope scanning against C ε2 To increase epitope diversity, Claude subsequently elected to perform design with randomized epitope selection: a query residue is randomly selected from the global scope from which an epitope is then constructed as the set of all in-scope residues whose alpha carbon lies within a 12 Åcutoff of the query. We henceforth refer to this approach as epitope scanning . Design is then carried out as in the previous phase, and against the C ε2 dimer, with the additional repulsive loss term applied to out-of-scope residues on C ε2. In total, this round generated a total of 3600 distinct trajectories evenly divided among the same twelve frameworks shown in Table 2. The top 75 designs by 10-seed mean ipTM were subsequently nominated for further iterative refinement totaling 50 additional H100 hours, and 23 (a) 0.0 0.2 0.4 0.6 0.8

footprint Jaccard to the B1 consensus epitope

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 mean binder-pair ipTM (10 seeds) design against the B1 consensus scope

(b) 0.0 0.2 0.4 0.6 0.8

footprint Jaccard to the B1 consensus epitope

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 mean binder-pair ipTM (10 seeds) randomised epitope conditioning, C ε2 scope

footprint Jaccard to the B1 consensus epitope

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 mean binder-pair ipTM (10 seeds) unconditioned design, C ε2 dimer

(d) designs mean B1 (reference anchor)

Figure 12. The C ε2 design scope and the three design rounds against it. (a) The C ε2 inside-face scope (green) and the nine-residue B1 consensus epitope (blue) on the receptor-bound IgE Fc of PDB 7SHT, and with all domains rendered as surfaces. (b–d) Mean binder-pair ipTM over the ten critic seeds against the footprint Jaccard to the B1 consensus epitope (mean over the Critic A seeds), one point per unique design, with marginals: (b) design against the B1 consensus scope, (c) epitope scanning over the C ε2 scope, (d) unconditioned design against the C ε2 dimer. 24 Framework FR1 FR2 FR3 FR4 Native CDR1 / 2 / 3 Sampled lengths

Isecarosmab DVQLVESGGGVVQPGGSLRLSCAAS MGWFRQAPGKEREFVAG YYVDSLKGRFTISRDNSKNTVYLQMNSLRPEDTALYYC WGQGTLVTVSS 8 / 8 / 17 7–9 / 8–9 / 14–20 Gocatamig2 EVQLVESGGGLVQPGGSLTLSCAAS LAWYRQAPGKKRELVAG VYMDSVKGRFTISRDNAKNSVYLQMNSLRAEDTAVYYC WGQGTLVTVSS 8 / 7 / 11 7–9 / 7–8 / 8–14 Gontivimab DVQLVESGGGLVQAGGSLSISCAAS LGWFRQAPGKEREFVAA IGPPNVEGRFTISRDNAKNTGYLQMNSLAPDDTAVYYC WGRGTQVTVSS 8 / 8 / 19 7–9 / 8–9 / 16–22 Vobarilizumab EVQLVESGGGLVQPGGSLRLSCAAS MAWYRQAPGKGRELVAG SYADSVKGRFTISRDNAKNTLYLQMNSLRPEDTAVYYC WGQGTLVTVSS 8 / 7 / 15 7–9 / 7–8 / 12–18 Brivekimig1 DVQLVESGGGVVQPGGSLRLSCAAS KGWFRQAPGKEREFVAA SYADSVKGRFTISRDNSKNTVYLQMNSLRPEDTALYYC WGQGTLVTVSS 9 / 8 / 16 8–10 / 8–9 / 13–19 Erfonrilimab (VHH2) QVQLVESGGGLVQPGGSLRLSCAAS MGWFRQAPGKGLEGVAA YYADSVKGRFTISRDNSKNTLYLQMNSLRAEDTAVYYC WGQGTLVTVSS 8 / 8 / 21 7–9 / 8–9 / 18–24 Produvofusp QVQLVESGGGLVKPGGSLRLSCAAS MGWFRQAPGKEREFVSA YYADSVKGRFTISRDNAKNSLYLQMNSLRAEDTAVYYC WGQGTLVTVSS 8 / 8 / 22 7–9 / 8–9 / 19–25 Sonelokimab1 DVQLVESGGGLVQPGGSLRLSCAAS VGWFRQAPGKEREFIGA YYAVSEKGRFTISRDNSKNTLYLQMNSLRPEDTAVYYC WGQGTLVTVSS 8 / 8 / 16 7–9 / 8–9 / 13–19 Podentamig1 EVQLVESGGGLVQPGRSLTLSCAAS YGWYRQAPGKQRELVAA LYADSVKGRFTISRDNAKNSIYLQMNSLRPEDTALYYC WGQGTQVTVSS 8 / 6 / 14 7–9 / 6–7 / 11–17 Lunsekimig2 EVQLVESGGGVVQPGGSLRLSCAAS LYWYRQAAGIERELIAS NYVDSVKGRFTISRDNSENTMYLQMNSLRAEDTGLYYC WGQGTLVTVSS 8 / 7 / 11 7–9 / 7–8 / 8–14 Erfonrilimab QVQLVESGGGLVQPGGSLRLSCAAS MAWFRQAPGKERERVAK YLADSVKGRFTISRDNSKNTVYLQMNSLRAEDTAVYYC WGQGTLVTVSS 8 / 8 / 21 7–9 / 8–9 / 18–24 Lofacimig QVQLVESGGGSVQAGGSLRLSCAAS MAWFRQAPGKEREGVAA SYADSVKGRFTISKDNSKNTLYLQMNSLRAEDTAVYYC WGQGTLVTVSS 9 / 7 / 23 8–10 / 7–8 / 20–26

