Gilead, AbbVie AI Push Hits Bottleneck as Bio-Pharma Testing Lags Behind

生物医药测试滞后，吉利德与艾伯维AI推进遭遇瓶颈 AI分子设计的加速引发了巨大的下游运营瓶颈，人体临床试验与转化生物学仍是药物研发中的限速步骤。 尽管生成式AI能在一天之内设计出全新药物分子，但将这些分子转化为获批疗法仍面临耗资数十亿美元、历时数年的临床开发瓶颈。与会专家探讨了多项高价值应用，包括AI辅助统计编程、基于图像的肿瘤负荷分析，以及用于精简罕见病试验对照组的数字孪生建模。他们还指出了关键的运营障碍，强调90%的传统制药工作流需要进行流程再造而非强制自上而下的指令，同时还需保障全球算力公平获取并推行与风险相适应的监管策略。 尽管 AI 已经能够在一天内完成新分子设计，但将药物推进至 FDA 获批仍需耗费超过 10 亿美元与 10 年以上时间，弥合分子发现与临床交付之间的鸿沟是当前核心挑战。 (0:04) Parexel 借助 AI 将监管申报文件撰写耗时缩减约 60%、统计数据集构建耗时缩减 50%，并利用数字孪生构建虚拟对照组以缓解 ALS 等罕见病试验的招募难题。 (22:15) AI 在上游靶点筛选与蛋白质折叠领域的加速引发了研发管线激增，但动物模型向人体首次试验的临床转化断层，依然是制约新药落地最关键的限速步骤。 (29:02) 专家指出，将传统的确定性计算机软件验证（CSV）生搬硬套到非确定性 AI 模型是导致项目搁浅的常见误区，企业在落地时还需应对生产力初期下滑的 J 曲线与 Token 成本控制。 (38:00)

While artificial intelligence has dramatically accelerated upstream molecular discovery, biopharma leaders emphasize that overcoming the decade-long, billion-dollar clinical development bottleneck requires rigorous operational re-engineering, specialized compute infrastructure, and nuanced regulatory navigation. The Downstream Clinical Bottleneck (0:04) Moderator Bryson Tombridge and Gilead's Patrick Loerch explain that although AI can now design novel compounds in a single day, taking a therapy through clinical trials to FDA approval still exceeds $1 billion and 10 years. Upstream breakthroughs in protein folding, virtual screening, and docking create a preclinical 'tsunami' of candidate molecules that downstream clinical pipelines are unequipped to handle. Loerch, Aman Thukral of AbbVie, and Tala Fakhouri of Parexel agree that software cannot bypass fundamental biological rate limits: translating results from animal models into human safety and efficacy remains the primary barrier to clinical success. Streamlining Operations and Digital Control Arms (22:15) To absorb the influx of drug candidates, organizations focus AI deployments on high-friction operational phases and trial design. Fakhouri notes Parexel achieved timeline reductions of 60% in drafting regulatory documents and 50% in generating submission-ready statistical datasets. In clinical research, digital twins using statistical methods like PROCOVA allow teams to model placebo trajectories using historical and Phase 2 data, significantly reducing required control group sizes in rare diseases like ALS. Thukral highlights operational automation at AbbVie, where administrative labor like electronic health record source data verification and compliance auditing can be shifted toward digital clones so monitors can focus on investigator relationships. Compute Infrastructure and Global Representation (20:56) Access to high-performance computing has become a strategic requirement for life science biotechs. Nebius's Laura Boykin Okalebo highlights that demand for GPUs vastly exceeds supply, noting an epigenomics dementia partner scaled from 16 Nvidia H200 GPUs in December 2024 toward an expected 1,000 Vera Rubin GPUs by late 2026. She notes that venture capitalists now evaluate whether startups possess concrete compute commitments before funding. Furthermore, infrastructure must co-locate encrypted patient storage with GPU clusters to avoid cross-border data latency. Okalebo argues against allocating resources strictly to the highest bidder, pointing to initiatives supporting diverse global research, such as a Rwandan team training breast cancer diagnostic models tailored to denser tissue in women of African descent. Regulatory Realities and Process Overhauls (38:00) Fakhouri, former FDA lead for AI policy at CDER, clarifies that early-stage discovery and internal operational workflows sit outside direct FDA oversight. She cautions that misapplying legacy 1980s Computerized Software Validation (CSV) protocols designed for deterministic software to non-deterministic AI models has prematurely killed viable projects. Operationally, Thukral cautions organizations to anticipate an initial J-curve drop in productivity, token economic overhead, and potential hallucinations. Integrating AI requires re-engineering business workflows, as up to 90% of legacy pharma processes are incompatible with AI out of the box, necessitating collaborative frameworks between domain experts and machine learning engineers over top-down directives. The panel underscores that raw computational speed does not automatically yield clinical velocity. Until fundamental biological translation hurdles are solved and enterprise business processes are restructured, AI's chief clinical contributions remain bounded by operational automation, trial design optimization, and equitable infrastructure access.

