基于 Metix AI 8.6 亿+ 全球人才库，对 35 家 AI 药物发现公司与大药企 AI 组做全量画像：会折叠蛋白的 ML 人在谁手里、干实验与湿实验双背景的稀缺画像在哪、AlphaFold 人脉网外溢去向、大药企 AI 组的工程师层组织拼图，以及计算生物领域的华人维度分析。Drawing on Metix AI's 860 million+ global talent pool, a full-coverage profile of 35 AI-drug-discovery companies and big-pharma AI groups: who holds the ML talent that can fold proteins, where the scarce dry-lab × wet-lab crossover profiles sit, where the AlphaFold network is spilling over to, an engineer-level org map of big-pharma AI groups, and a Chinese-talent lens on computational biology.
以下数字为 Metix AI 数据库口径（数据截至 2026 年上半年），统计对象 = 当前在职于 35 家目标公司、坐标美国、英国、瑞士、丹麦、法国、德国、加拿大、瑞典的人才。大药企口径取其 AI/计算子集。The figures below follow Metix AI database methodology (data through H1 2026). The population = talent currently employed at the 35 target companies and based in the United States, United Kingdom, Switzerland, Denmark, France, Germany, Canada, or Sweden. For big pharma, the scope is limited to the AI/computational subset.
15 家大药企 AI/计算组合计 3,332 人，是 20 家 AI-native 创业公司（1,042 人）的 3.2 倍。单一最大 AI 组是 Genentech（442 人，含 Prescient Design 抗体设计团队）。对买方的含义：想批量找 AI 制药人才，主战场是 incumbent 药企的计算组，而非只盯明星初创。The AI/computational groups at 15 big-pharma companies total 3,332 people, 3.2x the 1,042 across the 20 AI-native startups. The single largest AI group is Genentech (442 people, including the Prescient Design antibody-design team). For buyers: to source AI-pharma talent at scale, the main battleground is the computational groups inside incumbent pharma, not just the marquee startups.
55.6% 的计算人才同时具备湿实验（生物/生化）与干实验（ML/编程）背景，说明这个领域天然跨学科。但真正稀缺的是蛋白结构/设计方向的专精者：全样本仅55.6% of computational talent has both a wet-lab background (biology/biochemistry) and a dry-lab one (ML/programming), confirming this field is inherently interdisciplinary. But the truly scarce profile is the specialist in protein structure/design: across the full sample, only 125 人（2.9%）125 people (2.9%)明确从事蛋白结构预测或蛋白/抗体设计。这 125 人是 AlphaFold 时代最被争抢的核心画像。 work explicitly on protein structure prediction or protein/antibody design. These 125 are the most fought-over core profile of the AlphaFold era.
可见档案中有 DeepMind / Isomorphic 履历的 56 人，其中 49 人现在 Isomorphic（DeepMind 整建制分拆），少数外溢到 Xaira、Latent Labs、insitro。Isomorphic 现任技术人才任期中位仅 11 个月、54.1% 入职不足一年，是高速扩张中的「新军」。Of the visible profiles, 56 have a DeepMind / Isomorphic history; 49 of them are now at Isomorphic (a wholesale DeepMind spinout), with a few spilling over to Xaira, Latent Labs, and insitro. Isomorphic's current technical talent has a median tenure of just 11 months, and 54.1% have been there under a year, a fast-scaling "new guard."
不同于纯 AI（湾区独大），AI 制药人才分布在美国各地（571 湾区 / 478 波士顿）、英国（Ast 等）、瑞士（257 人，Roche/Novartis）、丹麦（169 人，Novo Nordisk）。跨地域、跨时区寻访是这个赛道的常态。Unlike pure AI (where the Bay Area dominates), AI-pharma talent is spread across the US (571 Bay Area / 478 Boston), the UK (AstraZeneca and others), Switzerland (257 people, Roche/Novartis), and Denmark (169 people, Novo Nordisk). Cross-region, cross-time-zone sourcing is the norm in this space.
以下基于 2025-2026 公开信源逐条核实（完整来源见研究底稿），只保留影响人才决策的事实。金额为报道口径。The points below were verified one by one against 2025-2026 public sources (full sourcing in the research memos), keeping only the facts that affect talent decisions. Dollar figures follow the reported amounts.
2025 年 AI 药物发现 VC 约 110 亿美元（348 轮，DealForma 口径）。Isomorphic 2026-05 再融 21 亿美元（生物科技史上第二大轮）；Chai Discovery 2025-12 融 1.3 亿美元、估值 13 亿；Xaira 出道即 10 亿美元、估值约 40 亿；Generate Biomedicines 与 Eikon 2026-02 双双 IPO（合计募资约 7.8 亿）。资金面明确回暖。AI drug discovery drew roughly $11 billion in VC in 2025 (348 rounds, per DealForma). Isomorphic raised another $2.1 billion in May 2026 (the second-largest round in biotech history); Chai Discovery raised $130 million in December 2025 at a $1.3 billion valuation; Xaira launched with $1 billion and a ~$4 billion valuation; Generate Biomedicines and Eikon both IPO'd in February 2026 (raising roughly $780 million combined). The funding environment is clearly recovering.