Table 2. VHH frameworks used for de novo anchor design. Framework regions FR1–FR4 (IMGT) of the twelve therapeutic VHH scaffolds; the three CDR loops between them are designed from random logits with lengths drawn uniformly from the sampled windows (CDR1 / CDR2 / CDR3). Native loop lengths are reported for reference. yielding an average of 20 cycles of refinement. Results are reported in Figure 12, including both 10-seed mean critic confidence and epitope Jaccard to the epitope from the previous round.

Unconditioned design against C ε2 Finally, Claude elected to run a final, unconditioned design run against the C ε2 scope, this time without any repulsive loss term. This round generated a total of 7194 design variants across the 12 VHH frameworks. Further experimental details can be found in Table 3, and results are reported in Figure 12.

B.3 Computational design against C ε3–C ε4

Linker- and epitope-related filtering procedure The flexible positioning of IgE C ε2 relative to C ε3–C ε4makes it hard to gauge the geometric compatibility of an anchor placed on the latter with a disruptor on the former on tandem or Fc-fusion linker geometry, as doing so would require more precise modeling of domain rearrangement and linker geometry to accommodate cis binding of the resulting biparatopic. In contrast, the comparatively rigid C ε3–C ε4 varies largely between the open and closed conformations described in Section 2, so that direct screening for linker compatibility is more straightforward. For this reason, Claude formulated a screening algorithm for determining the compatibility of a choice of anchor with a C-terminal IgG1 fusion, as in [41], and together with disruptor sdAb026 (via PDB 5NQW). The procedure involves first performing a rigid alignment of both the anchor arm as well as one of the two symmetrically related sdAb026 chains in 5NQW to PDB 2Y7Q. Then, for each of linker lengths (G 4S) 1 (n = 4 ) and (G 4S) 2 (n = 8 ), we construct circular shells centered at the respective VHH N-termini, and with inner and outer radii equal to 0.8 × L and 1.2 × L for L ≜ 0.8 × 3.8 × n. A 1 Åcircular grid is then constructed about the midpoint of the N-termini, and a grid search is performed for apices which satisfy strict clash criteria. Namely, the shortest clash-free paths are identified via Dijkstra’s shortest-path algorithm on a 1.5Å cubic lattice, with lattice points within 3 Åof IgE Fc, receptor, or arm chains excluded. A final check is applied by screening for clashes involving a rigid-body cylindrical approximation of the IgG1 Fc. If such a non-clashing path exists, the anchor is said to be linker-compatible. We additionally apply a secondary, epitope-related filter removing any design which overlaps with the respective residue footprints of sdAb026 or Fc εRI α.

Epitope scanning against C ε3–C ε4 In line with de novo design against C ε2, Claude similarly pursued design against C ε3–C ε4 with an identical epitope scanning procedure to the one described in Appendix B.2, and with a similar repulsive avoidance loss intended to exclude residues within the footprints of sdAb026 and Fc εRI α. A total of 1584 design trajectories were run, after which the top-50 designs by mean critic ipTM passing both linker compatibility and epitope filters were selected for further refinement in the same manner as described in the C ε2 epitope scanning phase, significantly diversifying the set of high-confidence, filter-passing designs, and as shown in Figure 13 (top-right, middle-left).