在药物研发领域，尽管人工智能已能在一天内完成分子设计，但新药落地仍受制于超过10亿美元投入与10年以上的临床开发周期。来自生物医药、监管咨询与计算架构领域的专家指出，AI的战略重心正从前端分子生成转向临床转化提效，但在解决临床试验瓶颈的同时，行业必须应对算力短缺、监管验证误区以及高达90%流程需重构的组织阻力。 研发断层与转化瓶颈：AI提速对临床承载力的冲击 (29:02) 吉利德科学（Gilead Sciences）高级副总裁Patrick Loerch指出，AI驱动的虚拟筛选、靶点优化、分子对接和蛋白质折叠技术大幅压缩了前期探索周期，形成了一场涌向临床开发后端的“研发海啸”。然而，Tono Health首席执行官Bryson Tombridge与Parexel副总裁Tala Fakhouri均强调，人体生物学规律无法被人为任意加速，临床管线无法简单实现等比例扩容。特别是在转化医学阶段，动物模型在安全性与毒性预测上与人体早期试验的低关联度，构成了纯算法无法即刻攻克的核心限速瓶颈。 为了量化AI对科学研发的实质贡献，团队正从技术与监管成功率（PTRS）与运营效率两个维度展开追踪。Loerch解释称，科学层面的核心在于提升各阶段关卡决策的确定性，而运营层面则设定了具体指标，例如将单项任务耗时缩短15%，或将整体研发通量提高20%。多位专家达成共识：若不针对前置数据清洗、代码编写与转化分析进行端到端自动化改造，前端激增的候选化合物只会加剧临床试验准入阶段的拥堵。 临床运营提效与数字孪生：从自动化到合成对照组 (22:15) 在临床试验执行层面，生成式AI与专用算法正显著压缩关键路径。AbbVie数字化运营全球负责人Aman Thukral表示，大语言模型已将特定临床任务交付周期从4至6周压缩至单日甚至数小时，但他强调单纯的高速必须结合“人机协同”监督，以确保符合FDA申报质量，实现走向正确监管靶向的“有效速度”（Velocity）。在Parexel的实践中，监管申报文档撰写耗时缩短约60%，符合监管级标准的统计数据集构建耗时缩短约50%。同时，吉利德通过部署AI辅助集成开发环境（IDE）自动化生成统计代码，并在肿瘤学中利用AI快速分析医学影像以评估肿瘤总负荷（超越传统RECIST标准），支持各阶段的快速决策。 数字孪生（Digital Twins）技术则开辟了试验设计与现场监查的新范式。Fakhouri指出，针对肌萎缩侧索硬化症（ALS）等招募极为艰难的罕见病，Parexel利用基础模型结合历史对照数据的PROCOVA统计方法模拟患者病情轨迹，在保持严格统计效力的同时显著缩减关键试验所需的安慰剂对照组规模。Loerch补充称，数字孪生还可提前精准预估效应量，避免因对照组表现超出预期而导致试验效力不足。Thukral则提出构建临床监查“数字分身”，自动化接管源数据核查（SDV）、药物管理与合规审计等繁复工作，释放人力专注临床关系维护。 算力底座与全球包容性：架构就近与生物多样性 (20:56) 生命科学AI算法的爆发式演进对底层计算基础设施提出了严苛要求。Nebius解决方案合伙人Laura Boykin Okalebo指出，当前生命科学领域的GPU算力需求已远超供应，风险投资机构在评估早期生物科技团队时，已将是否锁定高等级算力作为关键前置指标。以其合作伙伴Prim利用表观基因组学开展痴呆症研究为例，其实际计算规模正计划从2024年12月的16张NVIDIA H200 GPU扩展至2026年12月的1,000张NVIDIA Vera Rubin GPU，企业需在H100与Blackwell等多样硬件间保持灵活部署能力。 伴随算力扩展的是数据中心合规架构与全球样本代表性挑战。Okalebo强调算力分配应摒弃“价高者得”的单一逻辑，Nebius通过扶持机制覆盖了来自卢旺达、哥伦比亚等地的650多个申请团队；例如卢旺达团队针对非洲裔女性普遍更致密的乳腺组织特征专门训练影像模型，以消除算法偏差。而在硬件物理架构上，为满足HIPAA等全球医疗数据驻留与隐私法规，企业必须将高带宽加密存储与GPU计算集群部署在极近的物理空间内，否则跨境网络延迟将彻底抵消高性能芯片的加速收益。 监管范式与组织重构：破除CSV旧制与滴漏式陷阱 (38:00) 在监管合规层面，行业对于监管边界存在广泛误区。曾主导FDA药物评价和研究中心（CDER）AI政策制定的Fakhouri澄清，早期药物发现与药企内部运营工具并不在FDA的直接管辖范围之内。她指出，行业中一个普遍且代价高昂的错误是强行将20世纪80至90年代针对确定性软件设立的传统计算机化软件验证（CSV）框架，套用在非确定性的机器学习模型上，这导致了数百个极具潜力的AI项目被内部合规扼杀。实际上监管机构在确保数据隐私（如GDPR）的前提下，更倡导基于风险层级的弹性验证框架。 在组织落地维度，Thukral警示企业在引入AI时必须预留应对生产力短暂下滑的“J曲线效应”缓冲期，妥善管控模型幻觉并建立细致的“Token经济学”成本核算机制。更为关键的是，药企高达90%的原生业务流程无法与AI即插即用，单纯依靠管理层下达指令的“滴漏式AI”（Trickle-down AI）必然导致一线抗拒与产出受阻。专家建议建立正式的“平民开发者”（Citizen Developer）机制，将深谙实际临床痛点的业务专家与数据工程师深度结对，通过重构业务底层逻辑实现AI工具的精准赋能。 本场讨论表明，AI对制药产业链的重塑正在从前端化合物合成拓展至全流程临床运营，但纯计算工具无法直接逾越动物模型向人体转化的生物学鸿沟。行业未来的核心挑战不仅在于算法本身的演进，更取决于能否建立匹配合规标准的本地化低延迟算力底座、摆脱陈旧的软件验证框架，并彻底重构制药企业内部高达90%与AI脱节的传统业务流程。

## Background 本内容取自 YouTube 频道 Ai4 于 2026 年 8 月 22 日发布的研讨会视频《从分子到市场：AI 如何重塑药物研发》。页面介绍显示，本次研讨会由 Tono Health 联合创始人兼 CEO 主持，邀请了来自吉利德（Gilead）、艾伯维（AbbVie）、Parexel 以及 Nebius 的多位药企高管与 AI 科学家，共同探讨机器学习在靶点识别、分子设计、临床试验模拟等药物研发全流程中的应用，并分析当前面对的实际挑战。 ## TL;DR 研讨会聚焦于如何利用 AI 跨越从“生成分子”到“临床落地”之间的巨大鸿沟。嘉宾们深入探讨了 AI 在科学发现（如提升技术与监管成功率 PTRS、肿瘤负荷评估）与运营效率（如自动化撰写临床研究报告 CSR、代码生成）两大维度的落地实践。讨论强调了从单纯追求速度（Speed）转向正确方向（Velocity）的重要性，并指出了跨学科语言障碍、监管合规性、GPU 算力分配以及业务流程与团队转型等核心挑战。 ## Atmosphere & style 这场研讨会充满前沿科技感与严谨的工业界现实主义色彩。嘉宾们兼具深厚的学术背景与丰富的药企实战经验，交流氛围专业、高效且富有深度。探讨过程中既有对物理 AI、自驱实验室和算力民主化等突破性技术的兴奋与热情，又始终保持着对数据质量、临床监管风险及组织转型阻力的清醒审视，给人一种踏实而深刻的行业洞察感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=Uq0Wy6uZkuA&t=4s) 虽然 AI 已能在极短时间内设计出全新分子，但从分子设计到临床上市（Bedside）仍面临高成本与长周期的研发鸿沟。 - [00:04](https://www.youtube.com/watch?v=Uq0Wy6uZkuA&t=4s) 一款 FDA 批准上市的新药研发成本通常超过 10 亿美元，且耗时往往长达 10 年以上。 - [05:11](https://www.youtube.com/watch?v=Uq0Wy6uZkuA&t=311s) 使用大语言模型（LLM）等技术虽然大幅提升了数据处理的速率（Speed），但若缺乏人类把关（Human-in-the-loop），可能偏离正确方向（Velocity）并损害提交给监管机构的数据质量。 - [05:11](https://www.youtube.com/watch?v=Uq0Wy6uZkuA&t=311s) AI 在药物早期研发阶段引发了候选化合物的“分子海啸”，但受限于物理运行与科学决策，后续临床试验阶段依然是核心瓶颈。 - [05:11](https://www.youtube.com/watch?v=Uq0Wy6uZkuA&t=311s) 物理 AI 与自动驾驶实验室（Self-driving labs）通过自动化机器人接管繁重实验工作，重塑了传统实验室流程。 - [05:11](https://www.youtube.com/watch?v=Uq0Wy6uZkuA&t=311s) 跨学科团队（工程、计算生物学与生命科学）之间的“共同语言缺失”是当前 AI 落地生物医药最大的协同障碍之一。 - [10:52](https://www.youtube.com/watch?v=Uq0Wy6uZkuA&t=652s) AI 在药物研发全生命周期中的价值衡量主要分为“科学决策（如提高技术与监管成功概率 PTRS）”与“运营效率（如缩短流程耗时与扩大规模）”两大维度。 - [14:46](https://www.youtube.com/watch?v=Uq0Wy6uZkuA&t=886s) 艾伯维（AbbVie）利用 AI 重点优化 IND 到 NDA 之间的关键路径，包括数据快速分析、数据库锁定后的临床研究报告（CSR）自动撰写以及 eCTD 编译。 - [14:46](https://www.youtube.com/watch?v=Uq0Wy6uZkuA&t=886s) 吉利德（Gilead）通过 AI 图像识别评估肿瘤总体负荷（Total Tumor Burden），并使用 AI 辅助代码编写（AI-assisted Coding）提升数据分析效率。 - [14:46](https://www.youtube.com/watch?v=Uq0Wy6uZkuA&t=886s) 算力需求呈现指数级增长，如研究失智症的生物技术公司 Prim 在两年内对 GPU 的需求从 16 张 H200 激增至 1000 张 Vera Rubin 显卡。 - [20:56](https://www.youtube.com/watch?v=Uq0Wy6uZkuA&t=1256s) 通过云端算力赋能缺乏硬件资源的全球团队，能够保障 AI 训练数据的多样性与包容性。 - [20:56](https://www.youtube.com/watch?v=Uq0Wy6uZkuA&t=1256s) Parexel 利用 AI 将监管申报文件撰写时间缩短了 60%，并将生成监管级数据的时间削减了 50%。 - [20:56](https://www.youtube.com/watch?v=Uq0Wy6uZkuA&t=1256s) 在渐冻症（ALS）等罕见病临床试验中，利用 AI 打造数字孪生（Digital Twins）预测疾病进展，可显著减少所需的安慰剂组患者数量。 - [41:20](https://www.youtube.com/watch?v=Uq0Wy6uZkuA&t=2480s) 推行 AI 落地必须重构匹配的业务流程并带有员工共同转型，因为高达 90% 的传统流程无法直接与 AI 技术相匹配。 ## Detailed walkthrough 以下是针对视频《从分子到临床：AI 如何重塑药物研发》（Part 1）的详细总结与深度梳理： ### [00:04](https://www.youtube.com/watch?v=Uq0Wy6uZkuA&t=4s) 研讨会背景与引言：药物研发的核心瓶颈 主持人 Bryson Tombridge 开场指出，AI 正在重塑全球各个行业，生物医药领域更是发生了永久性的变革。虽然如今凭借 AI 可以在一个下午内设计出全新分子，但这并不意味着人类彻底攻克了疾病研发难题。 在与一位优秀的斯坦福博士后兼成功创业者交流后，Bryson 意识到从“设计出分子”到“最终送达患者床边（Bedside）”之间，仍存在巨大的鸿沟与技术壁垒。 * **研发成本高昂**：每一款最终获得 FDA 批准上市的药物，从头到尾的研发成本均超过 10 亿美元。 * **研发周期漫长**：从早期探索到获批上市，整个流程往往需要 10 年以上。 * **核心使命**：如何利用 AI 缩短这一巨大鸿沟，将先进的分子与创新概念真正转化为能救治患者的临床药物，是当今医学界乃至整个 AI 领域最重要的应用课题。 ``` [分子设计 (几天/下午)] ───► [临床前/动物实验] ───► [临床试验 ( Phase 1-3)] ───► [FDA批准上市] └──────────────── 多年耗时与十亿美金瓶颈 ────────────────┘ ``` --- ### [01:57](https://www.youtube.com/watch?v=Uq0Wy6uZkuA&t=117s) 嘉宾自我介绍与跨学科背景 主持人 Bryson Tombridge（Tono Health 联合创始人兼 CEO）介绍了自己的背景——Tono Health 作为全美最大的专科皮肤科平台，与众多顶级医疗系统、癌症中心以及早晚期药企合作开展临床试验。随后，四位来自不同领域的重磅嘉宾依次进行了自我介绍： * **Tala Fakhouri**（Parexel AI 与监管策略首席官）：曾任职于 FDA，领导 FDA 药物审评中心（CDER）的 AI 政策制定；目前在全球合同研究组织（CRO）Parexel 负责临床试验监管与 AI 策略。 * **Laura Boykin Okalebo**（Nebius 生物信息学与云基础设施专家）：计算生物学家出身，曾在撒哈拉以南非洲推广便携式基因组学；她指出在基层科研中“计算资源（Compute）”往往比科学本身更容易成为瓶颈，现专注于帮助全球医疗及生命科学团队部署 AI 计算集群。 * **Patrick Loerch**（Gilead Sciences 临床数据科学高级副总裁）：生物统计学博士出身，负责吉利德（Gilead）临床部门旗下的所有数据科学与分析职能，涵盖临床药理学、定量药理学（Pharmacometrics）、生物统计、真实世界证据（RWE）及整体 AI 战略部署。 * **Aman Thukral**（AbbVie 临床系统与数字运营全球负责人）：药理学家出身，在行业内拥有 19 年经验（在艾伯维任职 11 年），目前全面领导艾伯维（AbbVie）在临床开发中的 AI 落地实施。 --- ### [05:11](https://www.youtube.com/watch?v=Uq0Wy6uZkuA&t=311s) AI 带来的兴奋点与现实挑战（Speed vs. Velocity） 在首轮“头条观点”讨论中，嘉宾们分享了目前最令其兴奋的 AI 应用落地，以及当前依然严峻的行业阻碍： #### 速度与方向的权衡（Speed vs. Velocity） Aman Thukral 表示，大语言模型（LLM）等技术带来的数据处理速度和准确度令人惊叹——过去团队需要 4 至 6 周才能完成的交付物，现在仅需 1 天甚至 1 小时。但这种高效率也带来了隐患：**如果团队开始盲目接受 AI 生成的结果，而缺乏人类把关（Human-in-the-loop），可能会对提交给 FDA 等全球监管机构的数据质量造成致命打击**。 > Aman 提出了一对深刻的对比：“速率（Speed）是距离除以时间，而速度/矢量（Velocity）则是朝着正确方向的位移。我们在 AI 上已经获得了极高的‘速率’，但我们是否已经拥有了正确的‘方向’？我认为‘方向（Velocity）’仍有待达成。” #### 研发管线的“分子海啸”与临床试验瓶颈 Patrick Loerch 强调，AI 终于跨越了概念炒作阶段，进入了**可衡量价值**的落地期。吉利德的研究部门负责人形容 AI 让早期研发迎来了候选化合物的“分子海啸（Tsunami of potential compounds）”。然而真正的瓶颈依然在于后续的临床试验——临床试验三分之二的时间是靠物理运行消耗掉的，如何进行科学决策与高效运营依然是巨大挑战。 #### 自动化实验与全球化算力包容性 * **物理 AI 与自驱实验室**：Laura Boykin Okalebo 提到，生物智能化体（Bio-agents）与物理 AI 的结合正在重塑实验室，例如自动化机器人接管繁重枯燥的移液（Pipelining）工作，打造“自动驾驶实验室（Self-driving labs）”。 * **算力民主化**：Nebius 发起了一项全球倡议，收到来自卢旺达、哥伦比亚等全球 650 个团队的申请。通过云端算力赋能缺乏本地硬件和电力资源的开发者，能够确保 AI 训练数据的多样性与包容性。 * **跨学科语言鸿沟**：Tala Fakhouri 指出，工程人员、计算生物学家与生命科学家的深度协同令人振奋，但由于过去生命科学界习惯于“闭门造车”，不同专业团队之间“缺乏共同语言”成为了目前最大的合作障碍。 --- ### [10:52](https://www.youtube.com/watch?v=Uq0Wy6uZkuA&t=652s) 药物研发全流程图谱与 AI 应用分类 针对从“分子到床边”的漫长链条，嘉宾们梳理了标准研发路线图，并将 AI 的应用归纳为两大核心维度： ``` +-----------------------------------------------------------------------------------+ | 药物研发全生命周期 | | [早期探索/靶点筛选] ──► [临床前/动物实验(PK/PD)] ──► [ Phase 1 / Phase 2 / Phase 3 ] | +-----------------------------------------------------------------------------------+ │ ┌───────────────────────┴───────────────────────┐ ▼ ▼ 【科学发现 (Scientific Insight)】 【运营效率 (Operational Efficiency)】 • 提高 PTRS (技术与监管成功概率) • 自动化代码生成 (AI-assisted coding) • 肿瘤总体负荷分析 (AI Resist 影像) • 数据库锁定与临床报告(CSR)自动化撰写 ``` #### 1. 监管视角的研发阶段 Tala 总结了 FDA 监管体系下的全流程环节： * **靶点发现（Discovery）**：新靶点识别与分子设计。 * **非临床/临床前（Non-clinical/Preclinical）**：动物实验、药代动力学/药效学（PK/PD）建模。 * **临床研究（Clinical Research）**：Phase 1（安全性评估）、Phase 2（剂量与疗效探究）、Phase 3（确证性临床及对比标准治疗）。 * **上市后监测与生产（Post-market Safety & Manufacturing）**：不良反应追踪与制药工艺优化。 #### 2. 价值衡量的两类核心指标 Patrick 提出，评估 AI 赋能价值主要依据两个抓手： * **科学决策维度**：以 **PTRS（Probability of Technical and Regulatory Success，技术与监管成功概率）** 为核心。AI 帮助提升每一步科学决策的信心，确保进入下一阶段的化合物更具可行性。 * **运营效率维度**：以**时间缩短与规模扩大**为指标。例如目标是将管线中两个特定节点之间的操作时间缩短 15%，或在不增加成本的前提下将处理规模提升 20%。 --- ### [14:46](https://www.youtube.com/watch?v=Uq0Wy6uZkuA&t=886s) 实战案例研究与核心 KPI 在谈及各家企业具体的 AI 落地实践与衡量指标时，嘉宾们分享了多个生动的真实案例： #### 艾伯维（AbbVie）：缩短关键路径（IND 到 NDA） Aman 表示，公司的核心目标是**缩短 IND（新药临床试验申请）与 NDA（新药上市申请）之间的时间**，重点优化关键路径上的节点： * **系统建立**：利用 AI 加速临床数据采集系统的搭建。 * **数据分析**：在“最后一位受试者最后一次随访（LPLV）”结束后，实现数据的快速分析。 * **自动撰写**：数据库锁定（Database Lock）完成后，利用 AI 快速撰写临床研究报告（CSR）并编译通用电子文档（eCTD）。 #### 吉利德（Gilead）：肿瘤影像分析与代码生成 Patrick 介绍了两个已经落地的典型场景： * **基于 AI 的肿瘤负荷评估（Total Tumor Burden）**：在肿瘤学临床试验中，传统依靠 RECIST 标准评估肿瘤大小。吉利德引入 AI 图像识别技术，在拿到患者影像数据的瞬间快速计算总体肿瘤负荷，帮助团队在面对巨大资金投入时，更快、更精准地做出是否推进到下一阶段临床的决策。 * **AI 辅助代码编写（AI-assisted Coding）**：制药行业存在海量的一次性数据清洗与分析代码需求。通过部署 AI 赋能的 IDE 编程环境，大幅提升数据统计分析代码的编写与验证效率。 #### Nebius：算力基础设施与需求爆发 Laura 透露了算力市场的严峻现状——GPU 需求远超供给，算力云服务商必须做出战略抉择。 * **算力扩展案例**：一家名为 Prim 的生物技术公司（利用表观基因组技术研究失智症/认知障碍），在 2024 年 12 月仅试用 16 张 NVIDIA H200 GPU；而到了 2026 年 12 月，其需求已大幅增加至 1000 张 NVIDIA Vera Rubin 显卡。这一指数级增长成为了基础设施层衡量 AI 行业爆发的直观 KPI。 * **核心应用与指标速查列表**： * **IND (Investigational New Drug)**：新药临床试验申请，标志着药物从实验室走向人体临床。 * **NDA (New Drug Application)**：新药上市申请，提交给 FDA 等监管机构进行最终审批。 * **RECIST**：实体瘤疗效评价标准（Response Evaluation Criteria in Solid Tumors）。 * **CSR (Clinical Study Report)**：临床研究报告，数据库锁定后需撰写的核心监管文件。 * **eCTD (electronic Common Technical Document)**：电子通用技术文档，用于向全球监管机构提交申请的统一格式。 <ElicitationsGroup message="想要深入探讨这个话题的后续内容，您可以选择："> <Elicitation label="了解后两部分记录" query="请继续提供该研讨会 Transcript Part 2 和 Part 3 的详细记录与分析。"/> <Elicitation label="深入分析 AI 撰写 CSR 的监管风险" query="AI 自动撰写临床研究报告（CSR）和编译 eCTD 时，存在哪些合规与监管风险？FDA 如何审查这类数据？"/> <ElicitationsGroup/> ### [20:56](https://www.youtube.com/watch?v=Uq0Wy6uZkuA&t=1256s)### [020:56] 搭建全球包容性 AI 生态与临床高效落地 Laura Boykin Okalebo 表示，Nebius 最先进的垂直领域是 Healthcare and Life Sciences（医疗健康与生命科学），旗下包含生物医药、数字医疗、医学影像、医疗器械，以及一个旨在推动全球包容性的特殊项目。她分享了一个感人案例：来自卢旺达的一位合作伙伴，因姐姐患乳腺癌离世，致力于改善非洲裔女性致密性乳腺的影像诊断模型。即便将计算资源分配给这些项目，其他合作的大企业也非常认同这种跨国合作生态。Laura 强调，未来 GPU 算力终将不再紧缺，因此建立正确的合作伙伴关系、共同推进新药研发才是长远之计。 Tala Fakhouri 随后分享了 Parexel 在 AI 应用上的多个具体降本增效案例：在监管申报文件撰写（Regulatory Submissions）方面，AI 成功缩短了 60% 的撰写时间；在统计引擎和合规数据集中，AI 将生成“监管级”（Regulatory Grade）数据的时间削减了 50%。而最令人兴奋的是将 AI 用于证据生成（Evidence Generation）以挽救生命，例如在渐冻症（ALS）等罕见病临床试验中，利用 AI 打造数字孪生（Digital Twins）来预测患者在未接受治疗时的疾病进展，从而大幅减少安慰剂组（Placebo Arm ### [041:20] 推动业务流程与人员转型，避免 AI 落地阻力 在讨论 AI 落地应用时，专家们强调了高层支持与流程重塑的关键作用。例如，从高层确立明确的目标——比如要求“每一位统计程序员在年底前都要使用支持 AI 的集成开发环境（IDE）”，能为技术普及提供强有力的推力。然而，无论是初创公司还是大型制药企业，在深入推行 AI 时都不应低估阻力。现实中甚至有高达 90% 的传统流程无法直接与 AI 技术匹配。因此，企业不仅要关注技术和决策层面，更要重新审视业务流程以及员工的具体工作方式。“如果你不带上员工一起转型，这一切努力都无法取得任何实效。” - 组织在推行 AI 技术时，必须同时重构匹配的业务流程，否则大多数现有流程难以与 AI 兼容。 - 高层领导确立清晰的工具落地目标（如程序员使用 AI IDE）是推动组织转型的有效抓手。 ### [042:01] 提前规划算力资源：研发计算等同于临床试验 Nebius 的 ## Notable quotes > [05:11](https://www.youtube.com/watch?v=Uq0Wy6uZkuA&t=311s) 速率（Speed）是距离除以时间，而速度/矢量（Velocity）则是朝着正确方向的位移。我们在 AI 上已经获得了极高的‘速率’，但我们是否已经拥有了正确的‘方向’？我认为‘方向（Velocity）’仍有待达成。 — Aman Thukral > [05:11](https://www.youtube.com/watch?v=Uq0Wy6uZkuA&t=311s) 如果团队开始盲目接受 AI 生成的结果，而缺乏人类把关（Human-in-the-loop），可能会对提交给 FDA 等全球监管机构的数据质量造成致命打击 — Aman Thukral > [05:11](https://www.youtube.com/watch?v=Uq0Wy6uZkuA&t=311s) 分子海啸（Tsunami of potential compounds） — Patrick Loerch > [01:57](https://www.youtube.com/watch?v=Uq0Wy6uZkuA&t=117s) 在基层科研中“计算资源（Compute）”往往比科学本身更容易成为瓶颈 — Laura Boykin Okalebo > [05:11](https://www.youtube.com/watch?v=Uq0Wy6uZkuA&t=311s) 因为过去生命科学界习惯于“闭门造车”，不同专业团队之间“缺乏共同语言”成为了目前最大的合作障碍。 — Tala Fakhouri > [41:20](https://www.youtube.com/watch?v=Uq0Wy6uZkuA&t=2480s) 如果你不带上员工一起转型，这一切努力都无法取得任何实效。 ## Mentioned Bryson Tombridge, Tala Fakhouri, Laura Boykin Okalebo, Patrick Loerch, Aman Thukral, FDA, CDER, Parexel, Nebius, Gilead Sciences, AbbVie, Tono Health, Prim, NVIDIA H200, NVIDIA Vera Rubin, RECIST, IND, NDA, CSR, eCTD, PTRS ## Video description Patrick Loerch - Senior Vice President, Clinical Data Science, Gilead Sciences Aman Thukral - Director and Global Head, Clinical Systems and Digital Operations, AbbVie Tala Fakhouri - Vice President Consulting, AI & Digital Policy, Real-World Research, Parexel Bryson Tombridge - Co-Founder & CEO, Tono Health Laura Boykin Okalebo - Life Sciences Solutions Partner, Nebius This panel brings together pharma leaders and AI scientists to discuss the evolving drug discovery pipeline. Learn how machine learning is being used to identify new targets, design molecules, and simulate trials—all in less time and at lower cost. Panelists will share what’s working now and what hurdles still exist in AI-driven innovation. [Source](https://www.youtube.com/watch?v=Uq0Wy6uZkuA)

[5:11] Aman Thukral: 速率（Speed）是距离除以时间，而速度/矢量（Velocity）则是朝着正确方向的位移。我们在 AI 上已经获得了极高的‘速率’，但我们是否已经拥有了正确的‘方向’？我认为‘方向（Velocity）’仍有待达成。 [5:11] Aman Thukral: 如果团队开始盲目接受 AI 生成的结果，而缺乏人类把关（Human-in-the-loop），可能会对提交给 FDA 等全球监管机构的数据质量造成致命打击 [5:11] Patrick Loerch: 分子海啸（Tsunami of potential compounds） [1:57] Laura Boykin Okalebo: 在基层科研中“计算资源（Compute）”往往比科学本身更容易成为瓶颈 [5:11] Tala Fakhouri: 因为过去生命科学界习惯于“闭门造车”，不同专业团队之间“缺乏共同语言”成为了目前最大的合作障碍。 [41:20] 如果你不带上员工一起转型，这一切努力都无法取得任何实效。

[22:15] Tala Fakhouri: Think about rare diseases like ALS, conditions where it's very difficult to actually find patients and then enroll them in placebo. We can use AI... to model what a patient's response would be had they not taken the drug. [43:56] Laura Boykin Okalebo: I dare people to email me. I will email every single one of you back. I'm legit. I'm old school.