结构预测（AlphaFold 3 闭源 → Boltz-1/2、Chai-1、ESM3 开源爆发）→ de novo 蛋白/抗体设计（RFdiffusion、Chai-2 命中率近 20%、较旧法约 100 倍）→ 分子生成（Boltz-2 联合建模结构+亲和力，逼近 FEP 但快千倍）→ virtual cell / 表型世界模型（Arc State、CZI、Xaira）。每一级都对应一类新的稀缺人才画像。Structure prediction (closed-source AlphaFold 3 → an open-source surge of Boltz-1/2, Chai-1, ESM3) → de novo protein/antibody design (RFdiffusion, Chai-2 with a hit rate near 20%, roughly 100x older methods) → molecular generation (Boltz-2 jointly modeling structure + affinity, approaching FEP but a thousand times faster) → virtual cell / phenotypic world models (Arc State, CZI, Xaira). Each stage maps to a new scarce talent profile.
Anthropic 2026-04 以 4 亿美元收购 Coefficient Bio（团队来自 Genentech Prescient Design）；EvolutionaryScale 2025-11 被 CZ Biohub 收购（Alex Rives 任 Biohub 科学负责人）；NVIDIA 与 Lilly/Roche/CZI 深度绑定，OpenAI 投资 Chai。生物专有数据成为通用 AI 公司的新护城河，直接抬高对「懂生物的 ML 人」的争夺。Anthropic acquired Coefficient Bio (a team from Genentech's Prescient Design) for $400 million in April 2026; EvolutionaryScale was acquired by CZ Biohub in November 2025 (Alex Rives became Biohub's head of science); NVIDIA is deeply tied to Lilly/Roche/CZI, and OpenAI invested in Chai. Proprietary biological data has become the new moat for general-purpose AI companies, directly intensifying the fight for ML people who understand biology.
截至 2025 年底，AI 发现的药物尚无一款获批。第一代 AI biotech 在收缩：BenevolentAI 裁员、估值跌至约 1.45 亿；Atomwise 2025-10 重组为 Numerion；Recursion 合并 Exscientia 后砍管线控成本。新贵巨额融资与「尚未兑现获批药」并存，是这个赛道当前的真实状态。As of the end of 2025, not a single AI-discovered drug has been approved. The first generation of AI biotechs is contracting: BenevolentAI laid off staff and saw its valuation fall to about $145 million; Atomwise reorganized into Numerion in October 2025; Recursion cut its pipeline to control costs after merging with Exscientia. Massive funding for the new entrants coexisting with no approved drug to show for it yet is the real state of this space today.
统计对象 = 4,374 名计算/ML 人才（蛋白结构/折叠、蛋白/抗体设计、分子生成/化学、靶点/系统生物、ML 平台、计算/ML 科学）。Population = 4,374 computational/ML professionals (protein structure/folding, protein/antibody design, molecular generation/chemistry, target/systems biology, ML platform, and computational/ML science).
读数：单一最大 AI/计算组是 Genentech（442 人，含 Prescient Design 抗体设计），其后是 AbbVie、AstraZeneca、Roche、Sanofi、Merck、Novo Nordisk 等大药企。15 家大药企合计 3,332 人，远超 20 家 AI-native 创业公司合计 1,042 人。AI-native 中 Recursion（211）、Isomorphic Labs（187）、AbCellera（144）、Schrödinger（138） 规模领先，多数蛋白设计创业公司（Cradle/Latent/Profluent/Nabla/Dyno/Chai）团队精小（10-30 人量级，可见档案少）。Reading: the single largest AI/computational group is Genentech (442 people, including Prescient Design antibody design), followed by big-pharma players such as AbbVie, AstraZeneca, Roche, Sanofi, Merck, and Novo Nordisk. The 15 big-pharma companies total 3,332 people, far exceeding the 1,042 across the 20 AI-native startups. Among the AI-natives, Recursion (211), Isomorphic Labs (187), AbCellera (144), and Schrödinger (138) lead on scale, while most protein-design startups (Cradle/Latent/Profluent/Nabla/Dyno/Chai) run lean teams (on the order of 10-30 people, with few visible profiles).
读数：双背景浓度与公司类型相关但不绝对。浓度最高的是把湿实验闭环做进平台的公司（Bristol Myers Squibb 74.6%、Novartis 74.3%、AbCellera 72.2%、AstraZeneca 71.1%）；偏纯 ML、湿实验信号弱的多是纯基础模型/蛋白语言模型创业公司（Genesis Therapeutics 30.4%、Isomorphic Labs 33.2%、Schrödinger 44.9%、insitro 53.6%）。寻访启示：要「能直接读懂湿实验数据的 ML 人」去前者挖，要「纯算法/基础模型人」去后者挖。Reading: dual-background density correlates with company type but not absolutely. Density is highest at companies that build the wet-lab loop into their platform (Bristol Myers Squibb 74.6%, Novartis 74.3%, AbCellera 72.2%, AstraZeneca 71.1%); the more pure-ML, weak-wet-lab-signal companies are mostly pure foundation-model / protein-language-model startups (Genesis Therapeutics 30.4%, Isomorphic Labs 33.2%, Schrödinger 44.9%, insitro 53.6%). Sourcing takeaway: to find ML people who can read wet-lab data directly, recruit from the former; for pure-algorithm/foundation-model people, recruit from the latter.
