礼来的 AI 下注在赌什么 What Lilly Is Really Buying in AI-Bio Deals

2026-06-28 · ai-bio, pharma, drug-discovery, gene-editing, dealmaking

2026 年 3 月到 5 月，AI-bio 的融资和合作窗口突然变得很集中：礼来先与 Insilico Medicine 达成最高约 27.5 亿美元的全球研发合作，又与 Profluent 达成最高 22.5 亿美元的 AI 设计重组酶合作；同期，曾与礼来签过最高 17 亿美元合作的 DeepMind 系 Isomorphic Labs 完成 21 亿美元 B 轮融资。

把这些事件放在一起看，真正的问题不是“大药厂是否相信 AI 已经能稳定造药”，而是资本正在为药物研发链条里的哪一段提前买单。

礼来的钱先买的是入口，不是临床胜利

Insilico 这笔交易里，明确的 upfront 是 1.15 亿美元；27.5 亿美元是潜在总额，主要来自后续研发、监管和商业 milestone，并附带销售 royalty。交易对象包括一组仍处在临床前阶段的口服小分子项目，以及多个由礼来选择靶点的合作研发项目。

Profluent 这笔更靠前。Profluent 会获得 upfront 和承诺研发经费，但 22.5 亿美元上限主要取决于 development 和 commercial milestone。合作内容不是成熟药物资产，而是 AI 设计的 site-specific recombinase，用于遗传医学里的大尺度 DNA 编辑。

所以，headline number 不能等同于“今天到账的钱”。它更像一种风险分配结构：礼来用较小的确定性成本买入平台入口，把真正的大额支付留给后续验证。

AI 被定价的位置在临床前

这几笔钱都没有直接买“临床成功”。Insilico 更接近小分子资产和靶点发现平台，Profluent 更接近基因编辑工具平台，Isomorphic 则代表 AI drug design engine 本身被资本市场重新定价。

它们共同指向一个变化：大药厂开始愿意为临床前阶段的效率、搜索空间和候选项目数量付费。AI 的价值暂时不是替代临床试验，而是改变进入临床前的方式：更快提出候选分子，更系统地优化蛋白或编辑器，更早形成可以被大药厂评估、选择和锁定的项目。

Profluent 这笔是 modality platform 的下注

Profluent 交易值得单独看，因为它不是普通的小分子 AI discovery 故事。传统 CRISPR、base editing、prime editing 都已经打开了基因编辑空间，但大段 DNA 插入、替换和更复杂的 genome rewriting 仍然很难。

Profluent 的主张是，用生成式模型设计 recombinase，去解决更大尺度、更定点的 DNA 编辑问题。如果这条路成立，价值不是单个项目，而是一组疾病和一组编辑场景。

但这里的风险也更前置。AI 设计出的编辑器在细胞或早期实验中有功能，不等于它能在人体内安全、有效、可递送、可制造。OpenCRISPR-1 的 Nature 论文证明 AI 生成的基因编辑器可以在实验系统中表现出功能，但它不是临床疗效证明。

“GPT-1.5” 是行业叙事，不是临床证据

Profluent CEO 兼联合创始人 Ali Madani 在 2026 年 6 月 18 日 Air Street Capital 的访谈中，把 AI x biology 形容为仍处在 “GPT-1.5 era”。他也在 X 上预告这场访谈，提到会讨论为什么他这样描述 biology 当前阶段。

这个比喻有启发性，但要把它放在正确位置：它首先是创业者对行业阶段的判断，不是科学或临床证明。它真正说明的是，AI-bio 可能还很早，但已经早到足以让大药厂付钱买入口。

换句话说，礼来不是等到“GPT-5 级别”的 AI-bio 完全成熟才行动。只要模型已经能产生可测试、可合作、可授权的候选工具或候选分子，大药厂就会提前锁定选择权。

milestone 暗示风险还没有消失

milestone 的存在本身就在提醒我们：风险没有消失，只是被推迟结算了。AI 可以降低搜索成本，提高设计迭代速度，扩大候选空间，但它不能自动解决人体生物学、递送、免疫原性、长期安全性、制造一致性、临床终点和监管接受度。

所以接下来真正要看的不是 headline number，而是 milestone 能否被触发：这些 AI-originated 或 AI-enabled 项目能不能持续进入 IND、Phase 1、Phase 2，并在人体读数里显示出比传统路径更好的速度、质量或成功率。

我的判断

礼来的 AI 下注，本质上是在用 upfront 和研发经费买临床前入口，用 milestone 管住下游风险。Insilico 代表小分子与靶点发现，Profluent 代表基因编辑工具平台，Isomorphic 代表 AI drug design engine 的资本定价。

这说明 AI for Science 已经进入大药厂 BD 预算，不再只是研究展示。但它还没有证明药物研发的最终成功率已经被系统性改写。更稳的结论是：AI 正在改变临床前资产形成的速度和广度；最终是否改变药物研发经济学，还要等 clinical readout 结算。

主要信源

From March to May 2026, the AI-bio financing and partnership window became unusually dense: Lilly entered a global R&D collaboration with Insilico Medicine worth up to about $2.75 billion, then signed a partnership with Profluent worth up to $2.25 billion for AI-designed recombinases. In the same period, Isomorphic Labs, the DeepMind-linked AI drug design company that had previously signed a potential $1.7 billion collaboration with Lilly, raised a $2.1 billion Series B.

