Boston Children's Uses OpenAI o3 to Uncover Rare Disease Diagnoses

波士顿儿童医院利用 OpenAI o3 助力罕见病诊断 深度推理 AI 模型通过检索被忽视的医学文献以识别被遗忘的基因突变，从而缩短长达十年的确诊漫长历程。 专家组讨论了一项合作研究，该研究将 OpenAI o3 Deep Research 应用于 376 例未确诊的罕见病病例，通过综合科学文献并分析复杂的基因变异，成功获得了 18 例新诊断。一位患者倡导者分享了他长达十年的求医确诊历程，阐明了获得明确诊断如何为治疗方案、家庭生育规划以及病友互助带来根本改变。研究人员解释了 AI 如何通过筛选海量全基因组测序数据集并深入检索历史文献，成为人类遗传学家的分析倍增器。在 OpenAI 基金会资助的支持下，该团队旨在通过自动化、定期的重新分析工具，将这一工作流程推广普及到顶尖医疗中心之外。 波士顿儿童医院曼顿罕见病研究中心、哈佛大学与 OpenAI 展开合作，利用 AI 工作流重新分析了 376 例罕见病疑难病例，成功得出 18 个全新确诊诊断。 (0:09) 目前基因诊断的核心瓶颈已从测序技术转向对数百万基因突变的手工解读，而 OpenAI o3 Deep Research 作为“人机协同”工具被引入，自动化合成文献并生成有据可循的诊断假说。 (5:48) 该 AI 分析流程在经过提示词迭代以过滤测序假象后，在已知确诊的验证病例中达到了 80% 至 90% 的诊断准确率。 (21:42) 在一次重大突破（曼顿第 151 号病例）中，AI 深入挖掘了人类分析师常会错过的 30 页深层历史文献，成功锁定了关键致病候选基因 S1PR1。 (23:24) 在基准验证测试中，该模型还成功识别出一位患者同时患有的两种并发遗传疾病，而人类分析师往往在找到第一种致病突变后便停止了排查。 (35:23)

Researchers from Boston Children's Hospital's Manton Center and OpenAI demonstrated that pairing reasoning models with expert clinical validation can address the analytical bottleneck in rare disease genetics, establishing 18 new diagnoses across 376 unsolved cold cases. By automating deep literature synthesis and phenotype-to-variant mapping, the workflow surfaced overlooked historical findings and untangled complex dual-gene conditions. The Analytical Bottleneck in Rare Disease Genetics (5:48) Hundreds of millions of families globally face rare diseases, enduring an average diagnostic odyssey of six to seven years. Next-Generation Sequencing has significantly lowered testing costs—often below $1,000 for Whole Genome Sequencing, compared to standard MRIs—yet it has shifted the medical bottleneck from data acquisition to data interpretation. A single patient's genome contains 3 billion base pairs and generates hundreds of thousands of variants, which initial filtering narrows to 1,000 to 10,000 rare candidates across 500 to 1,000 genes. Manual clinical evaluation requires cross-referencing these variants against 8,000 to 9,000 known disease genes, tracing inheritance models in family trios, and scouring medical literature for newly published discoveries. Because no human geneticist can master the entire catalog of disease literature, re-evaluating cold cases creates an unmanageable backlog for clinical teams, often leaving patients in diagnostic limbo for nearly a decade. AI Literature Synthesis and Multigene Diagnosis (23:24) To re-evaluate 376 stalled pediatric cases, researchers integrated OpenAI o3 Deep Research into a structured workflow that fed filtered variant lists and standardized Human Phenotype Ontology terms into the model. After prompt refinements trained the system to discount common sequencing artifacts like low read depth, the model achieved an 80% to 90% diagnostic accuracy rate on known control cases before being tasked with proposing candidate genes for human expert review. Beyond matching recently published disease genes, the model synthesized disparate historical literature to uncover novel associations. In Manton case 151, the AI connected a patient's vitiligo, pulmonary hypertension, and heart defect to the S1PR1 gene by surfacing a 26-year-old paper located 30 pages deep in search results. Furthermore, during benchmark testing, the model successfully identified a patient suffering from two distinct, co-occurring genetic conditions—a dual diagnosis previously missed by human reviewers who had stopped searching after isolating a single pathogenic variant. Clinical Care Shifts from Symptomatic to Targeted (10:44) Securing a molecular diagnosis shifts patient care from reactive, supportive measures—such as corrective surgeries for secondary skeletal complications—to targeted interventions. For the patient advocate Stav, reaching a precise diagnosis after nearly ten years enabled direct contact with biotechnology companies advancing candidate therapies specifically targeting his neuromuscular condition. Beyond experimental drug development, diagnostic clarity provides immediate psychological closure and actionable behavioral guidance, helping patients avoid physical activities that could exacerbate tissue damage. The clinicians also noted that molecular findings critically inform family planning through reproductive screening and frequently resolve multi-generational medical mysteries across extended families, such as revealing inherited cancer susceptibilities or explaining symptoms across multiple generations. Automated Reanalysis and Direct Data Integration (27:04) Because a patient's genetic sequence is static while biological literature expands continuously, automated AI tools make periodic, low-cost reanalysis of negative cases clinically viable. Through support from an OpenAI Foundation grant, the Manton Center is developing a secure upload platform to extend these computational workflows to external healthcare providers, complementing Manton’s model that allows patient families to self-refer directly. Researchers emphasized that the current study relied on significant upfront curation, filtering raw sequence data down to a candidate pool of 1,000 to 10,000 variants paired with discrete symptom codes. The future objective is end-to-end automated interpretation, where emerging domain-specific models like GPT Rosalind can ingest raw, uncurated electronic medical records, imaging, laboratory results, and whole-genome files directly. While AI-assisted literature synthesis successfully identified elusive diagnoses in complex cold cases, the current approach remains reliant on intensive human pre-filtering of genomic variants and final expert validation. Evaluating whether future models can reliably analyze raw, uncurated health records and full genomic files without generating misleading clinical leads remains an open empirical challenge.