读数：与纯 AI（湾区独大）截然不同，AI 制药人才地理高度分散。美国占约 60%（湾区、波士顿/剑桥、圣地亚哥、新泽西等多个药企集群），其余分布在英国（伦敦/牛剑）、瑞士（巴塞尔/苏黎世，Roche/Novartis）、丹麦（哥本哈根，Novo Nordisk）、法国（巴黎，Sanofi/Bioptimus）、德国（美因茨，BioNTech/InstaDeep）、加拿大（温哥华，AbCellera）、瑞典（哥德堡，AstraZeneca）。跨地域、跨时区是这个赛道寻访的常态，本地化触达比单点扎堆更重要。Reading: starkly unlike pure AI (where the Bay Area dominates), AI-pharma talent is highly dispersed geographically. The US accounts for about 60% (the Bay Area, Boston/Cambridge, San Diego, New Jersey, and other pharma clusters), with the rest spread across the UK (London/Oxbridge), Switzerland (Basel/Zurich, Roche/Novartis), Denmark (Copenhagen, Novo Nordisk), France (Paris, Sanofi/Bioptimus), Germany (Mainz, BioNTech/InstaDeep), Canada (Vancouver, AbCellera), and Sweden (Gothenburg, AstraZeneca). Cross-region, cross-time-zone sourcing is the norm in this space, and localized outreach matters more than betting on a single cluster.
读数：多数档案 title 写「computational scientist / ML scientist」未标细分方向，归入综合列。明确标注蛋白结构（19 人）或蛋白/抗体设计（106 人）的专精者全行业仅 125 人，是真正的稀缺核心；靶点/系统生物 ML（473 人）多在大药企与 Recursion/insitro。Reading: most profiles carry a title like "computational scientist / ML scientist" with no specified sub-discipline and fall into the general column. Specialists explicitly tagged to protein structure (19 people) or protein/antibody design (106 people) total just 125 across the entire industry, the truly scarce core; target/systems-biology ML (473 people) sits mostly in big pharma and at Recursion/insitro.
读数：两条主进水管。① 高校/科研院所是 AI 制药人才的第一来源（Genentech、Novo、Merck 都大量从学界直招 PhD），印证「学界 PI 与博士是核心供给」；② 大药企之间与 biotech 同业的横向流动密集。Isomorphic 的进水管以「其他企业」为主，实为 DeepMind 内部整建制转入（分拆属性）。AlphaFold 谱系外溢到非 Isomorphic 的可见人数很小，多数仍在母体内。Reading: two main intake pipes. ① Universities/research institutes are the leading source of AI-pharma talent (Genentech, Novo, and Merck all hire PhDs straight from academia in volume), confirming that academic PIs and PhDs are the core supply; ② lateral movement between big pharma and among biotech peers is heavy. Isomorphic's intake pipe is dominated by "other companies," which is really a wholesale internal transfer from DeepMind (its spinout nature). The visible AlphaFold-lineage spillover beyond Isomorphic is very small, with most still inside the parent.
读数：2025 年入职 1122 人，是 2023 年（698 人）的 1.6 倍，反映 2024-2025 融资回暖带动的稳步扩招（不像纯 AI 那样爆发式）。Isomorphic 等新贵的入职高峰集中在 2024-2025。Reading: 1122 people were hired in 2025, 1.6x the 698 in 2023, reflecting the steady ramp in hiring driven by the 2024-2025 funding rebound (not the explosive surge seen in pure AI). The hiring peak for newcomers like Isomorphic is concentrated in 2024-2025.
读数：右上「老兵区」= AstraZeneca（中位 37 个月、51.4% 过 3 年）、Roche，大药企沉淀深；左下「新军区」= Isomorphic（中位 11 个月、54.1% 入职不足一年），高速扩张中，蜜月期反向流动难、但 12-24 个月后进入第一波动摇窗口。Genentech、Merck、Novo 居中（24-28 个月）。Reading: the top-right "veteran zone" = AstraZeneca (median 37 months, 51.4% past 3 years) and Roche, where big pharma runs deep; the bottom-left "new-guard zone" = Isomorphic (median 11 months, 54.1% under a year), fast-scaling, where moving people in the honeymoon phase is hard but the first wobble window opens 12-24 months in. Genentech, Merck, and Novo sit in the middle (24-28 months).
这是本报告的差异化角度。「会折叠蛋白的 ML 人」= 同时懂湿实验（生物/生化直觉）与干实验（ML/编程）的复合人才。数据库口径下 2,432 人（占计算池 55.6%）具备双背景信号。This is the report's differentiating angle. "ML people who can fold proteins" = hybrid talent who understand both the wet lab (biology/biochemistry intuition) and the dry lab (ML/programming). Under the database methodology, 2,432 people (55.6% of the computational pool) carry a dual-background signal.
读数：双背景在 AI 制药是常态而非例外（55.6% 均值），这恰恰说明这个赛道的入场券就是跨学科。浓度最高的是把湿实验闭环做进平台的公司（Bristol Myers Squibb、Novartis、AbCellera、AstraZeneca）；纯基础模型 / 蛋白语言模型公司（Genesis Therapeutics、Isomorphic Labs、Schrödinger、insitro）的双背景比例低，团队更偏纯 ML。Reading: a dual background is the norm rather than the exception in AI pharma (55.6% average), which is precisely why being interdisciplinary is the ticket of admission to this space. Density is highest at companies that build the wet-lab loop into their platform (Bristol Myers Squibb, Novartis, AbCellera, AstraZeneca); pure foundation-model / protein-language-model companies (Genesis Therapeutics, Isomorphic Labs, Schrödinger, insitro) have a lower dual-background share, with more pure-ML teams.
