BIOTENDER
深度 · 2026.06.05
Chai Discovery · 大药企验证期

Chai-3 发布!直接导致辉瑞下场合作!
AI 抗体设计进入大药企验证期

BioTender 编辑部 · 源材料:Forbes(Amy Feldman, 06.04)、BusinessWire 通稿(Pfizer 06.04、Lilly 01.08、Chai-2 2025.06.30)、bioRxiv 2025.07 论文
Forbes Daily Cover: Why Pfizer And Eli Lilly Are Betting On This $1.3 Billion AI Drug Discovery Startup
Forbes Daily Cover, 2026.06.04 ─ Amy Feldman 撰稿。Chai 此时估值 $1.3B,文末披露在谈 $3.4B 新一轮。

6 月 4 日,Pfizer 与 Chai Discovery 签约的官方通稿和 Forbes 独家同步上线。这一天发生了三件事:

把这三件事连起来看,AI 抗体设计今年从"AI 公司各自演示"明确切换到了"大药企买单"的相位。

Chai-3 的"发布"方式很微妙

严格说,6.4 这天 Chai-3 没有"发布"。没有 paper,没有 preprint,没有代码,没有 model weights。Pfizer 通稿原话是 previously undisclosed model──一个之前未披露的模型。

Pfizer × Chai BusinessWire 通稿原文(2026.06.04)。原句:“Pfizer will be among the first pharmaceutical partners to secure access to Chai's previously undisclosed model, Chai-3.” 注意 “among the first”──暗示在 Pfizer 之前可能已有客户。
Pfizer × Chai BusinessWire 通稿原文(2026.06.04)。原句:“Pfizer will be among the first pharmaceutical partners to secure access to Chai's previously undisclosed model, Chai-3.” 注意 “among the first”──暗示在 Pfizer 之前可能已有客户。

公司放出来的全部数字只有三个:

  1. 命中率翻倍:相对 Chai-2 的 16–20%,Chai-3 翻倍到 ~32–40%。
  2. 亲和力紧 100×:抗体结合靶点的亲和力相对前代提升约两个数量级。
  3. 约 50% 的设计,亲和力达到已上市抗体药的水平。
Forbes 原文。两条关键陈述高亮:“in roughly half of cases, the molecules generated by Chai-3 bind to their targets as tightly as approved drugs”,以及 Meier 关于多特异性的展望──“In the future, we will modulate targets in more powerful ways.”
Forbes 原文。两条关键陈述高亮:“in roughly half of cases, the molecules generated by Chai-3 bind to their targets as tightly as approved drugs”,以及 Meier 关于多特异性的展望──“In the future, we will modulate targets in more powerful ways.”

这三个数字如果成立,意味着 Chai-3 把 affinity maturation──抗体药开发中一道独立的、传统上要花几个月的湿实验工序──压进了模型的 forward pass。

Chai-2 的论文(Zero-shot antibody design in a 24-well plate, bioRxiv 2025.07)当时震动业界,是因为它把 de novo 抗体设计命中率从 <0.1% 拉到了 ~20% 级别,给一个抗原 + epitope,从零生成全长 IgG1 单抗,不需要起始 binder。

Chai-2 上市通稿(BusinessWire 2025.06.30)。原句:“a remarkable antibody design hit rate close to 20%”,对比传统方法 “hit rates below 0.1%”。注意 Forbes 这次写的 Chai-2 数字是 16%,与通稿的 ~20% 来自不同 denominator。
Chai-2 上市通稿(BusinessWire 2025.06.30)。原句:“a remarkable antibody design hit rate close to 20%”,对比传统方法 “hit rates below 0.1%”。注意 Forbes 这次写的 Chai-2 数字是 16%,与通稿的 ~20% 来自不同 denominator。

但 Chai-2 给出的抗体大多在 nanomolar 区间,离上市药 sub-nM 到 picomolar 的标准还差一档,仍然要走传统优化──phage display、yeast display、定向进化──把亲和力磨上去。这道工序在抗体药开发里通常要花几个月,需要专门的湿实验团队,是 R&D 流程里最贵、最不可省的环节之一。