Design ablations against C ε3–C ε4 The low incidence of filter-passing, high-confidence designs prior to refinement in the previous phase prompted Claude to suggest a series of ablations applied to the design 25 Linkable and overlap-free Overlaps with disruptor arm

0.0 0.2 0.4 0.6 0.8 1.0 Critic B mean binder-pair ipTM (5 seeds) 0.0 0.2 0.4 0.6 0.8 1.0 Critic A mean binder-pair ipTM (5 seeds)

Cε3–Cε4 Epitope Scan Pre-Re fi nement (n = 1584)

exclusion-free, path-aware linkable (711) (straight-arm test: 764)

0.0 0.2 0.4 0.6 0.8 1.0 Critic B mean binder-pair ipTM (5 seeds) 0.0 0.2 0.4 0.6 0.8 1.0 Critic A mean binder-pair ipTM (5 seeds)

Cε3–Cε4 Epitope Scan Post-Re fi nement (n = 1247)

exclusion-free, path-aware linkable (428) (straight-arm test: 433) nominated for the plate (12)

0.0 0.2 0.4 0.6 0.8 1.0 Critic B mean binder-pair ipTM (5 seeds) 0.0 0.2 0.4 0.6 0.8 1.0 Critic A mean binder-pair ipTM (5 seeds)

Unconditioned C ε3–Cε4 Pre-Re fi nement (n = 2916)

exclusion-free, path-aware linkable (154) (straight-arm test: 193)

0.0 0.2 0.4 0.6 0.8 1.0 Critic B mean binder-pair ipTM (5 seeds) 0.0 0.2 0.4 0.6 0.8 1.0 Critic A mean binder-pair ipTM (5 seeds)

Unconditioned C ε3–Cε4 Post-Re fi nement (n = 3127)

exclusion-free, path-aware linkable (1319) (straight-arm test: 1463) nominated for the plate (3)

epitope scan rerun alternative design model avoid o ff

attract ×3 attract ×5 attract ×10 attract ×20 unconditioned dimer unconditioned monomer 0.0 0.2 0.4 0.6 0.8 1.0 mean binder-pair ipTM (10 seeds) ± SEM

47 23 17 18 14 29 21 24

all designs exclusion-free and linkable

epitope scan rerun alternative design model avoid o ff

attract ×3 attract ×5 attract ×10 attract ×20 unconditioned dimer unconditioned monomer 0.0 0.2 0.4 0.6 0.8 1.0 mean intra-seed footprint Jaccard ± SEM

47 23 17 18 14 29 21 24

all designs exclusion-free and linkable Figure 13. De novo anchor design against C ε3–C ε4. Top: a nominated anchor passing both linkability and epitope-overlap criteria, tethered to the proxy disruptor arm sdAb026 (red) by a (G4S) 1 linker proxy (black; arms drawn as shortest obstacle-free paths as described in Appendix B.3) on the 5NQW Fc C ε3– Cε4; an anchor with epitope (and steric) overlap with sdAb026; the epitope-scan pool before refinement, Critic A vs Critic B mean binder-pair ipTM (5 seeds), with exclusion-free and KIH-linkable designs in green. Middle: the same pool after refinement (plate nominees in red), and the unconditioned pool before and after refinement (plate nominees in red). Bottom: mean 10-seed binder-pair ipTM and mean intra-seed footprint Jaccard, both ± SEM and per ablation setting, with all designs (black) against the exclusion-free and linkable subset (green; subset sizes above the points). 26 Phase Method H100 h Settings Designs produced / nominated

