Gilead, Flagship Execs Warn AI Target Models Miss Human Drug Risks

吉利德与Flagship高管警告：AI靶点模型忽视人体药物风险 在生物制药行业将重心从狭隘的计算机模拟靶点结合转向人体转化药理学并打破专有数据孤岛之前，计算药物研发仍将面临90%的临床失败率。 专家小组探讨了为何AI目前在数据丰富的前期靶点发现中展现出强大效用，却由于专有临床数据壁垒以及高达90%的残酷临床试验失败率，导致其在研发下游有效性急剧下降。为实现投资回报最大化，发言嘉宾主张通过反向转化医学、临床药理学数字孪生以及前置证据规划，弥合临床前到人体的转化鸿沟。他们指出，成功落地需要打破企业内部壁垒，将计算科学家直接嵌入湿实验室并培养“双语”跨界研究人员，而非依赖孤立的AI卓越中心。最后，专家们告诫要警惕“闪亮玩具”陷阱，即切勿仅专注于优化计算机模拟的靶点结合亲和力，否则可能忽视整体成药性、毒性及外源免疫原性等风险。 尽管早期研究吸引了最多的算法关注，但加速开发阶段并提升技术与监管成功率（PTRS）能带来最高且立竿见影的运营杠杆效益。 (3:18) 反向转化医学（Reverse translational medicine）利用真实世界临床数据、电子病历（EHR）与生物标志物，将人体临床见解直接反哺至早期靶点识别，以定向实验取代数百万次盲筛。 (8:18) 随着研发管线向下游推进，AI 的有效性呈现单调递减趋势；进入临床试验的候选药物仍面临约 90% 的失败率，主因在于转化与临床数据被锁在企业专有数据孤岛中，严重阻碍了预测建模。 (13:06) 将生成式 AI 应用于综合证据规划（IEP）与监管呈交流程，可将通常长达 10 至 15 年的药物开发周期缩短至多两年。 (17:42) 自上而下的企业 AI 行政命令往往以失败告终；成功的整合需要跨学科沉浸式培训，例如让数据科学家轮流进湿实验房学习建库测序，并鼓励实验人员编写脚本。 (21:50)

Industry panelists argue that while artificial intelligence has accelerated early target discovery, its broader impact on drug development is constrained by proprietary data silos and the translational failure between preclinical models and human biology. Maximizing return on investment requires shifting focus toward downstream operational efficiencies, cross-disciplinary integration, and predictive clinical pharmacology. Data Asymmetry and the Monotonic Drop in AI Efficacy (13:06) Drug discovery has benefited from the commoditization of next-generation sequencing—which dropped from $1 billion and 10 years per genome to millions of publicly available datasets. However, Mickey Atwal and Youssef Idelcaid observe that AI’s predictive power monotonically declines as assets move downstream into clinical phases. While upstream target discovery enjoys an abundance of open data, late-stage clinical trial outcomes and translational datasets remain locked behind corporate intellectual property walls. This lack of pooled clinical training data contributes to a persistent 90% failure rate for candidates entering human clinical trials. Operational Leverage in Clinical Trials and Regulatory Workflows (17:42) Nirav Mehta emphasizes that operational development offers immediate financial and strategic leverage, with biopharma aiming to shave at least two years off the standard 10-to-15-year, multibillion-dollar development cycle. Rather than evaluating only the Probability of Technical and Regulatory Success (PTRS), enterprise biopharma is adopting Integrated Evidence Planning to assess commercial viability (PTR & CS) earlier. Generative AI and literature mining are deployed to optimize trial protocols against historical precedents—such as determining whether a 36-week endpoint can substitute for a 56-week MRI regimen—while mining regulatory datasets like the FDA's roughly 200 public Complete Response Letters (CRLs) to diagnose why prior drugs failed. Dismantling Silos Through Bilingual Talent and Shared Infrastructure (21:50) The panel rejects isolated corporate "Centers of Excellence," noting that top-down data science hires frequently fail when divorced from laboratory realities. Mickey Atwal advocates for mandatory cross-training: computational researchers rotate into wet labs to construct sequencing libraries and experience experimental noise firsthand, while experimentalists learn command-line programming. Youssef Idelcaid observes that converted life science PhDs often outperform pure mathematicians because they possess domain context, while warning against fragmented SaaS tools and clinical data leakage to Contract Research Organizations (CROs). Nirav Mehta adds that dedicated platform engineering is essential to establish enterprise knowledge graphs and multimodal foundations across R&D programs. Bridging Preclinical Translation and the Traps of In Silico Optimization (32:11) When considering where to deploy capital, panelists prioritize bridging the preclinical-to-human translational gap, where drugs effective in animal models routinely fail in human physiology. Mickey Atwal cautions against treating computationally optimized target binding affinity from structural models like AlphaFold as a surrogate for true drug developability, warning that molecules designed purely in silico risk poor bioavailability, toxicity, and unpredictable immunogenicity because they have never undergone human evolutionary selection. To mitigate clinical attrition in complex spaces like oncology and immunology, Youssef Idelcaid advocates using digital twins and early clinical pharmacology simulations to predict dosing and pharmacokinetics before entering expensive human trials. The fundamental barrier to AI in life sciences is not algorithmic sophistication, but the lack of unified clinical data and predictive human biology. Until organizations address the translational gap and avoid over-optimizing isolated benchmarks like binding affinity, the rate at which computationally designed molecules reach the market will remain uncertain.

生命科学领域的AI应用正经历从算法狂热向务实落地的转变，但其效能呈现从早期发现向临床阶段递减的明显断层。行业专家指出，破解药物研发高昂成本与约90%临床失败率的核心，在于打破研发数据孤岛、消除跨学科协作壁垒，并将优化目标从孤立的靶点结合力转向真实的临床成药性与逆向转化。 早期数据繁荣与临床转化的效能断层 (13:06) 药物研发长期以来类似于一条单向传递的流水线，涵盖疾病靶点发现、先导分子优化、试验用新药（IND）申报至人体临床试验。然而，AI对该流程的赋力存在显著的单调递减现象。上游的靶点发现由于二代测序技术的普及，单个人类基因组测序成本从早期的10亿美元和10年周期骤降至海量公开可用数据，形成了靶点资源的极大富集；但在下游，候选药物进入人体试验后仍面临约90%的高失败率。新药研发整体平均耗时15年，单药成本高达数十亿美元，历史临床成功率仅徘徊在8%至10%左右。 下游效能断层的根源在于数据可及性与数据孤岛。与公开透明的基因组学数据不同，后期的临床试验、药代动力学及转化医学数据因商业机密与知识产权限制，被制药企业严格封闭在私有孤岛中。这导致预测模型在进入人体试验后失去充足的训练数据支撑，难以准确预判药物在复杂人体系统中的真实疗效，造成早期发现的高产出无法有效转化为后期的临床成功。 警惕算法玩具：靶点结合不等于成药性 (38:26) 生命科学界容易陷入追逐最新AI工具的“糖果店”陷阱，将算法指标等同于临床成果。以AlphaFold为代表的蛋白质结构预测工具为从头分子设计带来了突破，但业界过度聚焦于在计算机模拟中最大化靶点结合亲和力。Mickey Atwal强调，结合亲和力易于计算和量化，但孤立地对其过度优化会导致严重的下游风险：高度针对靶点设计的分子可能生物利用度极低、具有严重毒性，甚至在人体中引发无法预见的不良反应。 更为严峻的挑战在于临床前的“转化鸿沟”。正如Atwal所言，科学界在小鼠模型中已攻克了大量癌症机制，但动物生理环境无法等同于人体。纯粹依靠计算机算法生成的从头分子（in silico molecules）在人类数百万年的演化历程中从未出现过，将其引入人体面临未知的免疫原性风险。若无法在临床前阶段系统评估药物在整个人体生物系统中的成药性（Developability），过早优化单一结合参数将使得新药在耗费巨资进入人体试验后依然难逃折戟命运。 逆向转化与运营提效：重塑临床监管流程 (8:18) 面对新药研发高昂的试错成本，业界正通过“逆向转化医学”（Reverse Translational Medicine）和运营优化重塑研发范式。Youssef Idelcaid指出，研发不再是单纯的单向线性传输，通过将真实世界数据（RWD）、电子病历（EHR）、保险理赔及生物标志物等临床端信息反哺至早期的靶点筛选，实验室得以从数百万次盲筛精简为数百至数千次靶向实验。例如，Exscientia利用AI设计的强迫症候选药物在不到12个月内即从发现阶段推入一期临床试验。同时，在肿瘤学和炎症等高度不确定的疾病领域，利用临床药理学数据构建数字孪生（Digital Twins）进行虚拟模拟，可在实际开展昂贵试验前优化给药剂量并预测不良反应。 在临床试验和申报阶段，AI带来的业务杠杆同样明确。Nirav Mehta指出，药企正从关注技术与监管成功率（PTRS）扩展至结合商业可行性的综合成功率（PTR & CS）。借助集成证据规划（IEP）与生成式AI文献挖掘，团队可以在试验设计之初对比竞品历史数据，例如评估将56周MRI影像终点缩减至36周的可行性，并高效生成电子通用技术文档（eCTD）。顾问团队测算，通过重塑临床试验与监管申报流程，企业有望从通常10至15年的研发周期中至少压缩出2年时间。 跨学科融合与平台工程：打破组织协同壁垒 (21:50) 大型制药企业推行AI常因自上而下的行政命令而受挫：管理层往往单方面招聘一批被隔离在独立部门的机器学习专家，导致算法团队脱离核心生物学问题。Atwal建议推行沉浸式协作，让计算机科学家深入湿实验（Wet-Lab）亲手制备单细胞测序文库并操作测序仪，以直观理解真实实验数据的噪声本质；同时举办编程培训，鼓励湿实验生物学家学习基础脚本。这种双向轮岗打破了“把数据扔过围栏”的传统断层，消除了将数据科学家视为单纯跑p值或火山图的“魔法师”迷思。 组织维度的割裂还存在于研发（R）与临床开发（D）之间。Idelcaid指出，吉利德（Gilead）最具生产力的AI研究员往往是具备生物学背景并自学AI的复合型人才，同时企业必须停止将临床试验数据（如CT影像）滞留在合同研究组织（CRO）的泄露行为。Mehta进一步强调，平台工程是消除摩擦的第三大支柱，企业应摈弃孤立采购零散SaaS工具的做法，转向建设能够统一管理监管文献挖掘、知识图谱及特定递送模态的企业级多模态数据基础设施，从而将科研人员从繁重的数据预处理泥潭中解放出来。 将AI深度融入生物医药研发的决定性瓶颈已非单纯的算法演进，而是如何跨越动物模型到人体的临床转化鸿沟，并在企业内部建立无缝连通的多模态数据与跨学科组织生态。未来的行业破局点属于那些能够确立清晰业务目标、打破数据壁垒，并将算法指标真正锚定于患者临床成药结局的组织。

## Background 该视频由行业会议频道 Ai4 于 2025 年 8 月 28 日发布在 YouTube 平台。内容为 BioLabs 区域总监 Adam Jenkins 主持的专家圆桌论坛，汇聚了来自 Flagship Pioneering、吉利德（Gilead）和德勤（Deloitte）的资深领袖，重点探讨 AI 如何赋能生命科学、加速药物发现与临床开发，以及在强监管环境中实现技术落地的最佳实践。 ## TL;DR 圆桌嘉宾深入解构了 AI 重塑生物医药研发的核心逻辑，指出行业需跳出局部指标，通过逆向转化医学与早期临床药理模拟推动全局优化以提高综合成功率（PTRS）。研讨指出当前 AI 渗透率在数据充沛的早期靶点发现中极高，但在临床试验阶段受制于数据孤岛而面临高达 90% 淘汰率的行业痛点。专家们特别警示了利用 AI 过度优化分子结合力而忽视人体成药性与免疫原性的潜在陷阱，并呼吁打破跨学科壁垒、推动数据科学家下沉湿实验以建立真正的研发生态协同。 ## Atmosphere & style 整个研讨会呈现出务实、犀利且充满自嘲与真知灼见的对话基调，完全摒弃了浮躁的技术炒作。嘉宾们在轻松幽默的互动中直言不讳，例如 Mickey Atwal 对“治愈小鼠癌症与人体转化鸿沟”及“AI 巫师迷信”的辛辣剖析，与 Youssef 和 Nirav 基于药企一线开发与工程底座的严谨沉淀形成了绝佳互补。现场交流节奏紧凑、观点交锋火花四溅，让观众深切感受到生命科学界在面对巨大研发不确定性时的清醒反思与务实探索。 ## Key takeaways - [03:18](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=198s) 药物研发不仅关乎前期的科学发现（R 端），在高度监管的开发（D 端）利用 AI 加速监管申报并提升技术与监管成功率（PTRS），同样具有巨大的临床与商业价值。 - [03:18](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=198s) 研发各环节割裂是导致药物研发高损耗率的核心原因，行业需要跳出局部思维，以‘从科学出发对患者产生积极影响’为全局优化目标，打通化学、生物与计算科学。 - [08:18](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=498s) 逆向转化医学借助 AI 挖掘临床试验、真实世界数据（RWD）及电子病历，反向指导早期靶点发现，结合‘人在回路’与‘实验室在回路’机制大幅缩减物理实验量与研发周期。 - [11:25](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=685s) 新药研发耗资数十亿美元、耗时约 15 年且临床综合成功率仅约 8%，企业必须在‘速度’、‘成本’与‘成功率’等互斥目标中明确单一首要优化方向来设计 AI 场景。 - [13:06](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=786s) AI 在药物研发全流程中的渗透率呈现单调递减规律：前端靶点发现数据充沛但面临排序难，而中后期临床与分子成药性数据因药企知识产权壁垒形成孤岛，导致临床预测能力极为薄弱。 - [17:42](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=1062s) 针对临床 90% 的失败率，药企最务实的切入点是利用生成式 AI 优化临床试验方案与加速监管申报（如自动化生成 eCTD 文档），力争在 10 年开发周期中节省至少 2 年时间。 - [21:50](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=1310s) 打破数据科学家与生物学家的沟通壁垒需要建立深度融合的一体化团队，通过强制跨界轮岗（让算法人员进湿实验、让生物学家学编程）消除‘AI 玄学巫师’迷信与独立部门割裂。 - [25:34](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=1534s) 药企传统上将临床外包给 CRO 导致影像等核心多模态数据严重滞留，同时生物医学科班出身后转型 AI 的跨界人才在连接研发与临床中发挥着无可替代的作用。 - [29:37](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=1777s) 底层软件工程与基础设施平台化至关重要，药企需搭建统一的多模态数据与知识挖掘底座，消除烟囱式单点工具，将科学家从繁琐的杂务泥潭中解放出来。 - [32:11](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=1931s) 从动物实验到人体临床的转化医学鸿沟是推高新药开发成本的主因，未来 AI 投资的核心爆发点在于利用早期临床药理模拟和数字孪生降低临床前资产转化风险。 - [37:09](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=2229s) 过度优化易量化的单一指标（如计算机模拟中的靶标结合力 Binding Affinity）容易陷入误区，忽视分子的体内成药性、代谢缺陷及毒性，最终导致临床转化失败。 - [40:53](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=2453s) 完全在计算机中从头设计的分子脱离了自然进化规律，在人体内可能面临未知的免疫原性与剧烈排斥反应，计算生物学必须将体内生物学环境与免疫安全性置于算法指标之上。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=4s) 研讨会开场与嘉宾背景介绍 主持人 Adam Jenkins（BioLabs 区域总监）宣布研讨会正式开始，并向在会议最后一天到场的观众致谢。他幽默地提到，虽然在场的绝大多数参会者主要来自 SaaS 或企业级软件领域，但今天将展开一场聚焦“AI 药物发现与开发”的深度对话，探讨 AI 在生物医药领域经历的发展历程、未来趋势以及真正落地的有效方案。 随后，三位来自顶尖药企、风险投资与咨询机构的行业专家依次分享了自己的跨界背景： * **Mickey Atwal**：曾任 Flagship Pioneering 旗下 Pioneering Intelligence 的首席数据与 AI 官，早年拥有物理与医学背景。他调侃自己当年放弃医学院转向学术界搞机器学习和大规模数据分析时“伤透了母亲的心，至今都没被原谅”。之后他投身工业界，出任再生元（Regeneron）数据科学负责人，全流程参与了从临床前到临床试验的药物开发，随后加入 Flagship 负责管理平台型创新公司投资组合，目前正在纽约市筹备创立自己的新公司。 * **Youssef Idelcaid**：吉利德（Gilead）AI 研究中心（ARC）负责人，拥有法国应用数学与计算机工程背景。他目前领导一支由工程师和各领域博士组成的多元化团队，专注于药物开发（即 R&D 中的“D”端），涵盖临床开发、监管申报和药物警戒（Pharmacovigilance）。他有着跨界经历，曾在欧莱雅（L'Oréal）从事美妆与护肤研发，之后转型进入生物技术领域，先后就职于基因泰克（Genentech）与吉利德。 * **Nirav Mehta**：德勤咨询（Deloitte Consulting）生命科学数字化与 AI 创新负责人，技术工程出身。他长期深耕药物研发与开发一线，亲历了从早期传统机器学习、自然语言处理（NLP）到如今生成式 AI（GenAI）彻底重塑行业能力的演进过程，目前带领团队专注于 AI 工程化及 GenAI 在生命科学领域的落地应用。 --- ### [03:18](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=198s) 重新定义药物研发：从高昂成本中心到全局优化 针对主持人提出“药物研发（Drug Discovery）对各自意味着什么，以及 AI 正带来何种冲击”的问题，三位嘉宾从开发端、生命周期与系统工程等不同视角展开了解构。 Nirav Mehta 率先从“D”（开发）端切入。他指出，尽管学术界与大众往往将目光聚焦于“R”（研究与发现）端，但药物开发处于高度严格的监管体系之下，开发端的加速同样具有不可估量的商业与临床价值——“只要能让药物提前一天上市送到患者手中，都是极其宝贵的”。AI 在这里最具吸引力的机会在于加速监管申报路径，并有效提升“技术与监管成功率”（PTRS, Probability of Technical and Regulatory Success）。 Youssef Idelcaid 则从自身长期跨越化学与生物研发的经历出发，阐述了对药物研发本质的理解。在他眼中，药物研发是跨学科团队协作孕育新化合物的漫长旅程起点，是人类健康进化的基石；但不可否认，它在传统认知中也是一个极其昂贵的“成本中心”，甚至带有一种“老派、传统（Legacy）”的色彩。许多年轻人刚毕业时甚至觉得 R&D 过于陈旧偏向学院派，而今天 AI 的核心价值正是打破这种老派模式，让探索过程更加敏捷、高效与现代化。 Mickey Atwal 提出了全局视角的反思。他认为“药物发现”对于不同岗位的人定义完全不同，而这恰恰揭示了药物研发面临的深层挑战。行业应该跳出局部，审视真正的全局优化目标： > "什么是我们试图解决的全局优化目标？是从科学出发，对患者产生积极的影响。" 在 Mickey 看来，正是因为研究和开发链路中各环节彼此割裂，才导致了药物研发极高的高淘汰率与损耗率（Attrition rates）。如今化学家、生物学家、遗传学家必须与计算科学家和机器学习研究者紧密结合，才能真正打通全局。 ```mermaid flowchart TD accTitle: Drug R and D Global Optimization and Integration subgraph Research["研究端 (R: Discovery)"] A["靶点发现与验证"] --> B["计算设计与分子筛选"] end subgraph Development["开发端 (D: Development)"] C["临床前与临床试验"] --> D["监管申报与药物警戒"] end subgraph Goal["全局优化目标"] E["提高 PTRS 并缩短上市周期\n最终惠及患者生命健康"] end Research -->|"前向转化"| Development Development -->|"逆向转化医学 (RWD/EHR/生物标志物)"| Research Development --> Goal Research --> Goal ``` --- ### [08:18](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=498s) 逆向转化医学、AlphaFold 突破与“人在回路”实验加速 当主持人深入追问“目前行业中哪些 AI 应用真正实现了落地提速”时，Youssef Idelcaid 重点解析了研究端与开发端之间存在的“灰色地带”——逆向转化医学（Reverse Translational Medicine）。 传统研发是一条从前向后的单向流水线（体外实验、计算机模拟到临床），但现代药企在“D”端积累了浩瀚的海量数据，包括真实世界数据（RWD）、临床试验数据、医保索赔数据（Claims Data）、生物标志物以及电子病历（EHR）。通过 AI 挖掘这些真实临床反馈，算法能够反向指导“R”端的早期探索，帮助科学家精准锁定全新生物通路与先导化合物（Hit），直接将原本需要筛选测试的数百万次物理实验大幅缩减至数千次甚至数百次。 在技术拐点上，Youssef 将 2015—2016 年及 DeepMind 开源 AlphaFold 视为里程碑式的觉醒时刻，早在 ChatGPT 爆发之前就证实了 AI 精准预测蛋白质三维结构的颠覆性潜力。自此之后，行业全面迈向了“人在回路”（Human-in-the-loop）与“实验室在回路”（Lab-in-the-loop）的协同研发范式，算法不断由湿实验数据喂养并即时迭代，实现了对科研人员的指数级赋能。 - **实战提速案例**：AI 药物研发公司 Exscientia 针对强迫症（OCD）设计的候选药物，依托算法辅助设计，从早期发现一路推进至一期临床试验（Phase 1），全程耗时不足 12 个月（行业传统周期通常需要 4 至 5 年以上）。 ### [11:25](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=685s) 临床运营与优化目标：在速度、成本与成功率之间权衡 在药物发现阶段之后，人工智能与机器学习正深度介入到临床与开发运营（Development Operations）的整个生命周期中。机器学习不仅能够精准预测和规划患者的招募周期（Enrollment Period），还能结合真实世界数据（Real-World Data, RWD）评估并预测临床试验的成功概率。此外，数字病理学（Digital Pathology）与数字孪生（Digital Twins）技术成为当下极具前景的领域——在进入成本高昂的体内（in vivo）试验之前，研究人员希望先在计算机模拟（in silico）环境中充分摸清分子的作用机制。 面对动辄耗资数十亿美元、耗时约 15 年且平均成功率仅为 8% 左右的药物研发流程，企业必须明确自己的“作战方向”和核心优化目标。由于无法同时兼顾多个互相冲突的目标函数（Objective Functions），团队必须在“速度（Speed）”、“成本（Cost）”或“成功率（Success Rate）”之间做出明确抉择，确立唯一的优化核心，并据此逆向推导需要落地的具体 AI 应用场景。 ```mermaid flowchart TD accTitle: AI in Drug Discovery and Development Optimization subgraph TargetOpt["设定核心优化目标（需三选一）"] O1["优化速度 (Speed)"] O2["优化成本 (Cost)"] O3["优化成功率 (Success)"] end subgraph AppOps["临床与运营阶段的关键 AI 实践"] A1["患者招募周期预测"] A2["真实世界数据 (RWD) 成功率预测"] A3["数字病理学与计算机模拟 (In Silico) 数字孪生"] end TargetOpt --> AppOps ``` - **药物开发宏观数据**：单款新药平均研发周期长达 15 年，资金投入达数十亿美元，最终临床综合成功率仅约 8%。 - **核心逻辑**：无法在多目标函数并存的情况下全面兼顾，必须明确单一首要优化指标（速度、成本或成功率）以设计 AI 用例。 --- ### [13:06](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=786s) 药物研发“列车”与 AI 渗透率递减规律：数据孤岛的困境 尽管近年来 AI 在生命科学领域备受瞩目，但机器学习在药物研发中的应用绝非新生事物。早在 2000 年代初期，行业就已经广泛利用无监督聚类（Unsupervised Clustering）等方法来解析人类肉眼无法处理的海量基因组学（Genomics）和转录组学（Transcriptomics）数据集，如今许多方法只是被冠以“AI”的全新标签。 如果将整个药物开发生命周期比作一列单向行驶的“列车”，其流程从确定疾病适应症和靶点发现（Target Discovery）开始，推进到先导化合物优化（Lead Candidate Optimization，包括小分子、抗体或基因治疗方案），随后经历堆积如山、繁琐至极的临床试验申请（IND Filing），最终迈入人体临床试验阶段。 ```mermaid flowchart LR accTitle: Monotonically Decreasing AI Impact Along the Drug Development Train subgraph EarlyStage["早期研发：数据丰富 / AI 渗透极高"] T1["靶点发现<br/>(Target Discovery)"] T2["先导物优化<br/>(Lead Optimization)"] end subgraph LateStage["中后期阶段：数据孤岛 / AI 渗透锐减"] T3["IND 申报与法规审批<br/>(Regulatory / IND)"] T4["临床试验<br/>(Phase 1 - 3 Trials)"] end T1 --> T2 --> T3 --> T4 ``` 然而，AI 在这列“列车”上的影响力呈现出**单调递减（Monotonically Decreasing）**的趋势：AI 在前端靶点发现中发挥了巨大的作用，但在先导化合物优化、临床试验以及真实世界证据的应用上，渗透率与效能却大幅下滑。造成这一强相关性的根源在于**数据可用性**： 1. **靶点发现领域数据充沛**：全基因组测序已完全商品化，从过去的“耗资 10 亿美元、耗时 10 年测序一个基因组”演变为如今可轻松获取并分析数百万个基因组。这甚至带来了“甜蜜的烦恼（Embarrassment of Riches）”——行业发现了海量靶点，却难以进行优先级排序。 2. **临床预测能力匮乏**：行业预测药物在人体内真实疗效的能力依然非常薄弱。其核心症结在于临床与分子成药性数据高度碎片化，各大型药企因知识产权（IP）壁垒将关键数据深锁在内部孤岛中，公开共享的数据极其稀少。 3. **临床高失败率**：药物进入临床试验后，面临高达 90% 的失败率。正如发言者调侃所言：“你必须是一个极其极其乐观的人，才会选择在一个明知有 90% 概率会失败的行业里工作。”如果缺乏充足的跨机构共享数据，机器学习模型便无法有效训练以预测一期、二期和三期临床试验的过关几率。总结而言，目前行业现状是“研究端表现极佳，开发端表现平平，临床端表现糟糕”。 --- ### [17:42](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=1062s) 行业落地痛点：利用生成式 AI 压缩开发周期与整合证据规划 由于药物研发周期极为漫长，传统的线性模式往往缺乏持续快速的产品生命周期反馈回路。针对临床高达 90% 的失败率，制药行业当前最务实、商业回报最明确的切入点，在于**如何利用 AI 从长达 10 年的开发周期中至少抢出 2 年的时间**，重点发力于临床试验方案优化与监管申报提速。 这一趋势推动了“整合证据规划（Integrated Evidence Planning, IEP）”的复兴。企业不再仅仅关注传统的科学与技术监管成功率（PTRS），而是将视野前移至“技术、监管与商业综合成功率（PTR & CS）”。通过在方案设计初期引入 AI，药企能够在竞品林立的红海市场中及早确立目标产品特性（Target Product Profile, TPP）并开展卫生技术评估（Health Technology Assessment, HTA）： | 优化维度 | 传统模式瓶颈 | AI 与生成式 AI 的赋能方案 | |:--- |:--- |:--- | | **临床方案设计** | 缺乏跨试验数据对比，检查流程冗长（如设定长达 56 周的 MRI 扫描监测） | 挖掘历史审批先例与监管路径，论证缩短至 36 周的可行性，大幅削减临床时间与成本 | | **非结构化信息提取** | 竞品情报检索、海量医学文献挖掘耗费大量人工 | 利用 GenAI 自动化解析非结构化文献、竞品动态及前沿临床数据 | | **注册申报加速** | 从最后一名受试者出组（LPO）及数据库锁定（Database Lock）到完成申报材料耗时长 | 自动加速 eCTD（电子通用技术文档）撰写与组卷，提升与监管机构（FDA 等）的沟通效率 | 通过在临床设计初期深度对齐资产战略（Asset Strategy），药企能够实质性压缩临床与审批周期，使生成式 AI 真正转化为提升研发效率的实用生产力。 ### [21:50](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=1310s) 打破壁垒：数据科学家下湿实验与打破“AI 巫师”迷信 在谈及 AI 智能体（Agents）和跨部门协同的问题时，主持人 Adam Jenkins 指出，生命科学领域的合作远比普通商业客服场景复杂得多，因为这本质上是两个完全说不同“语言”的群体在试图互相交流与学习——一方是掌握计算机科学与人工智能前沿工具的技术人员，另一方则是拥有深厚生物学与化学专业背景的科研人员。Adam 随即向 Flagship Pioneering 的 Mickey Atwal 提问：在实际组织中，究竟该如何将这两股力量揉合在一起？过去有哪些行之有效的方法，又有哪些彻底踩坑的教训？ Mickey Atwal 幽默地反问：“这事你打算留多少时间让我聊？”他直言，把两者预设为“两个不同群体”本身就道出了核心痛点。在理想状态下，组织本不该被人为割裂为两个对立的阵营，而应建立一个高度融合的跨学科一体化团队。然而在现实中，大量初衷良好、资金雄厚的大型跨国药企却一次次在这一步上摔得粉身碎骨。典型的失败剧本往往是自上而下的行政命令驱动：公司首席执行官在《纽约时报》上读到了 AI 的颠覆性报道，大腿一拍宣布“我们公司必须立刻搞 AI”，随后便大张旗鼓招聘了一批顶尖的数据科学家和机器学习专家。结果，这批人被单独塞进一个全新的 AI 部门，甚至被安置在独立的办公大楼里，日常工作中根本没有机会与生物学家碰头，更接触不到一线湿实验（Wet Lab）人员。这种模式注定失败，因为离真实场景过远的算法团队根本无法洞察真正亟待解决的科研痛点到底是什么。 ```mermaid flowchart TD accTitle: 跨学科研发协作壁垒与一体化闭环模式对比 subgraph Siloed["传统部门割裂模式（常见失败路径）"] direction TB A["管理层 AI 战略指令"] --> B["独立 AI/数据科学部门"] C["湿实验/生物学家团队"] -. "数据隔墙投掷 / 沟通脱节".- B B --> D["输出无实际生物学价值的模型"] end subgraph Integrated["一体化闭环模式（Mickey 的跨界实践）"] direction TB E["数据科学家"] <-->|"湿实验轮岗 / 理解数据噪声"| F["核心生物学实验室"] F <-->|"基础命令行与代码脚本培训"| E E & F --> G["高效率研发闭环与协同产出"] end ``` 为了打破这一僵局，Mickey 分享了自己在以往工作中推行的“强制跨界轮岗”机制。他半开玩笑地承认：“以前跟我干过的人现在大概恨死我了，但我当年确实强令不少数据科学家必须去湿实验室轮岗，亲自动手制备样本和产生数据。”如果团队成员正在从事大规模基因组学或单细胞测序数据的算法分析，他们就必须穿上实验服，亲自参与构建文库、将样本加载到测序仪上，亲身体会实验操作中的各种变异与噪声来源，从而消除将原始数据当成“上帝恩赐之物”的盲目崇拜心态。 与此同时，Mickey 也会戴上“教授帽”亲自开课，为湿实验科学家普及基础数据科学知识，鼓励他们“不要被电脑屏幕上的命令行终端吓退”，学会使用终端和基础脚本进行数据预处理。这种双向渗透极大地加速了研发反馈闭环，彻底终结了“隔着篱笆互扔数据包”的低效协作，也避免了把机器学习专家神化为“玄学巫师”的现象： > “把一堆混乱的数据硬塞给机器学习科学家，然后对他们说‘快给我算个 p 值、画个火山图（Volcano Plot）出来’，并把那张图当成某种神秘巫师赐予的预言，这简直是对高智商大脑的残酷浪费。” - **中心卓越中心（CoE）存废争议**：在听完 Mickey 对一体化组织的倡导后，主持人 Adam 调侃道：“你这是在暗示所谓的‘卓越中心’（Center of Excellence）终于该寿终正寝了吗？” - **Mickey 的犀利回应**：“要是你能举出一个真正成功的卓越中心范例，我就撤回我的观点。” --- ### [25:34](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=1534s) 历史鸿沟与组织割裂：从 CRO 数据孤岛到“懂科学的 AI 人才” 吉利德科学（Gilead）AI 研究中心负责人 Youssef Idelcaid 随后接过了话茬。他认为生命科学领域如今的技术与业务两极分化，既有历史原因，也有深层的组织架构原因。即便不谈 AI 技术专家与生物学家之间的代沟，制药工业内部本身就长期存在着“研究”（Research，R）与“开发”（Development，D）之间根深蒂固的双重孤岛。制药业是一个底蕴深厚但极其传统、惯性巨大的行业，亟需从流程层面构建跨阶段的整合数据底座。 Youssef 举例说明了目前行业中最离谱的痛点：CT 扫描和医学影像本应是临床前研究与临床开发阶段最具共通价值的核心多模态数据集，可令人难以置信的是，大多数传统制药公司内部甚至连一个统一存储影像数据的中心数据库都没有。原因在于药企往往将临床试验直接全包给合同研究组织（CRO），试验结束后大量极具价值的原始多模态数据便长年滞留在 CRO 那边，根本没有被回流至药企内部供进一步挖掘。 从技术演变的历史脉络来看，早在第一波计算机浪潮席卷生物医学领域时，生物信息学（Bioinformatics）与生物数学专家便已借助数字工具加速新药研发。然而，当后续“数据科学”浪潮掀起时，业界却人为将其包装为一门似乎脱胎于全新维度的学科（尽管其本质依然植根于传统的应用数学）。这种认知直接导致了内部鄙视链和“AI 鸿沟”的诞生——“你是干生信的，而我是干 AI 的”，这种各自为政的门派划分对行业毫无助益。 为了在吉利德 AI 研究中心（ARC）化解这种脱节，Youssef 采取了一种高度务实的人才策略：团队中的 AI 研究人员绝不局限于纯数学、统计学或计算机科学博士，反而招募了大量生物医学博士以及在实验室工作期间自学转型的生物学家： > “这些生物学科班出身、后来在科研实践或加入公司后转型做 AI 的人，往往是表现最出色的一批。因为他们天生具备领域专业知识（SME），能够毫无障碍地与 R 端的科研人员在同一语境下‘聊科学’，又能把这些认知无缝带入 D 端的临床开发。” 此外，Youssef 还呼吁企业彻底打破各个部门为了回答各自细碎业务问题而盲目采购独立 SaaS 软件的旧模式。多头采购只会建立起更多的信息真空和数据竖井。他主张依托现代 AI 智能体架构（Agentic Framework），打通从研发到临床的数据连续体，让全流程团队在统一的集成界面下共享数据与工具。 --- ### [29:37](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=1777s) 工程底座与平台化范式：从底层泥潭解放科学家 德勤（Deloitte）生命科学数字化与 AI 业务负责人 Nirav Mehta 对上述观点表示赞同，并将问题进一步下沉至平台与工程架构层面。他指出，探讨生物学家与数据科学家的分歧固然重要，但若缺少了强大的软件工程与基础设施支撑，一切设想都会沦为空谈。 在当前药企的实际运营中，科学家们耗费了极其惊人的精力去处理各种繁琐杂务（Grunt Work）。很多咨询客户之所以无法加快研发节奏，核心死穴就是根本没有搭建起能够支撑多模态数据（Multimodal Data）整合的基础平台。数据散落各处、标准各异，这就迫切需要一支具备“平台工程思维”的技术团队，替前端的生物学家和算法专家扫清这些底层路障。 Nirav 强调，现代生命科学 AI 的平台化范式应重点突破以下方向： - **科学知识挖掘与监管内容挖掘平台**：通过平台化解决方案统一梳理全球文献、实验记录与复杂的监管申报材料，为数据科学家和生物学家提供开箱即用的高质量知识起点，避免每个新项目都要重新做数据清洗与文献搬运。 - **孤立技术项目的整合与收拢**：业界正在加速清理过去散落在各管线中的烟囱式小系统，将零碎的单点工具合并为统一的企业级技术平台。 - **核心研发技术平台化的投资演进**：除了纯 IT 与数据平台，类似 mRNA 技术平台等生物学通用平台的整合也在提速。药企正越来越倾向于以治疗领域（Therapeutic Area）为战略支点，将底层计算平台与特定疾病领域的生物学管线进行深度绑定与资产集中配置。 --- ### [31:15](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=1875s) 千万美元投资轮盘：AI 落地最具爆发力的核心环节 随着圆桌讨论进入尾声，主持人 Adam 向三位嘉宾抛出了一个颇具考验的假设性投资问题：“在向现场观众开放提问前，请大家戴上投资人的帽子。如果我现在给你们每人 1000 万美元的真金白银，让你们对未来 5 到 10 年的整个药物发现与开发链条下注——你们认为将 AI 应用在哪个具体环节上，会带来最高的投资回报率（ROI）？你们愿意把筹码压在什么方向上？” Mickey Atwal 率先接过了这个极具挑战性的问题。他没有立即列出枯燥的算法模型，而是以生命科学界广为人知的一句自嘲作为切入点： > “作为科学家，我们在‘治愈小鼠身上的癌症’这件事上一直干得极为漂亮。但据我上次照镜子确认，我并不是一只小鼠。” Mickey 指出，整个行业过去在基础研究上投入了难以估量的精力与资金，在解析各种小分子和生物制剂对各类适应症的作用机制方面取得了海量成果，学术期刊上也充斥着大量令人赞叹的论文与动物实验数据。然而，如何跨越从动物模型到人体临床转化的鸿沟，正是整个新药研发链条中最具痛点、却也可能是 AI 能够带来最颠覆性突破的关键所在。 ### [32:11](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=1931s) 转化医学痛点与利用 AI 降低临床前资产风险 Mickey Atwal 深入分析了目前新药研发中最核心的损耗（Attrition）瓶颈：即便基础科学研究取得了无数诺贝尔奖与行业大奖，但体外与动物实验中的可重复科学结论，往往无法直接转化为人体生理学表现。人类对候选分子为何在人体中失败的预测能力极弱。这种基础转化科学的脱节将整个新药生命周期的开发成本推高到 20 亿美元以上，研发周期更长达 10 年之久。Mickey 强调，如果 AI 能够破局、有效降低从临床前资产走向临床阶段的风险，将对整个新药研发体系产生极为深远的影响。 ```mermaid flowchart LR accTitle: Preclinical to Clinical Translation Gap subgraph Preclinical["临床前研究 (Pre-clinical)"] A["体外/动物实验数据"] --> B["高亲和力分子设计"] end subgraph Bottleneck["转化瓶颈 (Translational Gap)"] B -.->|"无法直接预测人体生理学反应"| C{"高损耗率 & 失败"} end subgraph Clinical["临床转化目标 (Clinical)"] D["AI 模拟与数字孪生"] --> E["缩短周期 & 降低 20 亿美元成本"] end B --> D ``` - **研发成本与周期数据**：一款成功上市的药物生命周期研发成本已突破 20 亿美元，耗时超过 10 年。 - **AI 核心发力点**：攻克临床前到临床阶段的转化不确定性，理解小分子在人体内药效衰减的底层生理机制。 --- ### [33:01](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=1981s) 药企研发视角：业务目标函数与早期临床药理模拟 针对如何规划研发预算，Gilead 的 Youssef Idelcaid 结合正在制定的三年规划提出了“业务目标函数”（Business Objective Function）的概念。他指出，不应只从纯数学角度建模，而是必须绑定商业目标，即明确药物开发（Development）端究竟在优化什么。在药企竞速中，“上市时间”（Time to Market）的核心就是“申报提交时间”（Time to Submission），谁能针对同类靶点或适应症率先向 FDA 或 EMA 递交申请，谁就能抢占先机。 在具体疾病领域上，Youssef 指出肿瘤学（Oncology）与炎症性疾病（Inflammatory Disease）的可预测性明显弱于病毒学（Virology）。因此，他会优先针对这两个高难度领域，重度投资于早期临床药理学（Clinical Pharmacology）的模拟技术，包括药代动力学模拟（PK）、给药剂量优化（Dosing）以及数字孪生（Digital Twins）。通过在临床前及早期阶段以数字化方式预测人体反应，药企可以在斥资数十亿美元展开为期十年的临床试验之前把控风险，大幅压缩整体研发时间线。 - **高预测难度疾病领域**：肿瘤学和炎症性疾病相比病毒学更难预测，是 AI 与模拟技术的重点攻坚方向。 - **核心优化手段**：利用早期临床药理模拟和数字孪生，在前期海量验证给药剂量和人体代谢，从源头压缩上市周期。 --- ### [34:59](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=2099s) 行业生态协同：转化研究与监管透明度 Deloitte 的 Nirav Mehta 指出，从基础研究跨越到动物实验再到临床获取真实证据（Proof Points），是当前整个行业最亟需加速的环节，他希望未来五年这一转化瓶颈能被彻底攻克。此外，Nirav 特别强调新药研发是一个系统性生态工程，不仅涉及研发团队，还需要监管机构的深度参与。近期监管透明度正明显提高，例如 FDA 公开披露了 200 份完整回应函（CRLs, Complete Response Letters），详细列明了新药被拒批的具体原因。这类数据集对于药企从技术和合规双重視角排查隐患、缩短上市周期具有极高的价值。 当被主持人问及监管机构究竟是在“促进”还是“阻碍”AI 进展时，Nirav 认为目前下定论为时尚早。监管机构对于 AI 安全性的担忧完全合理且必要，目前整个行业都在积极探讨如何建立合规框架，当下的关键机遇在于推动监管指南尽快出台，以标准引导并加速 AI 的规范化应用。 - **生态各方协同**：打通转化医学链条需要研究人员、制药企业开发团队以及监管机构形成合力。 - **监管数据资产**：FDA 披露的 200 份 CRL 提供了极为宝贵的被拒批原因数据，帮助药企在临床前和临床设计中避坑。 --- ### [37:09](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=2229s) 现场问答：多模态技术与过度优化单一靶标结合力的陷阱 在观众提问环节，一名兼具生物学与基因组学背景的学者提出疑问：面对长达 10 到 20 年的生命周期，以及临床前研究从动物模型转向高异质性人类细胞的政策背景，新兴的多模态（Multimodal）AI 方法能否打破数据孤岛并加速转化？ Mickey Atwal 回应指出，整个 AI 行业常常像“糖果店里挑花眼的孩子”，每天都在追逐最新、最耀眼的工具，却缺乏明确的战略去聚焦真正重要的问题。许多机器学习科学家习惯于“过度关注可测量的指标，而非真正有意义的指标”（overindex on problems that are measurable rather than meaningful）。以分子设计为例，许多 AI 团队借着 AlphaFold 的热潮，全力利用生成算法在计算机中设计与靶标结合力（Binding Affinity）更强的分子。结合力数据易于获取和量化，但这种局限于单点的“过度优化”往往适得其反：结合力超强的分子可能会丧失成药性（Poor Druggability）、产生极高的人体毒性，甚至彻底破坏整个下游管线。 ```mermaid flowchart TD accTitle: Over-optimization in In Silico Drug Design subgraph ML_Perspective["机器学习视角的过度优化"] A["高亲和力结合 (Binding Affinity)"] --> B["易于获取的计算机生成数据"] end subgraph Downstream_Risks["下游忽略的生物学风险"] B --> C["较差的成药性 / 溶解度"] B --> D["不可预测的系统毒性"] B --> E["未演化过的人工表位引发免疫原性"] end C --> F{"导致临床转化彻底失败"} D --> F E --> F ``` - **指标错位风险**：强行最大化结合力数据容易掩盖分子毒性与代谢缺陷，导致后续临床开发全盘崩溃。 - **跨学科协同要求**：机器学习科学家必须深入理解底层生物学机制与临床可开发性（Developability），不能在局部指标上过早过度优化。 --- ### [40:53](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=2453s) 计算机生成分子的稳定性与人体现存免疫风险 针对 Nirav Mehta 提出的关键疑问——“究竟有多少通过计算机设计、理论上可行的分子最终能成为稳定且可商业化上市的药物”，Mickey 直言目前尚无确切答案，因为整个行业此前极少把精力放在“到底是什么定义了一款好药”这一根本问题上。对于大部分计算机科学家而言，结合力好就等同于好药，但这完全忽视了复杂的体内生物学环境。 Mickey 进一步发出警示，大量完全在计算机中从头设计的分子（*in silico* designed drugs）正在带来全新的未知风险。人类机体在数百万年的自然进化中从未接触过这类人工合成结构，这些高度异质性的分子极易在人体内引发未知的免疫原性（Immunogenicity）及剧烈的免疫副反应。Mickey 直言，面对那些完全由计算机生成、彻底脱离人类基因组自然演化规律的全新分子，他自己在将其注入人体时会持极其审慎的态度。 - **成药性未解之谜**：计算生成的分子能否在物理和生理层面保持稳定并成功商业化，目前行业尚无定论。 - **免疫原性隐患**：未经自然进化筛选的人工合成结构可能触发机体免疫排斥，AI 制药必须将免疫安全性置于算法指标之上。 ## Notable quotes > [03:18](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=198s) 什么是我们试图解决的全局优化目标？是从科学出发，对患者产生积极的影响。 — Mickey Atwal > [03:18](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=198s) 只要能让药物提前一天上市送到患者手中，都是极其宝贵的 — Nirav Mehta > [13:06](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=786s) 你必须是一个极其极其乐观的人，才会选择在一个明知有 90% 概率会失败的行业里工作。 — Mickey Atwal > [21:50](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=1310s) 以前跟我干过的人现在大概恨死我了，但我当年确实强令不少数据科学家必须去湿实验室轮岗，亲自动手制备样本和产生数据。 — Mickey Atwal > [21:50](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=1310s) 把一堆混乱的数据硬塞给机器学习科学家，然后对他们说‘快给我算个 p 值、画个火山图出来’，并把那张图当成某种神秘巫师赐予的预言，这简直是对高智商大脑的残酷浪费。 — Mickey Atwal > [21:50](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=1310s) 要是你能举出一个真正成功的卓越中心范例，我就撤回我的观点。 — Mickey Atwal > [25:34](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=1534s) 这些生物学科班出身、后来在科研实践或加入公司后转型做 AI 的人，往往是表现最出色的一批。因为他们天生具备领域专业知识，能够毫无障碍地与 R 端的科研人员在同一语境下‘聊科学’，又能把这些认知无缝带入 D 端的临床开发。 — Youssef Idelcaid > [31:15](https://www.youtube.com/watch?v=T-2RCp1qCoY&t=1875s) 作为科学家，我们在‘治愈小鼠身上的癌症’这件事上一直干得极为漂亮。但据我上次照镜子确认，我并不是一只小鼠。 — Mickey Atwal ## Mentioned Adam Jenkins, Mickey Atwal, Youssef Idelcaid, Nirav Mehta, BioLabs, Flagship Pioneering, Pioneering Intelligence, Regeneron, Gilead, L'Oréal, Genentech, Deloitte Consulting, DeepMind, AlphaFold, ChatGPT, Exscientia, FDA, EMA ## Video description Adam Jenkins, Regional Site Director, BioLabs Mickey Atwal, Former Chief Data & AI Officer, Pioneering Intelligence, Flagship Pioneering Youssef Idelcaid, Head of AI Research Center (ARC), Gilead Nirav Mehta, Life Sciences Digital & AI Leader, Deloitte Consulting LLP |Abstracts This session will explore how AI is transforming the life sciences and pharmaceutical industry, particularly in accelerating drug discovery, development, and clinical trials. Panelists will discuss how AI-powered tools are being used to identify new drug candidates, optimize clinical trial design, and reduce the time and cost of bringing new treatments to market. Learn from experts how AI is revolutionizing R&D processes, enabling personalized medicine, and improving patient outcomes. The discussion will also cover the challenges of implementing AI in highly regulated environments and best practices for overcoming these hurdles. [Source](https://www.youtube.com/watch?v=T-2RCp1qCoY)

[3:18] Mickey Atwal: 什么是我们试图解决的全局优化目标？是从科学出发，对患者产生积极的影响。 [3:18] Nirav Mehta: 只要能让药物提前一天上市送到患者手中，都是极其宝贵的 [13:06] Mickey Atwal: 你必须是一个极其极其乐观的人，才会选择在一个明知有 90% 概率会失败的行业里工作。 [21:50] Mickey Atwal: 以前跟我干过的人现在大概恨死我了，但我当年确实强令不少数据科学家必须去湿实验室轮岗，亲自动手制备样本和产生数据。 [21:50] Mickey Atwal: 把一堆混乱的数据硬塞给机器学习科学家，然后对他们说‘快给我算个 p 值、画个火山图出来’，并把那张图当成某种神秘巫师赐予的预言，这简直是对高智商大脑的残酷浪费。 [21:50] Mickey Atwal: 要是你能举出一个真正成功的卓越中心范例，我就撤回我的观点。 [25:34] Youssef Idelcaid: 这些生物学科班出身、后来在科研实践或加入公司后转型做 AI 的人，往往是表现最出色的一批。因为他们天生具备领域专业知识，能够毫无障碍地与 R 端的科研人员在同一语境下‘聊科学’，又能把这些认知无缝带入 D 端的临床开发。 [31:15] Mickey Atwal: 作为科学家，我们在‘治愈小鼠身上的癌症’这件事上一直干得极为漂亮。但据我上次照镜子确认，我并不是一只小鼠。

[3:18] Nirav Mehta: every day that we can get a drug faster to market in the hands of a patient is valuable. [11:25] Youssef Idelcaid: You have to define your battle in what's going on today with AI in biotech, and you have to choose what you are optimizing for. If you are optimizing for speed, for cost, or for success... you cannot have multiple objective functions. You need to determine one and then work your way back until you see what are the use cases that you can work on. [15:49] Mickey Atwal: Long story short: doing very well with research, not so great with development, terrible with clinical. [17:42] Nirav Mehta: Can we actually shave off among the 10 years at least 2 years on optimizing clinical trials and regulatory submissions? There are real practical areas where we can actually make progress. [21:50] Mickey Atwal: as if it comes from God. [25:22] Mickey Atwal: Give me an example of a good one and I'll respond in kind. [31:15] Mickey Atwal: last time I checked, I'm not a mouse. [38:26] Mickey Atwal: kids in a candy store