Forbes 对 Chai-2 的描述。“The model was able to spin out antibody designs from scratch 16% of the time, compressing months of work in a wet lab to two weeks.” 但下一句明说:分子能 bind 上,“they didn't always do it well”──仍要走传统优化。
Forbes 对 Chai-2 的描述。“The model was able to spin out antibody designs from scratch 16% of the time, compressing months of work in a wet lab to two weeks.” 但下一句明说:分子能 bind 上,“they didn't always do it well”──仍要走传统优化。

换句话说:Chai-2 把"找苗头"从月级别压到周级别,但下游"磨亲和力"这一段,AI 没动。Chai-3 直接把这一档补上了。

但披露形式本身是个信号。Chai-2 走的是学术路径:bioRxiv 论文 + 完整方法 + 公开数据集 + 第三方 critique(Engin Yapici 在 Medium 上把方法学缺陷一一拆过,结论仍是 holds up)。Chai-3 走的是商业路径:通过 Pfizer 通稿"被披露",所有数字都来自公司单方面说法。这是个明确的转向──Chai 不再依赖学术 channel 给自己的模型背书,而是把 MNC 买单本身当作背书。

学术验证和商业验证的优先级,被切换了。

Pfizer 为什么签──以及标题没说的那一半

Forbes 文中那句话值得 underline:是 Chai-3 让 Pfizer 决定签约。Dent 自己也提到,"That got the Pfizer team really excited"。

Forbes 原文。“Chai-3, which it claims is vastly superior to the Chai-2 one that put it on the map. 'That got the Pfizer team really excited,' says Dent.” 注意 “it claims”──Forbes 在做转述,没有为这个 superior 的判断背书。
Forbes 原文。“Chai-3, which it claims is vastly superior to the Chai-2 one that put it on the map. 'That got the Pfizer team really excited,' says Dent.” 注意 “it claims”──Forbes 在做转述,没有为这个 superior 的判断背书。
Forbes 给 Chai-3 与 Pfizer 关系的核心一句:“It was Chai-3 that convinced Pfizer to sign on.” 同段亦给出 100× 数据:“bind 100 times more tightly to their intended therapeutics targets, according to the company.”──“according to the company” 是 Forbes 留的免责。
Forbes 给 Chai-3 与 Pfizer 关系的核心一句:“It was Chai-3 that convinced Pfizer to sign on.” 同段亦给出 100× 数据:“bind 100 times more tightly to their intended therapeutics targets, according to the company.”──“according to the company” 是 Forbes 留的免责。

但要在背景里补一行:Mikael Dolsten──前 Pfizer 全球研发总裁,今年从 Pfizer 退休──现在是 Chai Discovery 董事会成员。Forbes 文章里还引了他对 Chai 商业模式的判断。

Mikael Dolsten。Forbes 这次的描述是 “retired from Pfizer as president of worldwide R&D and is now on Chai's board of directors”。Pfizer 历任 CSO 与全球研发总裁,2025 年从 Pfizer 退休加入 Chai 董事会。
Mikael Dolsten。Forbes 这次的描述是 “retired from Pfizer as president of worldwide R&D and is now on Chai's board of directors”。Pfizer 历任 CSO 与全球研发总裁,2025 年从 Pfizer 退休加入 Chai 董事会。

也就是说,Pfizer 和 Chai 不是"两个陌生组织最近撞上",是"老朋友在最近的产品上达成商业落地"。Dolsten 在 Chai 董事会内部能看到 Chai-3 的早期数据,他对 Pfizer 的影响力又能直接传导到采购决策。"Chai-3 直接导致 Pfizer 下场"是事实的一半;另一半是 Chai 在 Pfizer 内部本来就有一条预铺好的人脉通道。

这点不削弱 Chai-3 的技术意义,但会改写"标志性合作"的解读。一个由前研发总裁牵线的合作,复制成本极低;如果 Chai-3 真有那么强,没有 Dolsten 这层关系的药企本来也会签。所以接下来真正要观察的是:第二个、第三个没有"内部线"的 MNC 签不签。

另一个产品周期的细节:Pfizer 这次签的是 Chai-3 访问权;Lilly 1 月签的合作通稿措辞是"access to our core models"(当时旗舰 = Chai-2)+ Lilly 自有数据上的定制训练。两次合作的代际是错开的。