B1-epitope C ε2 design de novo gradient-based against the 9-residue B1 consensus epitope; subsequent iterative refinement expansion of the top 40 75 random start, all CDRs, 12 frameworks drawn uniformly, native CDR lengths; cosine 1.0 →0.01 over 150 steps; one interface group on the 9 consensus-epitope residues (min-kcolumn 5, 12 Å, weight 0.5) 1 224 / 4 Cε2 design de novo gradient-based against the whole C ε2 inside-face scope with epitope scanning; subsequent iterative refinement of 75 epitope-diverse nominees 210 same start, frameworks, schedule and losses; one random 10 Å epitope patch per design drawn from the inside-face scope (min-k5/ 5, 12 Å, weight 0.5) 3 600 / 6 Cε3–C ε4 design de novo gradient-based with epitope scanning and an avoid group; subsequent iterative refinement 195 scanning patch 10 Å per Fc copy (min-k5 / 5, 12 Å, +0 .5); avoid group on the Fc εRI αand sdAb026 footprints ( −0.5, 8 Å); expansion with SolubleMPNN v_48_020 at T0.10, 50 starts ×2 streams 2 831 / 12 Scope-specific design ablations paired ablations of the scanning workflow at matched design indices 210 baseline rerun; grouping and avoid-weight variants; attract-weight ladder ×1. . . ×20 ; an unconditioned arm (no scanning, no interface groups) 2 703 / 0 Ablation-refined Cε3–C ε4 design unconditioned gradient-based; the 8 best designs seed the hit expansion 270 972 designs ×3 rows; same schedule and losses 2 916 / 0 Ablation-refined C ε2design unconditioned gradient-based paired index-by-index with the Cε2 design round 270 2 400 designs ×3 rows reusing the earlier frameworks and CDR lengths 7 194 / 32 Cε3–C ε4 hit expansion fixed-template inverse folding of the CDRs of the 8 picks 80 SolubleMPNN v_48_020, T0.1, Cys excluded; template = best Critic A seed; 400 draws per start in strata of 4 / 8 / 16 / all CDR positions 3 119 / 3 Total 1 310 H100 h; 23 587 unique designs scored under both critics; 57 nominated for BLI.

Table 3. Anchor design: rounds of computation.

procedure, swapping the design model checkpoint, dropping the avoidance weight, varying the weight ap-plied to the interface geometry loss term, dropping conditioning altogether, and designing instead against a C ε3–C ε4 monomer. For each, 150 designs were run with the ablated settings, with results reported in Figure 13 (bottom row). Overall, and among other observations, we find that removing the repulsion loss raises the average design confidence as well as the average intra-design epitope Jaccard score above the baseline, albeit at the expense of a reduced incidence of filter-passing designs.

Unconditioned generation against C ε3–C ε4 Citing a desire for consistency with the C ε2 design cam-paign, higher intra-design epitope consistency, and the possibility of hit-expansion as a means to mitigate initially low filter-passing rates, Claude elected to carry out a final round of completely unconditioned de-sign against C ε3–C ε4 totaling 2916 design trajectories. Of these, a small subset ( n = 8 ) was advanced to hit-expansion and diversification via inverse folding with SolMPNN, yielding a refined pool of over 3000 additional designs. Pre- and post-refinement results are shown in Figure 13.

B.4 Conclusion and nomination

Claude finally nominated a total of 57 designs for experimental validation, including four designed against the B1 consensus epitope on C ε2, 38 from subsequent design against C ε2, and 15 against C ε3–C ε4. Further design and model details can be found in Table 3.

C Claude-driven methodological extensions

Throughout the campaign, we repeatedly leaned on Claude to enable rapid, on-the-fly conceptualization, implementation, and deployment of methodological extensions of open-source and internal design work-flows custom-tailored to accommodate structural, sequential, or other experimental priors specific to the campaign. Two such examples came up during disruptor redesign which we believe materially shifted our capacity to prosecute the campaign, and which we expect to be of further use within our internal design platform going forward. We describe each now as follows. 27 0.0 0.5 1.0 design trajectory progress 0.0 0.5 1.0 temperature

tempered redesign normal design atemperature schedules Sapex = 0.60, Tapex = 0.35 012345678substitutions vs sdAb026 010 20 30 40 50 designs bTempered redesign (C ε3×CDR1,3) 040 step 01T 012345678substitutions vs sdAb026 0246810 12 14 cTempered redesign (C ε3×CDR1,3) 0120 step 01T 012345678substitutions vs sdAb026 050 100 150 200 250 dRedesign (CDR2) 0150 step 01T cold (well-determined residue identity) hot (uncertainty over residue identity)

Figure 14. Tempered redesign. (a) Temperature schedules: normal design follows the cosine schedule from hot to cold; tempered redesign runs the same schedule from cold up to the turnaround Sapex in reference time and back, drawn at Sapex = 0 .60 (Tapex = 0 .35 ). (b) Substitutions per design in the ultra-low-edit sweep cell (C ε3 × CDR1,3 scope, Sapex = 0 .85 , N = 40 steps, 76 trajectories). (c) The chosen production cell (C ε3 × CDR1,3 scope, Sapex = 0 .60 , N = 120 steps, 38 trajectories). (d) Normal design of the CDR2 scope under the standard schedule ( N = 150 steps, 695 designs). Insets depict the respective temperature schedules.

epitope × all binder target residues i

binder residues j

sub-interface 1 sub-interface 2 target residues i

binder residues j

Figure 15. Left: vanilla epitope loss from Candido et al. [55] promotes contact between a specified epitope and the entire binder. Right: the granular interface geome-try loss promotes contact between multiple distinct epi-topes, each with respective, distinct paratope groups not equal to the entire binder.

