Recursion Cuts Drug Discovery Timelines to 18 Months With AI Platform

Recursion借助AI平台将药物研发周期缩短至18个月 工业化高通量表型组学与计算基础模型帮助生物医药行业取代假设驱动型研究，将靶点至IND的时间缩短一半以上。 Erik Douglas阐释了Recursion如何利用由自动化表型组学、多组学数据生成以及Boltz-2等基础模型驱动的迭代式TechBio平台，取代传统线性的制药摸索。在整合Exscientia之后，该平台在高效的“T型”研发漏斗中将广泛的计算机模拟（in silico）靶点探索与精准自动化化学相结合。通过引入因果AI来优化临床试验设计和患者分层，Recursion展现出显著的效率提升，包括在大约18个月内将项目从靶点发现推进至开展IND使能研究阶段。 Recursion 通过自动化湿实验设施每周进行超过 200 万次细胞成像实验，并依托与 NVIDIA 联合打造的 BioHive-2 超算训练掩码自编码器等深度学习模型，实现无偏表型组学驱动的药物研发。 (3:05) 公司将技术拓展至转录组学及活细胞明场表型等领域，并与 MIT 联合开源 Boltz-2 分子基础模型，以极低的计算成本实现媲美自由能微扰（FEP）模拟精度的分子结合亲和力预测。 (6:40) 整合 Exscientia 后，Recursion 结合表型筛选与生成式分子设计推出“T 型漏斗”模式，在早期进行大规模全景式干实验计算探索，取代了传统制药高成本的“V 型”研发路径。 (11:25) 通过因果 AI 算法优化临床试验设计并加速患者入组，Recursion 将内部管线从靶点确定到进入 IND 申报阶段的时间压缩至约 18 个月，显著快于行业约 42 个月的平均周期。 (15:58)

Erik Douglas, VP of Product at Recursion, argues that the pharmaceutical industry's declining productivity can be reversed by replacing slow, hypothesis-driven research with an industrialized 'TechBio' operating system combining automated wet labs, supercomputing, and machine learning foundation models. Industrializing Phenomic Data Generation and Supercomputing (3:05) Recursion's platform centers on high-throughput phenomics, primarily using multiplexed fluorescent cell painting and brightfield imaging to capture morphological shifts when cells encounter chemical compounds or genetic knockouts. Operating automated laboratories in Salt Lake City and the United Kingdom, the company generates more than 2 million experiments and biological images per week, creating high-dimensional fingerprints of biological activity without requiring predetermined molecular targets. To process this data volume, Recursion collaborated with NVIDIA to build BioHive-2, a supercomputer ranked among the top 70 globally. The infrastructure trains deep computer vision architectures, such as masked autoencoders (MAEs) where 75% of cellular imagery is masked during self-supervised training. According to Douglas, scaling these architectures from an 86-million-parameter model trained on a quarter of their dataset to a 370-million-parameter model trained on the full dataset yielded marked improvements in visual biological reconstruction and predictive accuracy. Multimodal Omics and Biomolecular Interaction Modeling (6:39) Beyond cellular imaging, Recursion incorporates multimodal biological measurements, including over one million sequenced transcriptomic profiles, label-free brightfield live-cell imaging, and early automated absorption, distribution, metabolism, and excretion (ADME) pharmacology assays. This observational phenotypic platform is paired with structural biomolecular modeling to evaluate drug candidates earlier in the discovery cycle. To predict binding affinities, Recursion co-developed and open-sourced Boltz-2 in collaboration with researchers at MIT. Douglas describes Boltz-2 as a foundation model designed to predict molecular interactions with accuracy comparable to Free Energy Perturbation (FEP) simulations. Because Boltz-2 runs at a fraction of the computational expense and time required by traditional FEP, it can be deployed systematically across entire compound libraries rather than serving as a late-stage screening bottleneck. The T-Shaped Funnel and Automated Chemistry (11:25) Following its late-2024 combination with Exscientia, Recursion integrated its phenotypic target deconvolution tools with Exscientia's automated precision chemistry and generative molecular design workflows. Douglas contrasts this combined operating model with traditional pharmaceutical 'V-shaped' funnels, which invest heavily in testing narrow, early hypotheses through expensive physical assays. Instead, Recursion uses a 'T-shaped' funnel. The broad horizontal bar represents broad in silico exploration across biological space using phenomics, transcriptomics, Boltz-2, and large language model literature mapping. By evaluating target genes against public biomedical literature, Recursion identified a cluster of high-relevance oncology targets with extensive internal experimental data that remain virtually unaddressed in published literature. Once high-probability therapeutic pockets are identified, the vertical stem of the funnel applies automated robotic synthesis and focused assays to optimize leads rapidly. Translational Causal AI and Compressed Timelines (15:58) Recursion applies causal AI to human genomic datasets and real-world clinical records to refine clinical trial designs, stratify patient cohorts, and optimize trial site selection. In rare disease indications, the company projects that algorithmic site selection can accelerate patient enrollment timelines by up to 50%. Douglas points to internal development metrics as empirical evidence of the platform's operational velocity. In one internal therapeutic program, the company advanced from initial target identification to IND-enabling (Investigational New Drug) studies in approximately 18 months, compared to the pharmaceutical industry baseline of roughly 42 months. The company attributes this compression to earlier computational target validation and fewer iterative chemistry cycles. Recursion contends that shifting from siloed, single-hypothesis biotech workflows to an iterative, platform-scale techbio loop can systematically de-risk drug development and reduce capital costs. The ultimate industry impact of this model depends on whether accelerated transitions from target identification to IND-enabling studies reliably translate into higher clinical trial success rates.