双背景虽多，但Dual backgrounds may be common, but 明确从事蛋白结构预测或蛋白/抗体设计的专精者全样本仅 125 人（2.9%）specialists explicitly working on protein structure prediction or protein/antibody design number just 125 across the full sample (2.9%)。其中 45.0% 量级持 PhD。这 125 人是 AlphaFold 时代各家争抢的核心：会用 RFdiffusion/ESM/Boltz 这类工具做 de novo 设计、又能与湿实验对接验证的人。对买方的含义：① 这是真正需要主动寻访（而非被动等投递）的画像；② 供给主要来自 UW 蛋白设计研究所（Baker 谱系）、DeepMind/Isomorphic、Meta FAIR 蛋白团队（已解散，外溢中）、以及少数高校 PI 实验室；③ 单是会 ML 或单是会生物都不够，必须双能。. Roughly 45.0% of them hold a PhD. These 125 are the core everyone fights over in the AlphaFold era: people who can do de novo design with tools like RFdiffusion/ESM/Boltz and also tie into wet-lab validation. For buyers: ① this is a profile that genuinely requires active sourcing (not passively waiting for applications); ② supply comes mainly from the UW Institute for Protein Design (the Baker lineage), DeepMind/Isomorphic, the Meta FAIR protein team (now disbanded and spilling over), and a handful of academic PI labs; ③ knowing ML alone or biology alone is not enough, you must have both.
数据库口径：35 家公司计算池中华人 676 人（高置信 623），占 15.5%。这一比例低于纯 AI Lab（因 AI 制药含大量欧洲与湿实验背景人才稀释），但在计算/ML 核心岗位华人是重要构成，且 PhD 率显著更高。外部基准：MacroPolo 口径，美国顶级 AI 研究者本科 47% 来自中国。Database methodology: across the 35 companies' computational pools, there are 676 Chinese professionals (623 high-confidence), 15.5% of the total. This share is lower than at pure AI labs (because AI pharma includes a large dilution of European and wet-lab-background talent), but in core computational/ML roles Chinese talent is a significant component, with a markedly higher PhD rate. External benchmark: per MacroPolo, 47% of top US AI researchers did their undergraduate degrees in China.
读数：华人浓度最高的是 Pfizer（21.9%）、Generate（21.1%）、Merck（21.1%）、Genentech（23.1%）、insitro（20.3%）。大药企 AI 组的华人密度普遍高于纯模型创业公司。绝对数上 Genentech（102）、Merck（52）领先。Reading: Chinese density is highest at Pfizer (21.9%), Generate (21.1%), Merck (21.1%), Genentech (23.1%), and insitro (20.3%). Chinese density at big-pharma AI groups is generally higher than at pure-model startups. In absolute terms, Genentech (102) and Merck (52) lead.
读数：华人计算人才 PhD 率 56.7%（全计算池 45.0%），显著更高，集中在蛋白结构/设计、靶点/系统生物与计算/ML 科学等核心研究岗，是这个赛道学术含量最高的人群之一。Reading: Chinese computational talent has a PhD rate of 56.7% (versus 45.0% for the whole computational pool), markedly higher, concentrated in core research roles such as protein structure/design, target/systems biology, and computational/ML science, one of the most academically credentialed groups in this space.
读数：华人主力来自清北复交、中科大、台大、港大等；全池海外院校 Top 反映八国分布特征（英美的剑桥/牛津/帝国理工、伯克利/斯坦福/MIT，欧陆与加拿大的苏黎世联邦理工 ETH、哥本哈根大学/丹麦科技大学 DTU、多伦多大学等）。寻访华人计算生物人才，清北复交校友圈 + 顶会（NeurIPS/ICML/RECOMB/MLCB）华人作者网络命中率最高。Reading: the bulk of Chinese talent comes from Tsinghua/Peking/Fudan/SJTU, USTC, NTU, HKU, and the like; the top overseas institutions in the full pool reflect the eight-country footprint (Cambridge/Oxford/Imperial and Berkeley/Stanford/MIT in the UK and US, and ETH Zurich, the University of Copenhagen / DTU, the University of Toronto, and others in continental Europe and Canada). To source Chinese computational-biology talent, the Tsinghua/Peking/Fudan/SJTU alumni network plus the Chinese-author networks at top conferences (NeurIPS/ICML/RECOMB/MLCB) deliver the highest hit rate.
领军学者Leading academics：Possu Huang 黄柏栩（Stanford 生物工程，蛋白设计，Baker 学术谱系）、Bo Wang 王波（University of Toronto / Vector Institute，单细胞 AI，并任 Xaira 生物医学 AI 负责人）、Le Cong 丛乐（Stanford，CRISPR-GPT）、Jian Peng 彭健（UIUC 教授，Earendil Labs / Helixon 创始人，与 Sanofi 合作）。: Possu Huang (Stanford bioengineering, protein design, Baker academic lineage), Bo Wang (University of Toronto / Vector Institute, single-cell AI, and head of biomedical AI at Xaira), Le Cong (Stanford, CRISPR-GPT), and Jian Peng (UIUC professor, founder of Earendil Labs / Helixon, partnered with Sanofi).