[0:04] Well, thank you everybody for coming on [0:05] the last day here. Um, appreciate um, [0:08] everyone coming to a uh, we'll call it a [0:11] drug discovery talk. We know that most [0:13] of the people who are here um, in the [0:15] conference are some type of SAS or [0:18] enterprise solution. So, um, thanks for [0:20] coming everyone. Uh, so we'll have a [0:22] really great conversation today about [0:24] just kind of AI drug discovery, where [0:26] it's been, where it's going, what we [0:28] think works. We have a great panel about [0:30] uh diversification in across multiple [0:32] parts of the the industry. Um so I'd [0:34] like to have everyone introduce [0:36] themselves first um just so we can kind [0:38] of get some background. Mickey. [0:39] >> Yeah, [0:40] >> thanks Mickey out and thanks for the [0:42] opportunity to speak here. U very quick [0:44] uh brief history. So spent a long time [0:46] in academia after I quit medical school. [0:50] You disappointed my mom. She never [0:51] forgave me. um spent quite a bit of time [0:54] developing machine learning algorithms, [0:56] analyzing large data sets, moved to [0:59] industry um joined regeneron as the head [1:01] of data science um you know involved in [1:04] all aspects of drug development from [1:06] pre-clinical to clinical trials and [1:09] recently moved to flagship [1:11] um overseeing a portfolio of a number of [1:13] different companies in the platform [1:15] space and now transitioning into [1:18] starting my own venture in New York [1:19] City. [1:21] >> Great. Yousef. [1:22] >> Hi everyone. Happy to be with you today [1:25] uh in this conversation. So Yousef Idel [1:27] Kaid um I'm applied mathematics and [1:30] computing uh engineer by training from [1:34] France. So um excuse my accent in [1:37] advance. Um and today I represent Gilead [1:42] where I run an AI research center uh [1:45] within drug development. So the D of [1:47] R&D. Um and we have a large remmit uh [1:51] going from clean dev clinical [1:53] development [1:55] uh through um uh submissions and also uh [1:59] pharmarmacco vigilance. Um I have a team [2:02] of you know multid-disciplinary [2:04] uh folks engineers and PhDs. Uh prior to [2:08] that I worked in retail actually I was [2:10] making makeup and skincare at L'Oreal [2:12] for a while. uh but then converted into [2:15] biotech starting with Genentech and [2:17] today at um as Gilead really looking [2:20] forward to this conversation. [2:23] >> Great. Excellent. Uh hi everyone. Thank [2:26] you for the opportunity. Um I'm Nva. I [2:28] lead all things innovation for our life [2:30] sciences practice specifically focus on [2:32] all things R&D. Uh my pedigree is more [2:35] technology. Um you know I grew up in the [2:37] technology circles and I I'm more an R&D [2:39] person by trade more than academia. So, [2:42] you know, I I I come from um you know, [2:44] more of um you know, hands-on work, [2:46] especially in the the development space. [2:48] Um you know, over the years, obviously, [2:51] we've made tremendous progress, you [2:53] know, going from the days of machine [2:54] learning and NLP and now Genai [2:56] completely transforming what we can do. [2:58] Um you know, and especially [3:00] opportunities in the development space [3:01] are are tremendous. Um you know I I lead [3:04] a team of of you know expert expertise [3:06] in in all things um you know AI [3:08] engineering as well as uh um you know [3:11] just geni based innovations um you know [3:14] and uh you know this optic is what I'm [3:16] very excited about talking about today. [3:17] Yeah. [3:18] >> Fantastic. So the first thing I kind of [3:20] want to touch on with all of us is uh [3:22] drug discovery is a large [3:24] all-encompassing term. Um it can mean a [3:27] lot of different things to a lot of [3:28] different people. Um it can be the [3:30] research side, it can be the development [3:31] side, it can be um applications uh for [3:34] government regulatory filing. Um so real [3:37] quick for each of you kind of what does [3:39] drug discovery mean to you? Um and how [3:42] do you see AI impacting um how you see [3:45] it? Uh Nurof, sorry. [3:47] >> Yeah. So I'll I'll start with the D side [3:49] of of R&D really because um you know [3:52] I'll I'll leave the R side to obviously [3:54] the the experts and you know everybody [3:55] can tell you you know with with drug [3:58] discovery and and development being such [4:00] a highly regulated space the biggest you [4:03] know opportunity is going to be on the R [4:04] side right as much as we can accelerate [4:06] but even on the D side uh on the [4:09] development side this tremendous [4:10] opportunity to accelerate and every day [4:13] that we can get a drug faster to market [4:15] in the hands of a patient is is valuable [4:17] Right. So, you know, with the uptick [4:20] that we are seeing and the the [4:22] opportunities with AI, especially in um [4:25] you know, creating or or accelerating [4:27] regulatory pathways um you know, [4:29] improving PTRS [4:31] um that's really where a lot of the [4:33] focus is and that's what I'm excited [4:34] about. [4:36] >> Yeah. You [4:37] >> Yeah. Um [4:39] same. Uh I think I I represent more the [4:42] D than R but um uh I did work uh most of [4:46] my career in R&D would it be chemistry [4:48] or now uh in biotech. So so what what um [4:53] uh I would say drug discovery uh [4:56] represents for me is um uh obviously [4:59] besides the R of R&D uh experimentation [5:03] I think um that's what what drug [5:05] discovery uh represents. Um uh for me [5:08] it's a a vast uh domain where uh you [5:12] have you know multid-disciplinary [5:15] uh teams coming together to give birth [5:17] to uh a new compound and I think that's [5:21] for me also um the starting point uh for [5:25] a long long long long journey until the [5:30] the the the medicine gets in in in in [5:34] the patients. So I think uh drug [5:37] discovery represents also for me um the [5:43] uh foundation uh for um you know uh [5:48] human health um I would say uh uh [5:51] improvement and and evolution uh but it [5:55] also represents um a cost center because [5:58] it's very very expensive uh to to run [6:02] the experimentations [6:04] uh and last I I think it does uh [6:06] represent or sound as a uh legacy kind [6:11] of thing. So I remember when I graduated [6:12] so there was no way for me to go to R&D [6:15] because R&D for me was like for you know [6:17] older people and you know like people [6:19] who are not um uh you know industrial uh [6:24] more closer to academia but I think um [6:27] it's it's definitely has that legacy [6:29] kind of piece of it and I'm hoping that [6:32] you're going to touch on that in the [6:33] conversation on how uh AI can you know [6:37] take that old school um uh approach [6:41] approaches and legacy approaches to [6:42] something a little bit more uh faster [6:44] and and modern. [6:47] >> And maybe yeah um you know I think that [6:50] the premise of this question is kind of [6:51] interesting. I mean if the panel was [6:53] significantly bigger you would have a [6:55] wide distribution of of answers. I mean [6:57] it's different things to different [6:58] people and I think there in lies in the [7:00] challenge in drug dis drug discovery and [7:03] development right because we're all kind [7:05] of focused on differenting aspects of [7:07] this and I think it's useful to just [7:09] step back and ask know what is the [7:11] global optimization function we're [7:13] trying to solve for you know it's you [7:15] know going from science to making a [7:18] positive impact on patients and you know [7:22] when we focus on different aspects of [7:25] that train going from research [7:27] development that's where I think you [7:29] know we're kind of we see the attrition [7:31] rates of drug development um so having [7:34] that overall picture is useful it's a [7:36] challenge especially when you have this [7:38] you know interdiciplinary environment [7:39] where you have um chemists biologists [7:42] geneticists now working with [7:44] computational scientists and machine [7:45] learning researchers [7:49] personally for myself I've come come at [7:51] this more from the research side from [7:53] from academia training in medicine and [7:55] in physics [7:56] um mostly focused on target discovery [7:59] and development of new drugs. Uh but my [8:02] experience in drug development in big [8:03] farmer has really opened up my eyes and [8:06] and you know and I could speak more [8:07] about this later you know on how the [8:09] development aspect has really had less [8:12] focus in terms of where AI can play a [8:15] significant role going forward. [8:18] >> Yeah. So I'll say making use of you kind [8:20] of uh preemptive my next question is [8:23] where are the opportunities right now [8:25] like what have you seen actually develop [8:27] and speed up the process of drug [8:29] discovery is it the research side is it [8:32] the D side of everything um we all know [8:35] clinical trials takes a significant [8:36] amount of time you we have not figured [8:38] out a way to do that quicker what have [8:40] you seen that actually works in the real [8:42] world um in terms of speeding up the [8:45] process [8:46] >> yourself [8:47] >> okay I I start. So I just want to also [8:50] uh make um uh a point uh that I actually [8:54] learned also the hard way um especially [8:56] in biotech or pharma overall. So there [8:59] is a gray zone between the D and the R [9:02] and that's the reverse uh you know [9:04] translational uh discovery or medicine [9:07] because