Recursion产品副总裁Erik Douglas指出，传统制药业长期面临研发周期拉长且成本攀升的逆摩尔定律困境；该公司正通过自动化湿实验、超算集群与多模态AI模型构建闭环操作系统，以无偏见的表型组学与计算驱动模式重塑新药研发流程。 高通量表型组学与超算支持下的工业化数据生成 (3:05) 传统生物学机制极其复杂，人类目前仅理解生物系统的一小部分。为了摆脱预先设定的单一靶点假设，Recursion以无偏见的高内涵荧光细胞成像（Cell Painting）为基础，通过染色细胞核与细胞骨架等结构，系统记录化合物或基因敲除引起的形态学高维指纹变化。这种自下而上的表型组学方法将观察延伸至活细胞在受到扰动时的真实反应，避免了早期研究对单一通路的过度依赖。 为了支撑深度学习模型的规模化训练，Recursion在犹他州盐湖城以及英国伦敦和牛津附近建立了高度自动化湿实验设施，目前每周产生超过200万次自动化实验与图像，平均每15分钟产生相当于一篇博士论文的数据量。在算力端，公司与NVIDIA合作部署了全球排名前70位的BioHive-2超级计算机，成为制药行业内算力最强的计算集群之一。借助海量自研图像数据，研究团队开发了掩码自编码器（MAE）等计算机视觉模型，自监督学习中遮蔽率达75%；模型参数量从8600万扩展至3.7亿后，图像重建与生物活性预测性能均展现出显著提升。 多组学整合与Boltz-2分子动力学模型的结合 (6:39) 在荧光细胞成像之外，Recursion进一步扩展为多模态、多组学平台。公司引入无标记明场成像以捕捉未经染料固定的活细胞动态行为，测序了超过100万份转录组样本，并结合早期高通量ADME（药物吸收、分布、代谢与排泄）药理与毒理分析，以及真实世界患者临床数据。这些多层次的数据流被整合为统一的生物表型全景图谱，使研究人员能够从不同生物学尺度同时观察细胞网络的变化。 在生物分子设计层面，Recursion与MIT研究人员合作开发了开源生物分子基础模型Boltz-2。与专注蛋白质三维结构预测的AlphaFold不同，Boltz-2专为小分子候选药物的结合亲和力与分子相互作用预测而设计。该模型在保持与黄金标准自由能微扰（FEP）模拟相当的预测精度的同时，大幅压缩了计算时间与算力成本，使得高精度的结构设计能够前置并直接应用于整个化合物库的早期筛选，消除了传统流程末端的高通量瓶颈。 整合Exscientia与T型发现漏斗重塑管线策略 (11:25) 2024年底，Recursion完成了对英国公司Exscientia的业务合并，将自身的自上而下表型筛选及靶点反卷积能力，与Exscientia的自动化合成化学和生成式分子设计流程进行了全面融合。合并后的操作系统打通了原子尺度预测模型、自动化机器人化学合成和表型验证环节，使生物学评估与化学优化在高度平行的闭环中持续迭代，打破了以往各研发阶段割裂移交的传统机制。 Douglas将这种研发逻辑概括为“T型”发现漏斗，用于替代传统制药高成本试错的“V型”漏斗。传统流程往往始于狭窄的假设，在向下推进时耗费巨额资金进行繁琐的物理筛选；Recursion的“T型”顶部则利用计算机技术在整个生物学空间内进行广泛探索，结合大语言模型解析文献、Boltz-2结合力预测与海量多组学扰动数据，锁定高潜力靶区。在肿瘤学领域，平台在大语言模型文献分析中不仅验证了已知的肿瘤靶点，还在公开发表文献极少但公司私有数据丰富的象限内发现了密集的基因簇，为开发同类首创（first-in-class）药物提供了明确切入点。 因果AI赋能临床转化与推进IND阶段的实证表现 (15:58) 该平台的计算与数据驱动策略不仅限于实验室早期的候选分子发现，还直接延伸至转化医学与临床试验运营。Recursion通过因果AI分析人类基因组学数据集与真实世界临床证据，优化临床试验方案设计、提升患者分层精度并改进试验机构的选择策略。Douglas指出，这一方案针对罕见病临床试验招募困难的行业痛点，预计通过算法优化试验中心选址可使患者入组时间缩短50%。 在开发周期的实质加速方面，Douglas以一项内部新药管线为例：该项目从最初的靶点识别推进至IND试验申报（IND-enabling studies）阶段仅耗时约18个月，而传统制药行业的平均周期约为42个月。这一提速得益于平台在研发早期对靶点假设的高效验证、减少先导化合物优化的迭代轮次，进而降低了跨周期的资金消耗与时间成本。 Recursion将自身定位为以平台技术为核心的TechBio企业，旨在通过自动化、大规模多模态数据与AI闭环来打破传统Biotech高度依赖个体科学家单一假设的高风险研发范式。不过，将计算机模拟预测和体外表型模型大规模转化为最终临床成功与上市药物，仍有待更多管线在后续人体临床试验中提供决定性的有效性验证。

## Background 本视频由人工智能行业大会平台 Ai4 于 2025 年 8 月 28 日在 YouTube 上发布。演讲由 Recursion 产品副总裁 Erik Douglas 主讲，视频简介指出，Recursion 凭借十余年积累的 65PB 专有多模态生物数据以及与 NVIDIA 联合打造的 BioHive-2 超级计算机，通过 Recursion 操作系统全面革新传统药物研发流程，以显著的效率与成本优势推进前沿候选药物管线。 ## TL;DR 演讲系统阐述了 TechBio 领军企业 Recursion 如何借助“数据+AI+算力”范式打破生物医药研发的成本与效率瓶颈。公司通过自动化实验室每周产出数百万次 Cell Painting 等高内涵表型实验，依托超算 BioHive-2 训练大规模自编码模型，并联合 MIT 开源了逼近自由能微扰精度的结合亲和力大模型 Boltz-2。结合对 Exscientia 分子设计能力的整合与因果 AI 临床试验优化，Recursion 打造了端到端的 T 型研发漏斗，成功将临床前推进周期从行业平均 42 个月压缩至 18 个月。 ## Atmosphere & style 演讲整体呈现出严谨、前沿且充满自信的工业科技质感。主讲人 Erik Douglas 叙述从容流畅，开场以代谢图谱与同行熟知的生物学复杂性引发共鸣，随后层层递进地展现海量高通量实验与超算模型的工业化威力。整场分享兼具深厚的数据说服力与宏大的变革视野，清晰传递出 TechBio 平台化范式颠覆传统“碰运气式”制药的笃定与振奋。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=6BqIMXr3TLg&t=4s) 面对传统药物研发面临的周期漫长与成本高昂等逆摩尔定律困局，Recursion 旨在将计算智能、高通量自动化湿实验与工程体系深度结合，系统性加速新药发现。 - [02:04](https://www.youtube.com/watch?v=6BqIMXr3TLg&t=124s) 采用基于表型组学（Phenomics）与 Cell Painting 荧光成像技术，打破传统单一靶点假设的局限，以无偏见（unbiased）的方式从全细胞形态学变化中发掘未知生物学通路。 - [04:16](https://www.youtube.com/watch?v=6BqIMXr3TLg&t=256s) 盐湖城自动化湿实验平台以“每 15 分钟产出一名博士论文数据量”的高通量节奏运转，每周生成超 200 万次实验与成像，为机器学习模型训练提供专有数据基础。 - [04:16](https://www.youtube.com/watch?v=6BqIMXr3TLg&t=256s) 表型掩码自编码器（MAE）展现出显著的模型缩放定律（Scaling Laws），联合 NVIDIA 打造的行业顶尖超级计算机 BioHive-2 为海量图像表征训练提供算力支撑。 - [06:38](https://www.youtube.com/watch?v=6BqIMXr3TLg&t=398s) 平台构建了多组学图谱，将无标记明场表型组学、超百万样本转录组学、高通量 ADME 药理毒理数据与患者真实世界数据（RWD）深度映射，贯通微观机制到临床表征。 - [08:23](https://www.youtube.com/watch?v=6BqIMXr3TLg&t=503s) 与 MIT 联合开源基础模型 Boltz-2，专注预测小分子与靶蛋白结合亲和力，在逼近自由能微扰（FEP）金标准精度的同时极大降低计算开销，支持早期高通量虚拟筛选。 - [11:02](https://www.youtube.com/watch?v=6BqIMXr3TLg&t=662s) 通过整合 Exscientia 的分子与化学设计能力，打通“自上而下”的表型机制解卷积与“自下而上”的原子级分子生成优化，实现全流程闭环研发。 - [11:02](https://www.youtube.com/watch?v=6BqIMXr3TLg&t=662s) 平台形成高度自进化的研发飞轮效应，每一个管线产生的生物与化学实验数据都会持续回流，反哺并提升底层 AI 模型的预测精度与设计效率。 - [13:08](https://www.youtube.com/watch?v=6BqIMXr3TLg&t=788s) 提出“T 型研发漏斗”取代传统押注单一靶点的“V 型漏斗”，在顶端以计算模拟和高通量实验广泛探索生物学地形，大幅降低前期试错成本。 - [13:08](https://www.youtube.com/watch?v=6BqIMXr3TLg&t=788s) 利用大语言模型与专有数据识别公开文献尚未充分探索、但与肿瘤高度相关的隐秘靶点，避开行业内卷并在未开发生物学生态位中建立先发优势。 - [15:58](https://www.youtube.com/watch?v=6BqIMXr3TLg&t=958s) 将因果 AI 应用于临床试验设计与患者招募，使中心入组预测速度提升 50%，并将靶点识别到 IND 赋能研究的临床前周期从行业平均 42 个月大幅缩短至 18 个月。 - [15:58](https://www.youtube.com/watch?v=6BqIMXr3TLg&t=958s) Erik 区分了 TechBio 与传统 Biotech 的根本范式差异：TechBio 依靠平台化数据与计算工具进行系统化、全景式的空间探索，而非依赖个别科学家的单一假设和运气。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=6BqIMXr3TLg&t=4s) 生物学的极致复杂性与传统制药面临的困局 Recursion 的核心使命是“解码生物学以从根本上改善人类生活”（decoding biology to radically improve lives）。然而，生命科学探索的道路并不平坦，生物学系统蕴含着超乎想象的复杂性。演讲者 Erik Douglas 在开场时展示了经典的柠檬酸循环（TCA 循环）代谢通路图，并与台下的生命科学和制药界同行互动打趣道，大家都深知“生物学绝不会轻易交出它的秘密”。目前人类对生物学运作机制的理解仅占极其微小的一部分。 与科技行业几十年来自豪遵循的摩尔定律（计算成本呈指数级下降、性能指数级上升）形成鲜明反差的是，生物医药行业正经历研发周期越来越漫长、研发成本越来越高昂的严峻困局。如今需要使用药物的患者群体极其庞大，面对未被满足的医疗健康需求，孤立的单点技术已无法逆转行业趋势。Recursion 致力于将计算智能体、高通量实验平台与全流程工程化体系紧密结合，借助不断跃升的技术平台能力，系统性地把更优质的药物以更快的速度推向市场。 ```mermaid flowchart LR accTitle: Recursion 平台研发飞轮 accDescription: 展示从高通量湿实验生成数据，经由 BioHive-2 超算训练模型，再到指导药物研发的递归迭代闭环。 