华人创办的 AI 制药公司Chinese-founded AI-pharma companies：晶泰科技 XtalPi（温书豪、马健、赖力鹏，2024 港交所 IPO）、深势科技 DP Technology（张林峰，前 Princeton，戈登贝尔奖）、华深智药 Helixon / Earendil Labs（彭健）。这些公司提供大量计算 + 湿实验复合岗位，是华人计算生物人才的重要去向。: XtalPi (Shuhao Wen, Jian Ma, Lipeng Lai, HKEX IPO in 2024), DP Technology (Linfeng Zhang, formerly Princeton, Gordon Bell Prize), and Helixon / Earendil Labs (Jian Peng). These companies offer large numbers of hybrid computational + wet-lab roles and are an important destination for Chinese computational-biology talent.
宏观背景Macro context：黄仁勋称全球约 50% AI 研究者是华人；MacroPolo 顶级 AI 研究者本科在华占比从 2019 年 29% 升至 2022 年 47%。蛋白结构/设计的深度学习方法与主流 AI 同源，华人在该方向的密度与整体 AI 一致。: Jensen Huang has said roughly 50% of the world's AI researchers are Chinese; per MacroPolo, the share of top AI researchers who did their undergraduate degrees in China rose from 29% in 2019 to 47% in 2022. The deep-learning methods behind protein structure/design share the same roots as mainstream AI, so the density of Chinese talent in this direction matches AI overall.
市场上完全没有工程师层版本的大药企 AI 组织图。本节用全量档案把 Genentech / Novartis / AstraZeneca 三家 AI 计算组还原到带队层与 IC 厚度。层级依据公开职位 title 归类，非官方组织架构，仅作团队梯队结构概览。公开版人名默认模糊。The market has no engineer-level version of big-pharma AI org charts. This section uses the full set of profiles to reconstruct the AI computational groups at Genentech / Novartis / AstraZeneca down to the team-lead layer and IC depth. Levels are classified from public job titles, are not official org structure, and serve only as an overview of team-tier structure. In the public version, names are masked by default.
Roche 旗下，含 Prescient Design 抗体/蛋白设计团队（lab-in-the-loop 概念发源地），负责人 Aviv Regev。是大药企内部最强的 AI 蛋白设计组之一，也是 Coefficient Bio 创始团队的来源。Part of Roche, it includes the Prescient Design antibody/protein-design team (the birthplace of the lab-in-the-loop concept), led by Aviv Regev. It is one of the strongest AI protein-design groups inside big pharma and the source of the Coefficient Bio founding team.
与微软合作建 Generative Chemistry pipeline，与 Isomorphic（6 项目）、Generate、Schrödinger 合作。双背景浓度全样本最高之一。Built a Generative Chemistry pipeline with Microsoft and partners with Isomorphic (6 projects), Generate, and Schrödinger. Among the highest dual-background densities in the full sample.
AI 药物发现布局深，与 Absci 等合作。任期最资深的大药企 AI 组之一（老兵密度高）。Deeply invested in AI drug discovery and partnered with Absci and others. One of the most tenured big-pharma AI groups (high veteran density).
按级别、方向稀缺度与履历强度筛出三组共 32 人。档案事实来自 Metix AI 数据库；标注「公开核实」者已对照 2025-2026 公开信源确认现职（前沿 Lab 档案更新滞后，公开信源优先）。本节人物均来自公开职业档案。A 组为创始人、高管与公开技术负责人，B 组为资深技术骨干，C 组为华人与稀缺方向画像。公开版人名默认模糊。Three groups totaling 32 people, screened by level, direction scarcity, and history strength. Profile facts come from the Metix AI database; those marked "publicly verified" have had their current role confirmed against 2025-2026 public sources (frontier-lab profiles lag in updates, so public sources take precedence). All figures in this section come from public professional profiles. Group A is founders, executives, and public technical leads; Group B is senior technical backbone; Group C is Chinese-talent and scarce-direction profiles. In the public version, names are masked by default.
AI 制药薪酬的结构性特征是「与前沿 AI Lab 抢同一批 ML 人，但给不出同等的钱」。数字为 2025-2026 市场口径，非个案承诺。The structural feature of AI-pharma compensation is "fighting frontier AI labs for the same ML people but unable to match the money." Figures follow 2025-2026 market data, not individual offers.