um if you think about the RS are [9:09] of course people uh running a lot of [9:12] experimentations uh would it be [9:13] virtually or digitally or physically [9:15] like uh in silicon in vitro um uh but [9:19] then I think um uh the the link uh [9:22] between the the D and the R is quite [9:25] interesting because in the modern world [9:28] uh with the amount of data we [9:29] accumulated on the D side like real [9:32] world uh data clinical data claims you [9:35] name it uh biomarkers um uh EHR so [9:40] insights from that uh world can actually [9:43] go in the other way in the opposite [9:46] direction to uh help uh people find new [9:50] pathways or just find the hit um and [9:54] experiment uh maybe um you know hundred [9:58] or you know even thousand instead of [10:00] millions of experimentations. So so [10:02] that's one. So I think quickly I think [10:06] to to understand where we are today I [10:09] think it's it's very important to just [10:10] follow the life cycle of a molecule. Um [10:13] I think um for me the real uh [10:17] transformation was like uh 2016 15 maybe [10:20] with alpha fold where for the first time [10:23] uh it was really fascinating to see uh [10:26] AI uh folding u you know proteins and [10:30] and that that was like a a revelation [10:33] that we are into something before even [10:36] chat GPT um so that's one and I think [10:39] it's continues especially that you know [10:41] the Deep mind team uh put that uh model [10:46] uh you know out there to to the public. [10:49] And since then we have seen an emergence [10:51] of um you know human um uh in the loop [10:55] or um uh uh uh lab in the loop as well [10:59] uh where machine learning um is uh [11:02] augmenting researchers uh to discover [11:05] new compounds uh uh very fast. And I [11:08] have this example with um company called [11:11] uh uh Exencia I think um they developed [11:16] an OCD um uh went from uh discovery to [11:19] phase one in less than 30 in less than [11:23] 12 months I think or average 12 months [11:25] instead of four to 5 years and that [11:28] tells you where we are today with with [11:30] with um uh drug discovery and then for [11:33] the D for me it's operations right like [11:35] you need to make sure that uh From the [11:37] moment you design the molecule, you have [11:40] uh AI helping for um um u indications or [11:45] even uh targets identification. Machine [11:49] learning is very good and AI is very [11:50] good doing that. And then you go to the [11:52] operations that's into the life cycle [11:54] where uh machine learning can uh predict [11:57] or forecast the enrollment period for [11:59] patients. And then you go into uh even [12:03] using real world data to predict um uh [12:06] the the success rate for for the [12:09] clinical trial. uh a lot of people use [12:11] AI for digital pathology which is uh I [12:14] think uh one of the most exciting things [12:16] and also digital twins because we do [12:18] want to know what is going on in silico [12:22] uh before we go into in vivo because you [12:24] know it costs billions of dollars to run [12:27] a clinical research or run a clinical [12:29] trial um it takes 15 years uh in average [12:33] u and then uh 8% success rate so um you [12:39] have to define your your battle um in [12:42] what's going on today with AI in in in [12:44] biotech and you have to choose what you [12:47] are optimizing for. If you are [12:48] optimizing for speed uh for cost or for [12:53] uh success because you cannot have um [12:55] you know multiple objective functions [12:58] you need to determine one and then work [13:00] your way um uh back until you see like [13:04] what are the use cases that you can work [13:05] on. [13:06] >> Yeah. Do you agree? Are there certain [13:08] parts of research that it um AI just [13:12] inherently plays a much stronger role in [13:14] its opportunity? Um or is it just a [13:17] broad application that we should be [13:18] thinking about? [13:20] >> Yeah. Um it's kind kind of interesting [13:22] know these AI conferences we we now have [13:24] this new focus on healthcare and life [13:27] sciences but not truth be told you know [13:29] we those of us who've been working in in [13:32] drug development we've been using [13:34] machine learning for decades now right [13:36] so um that's the way you know a lot of [13:39] targets are discovered by analyzing huge [13:42] genomic data sets or transctomic data [13:45] sets um things that can't be analyzed by [13:48] the human eye so uh we've been deploying [13:50] ing you know unsupervised you know [13:52] clustering methods back to the early [13:54] days you know early 2000s you know we [13:57] don't call it AI but now it is I guess [13:59] it does get rebranded as AI so machine [14:01] learning machine is not new um in that [14:03] space if you think about the whole drug [14:05] development life cycle it's really a [14:08] train those of you new to this um [14:10] environment you start from what's the [14:14] disease or indication we're focusing on [14:16] so and then I'm trying to understand [14:18] what target should we be focused on. So [14:20] discovery of targets, right? And then [14:22] once you've got a target, so this may be [14:25] a protein that's over abundant in a [14:28] certain tumor, for example, you may [14:30] then, you know, f then want to shuttle [14:33] your efforts to understanding what [14:35] molecules you want to develop to [14:37] actually treat that target or bind that [14:39] target or edit it in some way. So once [14:42] you've got an so this could be an a [14:44] small molecule or an antibbody or some [14:48] um you know genetic medicine approach [14:50] once you've got a good lead candidate [14:52] you know you then you know have to work [14:54] with you know uh directory folks to you [14:56] know get approval to um put this uh [14:59] medication into humans it's called an [15:01] IND application and that's a bit of a a [15:04] nightmare you know mountains and [15:06] mountains of paperwork [15:08] you know have that my worst enemy and [15:11] Then then you have the clinical trials. [15:12] So it's it's really a train. So you're [15:14] kind of shuttling you know this basket [15:16] over the fence one over the other. And [15:19] if you look at where AI has played a [15:22] significant role, it is monotonically [15:24] decreasing from target discovery to lead [15:27] candidate optimization to clinical [15:29] trials. Right? So AI has played a huge [15:32] role up front in early in the train and [15:36] it's gone down significantly as you get [15:38] to um you know clinical trial [15:41] optimization and build world evidence [15:42] data although we are seeing some [15:45] seedlings in in those um more recently [15:49] and why why is why is there a strong [15:53] correlation between AI you know usage [15:54] and where we are on the train it's [15:56] really comes down to the data [15:58] there's so much data for target [16:01] discovery that's available publicly [16:03] available. You know genome sequencing [16:06] has been commoditized. We've gone from [16:08] you know 10 years $1 billion for one [16:10] genome to now having millions of genomes [16:12] readily available for us to analyze at [16:15] scale and that's great. And so now we've [16:19] identified so many new targets over the [16:20] past 20 years for so many different [16:22] indications. In fact, I'll go one step [16:24] further. We probably have an [16:25] embarrassment of riches. We have [16:26] probably have way too many targets. um [16:28] we don't know how to prioritize them. Um [16:31] our ability to predict um which um [16:35] medications will work effectively in [16:36] humans is terrible, [16:39] right? And the reason being is because [16:41] we have very little data that is shared, [16:43] commoditized and publicly available. [16:45] It's all siloed within big pharma [16:47] companies, right? For, you know, for um [16:51] reasons that you can postulate regarding [16:53] IP. And then um once you've got a you [16:57] know a drug in clinical trials there is [17:00] a 90% failure rate. Now you have to be [17:04] really really an optimistic person to [17:06] work in an industry where you know [17:07] you're gonna be failing 90% of the time. [17:09] Right? And it's and this is this a [17:11] starting point once you've already done [17:13] all the hard you know research and [17:15] development to getting a drug and there [17:18] we have very little ability to predict [17:20] which drugs will pass muster in phase [17:22] one, phase two and phase three. And if [17:25] we could have the data available to us, [17:28] we could maybe utilize, you know, [17:30] machine learning predictive models to [17:32] better inform these um trials. But [17:34] that's where we're failing right now. So [17:37] long story short, doing very well with [17:38] research, not so great with development, [17:40] terrible with clinical. [17:42] >> Yeah. And and I think that kind of goes [17:44] to we're kind of building the cart [17:46] without having the directions in biology [17:49] in its own. We still don't understand [17:51] every how everything works. So, we're [17:53] trying to, you know, figure out um how [17:54] to build a car without the instructions. [17:56] Nur from a large, you probably see more [17:59] companies than we all do in terms of how [18:01] companies are um addressing both what [18:03] Yousef and Mickey had uh talked about. [18:06] How are companies in general just [18:08] dealing with the data that you need to [18:10] to inform these AI programs and how are [18:14] they actually implementing them? [18:15] >> Yeah, great great question. I mean, I I [18:17] just want to piggyback on on both the [18:19] themes that I heard, right? Two themes [18:20] are front and center. 