Read together, these events do not answer the question, “Does big pharma now believe AI can reliably make drugs?” The sharper question is: which part of the drug development chain is capital starting to price in advance?

Lilly is buying entry points, not clinical victories

In the Insilico deal, the clearly disclosed upfront payment is $115 million. The larger $2.75 billion headline figure is a potential total, mainly tied to future development, regulatory, and commercial milestones, with royalties on sales. The collaboration includes a set of still-preclinical oral small-molecule programs and several discovery programs around Lilly-selected targets.

The Profluent deal sits even earlier in the chain. Profluent will receive upfront and committed research funding, but the potential $2.25 billion value depends mainly on development and commercial milestones. The collaboration is not built around a mature drug asset. It is built around AI-designed site-specific recombinases for large-scale DNA editing in genetic medicine.

The headline number therefore should not be read as cash paid today. It is better understood as a risk-allocation structure: Lilly pays a smaller certain cost to secure platform access, while the largest payments remain contingent on later validation.

AI is being priced before the clinic

None of these financings directly buys “clinical success.” Insilico sits closer to small-molecule assets and target-discovery platforms. Profluent sits closer to a genome-editing modality platform. Isomorphic represents the capital-market pricing of an AI drug design engine itself.

Together, they point to a specific shift: big pharma is becoming willing to pay for speed, search-space expansion, and candidate generation before the clinic. The near-term value of AI is not replacing clinical trials. It is changing how projects enter preclinical development: proposing candidates faster, optimizing proteins or editors more systematically, and forming programs that pharma can evaluate, option, and lock up earlier.

Profluent is a bet on a modality platform

The Profluent deal deserves separate attention because it is not another ordinary small-molecule AI discovery story. CRISPR, base editing, and prime editing have already expanded the gene-editing toolbox, but large DNA insertions, replacements, and more complex genome rewriting remain difficult.

Profluent’s claim is that generative models can design recombinases that enable larger-scale, more site-specific DNA editing. If that path works, the value is not a single program. It is a set of diseases and editing contexts.

The risk, however, is also earlier. An AI-designed editor showing function in cells or early experimental systems does not mean it will be safe, effective, deliverable, and manufacturable in humans. The OpenCRISPR-1 Nature paper supports the idea that AI-generated gene editors can function in experimental systems, but it is not evidence of clinical efficacy.

”GPT-1.5” is an industry narrative, not clinical evidence

Profluent CEO and co-founder Ali Madani described AI x biology as still being in a “GPT-1.5 era” in a June 18, 2026 Air Street Capital interview. He also previewed that interview on X, noting that it would discuss why he characterizes the current stage of biology this way.

The metaphor is useful, but it needs to be placed correctly. It is first an entrepreneur’s view of the field’s stage, not scientific or clinical proof. What it does help explain is that AI-bio may still be early, but it is already early in a commercially actionable way.

In other words, Lilly is not waiting for a fully mature “GPT-5” version of AI-biology before acting. Once models can generate testable, partnerable, licensable tools or candidate molecules, big pharma has a reason to lock in option value.

Milestones show that the risk has not disappeared

The presence of milestones is itself the point. The risk has not vanished; it has been moved downstream. AI can reduce search cost, speed design iterations, and expand candidate space, but it does not automatically solve human biology, delivery, immunogenicity, long-term safety, manufacturing consistency, clinical endpoints, or regulatory acceptance.

The next signal to watch is therefore not the headline number. It is whether milestones are actually triggered: whether AI-originated or AI-enabled programs keep moving into INDs, Phase 1, and Phase 2, and whether human readouts show better speed, quality, or success rates than conventional discovery paths.

My read

Lilly’s AI bet is best understood as upfront and research funding used to buy preclinical entry points, with milestones used to control downstream risk. Insilico represents small molecules and target discovery, Profluent represents a genome-editing platform, and Isomorphic represents capital-market pricing of the AI drug design engine.

This means AI for Science has entered big pharma’s BD budget; it is no longer only a research demo. But it has not yet shown that the final success rate of drug development has been systematically rewritten. The more defensible conclusion is narrower: AI is changing the speed and breadth of preclinical asset formation. Whether it changes the economics of drug development still has to be settled by clinical readouts.