[0:04] Good morning everyone. AI is transforming  the world. Biotech and pharma is certainly [0:12] no exception for that. It's changed forever.  One example of that is with AI molecules can be [0:20] designed in an afternoon. That's amazing. This  solves so many problems for humanity, right? [0:29] Well, last month I was talking to a good friend  who's much smarter than I am. Uh, Stanford PhD [0:37] postback who recently sold an earlystage  biotech company for more money than I can [0:42] count. He told me the problem is not solved there.  That there's a lot more from molecule to bedside. [0:51] What is so interesting in some of our biggest  bottlenecks today is it cost over a billion [0:56] dollars for every treatment and medication that  is FDA approved going through the entire process [1:03] and that process can take over 10 years. This is  perhaps the biggest story happening in medicine [1:12] today is how do we close that gap so that  these new ideas these new molecules these [1:18] AI innovation [clears throat] innovations can  actually reach patients. I would even argue this [1:25] is the biggest most important application of AI  today. This session is the most important session [1:32] at AI4 because this is what impacts all of us and  humanity. Who cares all the other things that AI [1:40] might impact if humans are not able to understand  ourselves and cure ourselves? So, welcome to from [1:48] molecule to bedside. How AI is transforming um  clinical uh early stage development. I'm Bryson [1:57] Tombridge as uh I was introduced the CEO of Tono  Health. We are the largest specialty dermatology [2:03] platform in the country. Uh we partner with health  systems, cancer centers. We have the most advanced [2:09] AI platform being used in clinical practice  and dermatology today. And most important to [2:14] this conversation is we partner with both early  stage and late stage life sciences and pharma [2:20] companies on clinical trials. And the reason  I uh phrase it as such is because I was blown [2:27] away by the panel that AI4 brought together. They  brought together such an incredible group to walk [2:33] us through early stage development to bedside,  every stage of the journey from startups to late [2:38] stage. Um what is working, what is not working and  what needs to change. Um so without further ado, [2:46] let's jump in and meet our panel. Um Talia,  could you maybe uh introduce yourself in 20 [2:53] 30 second introductions and then we'll we'll go  down. Sure. Um, first to the folks in the back, [2:59] the timer is not working. We can talk for three  hours, but I just want you to know that it's [3:04] not working. Um, hi everyone. Tala Fakuri, chief  AI and regulatory strategy officer for Paracel. [3:12] Paracel is a clinical research organization  that runs clinical trials across the globe. [3:18] I've been with Paracel for a little over a year.  Before that, I was at the US FDA where I led the [3:23] development of AI policy on behalf of the  center for drugs. Hi everyone, my name is [3:29] Laura Boyinokaleo. I am a computational biologist  by training. I was an academic for a while and [3:37] um took a lot of portable genomics and compute  to subsaharan Africa. So have spent a lot of time [3:46] on the ground and it's never the science that I  found was the problem getting out in the field. [3:51] It was always the compute. So I have now joined  Nebus which is a really awesome AI infrastructure [3:56] company cloud company that helps healthcare  and life science people run their workloads [4:01] AI workloads. So I'm really happy to be shaping  that um with partners all around the globe. So [4:08] happy to be here. Hi everyone. Thanks for having  me. My name is Patrick Lurch. I am senior vice [4:14] president of clinical data sciences and what that  entails at Gilead, it's it's basically any data [4:20] science or analytics function that sits under the  clinical organization. So spanning from clinical [4:25] pharmarmacology, phicometrics, statistical  programming, biostats, real world evidence [4:30] as well as AI and the overarching AI strategy.  Um I was a biochemist once upon a time and not [4:36] a very good biochemist. So I moved into data  science, got a PhD in bioatistics. Hi guys, [4:42] my name is Aman. Uh I work for ABV. Uh by training  I'm a pharmacologist. Uh but I head our clinical [4:49] systems and digital operations. Uh and part  of my current role with this uh AI revolution, [4:56] I'm leading AI implementation uh for ABV in  clinical development. Our goal is to accelerate [5:04] uh clinical development. That's me. Amazing. Thank  you everyone. All right, let's start with the [5:11] headlines. So, let's keep this maybe to a minute  each right now and then we'll go a lot deeper in [5:16] the future. Um, what excites you the most? What  is the most exciting AI implementations or impact [5:23] you're seeing today? And then what are the hurdles  that you also see um still exist? Uh maybe Aman, [5:30] do you want to kick it off for us? Yeah,  excitement. Uh I'm in data science business. [5:36] This is my 19th year is in in industry and 11th  year within ABV. So I think the the kind of speed [5:46] and accuracy what we are dealing in data business  with the help of these uh modern large language [5:53] models uh I think I never imagined uh that uh you  know there were [clears throat] deliverables which [6:00] my team was delivering in four weeks 6 weeks uh  we are talking about one day or one hour nowadays [6:07] uh um that is both exciting and it scares me a  lot uh as Well, um what if my team starts taking [6:17] AI on its face value without true being human  in the loop because our deliverables directly [6:25] go to FDA or regulatory authorities around the  world. Uh can I put the right quality gates? Uh [6:32] so it's both exciting and scary. My one minute.  So velocity is the most exciting piece as well as [6:38] the hurdle being um with that velocity are the  gates and the safety protocols still in place. [6:43] You know um I think I'm glad you used the word uh  um in my experience. I usually differentiate speed [6:51] and velocity. Speed is obviously mathematically  distance divided by time. Velocity is exactly [6:59] the same distance divided by time but in the right  direction. We have achieved speed with AI. Have we [7:08] achieved the right direction, the right velocity?  I think uh velocity is yet to be achieved. Speed [7:14] is achieved. Interesting. Interesting. Patrick,  what are your thoughts here in one minute? Um [7:19] okay, we'll have time to go deeper. I just want  to get the headlines out of the way. I I think the [7:24] headline is we finally reached a point that we're  creating value and we can actually measure value. [7:29] Um, so I I think there's I was an AI naysayer  several years ago. I mean, there's hype and hype [7:36] and hype. I would say we've finally reached  the point where I'm seeing tangible examples [7:41] across the entire development organization and  in research where people are able to measure [7:47] value in terms of our core purpose. Moving  drugs through the pipeline with greater scale, [7:53] greater speed. we have this the head of research  likes to call it the tsunami of of uh potential [8:00] compounds that are coming from research. Um and  I think that gets to the second part of your [8:04] question that the big hurdles um you still have  to run a clinical trial you know and twothirds of [8:10] the time of a clinical trial is actually running  the clinical trial. So I think we will we will get [8:15] into it but there's the operational and sort of  the scientific decision making aspect of that. So, [8:20] I think there's hurdles along the way, but it's it  is a really exciting time to see impact. Amazing, [8:25] Laura. Yeah, sure. I think that um for me, the  most interesting thing that's happening right now [8:32] is watching like some of the agents infiltrate  into the the lab, let's say. You know, anybody [8:39] who's done any molecular lab work knows that pipe  heading is just not really that much fun. Um so [8:46] it's really interesting to see like these biono  agents being deployed with the you know we call [8:54] it self-driving labs you know with the physical  AI so that intersection is really interesting and [9:00] I think the biggest challenge another thing that  excites me is we are able to create an ecosystem [9:08] of global scientists right so at Nebius we did  this call and 650 people applied and we were [9:16] able to incorporate and include teams from Rwanda  and teams from Colombia and teams from all around [9:21] the world because as we develop the models we need  to make sure that we're being inclusive with the [9:26] data that's in there so it applies to everyone.  So it's a really fun time to be working out how [9:33] to do this globally because a lot of places don't  have the advantage of having power or resources [9:39] to run on-prem things. So being in this really  interesting space of cloud computing, developing [9:45] that globally with all the voices is really  fun and cool and a challenge as well because [9:51] we need to all participate in that. Wow. What a  scientific pioneer. Um Tala, would you like to [9:57] uh Sure. wrap this this introduction up. C can I  ask a question of the audience? How many of you [10:03] are in the life sciences? Yes. Um, one of the most  exciting things to me now I saw about maybe like [10:13] 60 70% life sciences is um the collaboration  the multidisciplinary teams that we're seeing [10:22] form where you have engineers, data scientists,  computational biologists working closely with the [10:27] life scientists because for the longest time um  life scientists think that they know the answer [10:33] to everything and we don't. That's why drug  development generally is very inefficient. So [10:40] this work across disciplines is what excites me,  but it's also where a lot of the hurdles are. We [10:45] don't speak the same language. Sure. Um speaking  of getting everyone on the same page and aligned. [10:52] um as we go in and we double click on all the  different steps from molecule to bedside um could [10:58] you maybe give us a road map of like high level  some of the things that need to happen and then [11:03] uh Patrick maybe you can jump in here and talk  about uh the buckets of how how you think about [11:08] AI and then we'll have common definitions that as  a group we can use for for future conversations. [11:14] Um so in 2019 I was at the FDA. I was just  entering FDA and was asked to do some work for AI [11:23] policy and I was like why why do we need AI policy  and part of it was you know it was the era of the [11:29] hype. So one of the first things in my previous  job that we did was start to track the number of [11:36] submissions that FDA was seeing across all stages  of drug development. So it was quantifying the [11:43] number of submissions that are in discovery like  the keynote speaker this morning about you know [11:48] discovering new targets new um therapeutics then  entering nonclinical or preclinical so all the [11:56] animal testing PKPD then clinical research then  postmarket safety and then also in manufacturing [12:04] and it became very clear at that time that it's  not hype there were AI uses that were submitted [12:10] to the regulators across every stage of drug  development including in areas that FDA does [12:16] not regulate like discovery. So across that entire  spectrum from you know molecule to market AI can [12:25] be used in two ways and I'm going to pivot to  you to talk about them but generally scientific [12:31] discovery which is very exciting. We're becoming  much smarter but also operational efficiency. [12:37] All the things that we do really slowly really  inefficiently can now be accelerated with AI. [12:44] Yeah. and maybe um taking it just one level down.  