10% success rate. [18:23] It's appalling, right? 90% failure rate. [18:25] Second is um you know this the cycles [18:29] are so long that we lose focus on the [18:31] fact that at the end of the day we're [18:32] focusing on on product and product life [18:34] cycles, right? And the feedback back and [18:36] this continuous evolution cycle doesn't [18:38] happen just because the the the time [18:40] spans are so long. You have a one shot. [18:43] It's a linear process, right? As as [18:45] Mickey said. So um you know the key [18:48] things that that at least we are seeing [18:50] is you know on the research space [18:52] obviously you know there's there's you [18:54] know a lot of progress has been made [18:56] obviously in very targeted lab scenarios [18:59] you know um um uh biotechs etc. [19:03] But a lot of the focus is now on you [19:06] know where where there are real [19:08] possibilities and business cases are [19:10] around can we actually shave off among [19:13] the 10 years at least 2 years on [19:15] optimizing clinical trials and [19:17] regulatory submissions right there are [19:19] real practical areas where we can [19:22] actually make progress. uh integrated [19:24] evidence planning is is uh you know [19:26] getting a resurgence just because of the [19:28] possibilities and the pace at which we [19:29] can actually start to right up front [19:32] right I know I talk about PTRS but in [19:34] all reality is it's PTR and CS it's [19:37] commercial success as well right and you [19:39] know a lot of the focus now becomes how [19:42] do we actually right up front at the [19:44] time of clinical trial not just look at [19:46] the scientific side of things but you [19:48] know can we look at regulatory pathways [19:51] do I really need to spend 56 6 weeks of [19:53] MRI scans or can I can did somebody else [19:56] get an approval by just doing it in 36 [19:58] weeks? Can I actually save time in my [20:00] clinical trials? Right? Can I save cost [20:02] on my clinical trials? Um, you know, are [20:04] there regulatory pathways that I can I [20:06] can I I can be proactive about? um you [20:09] know uh can I actually know the target [20:11] market segment in a incredibly crowded [20:14] market that I can actually play in for [20:17] commercial success and actually doing a [20:20] lot of that planning upfront with [20:21] available data with integrated evidence [20:23] planning uh creating target product [20:26] profiles doing a you know health tech [20:29] assessment up front. These are all real [20:31] opportunities to essentially you know [20:33] make sure that your development [20:35] strategies and plans you know as you're [20:38] going through designing your clinical [20:39] trials um you know etc are are well [20:42] aligned with the asset strategy that [20:44] you're trying to design and that itself [20:46] has the potential to shave off quite a [20:48] bit of that that life cycle right um and [20:51] so real practical areas where we are [20:54] seeing a lot of resurgence just because [20:56] of the ability of genai to start parsing [20:58] through through you know unstru [20:59] structured data, competitive [21:01] intelligence, literature mining. These [21:03] are all things that are are real. Um, [21:05] you know, other areas that are are [21:07] really coming up, you know, from from [21:09] last patient out, you know, from [21:10] database lock, how quickly can I get an [21:12] ECTD filing done? How can I effectively [21:15] engage with with um, you know, agencies, [21:17] these are all areas where there's a lot [21:20] of focus right now that that pharma [21:22] companies are taking. [21:23] >> Yeah. What I what I loved about which [21:25] job is you touched on all the different [21:26] ways in which we can utilize it. Um [21:29] something that's going to be a little [21:30] bit contentious to some is that um we [21:33] are dealing with two very high um [21:38] I'll say very deep knowledge bases of AI [21:41] on one side and then biology chemistry [21:45] physics another very deep educational [21:48] way where you know people are very [21:50] ingrained in what they do um some people [21:53] when they talk about agents and they [21:54] talk about dealing with customers on our [21:56] side that's simple that sounds fantastic [21:58] to deal with. Um, we're dealing with two [22:01] two different groups that are trying to [22:03] interact with each other and learn from [22:04] each other, um, who don't always know [22:07] how to speak, um, the other person's [22:10] language. How are the three of you [22:13] dealing with the computer science AI um, [22:18] education portion [22:19] >> and integrating that with the biological [22:22] chemical knowledge base that it is? How [22:25] are you intertwining those two groups? [22:27] How are you making those work? Is there [22:29] anything that has worked or has not [22:30] worked, Mickey? [22:32] >> Yeah. So, the researcher [22:35] >> Yeah. How much time do you have? Right. [22:37] I mean, I I think, you know, just, you [22:40] know, the way you phrase the question [22:42] and outlines the real challenge. I mean, [22:43] it's two different groups, right? And it [22:46] really shouldn't be two different [22:47] groups. It should be one harmonized um, [22:50] you know, interdicciplinary [22:51] organization. And I've seen this, you [22:53] know, you know, fail time and time again [22:57] in in in in [22:59] well-meaning large organizations. Um, [23:02] typical story, you know, there's a [23:04] mandate from top above and, you know, [23:06] someone says, "Oh, we got to have AI in [23:08] organization, right? The CEO, you know, [23:11] reads a New York Times, this is what we [23:13] should be focusing on, right?" So, they [23:14] hire a bunch of data scientists, machine [23:16] learning scientists. um they're often in [23:19] their own department, sometimes in [23:21] different building, right? They never [23:23] get to speak to biologists. They never [23:25] get to speak to work with wet lab folks. [23:27] And that always fails, right? Because [23:30] you're just not understanding what the [23:31] right problems you are need to be [23:34] focused on. where you know um you know [23:37] I've tried to change um uh the culture [23:41] um somewhat is having an integrated [23:42] group where um the data scientists are [23:44] imshed with the biologists the ones who [23:47] actually generating the data. So in my [23:49] in my past life um they probably hate me [23:52] now but you know I actually required you [23:54] know a lot of the data scientists to [23:55] actually rotate in the wet lab and [23:58] actually generate their own data. So you [23:59] know if they're working on analyzing [24:01] large scale genomic data or single cell [24:03] sequencing data you know they would have [24:05] to rotate and actually generate those [24:07] libraries and you know and load them up [24:10] onto the sequencer and understand you [24:12] know what's the sources of variability [24:14] um and you know don't treat the data as [24:16] if it comes from God and vice versa. [24:18] I've had to educate and give lectures to [24:21] the experimentalists on basic data [24:24] science. you know, don't get scared of [24:25] the, you know, the the command shell uh [24:28] terminal on your on your computer screen [24:29] and use that for basic data scripting. [24:32] Um, that's been enormously productive [24:34] for us. So, it allows us to close that [24:36] the feedback loop much more efficiently. [24:38] Um, you're not throwing, you know, the [24:40] basket of data over the events. People [24:42] are able to have meaningful [24:43] conversations with each other without, [24:45] you know, treating machine learning [24:47] scientists as if they're wizards, right? [24:50] I mean it's a horrible waste of a brain [24:52] to give you know them a bunch of data [24:54] and say well give us a p value and a [24:56] volcano plot right and treat that as if [24:59] you know it was given to you by you know [25:01] a mystical wizard right knowing [25:04] you know all the issues that come right [25:05] that can come with an analysis of data. [25:08] So having an integrated group um it does [25:10] take a lot of work. You do have to put [25:12] your professor hat on and you know you [25:14] know and dedicate time to pedagogy [25:16] education workshops um but you know I [25:20] think the dividends pay off in the end. [25:22] >> Are you trying to say that centers of [25:23] excellence are here to die finally. [25:28] >> Give me an example of a good one and [25:30] I'll respond in kind. Um Ysef [25:34] >> yeah sorry this is for me uh twofold. Uh [25:39] we have a a historic um reason um for [25:43] why we have that you know um by polarity [25:47] or something like that. Um and also uh [25:50] organizational. So