波士顿儿童医院曼顿罕见病研究中心与OpenAI合作，利用o3 Deep Research等AI推理模型重分析了376例未确诊罕见病病例并发现18个新诊断，证明AI驱动的文献综合与表型关联能有效破解基因诊断的临床瓶颈。 罕见病诊断困境与基因组数据瓶颈 (0:09) 全球有数以亿计的家庭受罕见病困扰，患者获得确诊的平均周期长达6至7年。波士顿儿童医院的Alan Beggs博士指出，人类基因组包含约30亿个碱基对和2万个基因，目前已知有8000至9000个基因突变可致病。新一代测序技术虽能快速完成全基因组或外显子测序，但也会产生数以百万计的变异位点。临床瓶颈已从“数据生成”彻底转向“数据解读”，即如何从庞大的良性变异背景中筛选出唯一的致病突变。 患者代表Stav的经历印证了这一漫长过程：他青少年时期出现无法解释的身体疼痛与活动受限，医疗团队花费3至4年才确认其为神经系统而非肌肉骨骼病变，随后又历经5年基因检测才锁定病因，前后历时近十年。确诊后，他不仅获得了避免加重病情的科学生活指导、进入同质病友社群，还成功对接针对其突变的靶向药物研发管线，并能在现代生殖技术辅助下完成优生规划。 人工重分析的局限与AI推理嵌入 (7:51) 曼顿基因发现核心实验室科学总监Katherine Brownstein（后文亦提及其合著者Katherine Helming）详述了人工分析的巨大负担。遗传学家需要重评已知致病基因、分析父母与患儿的“三人组”数据以寻找新生突变或隐性遗传模式、排查染色体结构变异，并在人群数据库中过滤常见变异。研究人员往往需要耗费数小时比对最新医学文献以验证基因与表型的关联，且频繁遭遇死胡同。 针对在曼顿中心沉淀10至15年、人力难以逐一复核的积压“冷案”，OpenAI研究员Suyash Shringarpure等人设计了人机协同流程。全基因组测序的约50万行变异数据经基础过滤后仍保留数千至1万个独特变异，AI模型结合标准化的人类表型本体术语（HPO）与海量文献，自动进行变异假设生成与证据引用，将候选范围收窄至2到6个高置信度位点交由专家最终裁决。经工程迭代纠正测序深度不足等技术假阳性后，该流程在已知确诊的基准验证中达到了80%至90%的准确率。 深层文献挖掘与复杂多基因识别 (23:24) 在376例积压病例测试中，AI不仅找出了案例归档后才被医学界发表的已知致病基因，还展现了跨文献的深度推理能力。在曼顿第151号病例中，患者兼具白癜风、大动脉转位及肺动脉高压，AI出人意料地锁定了S1PR1候选基因。该推论直接溯源至约26年前（1998–2000年间）探讨S1PR1与白癜风病理关联的论文，其检索深度达到常规研究人员难以查阅的数十页之后，促成了研究团队与最初克隆该基因的学者就近展开功能性合作验证。 此外，在20例对照验证测试中，AI在19例中与专家共识一致；而在最初判定AI“失误”的1例中，复核发现该患者实际上同时携带两种独立且并存的遗传病，而非单一表型扩展。人类分析师在找到首个致病突变后往往终止分析，而AI无偏差地评估了全谱数据，展现了其理清复杂双重诊断的潜力。 早期全基因组测序与普惠临床推广 (27:04) Alan Beggs博士强烈呼吁早期应用全基因组测序（WGS），指出目前全基因组测序成本已降至1000美元以内，显著低于传统MRI检查，在新生儿重症监护中可在数天至数周内出具结果，世界纪录更是压缩在24小时内。Beggs批评了商业保险对早期测序的报销迟滞，强调数字化的基因组数据是一次性资产，能随AI算法和医学知识的更新进行低成本的终身定期重分析。 为避免先进技术仅局限于顶尖三级医疗中心，曼顿中心在OpenAI基金会的资助下正在开发安全、易用的外部上传平台，支持外部临床医生与研究人员提交数据。该中心同时支持患者家庭通过电话或邮件直接自荐，免去繁琐的转诊门槛，力求将一线临床遗传学家从繁复的基础比对中解放出来，集中精力攻关其余95%极其复杂且尚未解决的医学谜团。 本项研究证明了推理模型在经过结构化过滤的基因与表型数据上展现出强大的文献合成价值。研究团队指出，目前流程仍依赖人工对基因组与病历进行前期清洗筛选，未来的长期挑战与探索方向在于跳过人工过滤，直接端到端输入未经加工的完整电子病历、生化检验、医学影像与全基因组大文件，依靠更专精的生命科学模型（如GPT Rosalind）完成全自动化分析。

## Background 本内容由 OpenAI 于 2026 年 8 月 6 日在 YouTube 官方频道发布。视频记录了 OpenAI 论坛的一场深度研讨会，重点聚焦波士顿儿童医院曼顿孤儿疾病研究中心（Manton Center for Orphan Disease Research）如何借助 OpenAI o3 Deep Research 重新分析疑难儿科病例并挖掘关键致病线索，展现了在确保临床医生居于决策核心的前提下，AI 加速罕见病诊断与前沿科研的巨大潜力。 ## TL;DR 研讨会围绕罕见病患者长达数年的“诊断奥德赛”困境展开，介绍了波士顿儿童医院与 OpenAI 联合开展的一项先导研究。面对 376 例积压多年的疑难悬案，研究团队利用具备长文献检索与假设推理能力的 OpenAI o3 Deep Research 模型，成功在 18 例病例中挖掘出全新或意料之外的致病证据（包括重新关联沉睡 26 年文献的 S1PR1 基因及拆解并存的双重致病突变）。讨论深入剖析了从 30 亿碱基对海量数据中“大海捞针”的技术挑战、人机协同验证机制、降低罕见病分析门槛的普惠愿景，以及未来构建自主常驻后台 Agent 持续重分析病例的临床蓝图。 ## Atmosphere & style 整场对谈兼具严谨的学术硬核感与深沉的人文关怀，开篇由亲历误诊漫长求医的患者现身说法，瞬间奠定了真实而略带沉重的情感基调。随着临床遗传学家与 AI 研究员展开技术对谈，现场氛围转向科研攻坚的理性与振奋，尤其在讲述“翻阅 26 年前冷门论文并偶然发现克隆该基因的先驱就在隔壁大楼”时，充满了科学发现特有的戏剧张力与顿悟时刻。主持人的穿针引线与突发手机铃声的幽默调侃，又为严肃沉重的医学话题增添了温暖松弛的底色，让观众真切感受到技术突破正在为绝望中的罕见病家庭点亮切实希望。 ## Bulletin - [00:09](https://www.youtube.com/watch?v=SJJBPwlPpww&t=9s) 波士顿儿童医院曼顿中心与 OpenAI 合作，利用 OpenAI o3 Deep Research 重析 376 例积压多年的疑难罕见病病例，成功挖掘关键证据并直接促成 18 例新确诊。 - [16:08](https://www.youtube.com/watch?v=SJJBPwlPpww&t=968s) 系统通过拒绝黑盒推理、要求模型输出附带同行评审文献与分子机制的证据链，并将每例患者成千上万个候选突变精准收敛至 2 到 6 个高疑似位点供人类专家复核。 - [22:12](https://www.youtube.com/watch?v=SJJBPwlPpww&t=1332s) 在基准测试中，通过针对测序深度等生信陷阱持续优化 Prompt，模型在已知病例上的准确率提升至 80%–90%，甚至成功纠正了一例人类专家误判为单一突变拓展的双重遗传变异病例。 - [33:25](https://www.youtube.com/watch?v=SJJBPwlPpww&t=2005s) 尽管全基因组测序成本已降至 1,000 美元以下甚至数百美元，但商业保险报销阻力依然是限制早期普及的核心瓶颈；OpenAI 基金会正资助开发通用低门槛工具，推动基层医疗定期自动重分析。 ## Key takeaways - [00:09](https://www.youtube.com/watch?v=SJJBPwlPpww&t=9s) 罕见病影响全球数亿家庭，患者往往面临长达 6 至 7 年的“诊断奥德赛”，不仅身心饱受折磨，传统的支持性疗法也无法直击病因。 - [04:01](https://www.youtube.com/watch?v=SJJBPwlPpww&t=241s) 确诊不仅能带来心理释然与病友社群归属，更是连接靶向新药临床试验、实现精准个性化治疗以及科学规划下一代优生优育的先决条件。 - [05:32](https://www.youtube.com/watch?v=SJJBPwlPpww&t=332s) 全基因组测序普及后带来的最大痛点是“大海捞针”：全基因组包含 30 亿个碱基对，常规过滤后仍残留数千个罕见突变，远超人类遗传学家对 8,000 多种已知致病基因的认知极限。 - [07:51](https://www.youtube.com/watch?v=SJJBPwlPpww&t=471s) 传统人工分析需耗费大量时间逐个核查人群数据库并精读基础科研文献，往往推演数小时后才发现假说错误，形成了繁重的智力消耗与临床时间瓶颈。 - [16:08](https://www.youtube.com/watch?v=SJJBPwlPpww&t=968s) 人机协同工作流的核心是“可解释的证据链”而非黑盒判断：AI 输出候选基因短名单及关联文献，最终由人类遗传学家把关决定是否启动湿实验验证和出具诊断。 - [23:24](https://www.youtube.com/watch?v=SJJBPwlPpww&t=1404s) AI 具备深度跨学科文献检索与跨周期推理能力，如在 S1PR1 病例中自动挖掘出 26 年前的冷门基础研究，并严密构建出白癜风、大动脉转位与肺动脉高压的分子通路假说。 - [27:02](https://www.youtube.com/watch?v=SJJBPwlPpww&t=1622s) “人类基因组数据恒定不变，而医学知识每天都在更新”，低成本 AI Agent 在后台常驻并定期对沉睡的历史未解基因组数据进行自动化重分析具有巨大临床潜力。 - [29:26](https://www.youtube.com/watch?v=SJJBPwlPpww&t=1766s) AI 与临床专家的最佳分工在于消化中低复杂度疑难病例，从而解放人类顶尖专家的认知带宽，去集中攻坚剩下的 95% 极端疑难未解病例。 - [33:25](https://www.youtube.com/watch?v=SJJBPwlPpww&t=2005s) 全基因组测序（WGS）成本已大幅低于常规 MRI 检查，且已在 NICU 等重症场景实现数天甚至 24 小时内的极限确诊，阻碍其早期广泛筛查的主要障碍在于保险报销体系滞后。 - [35:23](https://www.youtube.com/watch?v=SJJBPwlPpww&t=2123s) 面对多系统复杂表型，AI 展现出卓越的去偏见解耦能力，能够有效识别并分离并存的双重致病基因变异，克服人类专家容易陷入的“单一致病表型拓展”思维定式。 - [36:47](https://www.youtube.com/watch?v=SJJBPwlPpww&t=2207s) 医学 AI 的终极演进方向是从当前依赖人工清洗降维的提示词阶段，跨越至能够直接消化未经清洗的原始电子病历（EHR）与全基因组文件的零人工干预端到端深度推理。 - [38:03](https://www.youtube.com/watch?v=SJJBPwlPpww&t=2283s) OpenAI 基金会已向波士顿儿童医院提供专项科研资助，并正在开发专为深层生命科学推理与文献合成定制的下一代模型（如 GPT Rosalind），未来将重点攻关罕见病与罕见癌症。 ## Detailed walkthrough ### [00:09](https://www.youtube.com/watch?v=SJJBPwlPpww&t=9s) 合作背景与罕见病诊断困境 这场 OpenAI Forum 研讨会聚焦于波士顿儿童医院（Boston Children's Hospital）与哈佛大学旗下的曼顿孤儿疾病研究中心（Manton Center for Orphan Disease Research），同 OpenAI 之间开展的一项突破性研究合作。在全球范围内，有数以亿计的家庭正与罕见病进行着无声的抗争。对这些家庭而言，获取确切诊断的过程异常艰辛，平均耗时往往长达六至七年之久，这一过程在医学界常被称为漫长而痛苦的“诊断奥德赛”（Diagnostic Odyssey）。 在此次联合研究中，研究团队针对 376 例疑难病例展开分析，借助 AI 驱动的研究工作流成功筛选出关键证据链，最终促成了 18 例罕见病的确诊。本场对谈邀请了该研究的多位核心参与者：首先是曼顿中心的研究参与者 Stav Rones，他将亲身讲述自己寻求确诊的心路历程；随后，三位论文共同作者登台探讨技术与临床细节——曼顿中心基因发现核心实验室（Gene Discovery Core）科学总监 Catherine Brownstein 博士（专注于基因组分析与诊断壁垒攻坚）、在人类遗传学与罕见病领域深耕数十年的曼顿中心主任 Alan Beggs 博士，以及 OpenAI 机器学习研究员兼统计遗传学家 Suyash Shringarpure（致力于跨越 AI、基因组学与计算生物学边界的研究）。他们将共同剖析罕见病诊断的顽疾所在、AI 系统的运作机制、研究发现、当前局限，并探讨如何让人类遗传学家和临床医师始终居于诊疗的核心决策地位。 - **全球罕见病现状**：波及全球数亿家庭，平均确诊周期长达 6 至 7 年。 - **核心研究成果**：在涵盖 376 个疑难未解病例的队列中，AI 驱动的工作流挖掘出关键证据，最终直接促成 18 例罕见病的新确诊。 ### [01:43](https://www.youtube.com/watch?v=SJJBPwlPpww&t=103s) 患者故事：Stav Rones 漫长确诊的体育少年 对谈首先将话筒交给了患者 Stav Rones。Stav 在波士顿地区出生并长大，从小就非常热爱体育运动，广泛涉猎棒球、篮球和赛艇等多项项目。然而，当他在十几岁步入更高强度的竞技体育阶段时，身体开始出现一些异样——他经历了同龄孩子从未遭遇过的持续疼痛以及身体活动受限。这种无法查明原因的身体失控让他陷入了极度的沮丧与迷茫。 ```mermaid flowchart TD accTitle: 罕见病诊断的漫长探索历程 accDescr: 展示患者从青少年时期发病，经由运动损伤误诊、神经系统病因排查、遗传学推断，最终通过全基因组测序获得明确确诊的漫长周期。 subgraph S1["初期症状与误区"] A["青少年时期发病<br/>剧烈疼痛与活动受限"] --> B["初诊阶段<br/>误诊为常规物理或骨骼肌肉损伤"] end subgraph S2["系统性病因定位"] B --> C["神经系统归因<br/>耗时 3 至 4 年厘清神经病理表现"] C --> D["遗传病因确认<br/>耗时 5 年以上推断存在遗传突变基础"] end subgraph S3["分子水平终极确诊"] D --> E["全基因组测序 NGS<br/>在曼顿中心协助下最终确诊罕见病"] end ``` 为了寻找答案，Stav 开始四处求医，经历了一次又一次的专业检查，甚至包括最初被证明完全错误的初步诊断。他解释说，如果是常见疾病，其体征和表征非常明确，医生很容易识别；但罕见病由于病例极少、临床表征多样且文献记载稀缺，想要将外在的躯体症状与底层的潜在病变关联起来极为困难。医疗团队光是排查并确认他的问题并非单纯的骨骼肌肉损伤、而是源于神经系统病变，就花去了整整三到四年；而从意识到病灶位于神经系统，再到最终推论出这是某种遗传基因缺陷并完成全基因组测序，整个过程耗时超过五年。如今，Stav 生活在佛罗里达州迈阿密，成为了一名软件工程师。 ### [04:01](https://www.youtube.com/watch?v=SJJBPwlPpww&t=241s) 确诊的意义：社群归属、前沿疗法与生育规划 当主持人询问完成全基因组测序并最终拿到明确诊断为人生带来了什么改变时，Stav 坦言，精准确诊彻底重塑了他的生活轨道。首要的改变在于心理与生活节奏的重建——在未知病因的情况下，患者往往会病急乱投医，甚至采取一些适得其反、对身体有害的盲目干预；而一旦对自己的疾病有了精准认知，就能够清醒地接纳自我，并学会在日常生活中科学适应与调整。 更为重要的是，确诊为患者推开了三扇原本关闭的大门： 1. **社群归属**：罕见病往往意味着极度的孤立感，确诊让患者得以融入由相同罕见病病友构成的互助群体，在精神和经验上不再孤军奋战。 2. **连接转化医学与新药研发**：许多生物科技公司和前沿科研团队正在针对特定罕见病靶点开发前沿疗法或特效药；只有在基因型完全确定的前提下，患者才有机会加入临床试验并争取潜在的治愈可能。 3. **精准家庭生育规划**：对于正处于人生黄金期的 Stav 而言，最大的释然来自于生育规划。明确了致病基因及其遗传模式后，他能清晰预判将该突变遗传给下一代的概率，并依托现代辅助生殖与产前基因筛查技术（如 PGD 等），在彻底规避疾病遗传风险的前提下孕育健康的下一代。 ### [05:32](https://www.youtube.com/watch?v=SJJBPwlPpww&t=332s) 诊断奥德赛：基因组数据的“大海捞针” 曼顿中心主任 Alan Beggs 博士接过话题，向大家阐释了为何像 Stav 这样的漫长遭遇在罕见病领域不仅绝非个例，反而是绝大多数患者都会面临的残酷现实。临床上将其形容为“诊断奥德赛”，其底层核心矛盾源于人类基因组本身的庞大尺度与现代测序技术产生的数据过载。 人类从父母双方各继承一套基因组，整体大小约为 30 亿个碱基对（Base Pairs），它们编码了大约 20,000 个不同的功能基因。在这 20,000 个基因中，目前已知异常时可导致人类疾病的基因数量在 8,000 至 9,000 个之间，且这一数字随着科研进展仍在不断攀升。此外，每一个基因内部又存在无数种不同的受损突变形式（如错义、无义、移码突变或剪接异常等）。过去，临床遗传学受制于检测手段，一次只能检测几个已知基因，犹如盲人摸象；而在近几十年，随着新一代测序技术（Next-Generation Sequencing, NGS）的普及，医学界终于能够一次性对全部基因甚至全基因组进行全覆盖测序。但这随之引发了新的次生难题：每做一次全基因组测序，系统输出的数据集就会包含数万乃至数百万个遗传变异位点。这正是医生们常说的“大海捞针”（Looking for needles in haystacks），而这正是引入 AI 协助专家团队快速锁定高疑似变异的关键契机。 - **人类基因组规模**：约 30 亿个碱基对，编码约 20,000 个基因。 - **已知致病基因库**：当前已明确发现约 8,000 到 9,000 个基因在异常突变时会直接导致疾病。 - **测序技术代际跃迁**：从昔日的单基因逐步测序，跨越到现代 NGS 生成包含成百上千万变异位点的大规模数据集。 ### [07:51](https://www.youtube.com/watch?v=SJJBPwlPpww&t=471s) 临床分析的时间瓶颈：繁琐的变异解读与 AI 的介入点 曼顿中心基因发现核心实验室总监 Catherine Brownstein 博士详细拆解了遗传学家在面对一例悬而未决的疑难病例时，究竟面临着怎样庞大而繁杂的时间瓶颈。面对一个未解病例，分析人员必须带着极其开放的心态从零推演。通常，首要步骤是排查已知相关基因，核实先前的临床分析是否有所遗漏。 ```mermaid flowchart TD accTitle: 临床疑难遗传变异分析流程 accDescr: 描绘遗传学家从已知基因排查、三人组测序、新发突变分析、隐性变异与拷贝数变异筛选，再到人群数据库比对和文献研读的全流程瓶颈。 subgraph G1["样本测序与遗传模式筛选"] M1["三人组测序数据 (Trio)<br/>患儿、父亲、母亲全套基因组"] --> M2{"模式分析"} M2 -->|"自发突变"| M3["新发突变 (De Novo)<br/>父母无突变，患儿独立突变"] M2 -->|"隐性遗传"| M4["双等位基因变异<br/>父母各为携带者"] M2 -->|"结构变异"| M5["染色体大片段改变<br/>缺失 (Deletion) 与重复 (Duplication)"] end subgraph G2["繁重的多数据库与文献筛查"] M3 & M4 & M5 --> V1["变异过滤：核对人群频率数据库<br/>排除常见良性变异多态性"] V1 --> V2["功能预测分析<br/>评估变异是否破坏蛋白质构象或功能"] V2 --> V3["前沿文献研读<br/>逐个基因检索全球最新研究报道"] end subgraph G3["临床决策瓶颈与 AI 介入"] V3 --> D1{"人工耗时评估"} D1 -->|"耗费数小时后推翻"| D2["否定假设，回到初始变异列表"] D1 -->|"锁定高置信度位点"| D3["进入临床遗传学表型印证"] D1 -.->|"大语言模型 (LLM) 赋能"| D4["自动化信息综合、透明证据聚合与快速推演"] end ``` 接下来，分析团队往往会扩展到家庭“三人组”（Trio，即患儿及其亲生父母）的全套基因组分析。首要的切入点是寻找所谓的新发突变（*de novo* 突变）——即父母双方均未携带，却在受精卵发育或胚胎发生早期自主出现的突变。Catherine 强调，绝大多数健康人体内其实都自然携带着大量*de novo*突变，甚至大片段的染色体缺失（Deletion）与重复（Duplication）在健康人群中也极为普遍；因此，不能一看到突变就判定其致病，必须小心翼翼地评估哪一个新发变异真正具有致病破坏力。除*de novo*突变外，分析人员还要寻找常染色体隐性遗传模式（父母各为杂合无症状携带者，患儿同时继承两个突变等位基因）或结构性变异。 锁定具有致病嫌疑的候选变异后，繁重的文献与数据库检索工作才刚刚开始。遗传学家需要逐一查验各大全球人类突变与人群频率数据库，核实该突变是否只是普通人群中常见的良性多态性，并使用计算生物学软件预测其是否会改变蛋白质的三维构象与生化功能。随后，他们必须深头扎进 PubMed 等海量生物医学文献库，阅读关于这些基因的所有基础科研文献。"有时你辛辛苦苦耗费好几个小时深度调研一个变异，最后却发现：'噢，等等，不可能是它。'"随后只能无奈地推翻重来，转入下一个候选位点。这种巨大的工作量让未解病例的诊断变成了一场漫长的智力消耗战，这也正是引入大语言模型（LLM）来大幅提速、理顺证据链并提升变异透明度的关键机遇所在。 ### [10:44](https://www.youtube.com/watch?v=SJJBPwlPpww&t=644s) 医疗系统的复杂性与患者家庭的漫长求医路 听完这段艰辛的变异解读流程，主持人感叹道，普通人若未曾亲历，很难体会到在正常生理运转背后，人类依靠何等幸运才避开了那些严重致病的自发突变；而对于那些不幸遭遇罕见基因突变的家庭而言，现实面临的医疗体系更是错综复杂到令人窒息。 罕见病患者往往需要在还没有明确疾病名称的情况下，奔波于神经科、骨科、内分泌科等多个专科之间，经历数不清的侵入性化验、影像学检查以及各种对症治疗。医疗团队往往只能疲于应付表层出现的各种症状，却始终无法指认幕后的真正元凶。Alan Beggs 博士随即针对这种极端复杂的就医困境，展开阐述这一艰苦探索之旅对患者个人、家庭以及长期提供支持的临床医护团队所带来的沉重心理与生理负担。 ### [11:15](https://www.youtube.com/watch?v=SJJBPwlPpww&t=675s) 从对症支持到精准医疗：找到致病基因的转折点 对于像 Stav 这样患有神经肌肉疾病的患者，传统的医疗体系通常只能根据患者肌无力的模式和严重程度进行分类。有的患者行走困难，有的患者甚至需要呼吸机维持生命，病情轻重不一。在目前的临床实践中，这类患者的大多数诊疗手段都停留在支持性治疗（supportive care）的层面。以神经肌肉疾病引起的上肢无力为例，脊柱失去支撑会导致严重侧弯，骨科医生能做的通常只有通过脊柱矫形手术在物理层面将脊柱撑直，却无法触及导致病症的核心基因突变或底层致病机制。 只有精准锁定引起肌无力的变异基因，医学界才有可能从根源上开创靶向疗法，实现真正意义上的个性化医疗（personalized medicine）——不再是像“头痛就吃阿司匹林”那样仅仅掩盖症状，而是直击病因。AI 工具在海量基因数据中迅速缩小排查范围、定位致病基因的能力，具有深远的影响。例如，Stav 在明确分子诊断之后，不仅获得了心理上的笃定，还直接对接了一家正在专门为其所患罕见病研发靶向药物的制药公司，这正是基因确诊所带来的最重大的临床转折。 - **传统支持疗法的局限**：对严重脊柱侧弯等并发症只能依靠骨科手术矫正，无法阻断病程进展。 - **个性化基因诊疗的前景**：从对症下药转向靶向致病基因，为患者打开参与针对性新药临床试验的大门。 --- ### [13:06](https://www.youtube.com/watch?v=SJJBPwlPpww&t=786s) 确诊时刻的震撼与释然：名字、社群与跨越三代的答案 当罕见病家庭经历了漫长的求医拉锯战，终于等来确诊结果的那一刻，往往是百感交集的“顿悟时刻”（Eureka moment）。很多家属在听到结果的第一反应通常是如释重负的感叹：“原来是这样，这终于解释得通了！”Alan 医生分享了一个附带发现（incidental finding）的典型案例：团队在为一名患病儿童进行 DNA 测序时，意外发现孩子携带了一个遗传自母亲的乳腺癌易感基因。当医生将结果告知那位身为护士的母亲时，她瞬间恍然大悟——她的母亲、姐姐和一位阿姨都曾患过乳腺癌，但整个家族过去从未意识到这背后竟存在遗传风险。 对于一小部分患者而言，确诊直接带来了针对性的治疗方案；而对所有患者来说，确诊赋予了疾病一个明确的名字、一个可以抱团取暖的病友社群，以及关于预后和未来的确定性，并为有需求的家庭提供了明确的生育规划依据。Manton 孤儿病研究中心的 Katherine 坦言，虽然自己极少直接面对患者，但少数几次与家属会面的经历都令人心碎且动容。罕见病患者所经历的漫长求医路无异于一场“奥德赛”（odyssey）：辗转于数不清的医生和专家门诊，做了一轮又一轮化验。她至今难以忘怀团队曾确诊的一位 91 岁高龄的老人——某种罕见病在一个家族中足足纠缠了三代人，直到第三代患者求医，团队才彻底破解致病基因并将结果反馈给整个家族。即便已是 91 岁高龄，老人在知晓答案时依然感到巨大的释然。然而，这种欣慰往往伴随着沉重的责任感：Katherine 坦言，每当她为破解一个疑难病例而振奋时，内心也会为那些依然未能解开的病例感到自责与煎熬，只能不断催促自己“立刻回到工作台前”。 - **确诊的多重价值**： - **医学层面**：少部分患者可直接获得特异性疗法，或提前防范附带发现的高危遗传病。 - **心理与社会层面**：结束多年无休止的求医流浪，融入同病相怜的患者支持社群。 - **家庭与未来层面**：获得明确的疾病预后指导，为后代的优生优育提供科学依据。 - **跨越代际的终点**：Manton 中心曾为一家三代遗传疑难病的 91 岁老人明确病因，终结了长达近一个世纪的困惑。 --- ### [16:08](https://www.youtube.com/watch?v=SJJBPwlPpww&t=968s) 人机协同工作流：o3 Deep Research 如何将黑盒转化为证据链 为了加速极其耗时的人工排查过程，波士顿儿童医院团队与 OpenAI 展开合作，将 OpenAI o3 Deep Research 模型引入研究工作流。研究并非让人工智能替代专家，而是让模型承担海量文献检索与假说生成的重任。在设计工作流时，OpenAI 团队的 Yash 与医学专家 Alan、Katherine 进行了深度探讨：明确临床专家需要怎样的提示词（prompt）、模型应输出何种结构化结果、分析复杂遗传表型时需规避哪些潜在陷阱，以及应当遵循哪些临床遗传学最佳实践。 该系统的核心在于**拒绝“黑盒式推荐”**。模型不会毫无缘由地指认某个基因为元凶，而是必须输出一份详实的“证据驱动型候选清单”：不仅列出候选基因和患者具体的变异位点，还附带支持该假设的同行评审文献与致病机制论述。最终的临床判断依然牢牢掌握在人类专家手中——像 Katherine 这样的遗传诊断学家根据模型提炼的高可读性报告进行交叉验证，进而决定是否启动后续的临床实验室验证，最终决定是否向患者出具诊断报告。 ```mermaid flowchart TD accTitle: AI辅助罕见病遗传诊断人机协同工作流 accDescr: 展示从输入基因变异与临床表型到o3推理及专家临床验证的全流程 subgraph InputStage["输入端：多维临床数据"] A["患者基因测序变异数据"] B["临床表型编码 (HPO Ontological Codes)"] end subgraph AIStage["AI推理端：OpenAI o3 Deep Research"] C["文献检索与知识交叉比对"] D["证据驱动型假说生成<br/>(剔除黑盒输出)"] end subgraph OutputStage["输出端：人机协同决策"] E["候选基因短名单<br/>(2至6个高置信度变异)"] F["遗传学家与临床医生人工审核"] G["后续实验验证与患者确诊"] end A --> C B --> C C --> D D --> E E --> F F --> G ``` --- ### [18:36](https://www.youtube.com/watch?v=SJJBPwlPpww&t=1116s) 积压十余年的“疑难悬案”：在 8,000 个基因与 30 亿碱基对中破局 这套工作流选择从最棘手的“疑难冷案”（cold cases）切入，这些病例历经人工数年排查均无功而返，部分患者甚至已经离世。Manton 中心积累了长达 10 到 15 年的历史测序数据，受限于当时有限的医学认知而未能确诊；尽管医学界每年都在发现新的致病基因，但面对浩瀚的数据积压，研究人员根本没有足够的人力和时间逐一回溯重析。 人类基因组拥有多达 30 亿个碱基对。在全基因组或外显子测序后，生成的原始数据表格通常多达 500,000 行，每一行代表一个基因变异，并包含 20 到 30 列注释信息（例如该变异在正常人群中的发生频率、影响的基因、变异类型以及生物信息学预测的致病性）。人类分析师的第一步通常是通过过滤算法剔除人群中的常见良性变异，但即便经过严格过滤，每个患者样本依然会残留数千甚至多达 10,000 个罕见突变，波及 500 到 1,000 个不同的基因。 现实的困境在于：人类专家的知识储备是高度细分的。Alan 专精于一小类基因，Katherine 熟悉另一类基因，全人类目前已知的致病基因超过 8,000 个，没有任何一位临床遗传学家能够完全掌握所有基因的最新研究动态。而具备长上下文与海量医学知识储备的推理模型恰好弥补了这一短板。研究人员将过滤后的基因变异与患者表型的标准化本体代码（例如代表肌无力的数值代码、因踝关节无力导致足下垂的特定代码）打包作为提示词输入模型，要求其根据医学文献将临床表型与基因变异进行交叉比对。模型最终将原本漫无边际的数千个候选范围压缩至 2 到 6 个高度可疑的变异位点。这极大地缩小了搜索空间，使原本不可逾越的“数据大海”变成了人类分析师可处理的聚焦范围。 - **基因组数据处理规模**： - 30 亿个碱基对 $\rightarrow$ 原始分析表格约 500,000 行变异数据，每行含 20 至 30 列注释。 - 常规人群频率过滤后，仍残留多达 10,000 个突变，涉及 500 至 1,000 个基因。 - **医学认知瓶颈**：人类已知致病基因逾 8,000 种，远超单个专家的记忆极限。 - **临床表型标准化**：利用数字本体编码（如特定表型特征编码）精准刻画患者临床特征。 - **搜索空间收敛**：AI 将候选变异大幅收敛至 2 到 6 个，使人工深度复核具备可行性。 --- ### [21:42](https://www.youtube.com/watch?v=SJJBPwlPpww&t=1302s) 从已知验证到未知突破：基准测试与提示词迭代 为了检验该方法的可行性，研究团队首先在“已知已解病例”（solved cases）上展开基准测试。在这些测试中，研究人员已经明确知晓致病基因和最终诊断，但在提示词中隐去标准答案，以此考察模型能否在盲测中独立推导出正确结果。 在初期测试过程中，团队经历了反复的工程迭代。他们针对模型的提示词逻辑进行了持续微调，重点教导模型规避分析复杂表型与复杂基因突变时的典型逻辑陷阱与误区，确保其推理链路严密，为后续攻克积压多年的未解疑难病例奠定了坚实的基础。 ### [22:12](https://www.youtube.com/watch?v=SJJBPwlPpww&t=1332s) 借助已知病例优化模型：从常见错误模式到 90% 准确率 在将 AI 模型应用于未知难题之前，研究团队采取的策略是先用已知确诊病例进行基准测试与调优。团队把这些已有明确结论的病例作为初始问题输入模型，观察它的判断准确率。随后，技术团队与 Alan Beggs 和 Katherine 紧密配合，复盘模型在判读时反复踩中的共性错误模式。人类遗传数据极其繁复，其技术层面的分析难点不仅困扰着人类分析师，大语言模型同样容易在类似的地方失误。 ```mermaid flowchart TD accTitle: 模型迭代与验证闭环 accDescr: 展示从输入已知病例、识别错误模式、针对性提示词调优，到准确率提升并进入临床分析的完整流程。 subgraph Iteration ["模型校准与迭代流程"] A["输入已知确诊病例"] --> B["识别共性错误模式\n(测序深度不足/假阳性)"] B --> C["注入针对性规则与 Prompt\n(提示规避特定技术陷阱)"] C --> D{"准确率是否达标？"} D -- "提升至 80%–90%" --> E["输出给人类遗传学家复核"] D -- "未达标" --> B end ``` 为了让模型学会避开这些技术陷阱，团队在提示词（Prompt）中加入了极具针对性的规则引导，例如明确告诫模型：“必须依赖足够高的测序深度（sequencing depth）才能确认变异的真实性，避免误将测序噪点当成致病变异”。经过这样反复的多轮定向调优，模型在已知测试集上的正确率一路上升到了 80% 到 90%。直到达到这个水准，团队才确信该输出已经具备临床实用价值，“值得让人类分析师投入宝贵时间去逐一复核”，而不是扔给他们一堆漏洞百出的无效推测。 - **验证策略**：以历史已确诊病例作为起点，对照基准真相评估模型的推断能力。 - **关键质控指标**：针对测序深度等底层生信指标建立过滤门槛，杜绝假阳性致病变异的干扰。 - **人机协同分工**：模型准确率跨过 80%–90% 门槛后，才正式接入临床专家的复核流水线。 --- ### [23:24](https://www.youtube.com/watch?v=SJJBPwlPpww&t=1404s) 挖掘 18 例新诊断与意外线索：S1PR1 基因与跨越数十年的科研巧合 在 Manton Center 重新分析的 376 例疑难病例中，OpenAI o3 成功挖掘出了 18 个全新或令人意料之外的明确诊断线索。Katherine 回顾了与模型协作的具体细节：最直观的情形是，某些变异本身是已知的致病突变，但在多年前初次检测该病例时，学术界尚未发现该基因与对应临床表型之间的关联。如今文献库已经更新，模型直接指出了这种已被证实的新对应关系，整个诊断“一目了然、毫不费力”（no-brainer），表型吻合得严丝合缝。 更令团队震撼的则是模型展现出的跨文献假设推理能力。在第 151 号历史疑难病例中（目前中心累计病例已突破 3100 例，这意味着该病例积压已久、历经数轮专家排查未果），患者同时罹患白癜风（vitiligo）、大动脉转位（transposition of the great vessels）与肺动脉高压（pulmonary hypertension）。o3 针对该突变给出了一个出人意料的候选基因：*S1PR1*。Katherine 起初对此半信半疑：“我当时心想：这算什么东西？不可能吧。” 然而，模型不仅精准提取了基因名，还进行了深度的全文献关联检索。它翻出了一篇发表于 26 年前的冷门论文，该研究早在当时就提出应当认真考虑 *S1PR1* 在白癜风病因机制中的核心作用，并且模型结合多篇研究构建了一条完整的分子通路推理链，论据极其扎实。更具戏剧性的是，Katherine 在 Google 上快速检索后赫然发现，**在 1988 年最初克隆出 S1PR1 基因的那位科学家，其实验室就在波士顿儿童医院的隔壁大楼里**。Katherine 随即发邮件询问对方看法，科学家立刻爽快回复：“好啊，你今天下午 4 点能顺道过来聊聊吗？” - **翻阅文献深度突破**：人类研究员在 PubMed 上翻阅 4 页文献往往就精疲力竭，而模型能够不知疲倦地检索挖掘到第 30 页甚至更深处的历史文献。 - **假设推理潜能**：模型不仅能匹配已知变异，还能依据突变类型与分子通路，跨学科将沉睡数十年的前沿基础研究与复杂的罕见病表型串联成全新假说。 --- ### [27:02](https://www.youtube.com/watch?v=SJJBPwlPpww&t=1622s) 打破医疗门槛：工具普及化与“不变的基因组，变化的新知识” 既然 AI 在疑难病诊断上展现出惊人的突破能力，接下来的核心瓶颈便是如何实现普及，而不是让先进诊疗手段局限在少数顶尖医疗中心内部。目前该系统虽然基于公开发布的 o3 模型研发，但对于基层机构而言，上传患者测序数据、确保合规安全以及环境部署仍然存在门槛。 为此，团队正借助 OpenAI Foundation 的资助，全力开发一套通用、易用且高度安全的临床分析工具，致力于消除医疗资源鸿沟。Katherine 强调，患者绝对不应该非得千里迢迢赶到顶级三级专科医疗中心（tertiary medical center），才能享受到对每一个突变变异的最前沿严谨评估。 ```mermaid flowchart LR accTitle: 基因组周期性自动化重分析流程 accDescr: 展示患者测序数据保持不变，随着全球医学知识的持续扩展，系统通过低成本周期性重分析促成新诊断。 A["恒定的患者全基因组数据\n(无需重复采样测序)"] --> C["低成本定期自动化重分析\n(OpenAI o3 驱动)"] B["持续爆炸式增长的全球文献库\n(新基因-表型关联不断涌现)"] --> C C --> D["精准识别新致病线索"] D --> E["向临床医生发送诊断提醒通知"] ``` 在患者权益倡导方面，Katherine 结合自身家庭的真实经历指出，很多时候传统医疗护理团队反而会无意中成为“看门人”（gatekeepers），阻碍罕见病家庭接触最新的科学研究与临床试验。而 Manton Center 始终保持开放机制，允许患者“自主转诊”（self-refer），家庭可以直接致电或发送电子邮件寻求评估。 对于临床医生与基层科研人员而言，该工具的核心价值在于大幅压缩时间成本。面对临床基因检测阴性的悬案，医生往往没有精力从零开始推倒重来。然而事实是：“**人类的基因组序列测定后或许不会改变，但全球医学知识每天都在持续更迭**”。只要能以低廉的计算成本实现定期自动重跑重分析，就能源源不断地从沉睡的陈旧数据中唤醒新的确诊希望。 - **去中心化诊疗愿景**：让基层诊所和偏远地区的罕见病家庭无需跨区域就医，即可获得顶级标准的变异位点再分析。 - **患者自主权**：绕过传统就医环节的信息层层阻断，推行受试者直接向科研项目自主报名的机制。 --- ### [29:26](https://www.youtube.com/watch?v=SJJBPwlPpww&t=1766s) 畅想 AI 重塑医学：后台常驻 Agent 与临床赋能 当被问及对 AI 变革医疗领域的长远期许时，三位嘉宾从技术自动化、临床决策和患者福祉三个维度给出了洞察： Suhas 展望了一种“全天候自主运转”的智能重分析生态。他设想在不久的将来，每当 PubMed 或全球医学期刊上线一篇新论文、涌现出任何新知识切片，底层的 AI Agent 就会自动触发后台流水线，对积压的未确诊病例库进行动态比对。一旦模型拥有足够置信度，确认新知识与某位患者的病变高度吻合，它便会自动通过电子邮件或系统弹窗向主治医生发出精准预警。这种模式能够极大解放人类专家宝贵的认知带宽：“在我们分析的案例中，AI 成功破解了 5% 的疑难杂症；而剩下的 95% 极端疑难案例，则能留给人类科学家集中精力攻坚。” Alan Beggs 对此深表赞同，并给出了具体数据：在 Manton Center 登记入组的 5,000 个家庭中，至今仍有大约 2,000 个家庭苦苦追寻不到确诊结果。AI 不仅能在已知知识范围内大幅提速诊断，更能自主推导新假说，协助医生跨入未知领域。面对 AI 是否会取代医生的疑虑，Alan 坦言自己完全不担心失业：“我们依然不可或缺，依然需要像我们这样的人类专业人员去解读信息、评估临床意义并制定下一步医疗方案，但 AI 无疑让整个流程的效率提升了数倍。” Katherine 则从情感与转化医学角度表达了对未来的憧憬。确诊仅仅是第一步，她梦想建立一个端到端的支持网络：从罕见病精准确诊，到无缝衔接个性化治疗方案、精准匹配最前沿的临床试验。她动情地表示：“面对能够帮孩子们确诊疾病、让饱受折磨的家庭生活重获希望的事情，又有谁能拒绝呢？” - **背景常驻智能体（Agentic Approaches）**：知识库更新驱动的静默自动重排查与精准提醒机制。 - **攻坚分流机制**：让 AI 自动化消化常规与中等复杂度的未确诊案件，使顶级遗传学家的智力资源聚焦于最具挑战性的 95% 未知盲区。 - **全链条闭环**：跨越从“表型匹配”到“病因确立”，再到“精准靶向治疗与药物试验入组”的医疗全生命周期。 --- ### [32:54](https://www.youtube.com/watch?v=SJJBPwlPpww&t=1974s) 资金与科研支持澄清，开启社区互动问答 在正式进入线上观众问答环节之前，主持人特意插入了一段机构运作与资金来源的说明：此前针对波士顿儿童医院 Manton Center 与 Suhas 的联合攻关项目，最初是由 OpenAI 官方直接提供技术与研究支持。随着机构组织架构的演进，团队目前所获得的用于推广普及该分析工具的专项资助，已经正式转入新成立的 OpenAI 基金会（OpenAI Foundation）旗下独立运行。 理清合作主体与项目支持体系后，研讨会正式进入社区互动环节，首位提问者是来自科学新闻界的技术撰稿人兼编辑 Elizabeth Stewart。 ### [33:25](https://www.youtube.com/watch?v=SJJBPwlPpww&t=2005s) 全基因组测序在早期诊断中的普及与经济壁垒 在问答环节中，来自 Innovate Bio 的参会者提出问题：如果在罕见病患者的诊断初期就全面开展全基因组测序（Whole Genome Sequencing, WGS）及分析，罕见病的确诊速度究竟能够提升多少？波士顿儿童医院的 Alan Beggs 毫不犹豫地用连串的肯定回答：“Yes, yes, and yes（绝对如此）”。他指出，当前罕见病医学界普遍面临的一大阻碍是商业保险公司的报销滞后。事实上，全基因组测序的检测成本目前已经降至 1,000 美元以下，甚至只需数百美元，这仅仅是医院常规核磁共振成像（MRI）费用的一小部分，而保险公司为高昂的影像学检查买单时往往毫不犹豫，却在基因检测报销上设立重重关卡。 Alan Beggs 进一步介绍了波士顿儿童医院参与的多项临床前沿计划，包括新生儿基因组测序，以及对所有病情严重到入住新生儿重症监护室（NICU）的患儿开展常规基因测序。得益于基因检测技术的飞速进步，如今明确诊断的周期已经从过去的数月乃至数年，大幅压缩至数天甚至数周之内。他提到，团队的部分同事此前甚至创下了“吉尼斯世界纪录（Guinness World Records）”，在短短 24 小时内就完成了从测序到确诊的全流程。尽管这种极限速度目前尚未成为日常临床的常规标准，但 Alan 强调，所有患有未确诊疾病或身体状况存在不明病因的患者，都理应接受全基因组测序。即便部分病例在当下无法立刻获得确切答案，这些宝贵的高精度遗传信息也将永久留存，随时可能在未来的医疗照护中发挥关键作用。 - 全基因组测序成本已降至 1,000 美元以下（部分仅需数百美元），远低于常规 MRI 检查费用，但保险报销阻力依然存在。 - 在新生儿重症监护室（NICU）及重症新生儿筛查中，常规基因确诊时间已缩短至数天至数周。 - 合作团队曾荣获吉尼斯世界纪录认证，实现了在 24 小时以内完成全基因组测序并锁定临床确诊。 --- ### [35:23](https://www.youtube.com/watch?v=SJJBPwlPpww&t=2123s) AI 跨体征关联与双重致病变异的精准识别 来自 Crossing Point IT Solutions 的 Jason Duca 提问：AI 是否有潜力将患者身上看似无关的多种健康异常联系起来，揭示出大多数临床医生可能会忽略的复杂诊断模式？研究人员对此给出了极其肯定的回答，并列举了最新的医学研究——例如仅通过视网膜或眼底的微小改变，就能预警并捕捉身体深处其他内脏器官的潜在病变，而高维度的 AI 跨学科模型正非常适合处理这类多系统映射任务。 为了说明 AI 洞察复杂病例的敏锐度，研究人员分享了该系统研发过程中的一个标志性转折点。当时团队为模型输入了 20 例在过去已被人类专家“成功确诊”的已知复杂病例以验证其性能。最终模型在其中 19 例给出了与专家一致的答案，但在唯独“答错”的那一例上，团队复核后却大吃一惊：原来出错的并非模型，而是人类医生。该患者体内实际上同时并存着两种独立的遗传变异与病理机制，而人类专家团队此前误以为所有临床表现都是单一致病变异导致的“表型拓展（Phenotypic Expansion）”，只关注了其中一个变异。AI 却不受思维定式干扰，成功同时捕获并分拣了这两个并存变异。这种强大的解耦能力，能够清晰理清复杂的临床表型究竟分别由哪些特定基因突变所引起，填补了人类医生在多重罕见变异重叠时的认知盲区。 ```mermaid flowchart TD accTitle: AI 解析多重基因突变机制 accDescr: 对比传统人类专家将所有症状归为单一变异表型拓展，与 AI 分离识别双重突变的过程。 