The way I like to think of it, research, you know, [12:50] from the pre-clinical space has to deliver three  things. They have to deliver an indication, [12:54] maybe a subopul that you're going to go after,  a compound, which is I think where a lot of [12:59] the advances in AI have happened, and a first in  human dose. Um, as that then comes into clinical, [13:07] we then effectively validate those between  the phase one trial looking at safety, [13:12] the phase two trial assessing dose and response  and efficacy and the phase three trial finally [13:16] producing the the data set that gets the compound  approved relative to standard of care. The two [13:23] angles that I like to look at those. So those are  sort of the scientific deliverables and decisions [13:27] you have to make. Operationally the question  is how can you do that as fast as possible with [13:32] the greatest scale you know so when we start  looking scientifically we measure the impact [13:38] of AI based on PTRS. So for the other 40% of you  that's probability of technical and regulatory [13:45] success. So it's how confident are you in the  scientific decision you're making to move to [13:50] the next stage and how confident are you that it  will be approved. And then operationally all of [13:56] our AI projects have goals like this will speed up  this task between these two points in the pipeline [14:02] by 15%. Or require you know 20% scale increase you  know through the use of AI. So it's as I said it's [14:12] very very measurable value that we've gotten to  at this point in terms of the use cases and that's [14:17] at least at Gilead that's really how we sort of  prioritize where we use it. Amazing. All right, [14:21] so I'm going to fly back up to 30,000 ft. And  my takeaway was there's really two breakdowns. [14:27] There's how can we use AI to make it more  operationally effective just to move faster and [14:33] then there's the clinical insights. What are the  clinical knowledge that we're gaining from using [14:39] AI? Um Aman, let's now jump into the real heart  of our conversation. And I think while everyone [14:46] came to this session, let's start using or let's  start discussing case studies, applications of AI, [14:53] KPIs like how are you actually seeing AI um impact  your your day-to-day and your work. Yeah. So like [15:01] Patrick mentioned uh once the molecule is shipped  from discovery or research to development team [15:11] uh we have set milestones or set journey phase  one phase two phase three um the program which [15:18] I'm working on is very specific accelerating time  between phases phase one phase two to phase three [15:26] um oh yeah and and within phases how we can um  reduce the time for deliverables. Overall the [15:38] goal is to reduce the time between IND and NDA.  IND is usually investigational new drug uh which [15:46] is filed to regulatory authorities once we start  clinical trials and NDA new drug application or [15:52] um is filed to FDA or regulatory agencies for the  drug approval. Now how we are doing it? Um we drew [16:01] the critical path and for every node between  the critical path we have decided that where AI [16:12] can be applied. So three or four examples are the  things which are on critical path. Can we set up [16:19] our systems faster to collect the data? That's  one thing. Second, can we set up our systems [16:25] to analyze the data faster? Once uh last subject  last visit is completed and last but not least uh [16:34] once your database lock is done, can you quickly  author clinical study report and compile an ECTD? [16:42] So that's how we have completed the journey  and the critical path items we're trying to [16:48] reduce the time between it. Now different pharma  companies are have different balance sheets. Uh [16:55] our goal as a company is not to focus too much  on um the return on investment. Our very hardcore [17:04] goal is reduce the time between the phases and  reduce the time within the phases. That's what [17:09] we are working on. Excellent. Uh Patrick, can you  speak a little bit to what real examples of what [17:15] you're doing? Yeah. No, so ROI is a good point.  So we we intentionally are not prioritizing like [17:21] cost savings. So yeah, I think similarly sort of  speed to market maybe [snorts] we have several [17:25] use cases. Before I describe them, there's two  lovely gentlemen right here in the front aisle, [17:29] Yousef and John. And everything I talk about  they actually do. I'm basically a talking head [17:35] up here at this point. So just clarifying  that. Um two maybe two different examples. [17:41] Um you know one I really like is in the imaging  space is people from the field know about resist [17:48] which is how you assess the basically efficacy  of an oncology treatment based on tumor size. So [17:53] one of the active engagements we have both  externally but also internally is building [17:57] out a total tumor burden based on AI and the  idea similar to what Aman said is really more [18:03] rapid decision making. you know, the moment we get  those images back, we can very quickly turn around [18:09] um a decision and a decision whether or not  to move to the next phase um which is a major [18:14] investment and probably a better decision. Yes,  some of us also think so. Um the other which may [18:21] be more familiar to everyone here who's not in  life sciences is just coding. We do a tremendous [18:27] amount of oneoff programming in pharma. So it's  single use writing code to format data, analyze [18:35] data. They have entire statistical programming  organizations that do this. So AI assisted coding, [18:41] AI enabled IDE is another area that we're  investing extremely heavily to just generate [18:47] code that can be validated much much quicker.  Once again, the goal is to move, you know, [18:52] the information through the pipeline to enable us  to make the next decision quicker. Great. Laura, [18:58] how are you um applying AI and what are some case  studies and KPIs and ways you're thinking about [19:04] it and measuring it? Yeah, you know, it's an  interesting question because the perspective um [19:09] that we have at Nebus and I think a lot of people  in the room probably have is access to GPUs is [19:15] hard to find, right? I mean that that's that's a  given. And um so we see this the supply the demand [19:24] is way outpacing the supply and so we have to make  decisions on who we are going to sell to right and [19:35] especially in this space we're going up against  so internally as a company we have to really [19:39] think about um you know what are the long-term  benefits of what are the people working on [19:45] um does this make sense to to build the ecosystem  around the cloud computing and so we like to build [19:54] with people and a metric is like okay so like we  have a a company called Prim and they came to us [20:01] and they started out with 16 H200s right that was  December 2024 they just wanted to give it a go [20:08] okay fine and then over time we've partnered with  them and now December 2026 they want a thousand [20:16] Vera Rubins right so that is a clear metric  that we're moving in the right direction and [20:22] partnering with, you know, a really important  company called Prim who is looking at dementia [20:27] and using the epiggenome to do that. So, it's  a really interesting space um trying to provide [20:33] the infrastructure for all of these cool things to  people that don't necessarily have the enterprise [20:40] um behind them with the the budget. So, it's it's  it's really fascinating. Can I uh push a little [20:46] harder there and ask you to explain more? Is there  an element of you're deciding what use cases and [20:52] what disease states that you as your organization  wants to prioritize or you just want to work with [20:56] good people or is it their chance of success or  how what is what are those frameworks for? Yeah. [21:01] So there's at Nebius we have the most um advanced  vertical is healthcare life sciences and we have [21:09] like five categories that we focus on and biioarma  digital health medical imaging medical devices [21:16] and then we also have this category of special  category where we really want to make sure that [21:21] we're being globally inclusive. Um so we so like  for example I'll tell you that one of our winners [21:27] in that category is from Rwanda. his sister died  of breast cancer and he wants to make sure that [21:33] and we know that African-American women have  denser breasts but do the images in the models [21:39] that are currently there actually incorporate that  so I feel really proud of that and then the bigger [21:44] companies that we work with are like oh how great  you have partnerships in Rwanda like so it's about [21:50] building the ecosystem and being and not just  saying highest bidder wins you know what I mean [21:55] like it's it's yeah because eventually GPUs will  be in supply right at some point and did we make [22:02] the right decisions with the partnerships to to  build this in a way that's going to push the the [22:08] the drug out you know what I mean so it's a it's  a fascinating place to be it is absolutely tala um [22:15] what would you say is uh a concrete example you  can give us of of applying AI yep I can uh give [22:21] a couple of examples from uh Perl where I work now  um operational efficiencies like the authoring of [22:29] regulatory submissions. That's a lowhanging fruit.  I think a lot of us are doing this and we're able [22:34] to quantify efficiency by about 60%. Right? 