subgraph LAB ["高通量自动化实验室"] A["高通量生物与化学实验<br/>(每周 >200 万次实验)"] B["多组学与高内涵成像数据<br/>(Cell Painting / 转录组 / ADME)"] end subgraph COMPUTE ["超算与模型训练"] C["BioHive-2 超算中心<br/>(与 NVIDIA 联合打造)"] D["多模态多尺度基础模型<br/>(表型 MAE / Boltz-2 等)"] end subgraph DISCOVERY ["药物发现与转化"] E["作用机制解析 & 分子设计"] F["临床转化与真实世界映射"] end A --> B B --> C C --> D D --> E E --> F F -. 递归迭代反馈.-> A ``` --- ### [02:04](https://www.youtube.com/watch?v=6BqIMXr3TLg&t=124s) 表型组学与 Cell Painting：从细胞形态学中发掘无偏见数据 为了从底层破解生物黑盒，Recursion 构建了一套“无偏见”（unbiased）的研发平台。平台的基石在于以空前规模在内部生成专为机器学习定制的标准化数据（fit-for-purpose data），输入超算集群训练模型，再以此指导新药研发，并通过递归循环（recursive cycle）持续自我进化。这一平台的起源可以追溯到大约十五年前团队对表型组学（phenomics）的突破性探索。 传统新药研发往往依赖于预设的单一靶点假说，而 Recursion 最初的立足点则是采用细胞组学成像（Cell Painting）技术。通过多通道荧光标记技术对活细胞进行多重染色，使细胞核呈现蓝色、细胞骨架呈现红色等，将细胞各亚结构的可视化动态尽收眼底。当对细胞施加化合物扰动或执行基因敲除时，细胞形态会发生极其精细的表型变化。这种基于无偏见表型观察的方法，能够在不预先假设靶点通路的前提下，系统性捕获药物处理与细胞形态变化之间的强相关性，从而发掘出科学文献中尚未记载的全新生物通路。 - **Cell Painting 多重荧光成像**：同时标记细胞核、细胞骨架及多种细胞器，捕获化合物或基因扰动带来的微观形态学变化。 - **无偏见发现逻辑**：打破传统单一靶点假设的局限，从全细胞整体响应反推生物学效应与作用机制。 --- ### [04:16](https://www.youtube.com/watch?v=6BqIMXr3TLg&t=256s) 规模化自编码模型与前沿湿实验基础设施 BioHive-2 为了使机器学习模型能够从海量表型图像中精准提取表征，Recursion 在实验通量上做到了极致。其位于犹他州盐湖城的高通量实验室日夜不停运转，以“每 15 分钟生成相当于一名博士毕业论文所需的数据量”（a PhD's worth of data every 15 minutes）的惊人速度，每周产出超过 200 万次自动化实验与高分辨率图像。 在模型算法层面，团队采用了掩码自编码器（Masked Autoencoder, MAE）架构对细胞图像进行表征学习。在将原始图像遮盖 75% 的极具挑战性的自监督预训练任务中，模型展现出了明确的缩放定律（Scaling Laws）：当训练数据翻倍、模型参数量从基线版本扩展至 3.37 亿参数时，图像重构与生物表征的预测质量获得了显著飞跃。为支撑这种超大规模数据的处理与模型训练，Recursion 联合 NVIDIA 打造了制药行业最为强大的超级计算机 **BioHive-2**，在全行业算力排行榜名列前茅（全球超级计算机前 70 强），为高通量数据挖掘提供了底层算力保障。 - **实验数据产出速度**：每周进行超过 200 万次高通量实验/成像，每 15 分钟产出相当于一名博士的数据量。 - **表型模型缩放表现**：在 75% 图像掩码测试下，更大的模型（3.37 亿参数）结合翻倍数据展现出远超 8600 万参数基线模型的表征重建精度。 - **全球研发与算力基地**：总部及核心湿实验室位于盐湖城，并在英国伦敦、牛津等地建有研发中心；与 NVIDIA 合作部署的 BioHive-2 构成了制药工业界第一算力底座。 --- ### [06:38](https://www.youtube.com/watch?v=6BqIMXr3TLg&t=398s) 多组学图谱整合与患者真实世界数据的闭环映射 在深耕高内涵荧光表型组学的基础上，Recursion 进一步将数据维度拓宽为多层次的多组学（Multi-omics）体系。除了固定染色的荧光图像外，团队还引入了明场表型组学（Brightfield Phenomics），无需荧光染料即可捕捉活细胞未受干预状态下的生物动力学行为与复杂应激反应。 与此同时，平台已完成超过 100 万份样本的转录组学（Transcriptomics）测序，并搭建了高通量 ADME 试验体系，以精确评估候选分子的药理学特性与毒理学风险。更关键的是，Recursion 正在将这些高维的微观细胞组学数据与来自医疗机构的患者真实世界数据（RWD）相连接，把表型组中探索到的关键基因与临床病患的具体疾病表型直接映射，构建起贯通微观机制与宏观临床的高维生物学图谱。 | 维度 / 组学类型 | 核心技术与数据规模 | 在平台中的主要价值与作用 | |:--- |:--- |:--- | | **明场表型组学** | 无标记活细胞成像 | 观测未染色活细胞的动力学形态变化与动态应激反应 | | **转录组学** | >100 万测序样本 | 测量基因表达谱变化，揭示转录调控与信号级联 | | **高通量 ADME** | 自动化药理与毒理筛选 | 评估化合物吸收、分布、代谢、排泄特征及早期毒性 | | **真实世界数据** | 临床患者队列与基因映射 | 将表型组学靶点直接关联至临床患者体验与疾病特征 | --- ### [08:23](https://www.youtube.com/watch?v=6BqIMXr3TLg&t=503s) 从表型自上而下到分子自下而上：开源结合亲和力大模型 Boltz-2 如果说表型组学代表了一种“自上而下”（top-down）、从宏观细胞表型反推机制的观察视角，那么分子设计则提供了“自下而上”（bottom-up）的生物分子结构设计能力。为了在分子层面精确解析药物与靶点的相互作用，Recursion 携手麻省理工学院（MIT）联合开发了全新的生物分子基础模型 **Boltz-2**。 与 AlphaFold 等主要聚焦于静态分子三维结构预测的模型不同，Boltz-2 的核心使命是直接预测**分子间的结合亲和力**（binding affinity），特别是小分子药物如何嵌入并结合至特定靶蛋白口袋。这一能力与表型组学形成了极强的协同互补：当表型实验发现未知活性的分子时，Boltz-2 能够通过快速模拟结合位点与相互作用构象，辅助阐明其分子作用机制（Mechanism of Action, MoA）。在计算性能上，Boltz-2 达到了传统计算化学中自由能微扰（FEP）这一“金标准”级别的预测精度，但计算速度提升了数个数量级且计算开销极低。这使得药物化学家不再需要等到先导化合物优化末期才进行极度昂贵的 FEP 计算，而是可以在研发全流程早期大规模调用。目前，Recursion 与 MIT 已将 Boltz-2 完全开源，向整个生物医药社区开放。 - **核心预测目标**：超越单纯的结构预测，专注攻克小分子与靶蛋白复合物的结合构象与结合亲和力模拟。 - **性能与效率突破**：在逼近自由能微扰（FEP）金标准精度的同时大幅压缩算力开销与模拟耗时，支持在前导期高通量筛选。 - **生态与社区贡献**：模型完全开源（Open Source），赋能全球药物设计人员加速机制解析与分子发现。 ### [11:02](https://www.youtube.com/watch?v=6BqIMXr3TLg&t=662s) 端到端药物研发全流程：从表型组学到 Exscientia 分子设计 Erik Douglas 深入阐释了 Recursion 如何将多种专有数据流与计算工具整合为一条完整的药物研发管道。研发流程由左至右推进：首先在早期利用表型组学（Phenomics）、转录组学（Transcriptomics）和表型解卷积（Deconvolution）提出初始假说并探索潜在的作用机制（Mechanism of Action, MoA），随后进入“设计工作流”（Design Workflow）。 在设计环节中，Recursion 于去年年底收购了总部位于英国的 Exscientia，将其全球领先的分子设计与化学设计工具引入平台。这使得公司能够将表型观测与先进的分子结构优化无缝结合。在设计工作流中，预测设计模型、原子级建模（Atomistic Models）以及自动化化学与生物学实验紧密协同。每一个在系统中推进的药物研发管线不仅能直接受益于底层 AI 模型的各项能力，同时其实验数据还会持续反哺底层模型，使后续的项目研发变得更加聪明与高效。当化合物通过候选物提名（Nomination）后，平台还会结合专用工具无缝衔接至后续的临床试验开发与运营中。 ```mermaid flowchart LR accTitle: Recursion 端到端药物发现与研发流程 subgraph Early["早期探索阶段"] A["表型组学 / 转录组学"] --> B["假说提出与机制解卷积"] end subgraph Design["设计与优化阶段 (Exscientia)"] B --> C["分子预测设计与原子建模"] C --> D["自动化化学与生物学实验"] D -- "数据反馈闭环" --> C end subgraph Clinic["临床阶段"] D --> E["候选物提名 (Nomination)"] E --> F["AI 驱动临床试验设计与执行"] end ``` - **全流程贯通**：从早期无偏见的表型组学发现，逐步过渡到原子级分子设计与临床验证。 - **战略整合**：通过整合 Exscientia 的分子设计能力，打通了“表型观测 + 精准化学优化”的双向闭环。 - **飞轮效应**：管线研发过程中产生的所有数据持续回流至模型，推动后续管线以更高智能化水平迭代。 --- ### [13:08](https://www.youtube.com/watch?v=6BqIMXr3TLg&t=788s) 探索生物学地形：T 型研发漏斗与肿瘤学未开发靶点 在寻找有效分子的“搜寻过程”（Hunting Process）中，传统药物研发往往采用相对随机或分散的试错方式。Recursion 则将生物学空间视作一个地形图（Biological Topography）：首先通过均匀分布的高通量扰动实验“照亮”特定区域周围的生物学景观，快速识别出最具潜力的“优质狩猎区”，随后集中算力与实验资源对这些高价值区域进行放大和深度优化。 基于这一理念，Erik 提出了“T 型漏斗”模型，并与传统研发中狭窄的“V 型漏斗”进行了鲜明对比。传统的 V 型漏斗通常始于科学家对某一特定靶点的局部洞见，一开始就将资源押注在极窄的路径上；而 Recursion 的 T 型漏斗在顶端保持极大的开放性，纳入海量假设并通过计算机模拟（*in silico*）与批量高通量实验进行筛选，从而避免在前期耗费昂贵且低效的传统实验成本。 以肿瘤学（Oncology）领域为例，演讲提到了当天上午 Geoffrey Hinton 教授在主题演讲中将肿瘤学视为 AI 极具前景的应用领域。Recursion 利用大语言模型（LLM）等工具，根据基因与肿瘤的相关性以及文献公开程度绘制了基因图谱。数据显示，除已被广泛熟知的常见癌基因外，图谱右上角分布着大量具有极高肿瘤相关性、但在公开文献中未被充分探索的“隐秘”靶点。凭借专有的生物学数据，Recursion 在这些未充分开发的独特靶点领域占据了显著优势。 ```mermaid flowchart TD accTitle: T型漏斗与传统V型漏斗对比 subgraph Traditional["传统 V 型研发漏斗"] direction TB T1["单一科学家假设 / 狭窄起点"] --> T2["昂贵的传统湿实验筛选"] T2 --> T3["逐步收窄管线 (高试错成本)"] end subgraph TechBio["Recursion T 型研发漏斗"] direction TB R1["海量假设输入 (无偏见顶端)"] --> R2["In Silico 模拟与高通量生物地形探索"] R2 --> R3["聚焦高价值未开发区域 (如右上象限新靶点)"] R3 --> R4["高效推进至临床阶段"] end ``` - **探索生物地形**：利用大范围无偏见扰动实验点亮未知生物图景，随后精准聚焦并优化。 - **突破文献盲区**：利用 LLM 与专有数据挖掘高肿瘤相关性但文献罕有记载的靶点，开辟全新研发生态位。 - **批量搜索模式**：将候选分子与靶点评估转变为大规模、并行的搜索和计算优化流程。 --- ### [15:58](https://www.youtube.com/watch?v=6BqIMXr3TLg&t=958s) 临床转化提速与 TechBio 驱动的科学范式变革 在推进项目进入临床阶段时，Recursion 将同样的计算严谨性与平台实力延伸至临床试验运营中。通过将因果人工智能（Causal AI）应用于人类基因组学，并结合真实世界数据（Real-World Evidence, RWE），平台实现了智能化的临床试验方案设计与患者招募优化。这有效解决了制药与生物科技领域长期面临的痛点——尤其是在罕见病患者招募困难且成本极高的场景下。 实际落地成果已经充分证明了该平台的提速实力： - **试验入组提速**：通过优化试验中心选择，患者入组预测速度提升了 50%。 - **临床前周期大幅缩短**：在某个具体项目中，从靶点识别（Target ID）推进到 IND（新药临床试验申请）赋能研究仅耗时约 18 个月，而行业平均耗时约为 42 个月。 在总结发言中，Erik 强调了“TechBio”（科技生物）与传统“Biotech”（生物科技）的核心区别：TechBio 是从“技术赋能的视角”出发，运用全套平台化工具大规模探索未知的生物学空间；而非传统 Biotech 那样依赖某位顶尖科学家提出单一最佳构想并在漫长路径中“碰运气”。加入公司约 5 个月的 Erik 表示，亲眼见证这种跨学科洞见与平台化工具的深度结合令人无比振奋。通过构建数据与模型相互增强的良性循环，TechBio 正在以更快的速度、更低的成本为患者带来同类首创（First-in-Class）与同类最优（Best-in-Class）的创新药物。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=6BqIMXr3TLg&t=4s) 解码生物学以从根本上改善人类生活（decoding biology to radically improve lives） — Erik Douglas > [00:04](https://www.youtube.com/watch?v=6BqIMXr3TLg&t=4s) 生物学绝不会轻易交出它的秘密 — Erik Douglas > [04:16](https://www.youtube.com/watch?v=6BqIMXr3TLg&t=256s) 每 15 分钟生成相当于一名博士毕业论文所需的数据量（a PhD's worth of data every 15 minutes） — Erik Douglas > [08:23](https://www.youtube.com/watch?v=6BqIMXr3TLg&t=503s) 自上而下（top-down）、从宏观细胞表型反推机制的观察视角 — Erik Douglas > [08:23](https://www.youtube.com/watch?v=6BqIMXr3TLg&t=503s) 自下而上（bottom-up）的生物分子结构设计能力 — Erik Douglas > [13:08](https://www.youtube.com/watch?v=6BqIMXr3TLg&t=788s) 照亮特定区域周围的生物学景观，快速识别出最具潜力的优质狩猎区 — Erik Douglas > [13:08](https://www.youtube.com/watch?v=6BqIMXr3TLg&t=788s) 传统的 V 型漏斗通常始于科学家对某一特定靶点的局部洞见，一开始就将资源押注在极窄的路径上；而 Recursion 的 T 型漏斗在顶端保持极大的开放性 — Erik Douglas > [15:58](https://www.youtube.com/watch?v=6BqIMXr3TLg&t=958s) 从靶点识别（Target ID）推进到 IND（新药临床试验申请）赋能研究仅耗时约 18 个月，而行业平均耗时约为 42 个月 — Erik Douglas > [15:58](https://www.youtube.com/watch?v=6BqIMXr3TLg&t=958s) TechBio 是从技术赋能的视角出发，运用全套平台化工具大规模探索未知的生物学空间；而非传统 Biotech 那样依赖某位顶尖科学家提出单一最佳构想并在漫长路径中碰运气 — Erik Douglas ## Mentioned Recursion, Erik Douglas, BioHive-2, NVIDIA, Exscientia, 麻省理工学院 (MIT), Boltz-2, Geoffrey Hinton, AlphaFold, Cell Painting, Masked Autoencoder (MAE), 因果人工智能 (Causal AI), 自由能微扰 (FEP) ## Video description Erik Douglas, VP of Product, Recursion Abstracts In order to use AI for drug discovery and development, you need a lot of high quality data. Recursion has been building these high quality biological datasets for over a decade – running millions of experiments per week using automation and robotics to capture images of cells in different states that are then used to train machine learning models. In all, Recursion has amassed 65 petabytes of proprietary data, including cell, chemical, animal and patient – all being processed by pharma’s most powerful supercomputer, BioHive-2, built with NVIDIA. In this talk, Erik Douglas, VP of Product, will share how Recursion is leveraging its multimodal, high quality data and the Recursion Operating System to transform traditional drug discovery, bringing significant gains in efficiency, speed and cost to deliver a pipeline of potential first-in-class and best-in-class medicines. [Source](https://www.youtube.com/watch?v=6BqIMXr3TLg)

[0:04] Erik Douglas: 解码生物学以从根本上改善人类生活（decoding biology to radically improve lives） [0:04] Erik Douglas: 生物学绝不会轻易交出它的秘密 [4:16] Erik Douglas: 每 15 分钟生成相当于一名博士毕业论文所需的数据量（a PhD's worth of data every 15 minutes） [8:23] Erik Douglas: 自上而下（top-down）、从宏观细胞表型反推机制的观察视角 [8:23] Erik Douglas: 自下而上（bottom-up）的生物分子结构设计能力 [13:08] Erik Douglas: 照亮特定区域周围的生物学景观，快速识别出最具潜力的优质狩猎区 [13:08] Erik Douglas: 传统的 V 型漏斗通常始于科学家对某一特定靶点的局部洞见，一开始就将资源押注在极窄的路径上；而 Recursion 的 T 型漏斗在顶端保持极大的开放性 [15:58] Erik Douglas: 从靶点识别（Target ID）推进到 IND（新药临床试验申请）赋能研究仅耗时约 18 个月，而行业平均耗时约为 42 个月 [15:58] Erik Douglas: TechBio 是从技术赋能的视角出发，运用全套平台化工具大规模探索未知的生物学空间；而非传统 Biotech 那样依赖某位顶尖科学家提出单一最佳构想并在漫长路径中碰运气

[0:04] Erik Douglas: decoding biology to radically improve lives [0:04] Erik Douglas: biology is not making it easy for us [0:04] Erik Douglas: Welcome our people [0:04] Erik Douglas: unbiased platform [3:05] Erik Douglas: a PhD's worth of data every 15 minutes [6:40] Erik Douglas: top-down [11:25] Erik Douglas: T-shaped [11:25] Erik Douglas: V-shaped [17:52] Erik Douglas: techbio [17:52] Erik Douglas: a very smart scientist think[ing] of her best idea and take[it] through the pathway and hop[ing] it works out [17:52] Erik Douglas: virtuous cycles