subgraph HumanExpert["人类医生传统诊断思维"] H1["复杂重叠临床表型"] --> H2["发现主导变异 A"] H2 --> H3["归因为变异 A 的表型拓展\n（忽略潜在并存病因）"] end subgraph AIModel["AI 深度交叉推理"] A1["多维度临床表型输入"] --> A2["深度关联分析与变异拆解"] A2 --> A3["变异 A: 对应表型分支 1"] A2 --> A4["变异 B: 对应表型分支 2"] A3 & A4 --> A5["双重致病机制精准确诊"] end ``` --- ### [36:47](https://www.youtube.com/watch?v=SJJBPwlPpww&t=2207s) 深度推理 vs 数据整合：AI 诊断的核心价值与演进路径 针对参会者 Svetlana Romanova 提出的关键问题——AI 在诊断突破中取得的巨大增益，究竟更多来自于多模态数据的整合与接入，还是模型自身的逻辑推理能力？与会专家进行了深入的技术拆解。在本次由 OpenAI o3 Deep Research 支持的先导研究中，两者都扮演了极为关键的角色。为保证早期测试的稳定性，研究人员实际上投入了大量精力对数据进行前期人工清洗与降维：并非直接把人类基因组包含的 30 亿个碱基对（Base Pairs）全部丢给模型，而是将其精简筛选为大约 1,000 到 10,000 个变异位点的候选集，同时对患者的体征与表型清单进行了标准化规范。 然而，团队对于医学 AI 的终极愿景是“零人工干预的全自动端到端分析”。在理想的未来模式下，临床医生无需再耗费大量时间去手动清洗基因组或整理表型关键词，而是直接把患者未经处理的完整电子病历（EHR）——包括检验报告、影像学原始数据等全部资料——连同未经裁剪的全基因组测序原始文件整体交付给大语言与推理模型。随着前沿推理大模型自主过滤噪音、多步综合逻辑推理与长上下文窗口能力的持续迭代，AI 将完全有能力在海量未清洗的生物信息中独立梳理出致病线索。 --- ### [38:03](https://www.youtube.com/watch?v=SJJBPwlPpww&t=2283s) 结语：迈向生命科学深度研究的 GPT Rosalind 与未来展望 讨论尾声发生了一段轻松幽默的插曲：现场突然响起了手机铃声，引来主持人打趣道：“难以置信，这个年代居然还有人真的在打电话”，随后笑称“这或许是后台在用音乐提醒我们该超时结束了”。在活跃的氛围中，主持人向波士顿儿童医院团队表达了诚挚的感谢，并动情地提到，包括自己的亲友在内，世界上有成千上万个家庭正深陷罕见病的困扰中，他们的生活与希望都高度维系于此类前沿生命科学与 AI 交叉研究的突破。 在技术总结中，主持人明确说明：尽管本次在波士顿儿童医院展开的临床回溯试验采用的是 OpenAI o3 Deep Research，但团队已经在紧锣密鼓地研发更新一代、专为更深层次生命科学研究而设计的通用推理与文献合成模型——例如专门以 DNA 双螺旋先驱 Rosalind Franklin 命名的 "GPT Rosalind"。与此同时，OpenAI 基金会（OpenAI Foundation）为波士顿儿童医院研究团队提供了专项资助赠款，全力支持这项未确诊罕见病研究的持续深化。OpenAI 论坛未来一年内将持续追踪 AI 与医疗健康深度融合的进展，并将攻关重点进一步延伸至罕见病与罕见癌症的诊断突破。 ## Notable quotes > [07:51](https://www.youtube.com/watch?v=SJJBPwlPpww&t=471s) 有时你辛辛苦苦耗费好几个小时深度调研一个变异，最后却发现：'噢，等等，不可能是它。' — Catherine Brownstein > [13:06](https://www.youtube.com/watch?v=SJJBPwlPpww&t=786s) 原来是这样，这终于解释得通了！ — 患者家属 > [23:24](https://www.youtube.com/watch?v=SJJBPwlPpww&t=1404s) 我当时心想：这算什么东西？不可能吧。 — Catherine Brownstein > [23:24](https://www.youtube.com/watch?v=SJJBPwlPpww&t=1404s) 好啊，你今天下午 4 点能顺道过来聊聊吗？ — 波士顿儿童医院科学家 > [27:02](https://www.youtube.com/watch?v=SJJBPwlPpww&t=1622s) 人类的基因组序列测定后或许不会改变，但全球医学知识每天都在持续更迭 — Catherine Brownstein > [29:26](https://www.youtube.com/watch?v=SJJBPwlPpww&t=1766s) 在我们分析的案例中，AI 成功破解了 5% 的疑难杂症；而剩下的 95% 极端疑难案例，则能留给人类科学家集中精力攻坚。 — Suyash Shringarpure > [29:26](https://www.youtube.com/watch?v=SJJBPwlPpww&t=1766s) 我们依然不可或缺，依然需要像我们这样的人类专业人员去解读信息、评估临床意义并制定下一步医疗方案，但 AI 无疑让整个流程的效率提升了数倍。 — Alan Beggs > [29:26](https://www.youtube.com/watch?v=SJJBPwlPpww&t=1766s) 面对能够帮孩子们确诊疾病、让饱受折磨的家庭生活重获希望的事情，又有谁能拒绝呢？ — Catherine Brownstein > [33:25](https://www.youtube.com/watch?v=SJJBPwlPpww&t=2005s) Yes, yes, and yes（绝对如此） — Alan Beggs > [38:03](https://www.youtube.com/watch?v=SJJBPwlPpww&t=2283s) 难以置信，这个年代居然还有人真的在打电话 — 主持人 ## Mentioned OpenAI, OpenAI Forum, 波士顿儿童医院 (Boston Children's Hospital), 曼顿孤儿疾病研究中心 (Manton Center for Orphan Disease Research), 哈佛大学, OpenAI o3 Deep Research, Stav Rones, Catherine Brownstein, Alan Beggs, Suyash Shringarpure, Elizabeth Stewart, Jason Duca, Svetlana Romanova, Rosalind Franklin, GPT Rosalind, OpenAI 基金会 (OpenAI Foundation), Innovate Bio, Crossing Point IT Solutions, S1PR1, PubMed, 吉尼斯世界纪录 (Guinness World Records) ## Video description For many families living with a rare disease, the hardest part is not having a name for what is happening. Even with modern genetic testing, about half of people with rare diseases remain undiagnosed. In this OpenAI Forum conversation, researchers from Boston Children’s Hospital's Manton Center for Orphan Disease Research discuss how they used OpenAI o3 Deep Research to reanalyze difficult pediatric cases and surface new leads for expert review. The conversation explores how AI can help scientists move faster, give families a better chance at answers, and support advanced medical research, while keeping geneticists and clinicians at the center of diagnosis and care. Join the OpenAI Forum to explore more conversations with experts working at the forefront of AI, science, and society: https://forum.openai.com/ [Source](https://www.youtube.com/watch?v=SJJBPwlPpww)

[7:51] Catherine Brownstein: 有时你辛辛苦苦耗费好几个小时深度调研一个变异，最后却发现：'噢，等等，不可能是它。' [13:06] 患者家属: 原来是这样，这终于解释得通了！ [23:24] Catherine Brownstein: 我当时心想：这算什么东西？不可能吧。 [23:24] 波士顿儿童医院科学家: 好啊，你今天下午 4 点能顺道过来聊聊吗？ [27:02] Catherine Brownstein: 人类的基因组序列测定后或许不会改变，但全球医学知识每天都在持续更迭 [29:26] Suyash Shringarpure: 在我们分析的案例中，AI 成功破解了 5% 的疑难杂症；而剩下的 95% 极端疑难案例，则能留给人类科学家集中精力攻坚。 [29:26] Alan Beggs: 我们依然不可或缺，依然需要像我们这样的人类专业人员去解读信息、评估临床意义并制定下一步医疗方案，但 AI 无疑让整个流程的效率提升了数倍。 [29:26] Catherine Brownstein: 面对能够帮孩子们确诊疾病、让饱受折磨的家庭生活重获希望的事情，又有谁能拒绝呢？ [33:25] Alan Beggs: Yes, yes, and yes（绝对如此） [38:03] 主持人: 难以置信，这个年代居然还有人真的在打电话

[1:43] Stav: When I was getting more competitive in my teenage years, I started to notice that I was having certain issues that other kids were not facing... It was pretty frustrating not knowing what the issue was. [10:44] Dr. Alan Beggs: So that explains it, that examines it... There is this type of Eureka moment for some. For a small subset of patients today, it leads to a specific therapy. For everybody... it leads to a name, a community. [33:25] Dr. Alan Beggs: yes, yes, and yes [37:58] Believe it or not, people still make phone calls occasionally.