uh like I said we're [25:53] you know before the uh the panel [25:56] organizational just because before we [25:59] talk about the AI people and the biology [26:01] and scientists there is already a [26:04] dichotomy between development and [26:05] research. I think first thing first um [26:08] you know uh from process perspective it [26:11] needs to be in pharma because it's an [26:13] old industry a um integrated model uh [26:18] that starts with what you I like what [26:19] you said about the data earlier uh [26:21] Mickey I think we need to have a uh [26:24] seamless way of making data accessible [26:27] for development people to know uh you [26:30] know the mechanism of um uh developing [26:33] different compounds uh and do shared uh [26:37] uh uh data for example um having a a [26:41] repository uh where you store um you [26:44] know uh CT scans and and imaging that's [26:48] actually a common data set that [26:50] development people and research people [26:52] can work on and believe it or not most [26:54] of the pharma people they don't even [26:56] have a repository because they go to CRO [26:59] they do their clinical trial and data [27:00] stays at the CRO level so they really [27:03] don't bring it back so that's number [27:05] Number two, I think the historic one is [27:07] when the first time we we we used [27:11] computers heavily was of course the [27:13] evolution of computer but also when we [27:15] had these bioinformaticians and [27:17] biomethematicians [27:19] uh starting to you know use uh digital [27:23] tools uh to to accelerate discovery. But [27:26] there was another wave um which is the [27:29] data science wave which was like oh [27:31] there is a new discipline called data [27:32] science though did exist like since uh [27:35] the creation of mathematics. So so there [27:39] is that that um dichotomy also between [27:43] you are bioinformatician I am AI so [27:45] there is that AI divide that is also not [27:48] helping today. Um and I think to solve [27:51] it of course you talked about center of [27:53] excellence uh but also um really um you [27:56] know by experience this is more [27:58] practical. Um my my AI researchers are [28:01] not all uh PhDs in mathematics or [28:04] statistics computer science. So I have [28:08] uh PhDs in um uh biomedical um even [28:12] biologists people who converted into AI [28:16] uh uh on the job um either in their lab [28:19] or after joining uh the company and [28:22] actually they are the best ones why [28:24] because they understand the subject [28:26] matter expertise they're able to talk [28:28] science uh freely and easily uh with the [28:31] R scientists on the R side uh and bring [28:34] it back to to the D and believe it or [28:36] not there are a lot of common projects [28:38] and use cases between the research [28:40] people and development people. So for me [28:43] it's a paradigm shift uh and um really [28:47] going uh or walking away from this model [28:49] also of SAS because if every single team [28:52] every single department who has a [28:54] business or research question to answer [28:56] goes and license the software uh it [28:59] doesn't help uh either because that's a [29:01] lot of silos and people work in a [29:03] vacuum. I think AI can help also [29:06] bringing people together around what I [29:08] call a unified experience and I think [29:10] everything now is becoming agentic and [29:12] if you have that like I said in the [29:14] beginning integrated framework for the [29:16] data making sure that your data is [29:18] accessible but but the right people but [29:20] doesn't matter where they sit uh and [29:22] give them the the the same tools to to [29:24] to create that continuum uh without that [29:28] I think yes like it's an old industry um [29:30] we will always have that divide that [29:32] other industry is fixed because they [29:34] have this integrated experience. [29:37] >> I think all all great points. I I'll [29:39] just build on that. Right. I think we [29:40] we're talking about the divide between [29:41] the the biologists and the the the data [29:44] scientists. I'll go a step further and [29:45] just also talk about about engineering [29:47] and platforms, right? That's another key [29:50] aspect. Um you know, the reason I why I [29:52] say that is there is just so much grunt [29:55] work around, you know, just what you [29:56] brought up, right? Repositories around [29:58] around, you know, multimodal data, [30:00] right? Um, you know, so many of our [30:03] clients don't have multimodal data [30:05] platforms set up in a way that can [30:06] really accelerate all things research. [30:09] It really needs to be backed up by [30:11] strong platform-minded engineering teams [30:12] that are able to get some of those out [30:14] of the way for for a lot of the [30:16] scientists and the the biologists, [30:17] right? Um, you know, the same thing as [30:20] as we look look towards, you know, the [30:22] rest of the continuum. um you know a lot [30:24] of the focus on uh scientific knowledge [30:26] mining, regulatory content mining. These [30:29] are all opportunities where taking a [30:31] platform-based approach, you know, [30:33] removing those hurdles and and [30:35] essentially giving both the the data [30:37] scientists as well as the the the [30:39] biologists a starting point is is is [30:42] really something that you know is now a [30:45] need of the hour to accelerate and we [30:47] are seeing a lot of movement in that [30:48] space to just sort of start to [30:50] consolidate you know siloed programs and [30:52] start thinking in terms of you know [30:54] where where we can accelerate with [30:55] platforms and you know I know we we talk [30:57] about it in terms of technology platform [30:59] The other aspect is also you know just [31:01] um you know um you know uh research [31:04] platforms right MRNA and others as [31:05] portfolios start to consolidate where do [31:08] companies want to focus and how do they [31:09] take a therapeutic area based approach [31:11] to essentially start investing is [31:13] another aspect of it. [31:15] >> Great. So one last question before we [31:17] kind of open up for questions from the [31:19] uh from the crowd. I want you to all [31:21] hypothetically put your investor hat on [31:23] um and say, you know, I give you $10 [31:25] million, you know, make a bet over the [31:28] next five years on what portion of that, [31:31] if you want to call it a train or um [31:33] part of the drug discovery, where do you [31:35] think we'll have the highest rate of [31:36] return in terms of utilizing AI, what [31:38] would you bet on it's going to work over [31:40] the next five, maybe 10 years, [31:43] >> Mickey? Yeah, I mean [31:47] we [31:48] as scientists know have done a really [31:50] great job of curing cancer in mice, [31:54] right? But last thing I last time I [31:56] checked, I'm not a mouse, right? And so [31:59] um know there's a lot of effort in you [32:01] know research in understanding the the [32:03] effect of molecules biologics um in [32:06] treating various indications and we have [32:09] fantastic you know publications and [32:11] science that's u reproducible the [32:14] attrition is it does not translate to [32:17] human physiology right we have very [32:19] little ability to predict why these [32:21] things fail right despite all the [32:23] successes and you know all the Nobel [32:26] prizes and all the awards that we've [32:27] given to basic science research. [32:29] Understanding why you know you know a [32:32] small molecule does not have the same [32:34] potency in in a human is really being a [32:37] big challenge and actually it has driven [32:38] up all the costs you know it's it's now [32:42] over $2 billion [32:44] um and over 10 years for you know the [32:46] life cycle of a successful drug. So if [32:49] AI can make a stab at that [32:53] understanding how to derisk pre-clinical [32:55] assets to a clinical sphere, I think [32:57] that would have a major major impact on [32:59] drug development. [33:01] >> Yeah. Uh this is a great question. So [33:04] it's timely also because I was just [33:06] working on my three-year plan uh and and [33:08] I have this principle um you know for me [33:12] to spend the 10 million. 