C.1 Targeted mutagenesis via tempered redesign

Our design platform consists of a gradient-based design model similar to that disclosed in Candido et al. [55], and which natively supports the redesign of particular residue subsets starting from scratch . In con-junction with Claude, we reasoned that by modifying the annealed temperature schedule to instead start at a low, near-zero temperature before rising to a predetermined apex, and then cooling back down to the starting temperature, we could realize a gradient-based mutagenic design scheme. In particular, recall the cosine temperature schedule

ρref : [0 , 1] → [Tmin , 1] , ρref (sref ) = Tmin + (1 − Tmin ) 12

�1 + cos( π s ref ), (3) so that the tempered redesign schedule ρtempered is realized as the composition ρref ◦ πtransform , with the tempering transform πtransform defined as

πtransform : [0 , 1] → [0 , 1] , πtransform (sreal ) =

(1 − 2 sreal (1 − Sapex ) sreal ≤ 12 ,

1 − 2 (1 − sreal ) (1 − Sapex ) sreal > 12 , (4) Here, Tmin denotes the minimum design temperature, and Sapex denotes apex reference time. Tempered redesign is then achieved by taking the initial sequence chosen explicitly rather than randomly and apply-ing Algorithm 11 of Candido et al. [55] with tempered temperature schedule ρtempered in place of ρref .

C.2 An improved interface geometry loss

An additional design modification whose rapid campaign integration was driven by Claude came in the form of a generalization of an existing epitope loss from Candido et al. [55], allowing for the specification of interface geometry at arbitrary granularity. This interface geometry loss enables the creation of multiple distinct interface sub-groups, and the specification of group-specific epitope and paratope residue sets each 28 Algorithm 1 Interface Geometry Loss

Require: distogram logits z∈RL×L×Bover Bdistance bins with midpoints δ1<· · · < δ B Require: groups g= 1 , . . . , G , each a tuple (Rg,Cg, k row g, k col g, d c,g , w g): Rgrow residues of the group’s block (the target residues) Cgcolumn residues of the block (the binder residues) kcol g∈N∪ { all }columns counted per row; krow g∈N∪ { all }rows counted per group (each drawn once per trajectory) dc,g contact cutoff (Å); wggroup weight Require: ε= 10 −8 1: p←softmax b(z)▷per-pair distance distribution pijb 2: L ← 0 3: for g= 1 , . . . , G do 4: Pij ←P b:δb<d c,g pijb for i∈ R g, j ∈ C g▷probability that the pair lies within dc,g 5: mij ← − log ( Pij +ε)▷pair measure 6: for i∈ R gdo 7: Ji←the kcol gcolumns j∈ C gwith the smallest mij ▷all of Cgif kcol g= all 8: ri←1 |J i| P j∈J imij ▷row score 9: end for 10: Ig←the krow grows i∈ R gwith the smallest ri▷all of Rgif krow g= all 11: ℓg←1 |I g| P i∈I gri▷group term 12: L ← L +wgℓg▷groups are summed, so no group can fill another’s kslots 13: end for 14: return L Reference epitope loss ≡G= 1 ,R1=epitope residues, C1=all binder residues, kcol 1= 5 ,krow 1= all ,dc, 1= 12 Å.

Figure 16. Pseudocode for the proposed interface geometry loss, generalizing the epitope loss of Candido et al. [55] to arbitrarily many interface groups realized as pairs of subsets of rows and columns within the larger distogram, with each containing distinct distance cutoffs and row- and column-count parameters

k{row ,column }

g

.with distinct distance cutoffs and contact parameters. We provide an overview of this modified interface geometry loss in Figure 16 as well as a demonstration as to the degree to which it can, in combination with tempered redesign, yield redesigns which more faithfully recapitulate the desired interface geometry when compared to the vanilla epitope loss.