| 群体Group | 总包/薪酬区间Total comp / pay range | 说明Notes |
|---|---|---|
| AI 制药 ML scientist（多数）AI-pharma ML scientist (most) | $98K-$176K base | ZipRecruiter / Takeda 口径，AI 药物发现 scientist 平均约 $123KPer ZipRecruiter / Takeda, AI-drug-discovery scientists average about $123K |
| AI 制药专业岗（高配）AI-pharma specialist role (high end) | $200K-$240K | D.E. Shaw Research 药物发现 AI/ML 数据科学家D.E. Shaw Research drug-discovery AI/ML data scientist |
| 前沿 AI Lab SWE 中位Frontier AI lab SWE median | $600K-$795K | levels.fyi，与 AI 制药差 3-5 倍Per levels.fyi, 3-5x the AI-pharma level |
| OpenAI / Anthropic 同级OpenAI / Anthropic equivalent level | $600K-$1.15M | 顶配研究员更高，名单制特殊包绕开职级Top-tier researchers go higher, with named special packages bypassing the leveling system |
| 中国 AI 药企回国包Return-to-China package at Chinese AI-pharma firms | 个案为准Case by case | 晶泰/深势/华深智药等提供计算+湿实验复合岗XtalPi / DP Technology / Helixon and others offer hybrid computational + wet-lab roles |
AI 制药与前沿 AI Lab 争抢同一批懂 ML 的人，但薪酬中位差 3-5 倍。结果是：AI 制药留不住纯算法明星（会被 OpenAI/Anthropic 挖走），只能靠「使命感（治病/诺奖级科学）+ 学术声望（与 Baker/Koller/Hassabis 共事）+ 股权上行 + 双背景门槛（纯 AI Lab 要不了懂湿实验的人）」差异化留人。这也是为什么这个领域以「主动寻访 + 高费率」为主，而非被动等投递。AI pharma and frontier AI labs fight for the same ML people, but the median pay gap is 3-5x. The result: AI pharma can't keep its pure-algorithm stars (they get poached by OpenAI/Anthropic) and must retain people differentially through "a sense of mission (curing disease / Nobel-grade science) + academic prestige (working alongside Baker/Koller/Hassabis) + equity upside + a dual-background bar (pure AI labs have no use for people who understand the wet lab)." This is also why this field runs on active sourcing + high fees rather than passively waiting for applications.
① 蛋白结构/设计专精者需主动寻访，且要用非现金优势（科学问题质量、湿实验闭环、平台数据）吸引；② 双背景人才是 AI Lab 难以吸引的「差异化画像」，应优先关注；③ 欧洲（瑞士/丹麦/英国）薪酬基准低于美国，是预算有限买方建计算组的洼地；④ 收缩公司释放的资深人才是当下少数「可立即谈」的来源。① Protein-structure/design specialists require active sourcing and must be attracted with non-cash advantages (quality of the scientific problem, the wet-lab loop, platform data); ② dual-background talent is a differentiated profile that AI labs struggle to attract and should be a priority; ③ European pay benchmarks (Switzerland/Denmark/UK) sit below the US, a value pocket for budget-constrained buyers building computational groups; ④ senior talent released by contracting companies is one of the few sources you can talk to right now.
来源：ZipRecruiter、levels.fyi、PwC 2025 AI 技能溢价、IntuitionLabs 生命科学就业报告（2025-2026 检索）。详见研究底稿。Sources: ZipRecruiter, levels.fyi, the PwC 2025 AI skills premium, and the IntuitionLabs life-sciences employment report (retrieved 2025-2026). See the research memos for details.
把地图变成动作：biotech 猎头看稀缺画像与窗口，HealthTech HR 看对标与防守，医疗 VC 看团队尽调信号。Turning the map into action: biotech recruiters look at scarce profiles and windows, HealthTech HR at benchmarking and defense, healthcare VCs at team-diligence signals.
① 蛋白结构/设计专精者（全样本仅 125 人）是高费率领域，适合主动寻访；② 干湿双背景（2,432 人）是 AI Lab 难以吸引的差异化画像，值得优先关注；③ 大药企 AI 组（3,332 人，占 73%）是人才的主要分布地，不应只关注明星初创；④ 任期窗口与公司类型叠加，可定位流动性较高的群体。① Protein-structure/design specialists (just 125 across the full sample) are a high-fee area well suited to active sourcing; ② wet-lab × dry-lab talent (2,432 people) is a differentiated profile that AI labs struggle to attract and deserves priority; ③ big-pharma AI groups (3,332 people, 73%) are where talent is mainly concentrated, so don't focus only on marquee startups; ④ overlaying tenure windows with company type lets you pinpoint the more mobile groups.
① 对照 3.7 任期基准给自家定位：新军（Isomorphic 等）防守留人窗口在 12-24 个月后；② 用非现金优势（科学问题、湿实验数据闭环、与顶级 PI 共事）弥补薪酬差距；③ 关注自家「双背景 + 任期过 30 月」人群（市场流动性最高的群体）；④ 华人计算人才 PhD 率 56.7%，校友链 + 顶会网络是高命中触达面。① Position yourself against the 3.7 tenure benchmark: for newcomers (Isomorphic and others), the retention-defense window opens 12-24 months in; ② use non-cash advantages (the scientific problem, the wet-lab data loop, working with top PIs) to close the pay gap; ③ watch your own "dual-background + over-30-month-tenure" cohort (the most mobile group in the market); ④ Chinese computational talent has a 56.7% PhD rate, and alumni chains + top-conference networks are the highest-hit outreach surface.
① AlphaFold 谱系外溢仍在早期（多数还在 Isomorphic 母体内），下一批 spinout 值得跟踪；② 学界 PI 下海是这个赛道创业主轴（Baker→Xaira、Koller→insitro、彭健→Earendil），盯紧顶级蛋白设计/virtual cell PI 的动向；③ 团队尽调可用本报告的双背景浓度与组织拼图判断「这支队伍是真双能还是纯算法」。① The AlphaFold-lineage spillover is still early (most remain inside the Isomorphic parent), so the next wave of spinouts is worth tracking; ② academic PIs going commercial is the main axis of startup formation in this space (Baker → Xaira, Koller → insitro, Jian Peng → Earendil), so watch the moves of top protein-design / virtual-cell PIs closely; ③ for team diligence, use this report's dual-background density and org maps to judge whether a team is truly dual-capable or purely algorithmic.