Lilly × Chai 通稿(BusinessWire 2026.01.08),Meier 引语。“Beyond providing access to our core models, training custom models trained on Lilly's data presents the opportunity to expand the boundaries of AI-enabled early-stage drug discovery and development.” 核心模型访问 + 定制训练──标准合同结构。
Lilly × Chai 通稿(BusinessWire 2026.01.08),Meier 引语。“Beyond providing access to our core models, training custom models trained on Lilly's data presents the opportunity to expand the boundaries of AI-enabled early-stage drug discovery and development.” 核心模型访问 + 定制训练──标准合同结构。

Chai 已经在走"按模型代际收费"的产品周期了。这是一个非常成熟的 SaaS 模式雏形,搬到 biotech 还不常见。

"大药企验证期"指的是什么

按 PitchBook 数据(Forbes 援引),2025 全球 AI 制药领域融资 $11.4B,比 2024 年的 $5.6B 翻一倍多;2026 年到目前已经 $5.5B,Isomorphic 一家 5 月就拿了 $2.1B

Forbes 引 PitchBook 数据:2025 年全球 AI 制药融资 $11.4B,2024 年 $5.6B,2026 YTD $5.5B;Isomorphic(Alphabet 衍生)5 月单轮 $2.1B。Dent:“we are in a completely different universe than we were a year ago.”
Forbes 引 PitchBook 数据:2025 年全球 AI 制药融资 $11.4B,2024 年 $5.6B,2026 YTD $5.5B;Isomorphic(Alphabet 衍生)5 月单轮 $2.1B。Dent:“we are in a completely different universe than we were a year ago.”

但融资规模不是验证。融资是 VC 在押注未来;验证是大药企自己掏钱用模型。当前公开的关键合作:

Chai 自己披露在谈 15+ 家药企,今年还希望签更多。

Forbes 原句:“Chai is in talks with more than 15 additional pharma companies, and hopes to sign more deals this year.” 这是 12 个月内最重要的观察指标──其中是否有“没有内部线”的 MNC 公开签约。
Forbes 原句:“Chai is in talks with more than 15 additional pharma companies, and hopes to sign more deals this year.” 这是 12 个月内最重要的观察指标──其中是否有“没有内部线”的 MNC 公开签约。

如果其中 1–2 家在年内落地,这就形成了一个"AI 抗体设计被 MNC 采购"的标准动作。

"验证期"的精确含义不是科学验证,是商业模式验证。 Lilly 自己的 CDIO Diogo Rau 3 月明确说过:第一个 AI 设计的药要到 mid-2030s 甚至 late-2030s 才能上市。也就是说:
Forbes 原文。Lilly CDIO Diogo Rau 3 月接受 Forbes 采访:“it would be 'mid-2030s, if not late-2030s' before any of its AI developed medicines are on the market. 'It's a big bet on the future,' he said.” 同时注意末句:Pfizer 拒绝就此合作发表评论。
Forbes 原文。Lilly CDIO Diogo Rau 3 月接受 Forbes 采访:“it would be 'mid-2030s, if not late-2030s' before any of its AI developed medicines are on the market. 'It's a big bet on the future,' he said.” 同时注意末句:Pfizer 拒绝就此合作发表评论。

这两件事之间有一个 5–10 年的时间差。在这个时间差里,AI 制药公司要么能不断释放中间数据让市场保持信心,要么会经历至少一轮估值修正。Chai 当前用"模型代际 + 合作签约"的节奏在填这个时间差,但能填多久取决于 Chai-3 之后还能不能有 Chai-4。

接下来 12 个月看什么

三件事:

  1. Chai-3 论文出不出──商业披露之后有没有学术披露。如果一直没有,市场会开始问数字的可信度。
  2. Pfizer / Lilly 项目的中间数据放不放──大药企拿到模型权之后,内部跑出来的有没有任何外部信号。
  3. 15+ 家在谈的有没有 1–2 家落地公开签约──尤其要看有没有跟 Chai 没有"内部线"的 MNC。

三件都发生,AI 抗体设计的叙事就从"商业验证"过渡到"科学验证";只发生第三件,那就是 hype cycle 的另一个 peak。

模型在变强是事实;模型在变强到能改写药企开发流程,还需要论文和数据来回答。

6.4 这天的故事,是商业先一步,到了。

BIO·TENDER · END OF DISPATCH