D Experimental methods

Protein expression and purification. VHH designs were expressed by cell-free protein synthesis (Gen-Script CFXpress PLUS E. coli CFPS kit, cat. RP-D00006-1) from 125 ng of linear DNA template carrying a T7 promoter, ribosome-binding site, the VHH with a C-terminal Twin-Strep tag, a stop codon and a T7 terminator, in the kit’s reaction buffer according to the manufacturer’s instructions. For the screens, reac-tions were used unpurified after dilution in PBS or Gator K buffer. For purification, variants were captured from the reactions on MagStrep Strep-Tactin XT beads (IBA Lifesciences, cat. 2-5090-010), washed with Buffer W (100 mM Tris-HCl pH 8.0, 150 mM NaCl, 1 mM EDTA) and eluted with Buffer BXT (Buffer W with 50 mM biotin). Protein concentrations were measured on a Qubit Flex fluorometer.

Bio-layer interferometry. All measurements were made on a Gator Plus instrument (Gator Bio) at 30 ◦Cwith shaking at 1000 rpm. Unless otherwise noted, all dilutions and buffer steps used Gator K buffer.

Competitive disruption assay, screen. Soluble His-tagged Fc εRI α ectodomain (Sino 13193-H08H-100) was loaded onto anti-His XT probes (Gator Bio) for 90 s, followed by 30 s of buffer. IgE (Abcam AB65866-1008) reconstituted to 100 nM was then bound to the receptor for 120 s, followed by 30 s of buffer. The probe was then moved for 120 s into the cell-free expression reaction of each design (diluted 3 × in PBS).

Competitive disruption assay, purified variants. Eight active designs, alongside sdAb026 and Y112A, were purified as above. They were tested at 25, 50, 100, 200 and 400 nM with omalizumab (MedChem-29 Express HY-P9950) in duplicate. The loading sequence was as in the screen, with phase boundaries at 0, 30, 90, 120, 150 and 180 s, followed by a 600 s exposure to the purified test article.

Analysis of the disruption traces. Each trace was decomposed into stacked receptor, IgE and VHH components. Let R(t) be the raw response and R0 the mean of the last ten reads before receptor loading. The receptor component was fitted on the buffer step after receptor loading, [t1, t 2], and extrapolated over the rest of the trace:

R(t) − R0 = AR e−kR(t−t1), t1 ≤ t ≤ t2. (5) The IgE component is the response above it, E(t) = R(t) − R0 − AR e−kR(t−t1), and its decay in the buffer step after IgE loading, [t3, t 4], was fitted in the same way, E(t) = AE e−kE(t−t3), giving ˆE(t), the IgE expected without a disruptor. On exposure to the VHH from t4, the rise to the first maximum of the trace, at tp, was assigned to VHH and non-specific binding and carried forward as the constant

V = E(tp) − ˆE(tp), (6) so that any decline below it reports IgE leaving the receptor. The fraction of IgE remaining, the IgE lost at the end of the exposure tend , and the apparent IgE off-rate kapp are then

f (t) = E(t) − V

ˆE(tp) , L = 1 − f (tend ) = E(tp) − E(tend )ˆE(tp) , E(t) − V = A e −kapp (t−tp), tp ≤ t ≤ tend ,

(7) with f (tp) = 1 . In the screen, t1, . . . , t 4 = 150 , 180 , 300 , 330 s and tend = 450 s. A design was called disruptive if

kapp > max

c∈ { E108A ,D110A ,Y112A ,sfGFP }

k(c)app . (8) For the purified variants the same decomposition was used, adapted to the missing final buffer step, and the fraction of IgE bound was read as f (t) at 600 s of exposure.

Binding assays. VHHs were captured on Strep-Tactin XT probes (Gator Bio) for 60 s from cell-free reac-tion mix, followed by 60 s of buffer, and exposed to IgE (Abcam AB65866-1008). The disruptor designs were exposed to 100 nM IgE for 120 s with 180 s of dissociation; the anchor designs were exposed to 560 nM for 180 s with 180 s of dissociation. Each trace was referenced to the mean of the last ten reads of the baseline before association, and Rmax is the peak response in the association phase. Each phase was fitted separately with a single exponential, R(t) = C + A e −kobs t for association and R(t) = R0 e−koff t for dissociation over the complete phase, with t measured from the start of each phase. For 1:1 binding at a single analyte concentration, kobs = kon [IgE] + koff , so the apparent association rate constant is

kon = kobs − koff

[IgE] , (9) and KD = koff /k on , reported only where both rates were resolved. Where koff falls at the lower bound of the fit it is reported as not resolved.