本报告的检索、画像、流动分析全部由 Metix AI 完成。可按同样口径为任意公司生成定制图谱：全量长名单导出、干湿双背景筛选、组织拼图、邮箱解锁与多渠道触达，并按「只为合格面试付费」计费。No interview, no charge.This report's search, profiling, and flow analysis were all done by Metix AI. We can generate a custom map for any company under the same methodology: full long-list export, wet-lab × dry-lab filtering, org maps, email unlock, and multi-channel outreach, billed on a "pay only for qualified interviews" basis. No interview, no charge.
8.6 亿+ 全球人才画像860 million+ global talent profiles4,374 人计算池 + 2,432 人双背景长名单4,374-person computational pool + 2,432-person dual-background long list蛋白设计画像分析Protein-design profile analysis只为合格面试付费Pay only for qualified interviewsAI-native（全量计算岗，20 家）：Isomorphic Labs、Recursion、Genesis Therapeutics、Iambic、Chai Discovery、EvolutionaryScale、Xaira、Generate Biomedicines、insitro、Cradle、Latent Labs、Profluent、Schrödinger、AbCellera、Absci、Insilico Medicine、Cellarity、Dyno Therapeutics、Nabla Bio、Bioptimus。大药企（取 AI/计算子集，15 家）：Genentech、Roche、Novartis、AstraZeneca、Pfizer、Merck、Eli Lilly、Novo Nordisk、Sanofi、GSK、Amgen、BioNTech · InstaDeep、Bristol Myers Squibb、AbbVie、Johnson & Johnson。地理范围 = 档案常驻地在美国、英国、瑞士、丹麦、法国、德国、加拿大、瑞典。AI-native (all computational roles, 20 companies): Isomorphic Labs, Recursion, Genesis Therapeutics, Iambic, Chai Discovery, EvolutionaryScale, Xaira, Generate Biomedicines, insitro, Cradle, Latent Labs, Profluent, Schrödinger, AbCellera, Absci, Insilico Medicine, Cellarity, Dyno Therapeutics, Nabla Bio, Bioptimus. Big pharma (AI/computational subset only, 15 companies): Genentech, Roche, Novartis, AstraZeneca, Pfizer, Merck, Eli Lilly, Novo Nordisk, Sanofi, GSK, Amgen, BioNTech · InstaDeep, Bristol Myers Squibb, AbbVie, Johnson & Johnson. Geographic scope = profiles based in the United States, United Kingdom, Switzerland, Denmark, France, Germany, Canada, or Sweden.
大药企体量巨大，本报告的「(AI/计算)」= 当前在该公司且 title/headline 可识别为 AI/ML/计算/数据科学相关者，是可识别子集而非全员，绝对数偏保守。AI-native 创业公司中蛋白设计团队精小（10-20 人），可见档案天然偏少。Big pharma is enormous, so in this report "(AI/computational)" = those currently at the company whose title/headline is identifiable as AI/ML/computational/data-science-related, an identifiable subset rather than the full headcount, making the absolute numbers conservative. Among AI-native startups, protein-design teams are lean (10-20 people), so visible profiles are naturally few.
计算/ML 人才池 = 蛋白结构/折叠、蛋白/抗体设计、分子生成/化学、靶点/系统生物、ML 平台、计算/ML 科学（综合）。湿实验/纯生物岗不计入计算池。干湿双背景 = 教育/技能/履历同时含湿实验（生物/生化/化学/实验）与干实验（ML/编程/计算/统计）信号，为概率判定。Computational/ML pool = protein structure/folding, protein/antibody design, molecular generation/chemistry, target/systems biology, ML platform, and computational/ML science (general). Wet-lab / pure-biology roles are excluded from the computational pool. Wet-lab × dry-lab = education/skills/history containing both wet-lab (biology/biochemistry/chemistry/experimental) and dry-lab (ML/programming/computational/statistics) signals, a probabilistic determination.
沿用五信号交叉验证（姓名族裔模型 / 汉字 / 中文 / 中国院校 / 多拼写姓氏库），分高/中置信，主口径 = 高 + 中。数据截至 2026 年上半年；档案更新存在滞后，代表人物已对照公开信息复核，2025-2026 的最新职位变动以公开信源为准标注。Uses the same five-signal cross-validation (name-ethnicity model / Chinese characters / Chinese-language text / Chinese institutions / multi-spelling surname library), split into high/medium confidence, with the primary scope = high + medium. Data is through H1 2026; profile updates lag, representative figures have been re-checked against public information, and the latest 2025-2026 role changes are annotated based on public sources.