[0:09] Welcome everyone. Thank you for joining [0:11] us again. Uh today's conversation looks [0:14] at a research collaboration between the [0:16] Manton Center for Orphan Disease uh [0:19] orphan Disease Research at Boston [0:21] Children's Hospital and Harvard and [0:23] ourselves at OpenAI. Um, so you should [0:25] know that hundreds of millions of [0:27] families worldwide struggle with rare [0:29] disease and the average time just to get [0:31] a diagnosis is six to seven years. So [0:34] it's it's a huge long struggle for them. [0:37] Um, and this research is super [0:39] important. So in this in this study [0:41] across 376 cases, an AIdriven workflow [0:44] surfaced evidence linkeds that led to 18 [0:48] diagnoses of rare disease. So we want to [0:51] dig into that. First, we're going to [0:53] hear from Stavronis, a Maten Center [0:55] research participant, about his search [0:58] for a diagnosis for his own rare [0:59] condition and the impact of getting [1:01] diagnosed. We're then going to hear from [1:03] three study co-authors. Dr. Katherryn [1:04] Brownstein is the scientific director of [1:07] the Maten C Center's gene discovery [1:09] corps where she focuses on genome [1:11] analysis and barriers to diagnosis. Dr. [1:14] Alan Begs is director of the Maten [1:15] Center and brings decades of experience [1:17] in human genetics and rare disease. and [1:19] Suya Shringerpur is a machine learning [1:22] researcher and a statistical geneticist [1:24] at OpenAI working across AI, genomics [1:26] and computational biology. So they'll [1:29] discuss why rare disease is still hard, [1:31] how this workflow operated, what the [1:33] study found, and what its limitations [1:35] were. Then we'll open up the [1:36] conversation to the community. So let's [1:39] start with Stav. [1:43] St. Hi, welcome. [1:45] >> Thank you. [1:47] So stop, you first noticed symptoms as a [1:49] teenager, but you spent years searching [1:51] for the right diagnosis, including an [1:53] initial diagnosis that turned out to be [1:54] wrong. Can you tell us first [1:58] about yourself? Um what where do you [2:01] sit? What do you do? And then we'll get [2:03] into the story of of your search for [2:06] diagnosis. [2:08] >> Yeah. Yeah, absolutely. Uh thanks for [2:10] having me. So my name is Stav Rones. Uh [2:12] was born and raised in the Boston area. [2:15] Um, growing up, uh, you know, I always, [2:17] uh, loved playing sports, different [2:19] sports, baseball, basketball, rowing, [2:22] and, um, when I was, uh, kind of getting [2:24] more competitive in my teenage years, I [2:26] started to notice that I was having [2:27] certain issues that other kids were not [2:29] facing. Um, and it was pretty [2:32] frustrating not knowing what the issue [2:34] was. I was getting certain pains, [2:35] limitations with movements. Um, and you [2:39] know, eventually tried to to seek [2:40] professional medical help. And um it [2:43] took a long time to even figure out that [2:45] it may be genetic related. Um and then [2:48] to get from even there to doing a full [2:50] genome testing and figuring out what [2:53] kind of uh rare disease it was was a was [2:55] another whole process. But uh currently [2:58] uh yeah I'm based out of Miami, Florida [3:00] and uh I'm a software engineer but um [3:02] yeah. [3:03] >> Yeah. So so started when you were a [3:06] teenager. How long did it take you to [3:08] get that the correct diagnosis? [3:12] It took a long time. Um I mean there are [3:14] a variety of different factors you know [3:16] like trying to correlate certain [3:18] physical issues with having a genetic a [3:20] rare especially a rare genetic disease [3:22] right. Um I mean certain things express [3:24] themselves that are common diseases it's [3:26] more easy to diagnose but with rare [3:28] diseases because there's so little [3:30] patients and so little understanding [3:32] about what the how it presents itself um [3:35] what kind of symptoms it took at least 3 [3:38] to four years to figure out that it [3:40] wasn't just a physical issue that that [3:41] it was neurological um and and from [3:45] knowing that it was neurological to get [3:47] to the point that it was genetic was a [3:49] wasful [3:51] process [3:52] at least five years. [3:54] >> Yeah. Wow. Um, so you got your genome [3:57] sequenced and then you got a diagnosis [3:59] and you got clarity. What did that [4:01] unlock for you? What was life like? Uh, [4:04] and what were you able to do after you [4:06] got clarity? [4:09] >> Yeah, absolutely. I mean just in terms [4:10] of um you know the the knowing exactly [4:15] what the issue is um just kind of helps [4:18] you situate yourself and and um [4:22] adapt to to knowing what you have [4:26] because if you don't know what the issue [4:27] is, you know, you might do things that [4:29] are harmful. [4:31] um and knowing exactly what you have to [4:35] um you can be part of a community of [4:37] others with the same thing and that's um [4:40] you know it's just a really nice to to [4:44] kind of understand and and be with [4:45] others who have the same thing. Um and [4:48] also really importantly um there are um [4:52] companies and there are teams working on [4:54] cures and solutions for these rare [4:56] diseases. So knowing exactly what you [4:58] have, you can actually um start to get [5:01] involved and potentially uh find you [5:05] know find cures uh for the river [5:08] disease. And then also um for me the [5:11] biggest things is when it came to family [5:12] planning. So knowing the implications of [5:16] um what having the disease does in terms [5:19] of uh when you go to have children uh [5:21] was very important for me and um using [5:23] modern technology how you can kind of um [5:26] you know get around that. [5:30] Yeah, that's amazing. Thank you for [5:32] sharing that story. Um, we're going to [5:35] bring on the the three study co-authors [5:38] we have here, Kathern Brownstein, Alan [5:40] Begs, and Su Shringerpur. Um, to kind of [5:43] widen what we're uh what we're [5:46] discussing here. So, I'd like to start [5:48] the panel with Katherine Allen Su. Um, [5:52] Allan, I I'd like to start with you [5:54] first. St. Search for an answer took [5:56] years. How common is an experience like [5:58] his and why can rare disease cases [6:01] remain unsolved for so long? Sure. Well, [6:04] it unfortunately it's way too common. [6:06] People often refer to this as a [6:09] diagnostic odyssey. Um folks know that [6:12] something's going on. In many cases now [6:15] in recent decades, we've realized that [6:17] many of these rare conditions have a [6:19] genetic basis. But our ability to uh [6:23] test all the genes that we have been [6:25] very difficult. To put the problem in [6:28] context, our genome, the DNA that we [6:31] inherit from each of our parents is [6:32] roughly three billion bases. And those [6:35] three billion bases encode something in [6:37] the order of 20,000 different genes. And [6:40] then each gene can have many different [6:43] ways that it is damaged. And we know now [6:46] of the number is constantly going up, [6:48] but but between eight and 9,000 [6:51] different genes that are known to be [6:53] associated with disease when they're [6:54] abnormal. [6:56] Um and [6:58] so patients in the past we only had the [7:01] ability to look at a few genes at a [7:03] time. What's changed is newer technology [7:06] generically called next generation [7:08] sequencing that allows us to sequence [7:10] everything. But then our data sets of [7:12] course contain upwards of hundreds of [7:15] thousands to millions if not more [7:18] genetic data points. And uh as we'll get [7:21] into what the ai allows us to do is help [7:24] hone in on the relevant ones. So this [7:27] has been for many families a long [7:29] process this so-called genetic or [7:32] diagnostic odyssey. Um and through uh [7:36] collaboration with research institutes [7:38] like us and other around the world uh [7:42] people eventually get an answer. [7:44] >> Yeah. So you're really looking for [7:46] needles in haststacks. [7:47] >> Yeah. That's what people often say. [7:49] Yeah. Um, thank you, Katherine. Um, you [7:53] said the bottleneck is time because [7:55] geneticists are sorting through varants [7:57] and databases and complex ambiguous [8:00] symptoms and new research. What does [8:02] that time bottleneck look like for you [8:04] when you're working on an unsolved case? [8:08] So when I don't know when no one knows [8:11] what's going on, you try to approach a [8:13] case with an open mind and really kind [8:16] of start at the beginning and look at [8:19] the known genes, see if something was [8:21] missed. Um then you broaden it out and [8:24] you if you have a complete trio that is [8:27] mother and father and child, you look at [8:30] what's not in the parents but they're it [8:33] but is in the child. So, it's an error [8:35] during development or at some point [8:38] along. And all of us have these things [8:40] that are called denovo mutations that [8:42] aren't in either of our parents, but [8:45] they're in us. And there it's a good [8:48] place to start to kind of look and see, [8:50] okay, is that what's going on here? Is [8:54] there was a denovo variant that could be [8:57] disease-causing? or you kind you can [9:00] look at things where there's each parent [9:02] is a carrier and the child got two uh [9:05] variants, one from each parent. Um and [9:08] then from then you go on and look at [9:11] large changes um in a person's genome [9:15] like deletions or duplications and it [9:18] sounds scary but like a lot of deletions [9:21] and duplications are in all of us and [9:23] we're totally fine. Um so uh because [9:26] there are so many genes, we have to um [9:31] look at [9:33] the ones that we think could possibly [9:35] cause disease and that takes a huge [9:37] amount of time. And then when we see [9:39] something that's interesting, we have to [9:41] look at all the databases to make sure [9:43] that it's not actually just a common [9:45] variant. Um or it's predicted to not do [9:50] something to the protein and be [9:51] disease-causing. And then we go to the [9:54] literature and read all about that gene. [9:57] Um, I'm kind of rambling here, but the [10:00] the idea is and the main idea is that [10:03] there's a huge amount of work that goes [10:04] into looking at each and every variant. [10:07] And then sometimes like after several [10:09] hours, you're like, "Oh, wait, no, no, [10:12] no. This can't possibly be it." And then [10:14] you're on to the next one. So it was a [10:16] real opportunity for uh uh an LLM to [10:21] help us a lot in this entire process. [10:25] >> Yeah. And bring transparency to that [10:27] data. That's amazing. And I'm I'll bet [10:29] many people in the audience are are [10:31] surprised like I am to think I'm [10:32] carrying a lot of denovo mutations [10:34] myself. And you thank goodness I just [10:37] got lucky that they didn't severely [10:38] affect the function of my body as they [10:40] have in in these rare disease families. [10:43] Um, [10:44] >> so I got lucky. Um, Alan, uh, health [10:48] care in this context is incredibly [10:50] complex. It's it just it's a kind of [10:53] journey that most people, thank [10:54] goodness, don't have to go on, but [10:56] there's specialists, there's lots of [10:57] tests, treatment for symptoms before [11:00] anybody names a condition. Um, can you [11:04] talk to me a little bit about what what [11:05] does this journey mean