60%  cut in uh the time that it takes to author. Then [22:43] also in statistical engines type of work, the  creation of data sets that are regulatory grade [22:49] ready for a submission, we're also able to cut the  time by 50%. The part that I'm most excited about [22:56] is the use of AI in evidence generation that  can also cut time and actually save lives. So, [23:02] we work with a lot of our partners to explore  if they can use a digital twin in their clinical [23:08] trial to reduce the number of patients needed in  a placebo arm. Think about rare diseases like ALS [23:15] conditions where it's very difficult to actually  find patients and then enroll them in placebo. [23:22] We can use AI. We could use foundational models to  model what a patient's response would be had they [23:29] not taken the drug and using statistical methods  were able to generate, you know, estimates and [23:36] um efficacy estimates that would be acceptable by  the regulator. So these are some concrete things [23:41] that we're doing. All right, let's double click  on digital twins. Um just last week a co-orker [23:47] of mine was telling me, "Oh my gosh, you're going  to AI4. there's Alex is like this uh who was just [23:53] the keynote in the the main stage before the  work he's doing is amazing digital twins is [23:58] changing everything and I said I actually don't  think I understand what a digital twin is and why [24:03] everyone's so excited so would you um expand there  a little go a little deeper I don't think Alex [24:08] is here but I don't think they do digital twins  it's mostly like discovery like discovering new [24:14] compounds um which is very important they've been  able to show that they can cut timelines there [24:20] Yes, digital twin means if I ask here what do  you think a digital twin is? We'll probably get [24:26] 10 different answers. Um digital twin the way  that we use it is a statistical method called [24:34] prova that enables you to again reduce the number  of patients needed in a placebo arm. Other people [24:41] think about digital twins as synthetic controls  or external controls using um EHR data that is [24:48] different and I'm kind of using the definitions as  accepted by the regulators. Again, I you basically [24:54] develop a foundational [snorts] uh model usually  for a specific therapeutic area. Then you take the [25:01] patients baseline information from a trial.  Ideally, it would be phase two. you start [25:07] modeling what their placebo response would be and  then you can use this information in a pivotal [25:13] trial to reduce number of patients needed in a  control arm. So reading between the lines there, [25:21] the common bottleneck is recruiting patients  for clinical trials and finding and recruiting, [25:26] which is also where AI can help, right? You could  um you could basically do deep dives of EHR data, [25:33] do EHR phenotyping, match patients to trials u  more efficiently and then um also utilize digital [25:42] twins and things like that. But the issue there  is the availability of data. Um data still sits [25:50] in silos which makes it very difficult to develop  some of these twins. It makes it still difficult [25:55] to do a lot of the EHR phenotyping needed for  recruitment. This is why other countries and [26:00] regions have a better uh time doing this than  in the US because our healthcare system is so [26:06] fragmented and our data is you want to jump in?  Oh no maybe a little caveat. I think it's not [26:11] just only about reducing the number of patient.  It's just also yeah being more accurate about the [26:16] effect size you're going to expect to see within a  trial. So it's you know a lot of time it's really [26:21] difficult without doing some sort of digital  twin or simulation or using real world data [26:26] um to actually estimate how many patients will  you need. Um a lot of studies is notorious the [26:32] placebo arm does better than you think it's going  to do. Um you know everyone smiles a little bit [26:38] on stage when I say that. So yeah, it is a it  is an extremely valuable approach, but I I do [26:42] agree with the comment. It just it is a term  that's losing its meaning a little bit. And so [26:48] whenever I hear digital twin, it's okay, digital  twins for what? Um we do the same thing for like [26:52] digital twins to improve the accuracy of first  human dose. Um and these sorts of things. Oh, [26:59] sorry. I'm used to a lavalier. So Aman, did you  have thoughts or want to jump in here? Yeah. No, [27:06] I think digital twin is a broad term. Uh I think  the example given uh by Patrick and other fellow [27:14] panelists uh uh we are envisioning digital  twin in clinical monitoring. What it means is [27:21] uh we send our sponsors clinical monitors to  these hospital sites to monitor the study and [27:29] they do four things. uh a lot of it is a very  operational and tactical work matching the [27:35] data from EHR in literal terms it is called SDV  source data verification then they [snorts] do lot [27:42] of drug accountability uh then they check lot of  compliance related documentation and last but not [27:49] least uh they which is the most important thing  they do they have this relationship with physician [27:56] or the principal investigator and site coordinator  and nurse We want our monitors to focus only on [28:03] the fourth item. The first three, the tactical  ones. Can a digital twin do it for you? It's I [28:11] think it's the one thing which we are considering  uh within clinical operations. But if I give you a [28:17] very loose example, if you want a digital twin of  yours, can you spend quality time with your kids [28:24] and spouse? all all of your other work gardening  cooking dishing can your digital twin do it I [28:32] think that's what we are trying to figure out the  quality time with same thing in clinical research [28:38] can the real scientific work or the relationship  work be done by actual humans and remainder of the [28:44] work can it be done by your digital clone or twin  is the ultimate goal and there are I think you [28:50] heard several applications like external control  arm which he mentioned and Patrick mentioned on on [28:56] the dose and I I mentioned on clinical monitoring.  Excellent. Uh Patrick, earlier you mentioned the [29:02] tsunami. Um maybe let's just go a little deeper  there. I think you you have everyone's attention [29:08] with that phrasing. Yeah, I I think um we  like most companies well was mentioned in [29:15] in the keynote and I will try to speak into my mic  more clearly. um we're just seeing the timelines [29:21] to research shrink which is a great thing and you  know over time you know being on the clinical side [29:27] of the house it it is also one of the biggest  stressors I have because you have you have AI [29:32] models that can actually significantly reduce  the time to actually getting things from you [29:37] know an initial idea through all of our compound  screening optimizing across every single variable [29:43] you would need to for that specific compound and  that indication actually showing it would dock [29:48] in the protein, we can actually do the protein  folding. So all of that work dramatically shrinks [29:53] down the pipeline and then we sit on the other  side of that and you know our objective is really [30:00] to try to do the same. Um, so I I would love a  future where everything bar the actual recruitment [30:06] and monitoring of the patients is basically  democratized and automated like because there's so [30:11] many manual steps involved, you know, getting to  the data, doing the mapping, doing the cleaning, [30:17] doing the querying, writing the code, you know,  doing ad hoc analyses and and these are basic [30:23] computer programming problems with you need the  scientific, biological, clinical expertise on top [30:28] of that. But if you know if we continue to move in  this direction ideally we can significantly shrink [30:34] the clinical timeline as well as the resources  so that we can all do more because I think you [30:39] know every decision we make to bring in a new  research compound is an investment decision and [30:46] we're we're always faced with more opportunities  than our resources allow us to handle. So, I mean, [30:52] one of the huge benefits that I get excited about  with AI is actually the ability to do more and [30:57] bring more, you know, medicines to patients who  need them quicker. Amazing. Can I be a little [31:02] snarky? Please do. That tsunami to me is like the  line today to get the badge of all these new newly [31:10] discovered compounds and molecules getting stuck.  Did you just call clinical registration? Yeah. [31:16] Yes. [laughter] Um it so I think this is the fear  at least also for CRO's of all of these compounds [31:24] coming in that they will eventually have to  be tested. There's not enough um patients. [31:29] There's not enough resources and the the clinical  development itself is still very slow because you [31:36] can only accelerate human biology so much. that  I think is the is the worry and this is where [31:43] um it's very important for glob drug development  to remain a very you know a global enterprise [31:50] with trying to find patients everywhere.  So safe to say uh oversimplification but [31:58] uh there's a lot more velocity and bottlenecks  have been removed on the front end but clinical [32:02] trials basically stay the same. Has any like how  is AI impacting clinical trials? They're faster. [32:10] Things are definitely getting faster but again  remember that um right now regulators expect [32:16] you to have data from humans. We can't just like  simulate everything. Maybe in the future we can. [32:22] So that is still going to be a rate limiting step  to some extent. And one one area where I also pay [32:32] a lot of attention to is um it's the translational  space. Our ability to translate the results that [32:38] we see in animal models. It it it's not great.  