| 公司Company | 在职画像Current Profiles | 计算池Computational pool | 双背景率Dual-background rate | 华人计算Chinese computational | 华人占比Chinese share | PhD 率PhD rate |
|---|---|---|---|---|---|---|
| Genentech (AI/计算)Genentech (AI/computational) | 444 | 442 | 50.7% | 102 | 23.1% | 48.6% |
| AbbVie (AI/计算)AbbVie (AI/computational) | 353 | 353 | 71.1% | 65 | 18.4% | 58.1% |
| AstraZeneca (AI/计算)AstraZeneca (AI/computational) | 335 | 332 | 71.1% | 31 | 9.3% | 53.6% |
| Roche (AI/计算)Roche (AI/computational) | 281 | 281 | 51.6% | 34 | 12.1% | 50.9% |
| Sanofi (AI/计算)Sanofi (AI/computational) | 252 | 249 | 26.9% | 35 | 14.1% | 17.7% |
| Merck (AI/计算)Merck (AI/computational) | 246 | 246 | 54.5% | 52 | 21.1% | 45.9% |
| Novo Nordisk (AI/计算)Novo Nordisk (AI/computational) | 243 | 243 | 66.7% | 21 | 8.6% | 46.1% |
| Johnson & Johnson (AI/计算)Johnson & Johnson (AI/computational) | 238 | 238 | 42.0% | 51 | 21.4% | 46.6% |
| Recursion | 612 | 211 | 54.0% | 12 | 5.7% | 44.5% |
| Novartis (AI/计算)Novartis (AI/computational) | 206 | 206 | 74.3% | 28 | 13.6% | 33.0% |
| Bristol Myers Squibb (AI/计算)Bristol Myers Squibb (AI/computational) | 189 | 189 | 74.6% | 47 | 24.9% | 42.3% |
| Isomorphic Labs | 334 | 187 | 33.2% | 16 | 8.6% | 44.4% |
| GSK (AI/计算)GSK (AI/computational) | 156 | 156 | 62.2% | 22 | 14.1% | 44.2% |
| AbCellera | 458 | 144 | 72.2% | 19 | 13.2% | 40.3% |
| Schrödinger | 668 | 138 | 44.9% | 25 | 18.1% | 41.3% |
| Eli Lilly (AI/计算)Eli Lilly (AI/computational) | 131 | 131 | 51.9% | 27 | 20.6% | 45.8% |
| BioNTech · InstaDeep (AI/计算)BioNTech · InstaDeep (AI/computational) | 126 | 126 | 27.8% | 9 | 7.1% | 29.4% |
| Amgen (AI/计算)Amgen (AI/computational) | 108 | 108 | 50.0% | 21 | 19.4% | 44.4% |
| Xaira | 158 | 74 | 67.6% | 13 | 17.6% | 66.2% |
| insitro | 230 | 69 | 53.6% | 14 | 20.3% | 42.0% |
| Iambic | 115 | 37 | 67.6% | 4 | 10.8% | 48.6% |
| Pfizer (AI/计算)Pfizer (AI/computational) | 33 | 32 | 65.6% | 7 | 21.9% | 37.5% |
| Cellarity | 88 | 26 | 69.2% | 2 | 7.7% | 42.3% |
| Genesis Therapeutics | 71 | 23 | 30.4% | 4 | 17.4% | 52.2% |
| Absci | 64 | 23 | 56.5% | 2 | 8.7% | 43.5% |
| Generate Biomedicines | 68 | 19 | 73.7% | 4 | 21.1% | 73.7% |
| Cradle | 42 | 19 | 26.3% | 0 | 0.0% | 26.3% |
| Bioptimus | 30 | 17 | 5.9% | 1 | 5.9% | 29.4% |
| Profluent | 24 | 13 | 69.2% | 1 | 7.7% | 61.5% |
| Dyno Therapeutics | 44 | 11 | 90.9% | 2 | 18.2% | 54.5% |
| Nabla Bio | 17 | 9 | 55.6% | 3 | 33.3% | 22.2% |
| Insilico Medicine | 21 | 7 | 42.9% | 0 | 0.0% | 28.6% |
| EvolutionaryScale | 8 | 6 | 33.3% | 1 | 16.7% | 50.0% |
| Latent Labs | 21 | 6 | 33.3% | 1 | 16.7% | 66.7% |
| Chai Discovery | 6 | 3 | 33.3% | 0 | 0.0% | 66.7% |
① 快照时效Snapshot currency：数据截至 2026 年上半年，近期人事变动存在滞后；代表人物已对照公开信息复核，长名单使用前建议二次确认。: data is through H1 2026, with recent personnel changes lagging; representative figures have been re-checked against public information, and we recommend a second confirmation before using the long list.
② 覆盖率Coverage：本报告基于公开职业档案聚合，大药企取可识别 AI/计算子集，小型蛋白设计创业公司（Chai/Latent/Cradle/EvolutionaryScale）团队精小、覆盖偏低；各项为数据库口径，宜与公司公开编制互为参照。: this report is built from aggregated public professional profiles, with big pharma counted as its identifiable AI/computational subset; small protein-design startups (Chai/Latent/Cradle/EvolutionaryScale) run lean teams with lower coverage. All figures follow the database methodology and should be cross-referenced with companies' public headcounts.
③ 职能与双背景为推断Function and dual background are inferred：基于 title/headline/技能/教育关键词；大量档案 title 写「computational/ML scientist」未标细分方向，归入综合列，故蛋白结构/设计专精者的绝对数是下限。: based on title/headline/skills/education keywords; many profiles carry a title like "computational/ML scientist" with no specified sub-discipline and fall into the general column, so the absolute number of protein-structure/design specialists is a lower bound.
④ 华人识别为概率判定Chinese-talent identification is probabilistic：主口径 676 人 = 高置信 623 + 中置信 53；使用西文名且无中国信号的华裔会漏检。: the primary scope of 676 = 623 high-confidence + 53 medium-confidence; people of Chinese descent who use Western names with no China signal will be missed.
⑤ 研究底稿Research memos：两份带全部信源 URL 的研究备忘录（行业格局 / 人才生态）与本报告同目录交付。: two research memos with all source URLs (industry landscape / talent ecosystem) are delivered in the same directory as this report.