[0:04] Thank you very much. It's so nice to be [0:06] with you. I am Eric Douglas from [0:08] Recursion and uh Tech Bio is [0:10] transforming drug discovery. The uh [0:14] overarching mission for recursion is [0:18] decoding biology to radically improve [0:20] lives. And biology is not making it easy [0:24] for us. Biology is incredibly complex. [0:28] biology is, you know, this example [0:30] diagram here. We've got our our citric [0:32] acid cycle. You might remember from [0:34] biology class way back when. I guess I I [0:37] should maybe a quick pause. Who how many [0:39] folks in the audience are uh life [0:42] science or pharma or that that sort of [0:44] side of the world? Okay. Well, great. [0:47] Welcome our people. um well then these [0:51] diagrams will not be news to you but you [0:53] do know that biology is uh is is not not [0:57] uh sharing the secrets easily. So we we [1:01] understand a very small fraction of how [1:03] biology works right now and what on the [1:07] tech side of the world that everybody's [1:09] benefiting from Moore's law we all know [1:11] well but on the pharma and biotech side [1:14] drugs are getting slower and more [1:16] expensive to develop and that's a that's [1:19] a that's a pain that's a problem as [1:22] people who will one day use drugs may [1:24] currently use drugs we need better [1:25] solutions for so many so so many [1:28] healthcare needs [1:29] And what we're trying to do at Recursion [1:31] is basically put these pieces together [1:34] to take advantage of the increasing [1:37] wonderful capabilities of the the [1:40] various elements of our technology [1:41] platform to be able to bring better [1:43] drugs to market faster [1:46] and uh technology is working faster [1:48] here. you know, is the way that agents [1:51] are being able to string together faster [1:53] as we put the pieces of these elements. [1:55] It's not just point solutions, but [1:57] actually being able to go through the [1:59] whole process that we're going to need [2:00] to actually do this work. So, at at [2:04] Recursion, the way we've approached it [2:06] is we've built a platform. We've built a [2:08] platform that the goal is to combine a a [2:13] unprecedented scale of internally [2:16] developed uh fitforpurpose data, feed it [2:19] into a a massive supercomput uh and [2:22] build better models to be able to [2:24] develop better medicines and do that in [2:27] an iterative cycle. Do that in a [2:28] recursive fashion such that each time we [2:31] run it, the the out outputs get better. [2:34] And we we talk about it being an [2:35] unbiased platform. And one of the [2:38] reasons we we often use that language is [2:40] that some of the initial insights came [2:43] from phenomics from the idea of just [2:46] looking at how cells behave. So the the [2:49] origin story for recursion is uh you [2:52] know like like with many biotech and and [2:55] pharma companies there was some some [2:56] initial insight some point solution that [2:59] our our founders and our team were going [3:01] after and in this case it was uh the [3:05] idea of cell painting using using [3:08] multipplexed [3:10] um imaging for understanding dynamics [3:13] within a cell. understanding if we could [3:15] treat a cell with something, looking at [3:17] morphological changes on various [3:19] elements of that cell and being able to [3:21] correlate that with what we were doing [3:24] to the cell. So that that aspect that [3:27] that assay called cell painting was was [3:29] basically the initial start of how we [3:30] were approaching things and it was very [3:32] different than how the rest of drug [3:34] development was being done. This was uh [3:37] 15 years ago or so in the early early [3:39] teens. [3:41] And you know these these multiplex [3:43] fluorescent images uh beautiful images [3:45] we could you know you get the you get [3:47] the nucleus in blue you get some [3:49] cytokeleton in red you get these all the [3:51] different elements of what's going on in [3:53] the cell. And when when we would treat [3:55] the cells with various things or when we [3:57] would uh do various knockouts to the [4:00] cell, you could start to see changes. [4:02] These morphological changes or very [4:04] subtle changes within the uh cell images [4:08] start started to produce some really [4:10] interesting data. And we we started [4:14] taking these images at a very high [4:16] scale. So right now we we take about a a [4:19] PhD's worth of data every 15 minutes or [4:22] two over two million of these [4:24] experiments or images per week in our [4:26] high throughput labs. And so that allows [4:29] us to start building some really [4:31] interesting models to be able to predict [4:34] what'll happen when we do some [4:36] perturbation to the cells. So in this [4:39] case, this is a a masked autoenccoder [4:42] approach on uh one of these cell images. [4:45] And with the original image um 75% [4:48] masked, we had a a smaller model, a [4:52] quarter of the images trained, 86 [4:53] million parameters, we see that as we [4:55] have more and more data going into [4:57] training these models, we get [4:59] conccomantly better outputs. [5:01] Unsurprisingly, but it's a real argument [5:04] for having a whole lot of data that [5:06] you're training these systems on. So in [5:08] this case it was twice the data 3x [5:11] bigger model and 30 37 million [5:13] parameters. And you can maybe tell in [5:16] this in in this lighting but uh you know [5:18] there's some papers online as well that [5:20] go into more detail on the specifics of [5:23] that u those those models for the the [5:27] automated cell imaging. But