E Extended characterization of disruptor designs

Table 4 collects, for every disruptor design and control, the results of the initial disruptor screen and a binding assay against free IgE (Appendix D). From the competitive disruption assay, we report the fraction of receptor-bound IgE lost by the end of the exposure (Equations (7) and (8)). From the binding assay of free IgE at 100 nM, we report the apparent rates kobs , koff and kon (Equation (9)). 30 Table 4. Disruption and binding assays of the disruptor designs and controls. Variants are called active in the disruption assay if they have an apparent IgE off-rate faster than that of every negative control; our primary endpoint (reported in the main text) is the percentage of receptor-bound IgE removed by the end of the exposure. For the binding assay, we report kobs and koff from single-exponential fits of the association and dissociation phases, and kon from Equation (9). n.r., not resolved (rate at the bound of the fit). Scope and method as in Section 3.2; Subs., substitutions relative to sdAb026. Designs D1–D91 are ordered by IgE lost, highest first, and V1–V8 mark the purified variants of Figure 5F–K.

Disruption assay Binding assay Design Scope Method Subs. Active IgE lost

(%)

kobs

(10 −3 s−1)

koff

(10 −6 s−1)

kon

(10 3 M−1s−1)sdAb026 control 36 13.38 89 133 E108A control 27 11.84 159 117 D110A control 16 11.40 137 113 Y112A control 10 11.49 571 109 D1 (V3) CDR2 redesign 6 ✓ 43 13.63 83 135 D2 (V1) Cε3 × CDR1,3 tempered redesign 2 ✓ 43 20.07 79 200 D3 (V5) CDR2 redesign 7 ✓ 39 11.90 86 118 D4 (V6) CDR2 redesign 6 ✓ 38 13.37 86 133 D5 (V7) CDR2 redesign 6 ✓ 37 15.67 108 156 D6 (V2) Cε4 × CDR3 inverse folding 2 ✓ 37 19.42 92 193 D7 CDR2 redesign 6 ✓ 33 13.47 79 134 D8 Cε3 × CDR1,3 tempered redesign 2 ✓ 32 15.16 89 151 D9 CDR2 redesign 5 ✓ 31 13.42 67 134 D10 CDR2 redesign 6 ✓ 31 13.78 106 137 D11 CDR2 redesign 5 ✓ 31 9.69 95 96 D12 CDR2 redesign 5 ✓ 30 14.41 60 144 D13 (V4) Cε3 × CDR1,3 inverse folding 2 ✓ 30 19.54 82 195 D14 (V8) Cε3 × CDR1,3 inverse folding 2 ✓ 29 20.44 111 203 D15 CDR2 redesign 5 ✓ 29 12.53 86 124 D16 CDR2 redesign 6 ✓ 28 11.87 85 118 D17 Cε3 × CDR1,3 tempered redesign 2 ✓ 28 14.94 90 149 D18 Cε4 × CDR3 tempered redesign 2 ✓ 28 14.21 87 141 D19 CDR2 inverse folding 8 ✓ 28 11.85 119 117 D20 CDR2 redesign 6 28 14.11 77 140 D21 CDR2 redesign 6 ✓ 27 13.51 78 134 D22 Cε3 × CDR1,3 inverse folding 3 27 16.84 115 167 D23 Cε3 × CDR1,3 inverse folding 3 26 15.56 123 154 D24 Cε3 × CDR1,3 tempered redesign 3 26 18.09 98 180 D25 Cε4 × CDR3 inverse folding 3 26 20.55 98 204 D26 Cε3 × CDR1,3 tempered redesign 3 26 13.69 105 136 D27 CDR2 inverse folding 8 26 13.94 97 138 D28 Cε4 × CDR3 tempered redesign 2 26 13.58 116 135 D29 CDR2 inverse folding 8 25 15.66 103 156 D30 CDR2 redesign 6 25 9.77 138 96 D31 CDR2 redesign 6 24 13.62 73 135 D32 CDR2 inverse folding 8 24 16.58 88 165 D33 Cε4 × CDR3 inverse folding 3 23 13.67 130 135 D34 Cε3 × CDR1,3 inverse folding 3 23 20.53 98 204 D35 CDR2 redesign 5 23 12.66 85 126 D36 CDR2 redesign 6 23 14.38 72 143 D37 CDR2 inverse folding 8 22 15.34 109 152 D38 CDR2 inverse folding 8 22 15.94 90 159 D39 Cε3 × CDR1,3 tempered redesign 3 22 18.97 87 189 D40 Cε3 × CDR1,3 inverse folding 3 22 19.46 102 194 D41 Cε3 × CDR1,3 tempered redesign 3 22 18.49 62 184 D42 Cε4 × CDR3 tempered redesign 3 21 14.32 86 142 D43 CDR2 inverse folding 8 21 14.76 98 147 D44 Cε3 × CDR1,3 tempered redesign 2 21 17.32 98 172 D45 Cε3 × CDR1,3 inverse folding 3 21 15.28 149 151 D46 CDR2 inverse folding 8 21 15.23 86 151 D47 CDR2 inverse folding 8 21 15.21 110 151 D48 Cε3 × CDR1,3 inverse folding 2 20 17.81 125 177 D49 Cε4 × CDR3 tempered redesign 2 20 14.54 126 144 D50 CDR2 redesign 5 19 9.36 107 93 D51 Cε4 × CDR3 tempered redesign 2 19 9.52 246 93 D52 Cε3 × CDR1,3 tempered redesign 4 17 16.28 124 162 D53 Cε4 × CDR3 inverse folding 3 17 19.34 53 193 D54 Cε3 × CDR1,3 tempered redesign 2 17 17.36 94 173 D55 Cε4 × CDR3 tempered redesign 2 17 14.42 116 143 D56 Cε3 × CDR1,3 tempered redesign 2 17 12.52 101 124 D57 CDR2 inverse folding 8 16 10.26 129 101 D58 CDR2 redesign 6 16 8.83 102 87 D59 Cε3 × CDR1,3 inverse folding 3 16 19.99 85 199