10 million is [33:14] not a lot for I want to what I want to [33:16] do uh but but you know supposedly so so [33:19] defining the the like I said the [33:21] business what I call the business [33:22] objective function not to make it 100% [33:24] mathematical adding the business to it [33:27] defining what are we optimizing for in [33:29] this process uh as far as I'm concerned [33:31] on the D side so I want to accelerate [33:33] the timeline uh to uh market and by uh [33:38] you know saying that time to market is [33:40] also time time to submission because [33:41] that's the race between pharma companies [33:44] who's going faster for similar [33:46] indications and sim similar biological [33:49] technologies to to FDA or AMA or [33:52] whatever regulation. So I think that [33:55] that lack of predictability in certain [33:57] tas for example oncology um inflammatory [34:02] uh disease are are less predictable than [34:04] verology. So I would definitely first [34:07] and foremost pick two areas in these two [34:10] uh and then I will invest heavily in uh [34:13] in the early uh part of clinical [34:15] pharmarmacology with simulations. It's [34:18] an amazing space. I'm you know a huge [34:20] fan of my colleagues who are part of [34:21] that clean farm organization. So I [34:24] always really give them shoutouts that [34:26] are scientists uh and that are very [34:29] analytical and they generate a lot of [34:31] data for example pap you know like a lot [34:33] of simulations dozing very important to [34:36] know that that these things um at scale [34:40] using simulation and digital twins I [34:42] think that's the future because you can [34:44] predict digitally and virtually uh what [34:47] can happen before you go spend billions [34:49] in a clinical trial for the last you [34:52] know the upcoming 10 It's really [34:53] shortening the the timelines by [34:56] investing early on. [34:59] >> Nur, what are you going to spend your 10 [35:00] million on? [35:01] >> Yeah, I I think um you know I agree with [35:05] both [35:06] Mickey and Sophie, but you know I think [35:08] if I look at where the success rate is [35:12] going back to Mickey's point, right [35:15] coming out of research, that's obviously [35:16] the biggest space right now where [35:18] acceleration is needed, right? [35:20] translational research going from from [35:22] animal trials into into getting real [35:25] proof points is is where you know a lot [35:27] of the focus is my hope is that over the [35:30] next 5 years that's something that that [35:32] we'll be able to overcome because I [35:35] think for all the other areas that we [35:37] are talking about I think a lot of [35:38] progress is being made on the D side [35:40] right um the other part is also I mean [35:43] it's not just on the research side as [35:44] well as pharma it's also regulators [35:47] right as we look at this end toend life [35:48] cycle um you the the transparency coming [35:51] out of the regulators is is really [35:53] improving. um you know if anybody is [35:56] tracking there were 200 you know uh CRLs [35:58] that were released by by the FDA [36:00] recently um you know on on you know uh [36:03] essentially the you know uh reasons for [36:06] rejection right really really important [36:08] data sets to say you know what is it [36:10] that'll take shorten the timelines in [36:12] terms of um making sure that not just [36:15] from a from a technical but regulatory [36:18] standpoint as well what would it take to [36:20] actually cut down some of the life [36:21] cycles to get these drugs faster to [36:23] market. So the transparency is improving [36:25] but it's we have to remember it's an [36:27] ecosystem so it'll take everyone to do [36:29] their bid everything from from [36:30] researchers to development you know [36:32] within malage pharma as well as [36:33] regulators right [36:34] >> yeah real quick nur do you think [36:35] regulatory is helping or hurting AI [36:38] progress in the field [36:40] >> uh I think it's early days to actually [36:42] you know sort of answer whether it's [36:44] helping or hurting I I do know that [36:45] progress being is made tremendously fast [36:48] right everybody's asking the right [36:49] questions to say hey what should we be [36:51] doing from a regulatory standpoint to [36:53] sort of you know accelerate the use of [36:55] AI. The concerns are real too, right? [36:58] And and not not unfounded. Um you know [37:01] so um I I wouldn't say it's helping or [37:03] hurting. I I think the the opportunity [37:05] is now to make sure that the guidance [37:06] are coming out as quickly as possible to [37:08] accelerate. [37:09] >> All right. Fantastic. So with that I [37:11] want to open up to the audience. Does [37:13] anyone have any questions for anyone on [37:14] the panel? Uh there is we have a [37:19] >> microphone as well. I know my question [37:21] I'm a biologist but also a genomic [37:23] scientist. I know when you consider the [37:26] timeline okay the life cycle it's when [37:28] we whenever we talk about a drug come to [37:31] in reality takes 10 to 15 years or 20 [37:33] years that includes also postmarket [37:36] surveillance. Uh given this time as life [37:39] cycle uh where do you think that for [37:42] example there is a potential uh option [37:44] to go to the preclinical area? uh [37:47] preclinical again it's again this mice [37:49] going to help but now people I mean [37:51] government has different policy well [37:52] let's use human cells to do the [37:55] preclinical study but again it is going [37:56] to add more variance so what do you guys [37:59] think about I think that is one of the [38:01] place where there should be more role [38:04] for AI you know this you talked about [38:06] genomics transcript everything uh what [38:09] do you guys think about the recent [38:11] multimodel approach people talks here [38:13] even where it could be kind of break the [38:16] sylloss uh so that way it could [38:18] accelerate the whole process in a [38:22] positive way. [38:23] >> I need to take a stab at that. [38:26] >> I mean I I think it's related to the the [38:28] questions that were raised earlier about [38:30] you know focus and honestly um I don't [38:33] think enough time is spent on asking [38:35] what are the right questions we should [38:36] be addressing. I think there's you know [38:39] um we're all like kids in a candy store. [38:42] Every day there's a new shiny tool that [38:44] emerges in the AI space and it's very [38:47] easy to get overwhelmed and unless you [38:50] have a clear AI strategy, right? You're [38:53] going to be diluting your efforts and [38:54] chasing every new shiny toy and not [38:57] actually moving the needle in any [38:58] meaningful aspect. Uh I'll give you an [39:01] example of, you know, where I think um [39:04] AI is perhaps the barrel of AI is [39:07] perhaps focused on the wrong target, [39:09] wrong problem or area. So um in drug [39:12] development um understanding how a [39:15] molecule or drug or biologic binds to a [39:19] target is a a critical step right and [39:22] there's a lot of um well-meaning [39:24] well-intentioned efforts by AI [39:26] scientists of designing new molecules in [39:29] silicon on a computer that have better [39:32] binding properties to a target and this [39:34] is leveraging over the recent successes [39:36] of alpha fold you know can we predict [39:38] the structure [39:39] these two molecules binding together. [39:42] It's a very noble and aspirational uh [39:45] goal, but it's actually not that not [39:47] even that all that useful in drug [39:49] development, right? And the reason and I [39:52] think it's because as machine learning [39:54] scientists, we kind of overindex on uh [39:57] problems that are measurable rather than [40:00] meaningful. [40:02] Binding data is relatively easy to [40:04] capture. it's readily available and so [40:06] you can design, you know, a really good [40:08] predictor or generative algorithm for a [40:10] better binder to a target. But once [40:14] you've optimized that problem, you've [40:16] now created further problems downstream [40:18] you haven't thought about. A better [40:20] binder may have poor dropback [40:23] properties. It may be more toxic, right? [40:25] So if you're overoptimizing on one [40:27] aspect of the train, which you know, [40:30] which sounds like a great goal, you lose [40:32] the whole train, right? So this is why [40:34] it's really important to have this you [40:35] know culture or where the ML scientists [40:38] really understand the biology and get [40:40] involved in understanding developability [40:42] in the clinical trial aspect not just [40:45] focus on their own specific problems. It [40:47] isn't you you really can't over if you [40:49] start overoptimizing too early you will [40:51] fail. Yeah, Mickey, quick question on [40:53] that, right? How in your mind, how many [40:56] computationally generated, you know, [40:58] molecules, right? Even though they're [41:01] theoretically possible to find actually [41:03] make it to stable drugs that are [41:06] actually marketable, you know, I mean, [41:08] AI can generate a lot of possibilities, [41:10] but how many of those are real stable [41:12] drugs that can actually be marketed? Um [41:14] so I mean great question and the answer [41:16] is we don't know yet because it's so [41:17] little um effort focused on [41:19] understanding what makes a a good drug [41:23] right and and for most machine learning [41:25] scientists in the space to them a good [41:27] drug is something that binds well there [41:29] this there's so many other factors that [41:30] are much more much more relevant in fact [41:32] I'll go one step further and say the [41:35] plethora of incilodes [41:38] drugs is actually creating a new problem [41:41] right these are drugs which the human [41:42] body has not seen in our millions of [41:44] years of evolution. So you're creating [41:46] possible, you know, imunogicity side [41:49] effects, immune reaction, you know, [41:51] which which isn't thought about. I mean, [41:53] I'll be very hesitant to put something [41:55] into my body that was designed in [41:57] silicone knowing that, you know, it's [41:59] very very foreign to the, you know, the [42:02] the evolution of the human genome.