it's just [5:29] really exciting with that that scale of [5:31] data to be able to have those [5:33] predictions and be able to understand [5:34] what the cell's doing. [5:36] And so we've [5:38] we've built this [5:41] oops we've built this [5:44] uh big lab in Salt Lake City. So we've [5:47] got companies based in Salt Lake City [5:48] with other another large site in London [5:51] and uh outside of Oxford and UK and and [5:54] a couple other offices. But really these [5:57] wet lab spaces that we have are doing [5:59] this high throughput chemistry and [6:01] biology to continue to feed the [6:03] algorithm development. And we're we're [6:05] taking these these systems are running [6:07] day and night collecting this massive [6:08] data and fe feeding it into our our [6:11] supercomputer. So we have we've [6:13] developed in partnership with Nvidia a [6:16] computer we call biohive 2 that is the [6:19] largest supercomputer in the pharma [6:21] industry and uh one of the I think in [6:24] last count top 70 in the world uh [6:27] supercomputer. So really exciting [6:29] compute power being applied to these [6:32] data problems or opportunities to [6:34] understand how to start to build these [6:36] models. And I I talked mostly about the [6:39] phenomics so far, but building on that [6:41] original phenomics insight, we've also [6:44] been [6:46] uh continuing to expand a variety of [6:49] different OMIX applications. So in this [6:51] case we've got our uh brightfield [6:54] phenomics on the left where we're able [6:57] to look at dynamics on of cell changes. [7:00] You know the cells in this case aren't [7:01] fixed. They don't have the fluorescent [7:03] dyes. So it gives you a different window [7:06] into the cell activity with some some [7:09] additional sort of complexity on the [7:11] cell behavior and response. Um we're [7:13] doing transcrytoics. We've got over a [7:16] million samples sequenced there. doing [7:19] uh high throughput admi assays to [7:21] understand the pharmarmacology and and [7:23] toxicology [7:24] and also real world data. So [7:26] partnerships with uh companies like [7:30] those shown to be able to understand how [7:33] we can connect back to the actual [7:34] clinical experience of these patients [7:36] and as we map to the the various genes [7:39] that we're exploring on the phenomic [7:41] side being able to connect that to [7:43] patient data and put those pieces [7:45] together in a really exciting way. [7:50] And so this is an an illustration. I [7:53] think this concept of integrating those [7:56] layers to build maps is really what [7:58] we're driving after. It's not just one [8:01] or the other genomics, transcrytoics, [8:03] etc. But ultimately trying to build [8:05] these highdimensional multi- uh sort of [8:09] data type models that give us these [8:11] really interesting insights into how the [8:14] cell's going to respond, how the cell is [8:16] is uh interacting with with these these [8:20] perturbations that we're providing. [8:23] And so most of what I've talked about so [8:25] far has been through the lens of the [8:27] observations, the biology basically top [8:30] down understanding how the cell [8:32] responds, understanding the phenomics, [8:33] taking images, looking at um maybe [8:37] unexpected things that happen. I the [8:40] original insight with phenomics was that [8:43] the concepts of uh what what the target [8:46] might be wasn't always necessarily what [8:49] happened. And so being able to to treat [8:51] the cell with a whole lot of things to [8:53] have a high very high throughput uh [8:56] approach to being able to treat the [8:57] cells and understand what they were [8:59] doing could give us new insights could [9:00] help identify new pathways that perhaps [9:03] weren't already identified in the [9:05] literature. And then another way into [9:07] the some of these questions is more uh [9:11] from a biomolelecular design [9:13] perspective. So some really interesting [9:15] work that we've uh shared lately um that [9:19] we've co-developed with some folks at [9:20] MIT is a new model we call Boltz 2. And [9:24] Boltz 2 is a a biomolelecular foundation [9:28] model for binding affinity. So you know [9:31] alpha fold and the like great models for [9:33] looking at molecular structure but in [9:36] this case we really also want to look at [9:38] how how the molecules are going to fit [9:40] together. we need to not just know [9:42] structure but to predict binding to be [9:44] able to predict you know for for a given [9:46] target uh what where is our especially [9:49] small molecule drug going to fit in [9:51] there and in in our case especially if [9:54] you're starting out with phenomic [9:56] observations being able to use models [9:58] like this to help elucidate what might [10:01] be the mechanism of action that can [10:05] simulate where the binding might might [10:06] happen and simulate what that that [10:08] interaction actually looks like gives us [10:10] a really powerful complimentary tool to [10:12] some of the phenomics work I've [10:14] described. And so Bolt Bolts 2 is a [10:16] really exciting model because it [10:18] operates at about the same level as the [10:21] the gold standard uh free energy [10:23] perturbation style uh simulations that [10:27] uh medical chemists might be used to [10:29] doing but much faster and much more um [10:33] uh computationally inexpensively. So [10:36] it's something that you don't have to [10:37] wait to run at the end when you've got [10:39] just a couple of molecules left. It's [10:41] something you can include earlier in the [10:43] process and use across this whole [10:45] pathway of uh understanding the the um [10:48] pipeline. [10:50] And I'll also say you know this this [10:52] model so we co-developed it with MIT and [10:54] it's now open source. So this is [10:55] something that you know we and they are [10:58] are are sharing with the community to [10:59] take things like this and make them [11:02] broadly available. And I think it also [11:04] it it emphasizes, you know, I think [11:06] helping helping a broad a broad variety [11:08] of companies have access to these kind [11:09] of tools is great. And also having the [11:13] proprietary sort of data streams that [11:14] companies like recursion and and some [11:16] others, you know, are developing is an [11:19] an important additional advantage that [11:21] um I think we're we're especially [11:22] excited about. [11:25] And so this is this is an illustration [11:27] of how some of these pieces fit [11:28] together. So we've got I've talked about [11:31] the early phenomics type work. Basically [11:34] from from left to right we're going [11:37] through the funnel. We're going through [11:38] the the drug development pathway. Uh [11:41] initially starting in this case starting [11:43] with with phenomics and transcrytoics [11:45] and deconvolution and taking an initial [11:49] hypothesis and thinking about um what [11:53] what might be the mechanism of action, [11:54] how we can start to optimize for that. [11:57] but then ultimately moving it into what [11:59] we call our design workflow that takes [12:02] the chemical design tools. And so, you [12:04] know, late last year, we uh combined [12:06] with a company called Exencia that's [12:08] based in the UK who's got some [12:11] worldleading tools for some of that [12:13] molecular design and bringing those [12:16] pieces together to be able to do the [12:17] phenomic observations and the molecular [12:19] design to optimize the chemistry for [12:22] these these molecules is a really [12:24] exciting place to be. And so in the [12:27] design workflow, we've got our [12:28] prediction design, we've got atomistic [12:30] models, automated chemistry and biology. [12:32] And in either case, as things are going [12:35] through these cycles each each time a [12:38] program is going through our system, [12:40] it's continuing it it it benefits from [12:42] all these pieces. You know, it's [12:43] touching all these models and it's, you [12:45] know, benefiting from the the [12:46] characteristics and capabilities of all [12:48] the models, but it's also feeding back [12:50] into the system. So the next program is [12:53] going to be that much smarter. [12:55] And then ultimately as we go through [12:57] nomination and design into final um [13:00] clinical uh clinical trials using [13:03] additional tools for the um clinical [13:05] trial development and operation. [13:08] So just to illustrate a little bit more, [13:10] you know, this this hunting process that [13:12] we're going through and whereas some [13:15] historical hunting has been sort of a a [13:18] a random approach or broadly distributed [13:20] across a space, in our case, if we're [13:23] trying to visualize this landscape of [13:25] the this say biological topography that [13:28] we're operating within, you know, if you [13:30] if you start by having a bunch of evenly [13:32] distributed experiments, evenly [13:34] distributed perturbations, [13:36] you now can illuminate the area around [13:38] those those experiments and now you know [13:40] where's better hunting where where [13:42] you're going to want to try to tune more [13:44] for and so then we can zoom in on those [13:47] particular spaces be able to much more [13:49] efficiently try to optimize these these [13:52] explorations. [13:54] So we talk about a T-shaped funnel and [13:57] we contrast that with the the V that you [13:59] know is typical of historical approaches [14:02] where you've got a particular insight [14:04] and you're trying to really optimize for [14:06] that insight. In our case, it is it's [14:08] much more we we bring a lot into the top [14:11] of the funnel. We try to be very open. [14:14] You know, I talked about the unbiased [14:15] platform. We start with a lot of [14:16] hypotheses and we do a lot of this [14:18] optimization in silic [14:21] have a lot of expensive experiments up [14:24] at the top of the funnel, but then as we [14:27] get closer move move things through the [14:30] process, we're able to uh do it more [14:32] efficiently by having the right [14:34] experiments at the right time. [14:37] So for for example in the oncology space [14:40] where we're doing a lot of work and uh [14:42] for folks who were at the keynote uh [14:44] with professor hinton this morning you [14:45] know it's very exciting that the [14:47] oncology space being the great hope for [14:49] AI applications uh we spent a lot of [14:53] time in oncology and and in this case [14:55] the the green dots represent um using [14:58] LLM to map various genes that that we've [15:01] looked at according to their uh oncology [15:04] relevance and than their basically [15:07] knownness in the literature. And in our [15:10] case, wanting to so you know we are [15:13] reassured to see that many common [15:16] oncology genes are represented in the uh [15:18] highcology relevant side of things, but [15:20] in that top right quadrant are a whole [15:22] bunch more genes that we've got great [15:25] data on that aren't well explored in the [15:28] public literature. And so that's a [15:30] that's a space where we're really [15:31] looking to play is finding those things [15:33] that we think we've got a unique [15:34] advantage in trying to bring bring to [15:37] bear. [15:39] And uh yeah, as as those things go into [15:42] the top of the funnel now, we're able to [15:43] more efficiently move them through [15:45] evaluating programs in bulk, making it [15:47] more of a search process, taking a whole [15:49] lot of these programs and being able to [15:51] efficiently operate them through that uh [15:54] evaluation and optimization. [15:58] And then as we get the get the uh [16:02] programs ready for for the clinic um we [16:04] bring the same sort of rigor and uh [16:07] platform prowess to thinking about how [16:10] we operate the clinical trials. So we're [16:12] we're using our causal AI for human [16:15] genomics. We're doing intelligent [16:16] clinical trial design and uh in [16:20] recruitment and the actual execution of [16:22] the trials. You know, obviously that's a [16:24] that's a real pain point for pharma and [16:26] biotech. Understanding how to actually [16:28] get the right medicine to the right [16:30] patient at the right time. In the case [16:32] of trials, especially where if it's a [16:34] rare condition, recruitment can be [16:36] incredibly hard and expensive. And for [16:38] us, working with both real world [16:40] evidence and uh all of our various [16:42] tools, we can move that much more [16:45] quickly through the through the [16:46] pipeline. So for example, 50% faster [16:49] enrollment projections by optimizing [16:51] sites. [16:53] And in one of our programs, we were able [16:55] to move from our target ID uh to the IND [16:59] enabling study in about 18 months, [17:01] comparing that with an industry average [17:03] of about 42. So we can see that it's [17:05] it's already working. You know, this [17:07] application, this idea of bringing these [17:09] tools to be able to more screen a whole [17:11] lot more things, look at a lot more [17:12] hypotheses, but then actually work that [17:14] through the whole process. We're we're [17:16] showing that we can do that more [17:17] quickly, not just in that example, but [17:20] in a a number of pieces of internal data [17:22] where we can optimize. We're able to [17:24] move quickly validate the hypothesis, [17:27] uh do more highly productive compound [17:30] design, spend less money, and go faster. [17:33] Really, all the all the things you're [17:34] looking for in your drug development [17:36] program. And it's by virtue of being [17:38] able to use these tools to optimize that [17:40] that whole process. [17:43] And so just as uh sort of as I as I come [17:46] to the end here um I I talked a little [17:49] bit about at the beginning how tech bio [17:52] as we call it is is revolutionizing the [17:55] approach to developing new drugs and we [17:58] we call it tech bio rather than biotech [18:00] because it's really from a from a tech- [18:01] enabled lens. You know we've got this [18:03] great suite of tools and we're using it [18:05] to explore the biological space rather [18:07] than starting with with one particular [18:09] biological niche we're operating in. And [18:12] so just kind of the the contrast of uh [18:16] having you know on on the traditional [18:18] side having a very smart scientist think [18:20] of her best idea and and take it through [18:22] the pathway and and hope it works out. [18:24] What we're really trying to do is build [18:26] those virtuous cycles, use the [18:28] technology, use the platform such that [18:31] we're we can do it more quickly and we [18:33] can hopefully get get better drugs to [18:35] market faster. So it's a it's an [18:37] exciting time in the space. Um, you [18:40] know, I I'm uh I I I just joined the [18:43] company about uh five months ago, and I [18:45] think it's just been thrilling to see [18:47] the the array of tools we have at our [18:49] disposal and and how we're bringing them [18:51] together in a lot of ways that are these [18:54] intersection of different types of [18:55] insights. Um, so yeah, I I'll be around [18:58] afterwards. I'd love to answer [18:59] questions, but uh thanks for your