31 Disruption assay Binding assay Design Scope Method Subs. Active IgE lost

(%)

kobs

(10 −3 s−1)

koff

(10 −6 s−1)

kon

(10 3 M−1s−1)D60 Cε4 × CDR3 inverse folding 3 16 18.69 91 186 D61 Cε3 × CDR1,3 inverse folding 3 16 17.15 84 171 D62 CDR2 inverse folding 8 16 15.64 107 155 D63 Cε3 × CDR1,3 inverse folding 4 16 14.15 196 140 D64 CDR2 redesign 6 16 12.02 42 120 D65 CDR2 inverse folding 8 15 13.51 86 134 D66 CDR2 inverse folding 8 15 15.96 65 159 D67 Cε4 × CDR3 inverse folding 4 14 14.52 566 140 D68 Cε3 × CDR1,3 inverse folding 3 14 19.80 89 197 D69 CDR2 inverse folding 8 13 15.16 91 151 D70 Cε4 × CDR3 tempered redesign 3 13 12.94 295 126 D71 Cε4 × CDR3 tempered redesign 2 13 11.58 309 113 D72 Cε4 × CDR3 inverse folding 3 12 14.35 305 140 D73 Cε3 × CDR1,3 tempered redesign 4 12 18.25 39 182 D74 CDR2 inverse folding 8 12 11.70 67 116 D75 Cε4 × CDR3 tempered redesign 3 12 12.11 1156 110 D76 Cε4 × CDR3 inverse folding 3 12 28.52 2919 256 D77 CDR2 inverse folding 8 11 15.39 62 153 D78 Cε4 × CDR3 inverse folding 2 11 16.20 222 160 D79 Cε4 × CDR3 tempered redesign 2 11 14.51 71 144 D80 Cε4 × CDR3 inverse folding 3 11 14.61 207 144 D81 CDR2 inverse folding 8 10 18.05 101 179 D82 CDR2 inverse folding 8 10 18.06 n.r. 181 D83 Cε3 × CDR1,3 inverse folding 2 10 28.20 n.r. 282 D84 Cε3 × CDR1,3 tempered redesign 3 9 25.87 n.r. 259 D85 Cε4 × CDR3 inverse folding 4 9 29.78 3556 262 D86 Cε4 × CDR3 tempered redesign 3 9 9.48 294 92 D87 Cε4 × CDR3 inverse folding 3 8 12.87 279 126 D88 Cε4 × CDR3 tempered redesign 3 8 14.62 99 145 D89 Cε4 × CDR3 inverse folding 4 8 19.87 3036 168 D90 CDR2 inverse folding 8 7 18.87 n.r. 189 D91 Cε4 × CDR3 tempered redesign 2 6 14.03 97 139