It's really quite poor. So I I think one big [32:46] opportunity that I'm always looking for is the AI  company that can improve our ability to translate [32:52] dose safety efficacy from the animal models into  humans because you can imagine the point of those [32:58] first trials is to make sure this is safe. Yeah,  you can't speed that up. And until we're able [33:04] to actually improve our first inhuman dose, our  therapeutic area window projections, the safety, [33:09] the offtarget effects, all of those things that  go into that, I think we'll we will continue to [33:13] be in a place where the clinical trial step is  is slower than a lot of people would like it [33:18] to be. I think this is the most important piece  of this panel. Like seriously, that translation [33:24] is the huge huge problem. Aman, do you have  a perspective on this? though I think uh like [33:32] uh Patrick mentioned biology and the translation  from preclinical to clinical results or simulation [33:40] if that can be expedited it is it's going to  be remarkable for the humanity but at our end [33:47] purely in clinical development we can only um move  the needle by few months but few months are also [33:55] uh going to benefit the humanity quite a bit but  if you want to move by few years then preliminary [34:02] biology or early biology how the translates from  pre-clinical to clinical I think that's going to [34:09] be the key or late rate limiting step excellent  Laura do you have that I wanted to say that um [34:15] our one of the things that we keep our eye on at  Nebus is the compliance issue right because you [34:21] guys have your internal systems that are all you  know HIPPA compliant and if there's human data but [34:27] then that's you know you have this big tsunami of  startups in the drug discovery space and they're [34:35] looking for a place to test their models and  how do they do that? And so one thing that is [34:42] um been interesting is making sure that a big  company like Nebius that we make sure that [34:48] compliance around human data and all of those  things is in place so that people can then you [34:54] know collaborate with the the bigger enterprise  companies. So that's been a really interesting [34:59] thing to keep like for me to keep my eye on is  like okay how do we help the non-enterprise people [35:07] that are coming up be in this space with the the  compliance issue. So that's been an interesting [35:13] part of and I think that the more we can help lift  some of that the faster things will go. Can we [35:19] double click on that a little bit because earlier  we talked about compute being the bottleneck for [35:23] these early stage biotech startups. Um do you also  see regulatory and compliance and like building [35:29] these global ecosystems as another bottleneck or  how can you speak a little more to that? Yeah. So, [35:34] you know, I mean, one of the interesting things  about um building clusters for AI is, you know, [35:44] this talking between the actual GPU and  where the data is stored, right? So, really [35:50] um making sure that the storage near the compute  is encrypted so that we tick the HIPPA compliance, [35:56] right? So there's a lot of configurations of data  centers that can happen and being mindful of that [36:04] helps things move faster, right? Like if you're  going to have to keep your data in a in a whole [36:11] another country than where you're trying to  do the compute, that's not your model. You're [36:14] not going to get any efficiency out of, you know,  getting the latest GPU. So I think it's designing [36:20] the data centers to incorporate encryption on the  near the compute is going to be really important. [36:27] So we try to influence and and let people  know at in our particular company like hey [36:33] let's make sure that we're designing and building  for life sciences and healthcare. So I think that [36:40] that's been an interesting navigation because us  doing our workflows is not the same as chat GPT, [36:46] right? Like how do the chips actually run as you  get more complex? Are they great for us or not in [36:52] this space? So yeah, great. Um Tala, we can you  wrap the can you bring this home for us in our [36:59] last we have about eight minutes left. Um how does  the FDA think about this? We now have clinical [37:05] trials, data compute happening globally. There's  more data than there's ever been before. There's [37:11] digital twins. Um, can you give us maybe some just  examples of how the FDA might be thinking about AI [37:17] in uh 2026? Sure. Disclaimer, I don't work for  the FDA anymore, so I'm not speaking on behalf [37:23] of the FDA, but I can um tell you what, you know,  FDA has said publicly repeatedly over the the past [37:30] several years. Um, generally regulators are very  open to the use of AI. most AI uses that we're [37:39] seeing right now in drug development actually  um are outside of the regulators oversight. So [37:45] again I mentioned discovery if you think about  authoring any application of AI that is purely [37:52] operational sits outside of regulatory oversight.  What I'm seeing now working with a lot of sponsors [38:00] is a lot of confusion about where FDA's oversight  sits. So there's many assumptions out there about [38:08] what regulators want to see which I think is  very problematic. So I've been preaching the [38:14] following for for the past uh year plus. Um  don't assume that your AI applications sit [38:23] within the regulatory oversight. do not use  uh traditional artifacts that were meant for [38:33] deterministic computer models of the 80s and  90s to try to validate your AI applications. [38:40] A lot of companies are spending so much time,  so much money trying to do things called CSV, [38:46] computerized software validation of their AI  applications, which is not what the regulators [38:51] ask for. So it's causing deaths of hundreds of AI  applications in drug development. Regulators are [39:00] generally open. You obviously have to be compliant  with certain regulatory requirements. In Europe, [39:06] you have GDPR if you're using patient data, for  example. Um but also follow the flexible risk [39:12] proportionate guidance that the FDA and others  have um have emphasized. Amazing. All right, [39:19] we have about 5 minutes left. So, let's um end  like we started, which is a massive topic in 1 [39:25] minute or less. Um wrap up um the key takeaway  that you would like the audience to take home [39:31] about AI or your perspective. Um Aman, maybe  you kick it off and we come down the line. [39:39] I think I will say that uh um you know you need to  prioritize what you want to focus on and if your [39:48] goal is uh productivity or efficiency what we have  observed it initially it goes down significantly [39:55] before it comes up. So um brace yourself for that  uh uh if ROI is purely your focus in our case [40:04] that was not the scenario. So that was the one  lesson we learned hard way. Uh second thing is [40:11] u um there are two or three things which you need  to be very mindful is uh in specifically clinical [40:17] development um AI hallucinations it it happens  a lot and then how do we manage the compliance [40:25] uh around it and last but not least uh uh one year  back we never considered our token usage at all. [40:35] we just went on um doing it but now all of us have  been taught a chapter on token economics which now [40:45] I get it at least uh so I'm very mindful of uh  how we design certain things and how we strive [40:52] for certain outcomes. Thank you Patrick. Um yeah  there's so one topic we didn't cover is I I think [40:59] the technology is proven I think finding the  right use case you know is a little bit of [41:04] a challenge I think what we're going through  within Gilead in our lessons learn it's it's [41:10] the execution piece I think you made the comment  about that the problem isn't the technology it's [41:14] not the AI it's the processes it's the change  management it's actually incorporating it and [41:20] getting really top level buyin that you know what  every statistical programmer camera as an example [41:25] is going to use an AI enabled IDE by the end  of the year. And and so I think the big thing [41:30] especially whether it's startups or whether you're  in a pharma adopting this is don't underestimate [41:36] the push back that you'll receive just when you  start digging in and and guaranteed 90% of your [41:42] processes don't work with AI. So just you really  have to think through not just the technology and [41:47] the AI and the decisions but the processes and  how the people actually work because if you don't [41:53] bring the people along with you know none of this  actually bears any fruit. So amazing Laura. Yeah [42:01] you know I think I would say plan your research  computing in advance right I think that that's [42:08] a huge thing. plan it like you would plan a  clinical trial because if you're in a startup [42:14] um a lot of the VCs are asking oh you might  have a great plan but if you have no access [42:19] to the compute that plan is going nowhere and it  is such a tight tight tight tight market to get [42:26] GPUs like start engaging with people early and be  open to different types of capacity of the Nvidia [42:35] chips right like you might want an H100 but you  know, one of the other blackwell chips might be [42:42] better. So, I think it's my take-home point is  the research computing planning is as important [42:47] as anything else in this day and age, be it from  startups to the enterprise. Like thinking about [42:54] that ahead of time is going to set you forward.  Bring it home for us. um don't do trickle down AI [43:03] where you know leaders and organizations just tell  everyone go use AI and then everyone is supposed [43:09] to figure it out or roll it out like an IT you  know roll out of a specific program there has [43:16] to be again this multidisciplinary collaboration  [snorts] um getting people who are actually doing [43:22] the hard work collaborate with data scientists  and engineers to figure out the best solutions [43:28] for the problems they have from my experience  this is how you get buy in if there's no process [43:33] to actually formalize citizen developer you end up  with a lot of very frustrated people being told to [43:40] use AI but they don't know what they're supposed  to do with it and then at the same time they're [43:46] being told to measure the return on investment  it's very confusing so get citizens in your [43:52] organizations involved in the building of tools  that are helpful for them amazing so this was a [43:57] fascinating conversation assuming that the people  in this room came to this session because they [44:02] like having these types of conversations and they  [snorts] want to connect with you for the rest of [44:06] the session uh or while we're here in Vegas. Um  obviously the AI4 app has a networking feature [44:13] um in our last maybe 40 seconds can we  just not to maybe I'm oversharing here but [44:20] um what is the preferred way for people to  reach out? Is it through the AI4 app? Are [44:24] you going to be around for the next day or two?  Um, how would you recommend people connect with [44:29] you here at the conference? I'll be at Starbucks  the whole time. I'm on the app. I'm on LinkedIn. [44:36] You can email me. It's laura.io nebius.com. Or  you can I'm not guys. Bring it on. I dare people [44:45] to email me. I will email every single one of  you back. I'm legit. I'm old school. [laughter] [44:52] I will also likely be at a Starbucks, but you  can also get me through the app. So I think [44:57] LinkedIn and app both are good for me. Thank  you. I'm on Instagram. Yeah, [laughter] guys, [45:04] we're in. We have to modernize ourselves.  Excellent. You will not find me at the [45:11] smoke fil slot machine. [laughter] Well,  that wraps up from molecule to market. [45:15] How AI is reshaping drug discovery.  Thank you everyone for your attention.