for patients and [11:08] the families and the care teams that try [11:10] to support them? like it sounds like a [11:12] very different experience than what than [11:13] what most of us are used to. [11:15] >> Well, so I mean for um someone like Stav [11:18] who has a neuromuscular condition um in [11:21] general uh patients have been [11:23] characterized by their pattern and [11:24] severity of weakness. Some people have [11:27] trouble walking, others need assistance [11:29] breathing. Uh can be mild, it can be [11:31] severe. And management and care in [11:35] medicine today is largely supportive for [11:38] uh patients like this. Um specifically [11:41] thinking about neuromuscular disease for [11:42] example um patients who h have [11:45] particular weakness of their upper body [11:48] can develop a very curved spine and so [11:50] an orthopedic [11:51] management could involve surgical [11:53] correction to keep the spine straight [11:55] and so on but there's really nothing [11:56] targeted at the gene or at the [11:58] underlying cause of the disease itself. [12:01] um through identifying the genes that [12:04] are abnormal and leading to the basis [12:06] for in this case weakness. Um there's [12:10] hope that we can start to develop an [12:13] approach that's targeted specifically [12:15] for that. And uh nowadays in medicine, [12:19] we talk about the concept of [12:21] personalized medicine. instead of just [12:23] simply giving somebody aspirin for a [12:26] headache, if we know exactly what's [12:28] causing it, that we may be able to [12:30] develop a genetically based approach for [12:32] that. So, the ability of [12:36] um an AI tool to help us identify and [12:40] hone in on the particular gene can [12:42] really be um very dramatic. for example [12:45] um I don't think he had a chance to [12:47] mention it but stav knowing his [12:48] diagnosis now um in addition to the [12:52] other benefits he cited now has the [12:55] ability to he's actually connected with [12:56] a company that's developing a drug [12:59] specifically for his condition [13:01] >> so this is I think one of the major [13:03] benefits that this genetic diagnosis [13:05] makes possible [13:06] >> yeah thank you um it so we're really [13:09] talking about that big inflection point [13:11] when they're going through this journey [13:13] and and they finally have a name for [13:15] something. You've probably been in the [13:17] room with some of those families like [13:19] and you know the impact that that that [13:21] diagnosis has on them. Can you talk to [13:23] me about that moment? What does it mean [13:24] for them? Um how does it feel? [13:28] >> Well, I can think of a few times when [13:30] somebody like it would tell them what we [13:33] found, they go, so that explains it, [13:36] that examines it. Um I'm actually going [13:39] to tell you about an incidental finding [13:42] with somebody uh where we had sequenced [13:44] their child's DNA because the child was [13:48] sick and we identified for example a a [13:51] susceptibility gene for breast cancer um [13:54] and it had been inherited from the [13:56] mother and when we returned this to them [13:59] um her f she was a nurse her first [14:01] reaction was oh my gosh and a light bulb [14:04] went off in her head and her sister and [14:06] her mother and an aunt had all had [14:09] breast cancer [14:10] >> and they had no idea it was genetic [14:11] invasives. [14:12] >> So there is this type of Eureka moment [14:16] for some. [14:16] >> Yeah. Um, for a small subset of patients [14:20] today, it leads to a specific therapy. [14:23] For everybody, as Do mentioned, it leads [14:25] to a name, a community. [14:28] >> Um, and it leads to some information on [14:31] your future on your prognosis and what [14:33] the future holds for you, as well as f [14:35] potential family planning for those who [14:37] choose to utilize it. [14:39] >> Thank you. And and Katherine, I I think [14:41] you've seen this, too. What what what [14:43] are those moments like when a family [14:45] learns really what's going on? [14:48] >> Well, I don't usually meet with the [14:50] families. Um only in a couple instances [14:53] I have um and they've asked to meet me. [14:56] Um and it's usually quite um emotional [15:00] just because, you know, they've [15:04] been on this odyssey. I mean, we keep [15:06] throwing around that word, but there's [15:08] no real substitute for it. It's been so [15:11] long and frustrating and for them and [15:15] going from doctor to doctor and [15:17] specialist to specialist and all these [15:18] tests and you know uh I like throwing [15:21] out the example that at the Manton [15:23] Center once we diagnosed I'll never [15:25] forget a 91year-old and it was a family [15:28] member who had a condition that was in [15:32] three different generations and finally [15:34] the third one we were able to figure out [15:36] what was going on and return the results [15:38] to the entire family and you know even [15:41] at 91 he was like oh finally we have an [15:43] understanding we we know what's going on [15:46] now um and it's a wonderful feeling but [15:49] I can't help but say that every time I [15:53] have that great feeling that we were [15:55] able to figure it out I feel really [15:58] upset with myself that I can't figure [16:00] out those other cases so um it's always [16:04] like okay get back to work [16:08] That's a heavy load. Um [16:12] so Yash um the team used OpenAI 03D [16:16] research as part of research workflow. [16:18] Uh first on the diagnosed cases where [16:20] they already knew the answer and could [16:21] test against that and then on some [16:23] unsolved cases. Uh what role did the IM [16:27] model play in the study and what parts [16:31] really needed that collaboration with [16:35] experts? Yeah, I think like Katherine [16:37] described her process of looking [16:38] exhaustively through genetic variants [16:40] that could cause the disease. We [16:42] wondered if we could sort of accelerate [16:44] that by having the AI models do the [16:46] literature search, identify hypothesis [16:49] and present that list to a person to [16:53] prioritize. So we focused on first [16:55] getting input from uh Alan Katherine on [16:58] like what is the prompt that the model [17:00] should get? What kind of output would [17:02] you like to see from the model? What are [17:04] the kind of pitfalls it should avoid [17:06] like analyzing complex data like [17:08] genetics and symptoms? What sort of best [17:12] practices should the model use? And then [17:14] at the end, like can you can you produce [17:17] like a a evidence-driven list that is [17:21] convincing to a person to say, "Oh, I'm [17:23] I'm not just a black box nominating a [17:26] gene and as a cause. I'm saying, oh, [17:28] here's the evidence that supports this [17:30] gene. Here's the variant that this [17:32] person has. Here's the literature that [17:34] supports that hypothesis. And then [17:37] rather than the model making a judgment [17:39] on its own, we had experts look at the [17:42] sort of list of hypothesis the model [17:44] produced and see if that made sense so [17:47] that this could then drive like follow [17:48] on testing and decide eventually if this [17:51] was a result we could return back to the [17:53] patients. [17:55] >> Got it. Um, so it sounds like the LLM [17:59] and the reasoning model are doing their [18:01] work of like bringing transparency into [18:04] this massive data, but and then they [18:05] come back and have to kind of give [18:07] something digestible [18:09] to a diagnostician like Katherine who [18:13] can actually validate what the model [18:15] surfaced. Is that right? [18:16] >> Exactly. There are many like like [18:18] Katherine said, denovo mutations and [18:19] other variants that could be [18:21] responsible. How do you make that list a [18:25] short one that a person can spend their [18:27] attention on and identify the most [18:29] promising ones? [18:30] >> Yeah. Yeah. Um, thank you. Uh, Alan [18:36] is sometimes called the hardest of hard [18:38] cases. It's like the cold cases that [18:40] some people think will never be solved [18:41] and may and sometimes the patients are [18:42] not even alive anymore. Um, why were [18:46] these cases such a meaningful test of [18:48] whether AI could help researchers find [18:51] new leads? Why why were these chosen? [18:53] >> They were difficult because we hadn't [18:56] been able to solve them manually. Um, [18:59] and there were large numbers of them [19:00] because we had 10 15 years worth of [19:03] accumulated data where in many cases we [19:06] had looked years ago with the knowledge [19:09] that was available at the time. We [19:10] hadn't made a diagnosis. Um but um it we [19:15] hadn't had a chance to go back and [19:16] review with with new information because [19:18] there are new disease gene discoveries [19:20] being made all the time. Let me just um [19:24] u explain a little bit about what Sash [19:26] was talking about in terms of the data [19:27] that go in. So three billion bases we [19:30] get a a data set that contains maybe [19:34] 500,000 rows. Each row is a different [19:36] genetic change in our gene. And then [19:39] there are maybe [19:41] 20 to 30 columns of information about [19:44] each of those genetic changes. How [19:46] common it is in the general population, [19:49] which gene it might impact, what the [19:52] nature of that impact is on the gene, is [19:53] it likely to be damaging or not, and so [19:56] on. Um, and then so that is our genetic [20:00] data. And typically a human analyst, the [20:03] first thing we do is we'll take that and [20:04] we filter it and remove all the changes [20:07] that are common in the population, but [20:09] we're still left with several thousands [20:11] up to 10,000 different genetic variants. [20:14] Um, and involving maybe [20:17] 500 to a thousand different genes. Now, [20:19] I know a lot about a small set set of [20:21] genes. So does Katherine, a different [20:23] subset. Neither one of us knows about [20:26] all 8,000 disease genes that are [20:28] characterized. But the LLM has [20:30] information about that. Um, so we [20:33] combine we send in that information with [20:35] our prompt together with metadata about [20:37] the patient which is typically numerical [20:41] ontological codes that define the [20:43] phenotype. So there's a numerical code [20:45] for weakness. There's a numerical code [20:48] for your foot drop because you have [20:50] weakness of your ankle and so on. And so [20:53] that combination of codes describes the [20:56] clinical presentation of the patient. [20:58] And what the prompt act the asks the LLM [21:01] to do is basically intersect these based [21:04] on the knowledge uh extent knowledge um [21:08] of which clinical presentations [21:11] correlate with changes in which genes [21:14] and so then its report to us nominates a [21:18] small focused number of things that are [21:20] possible. Of course this doesn't replace [21:22] the human diagnostician. It nominates [21:26] usually [21:27] anywhere from two to half a dozen [21:29] potential changes and maybe none of them [21:32] are the right answer but it very [21:34] dramatically lowers that universe of [21:37] information that our human analysts need [21:39] to think about. [21:40] >> Gives you a tractable search space. [21:42] Yeah. Interesting. Um so I think this [21:45] work started on solved cases. You were [21:47] testing uh uh the model against [21:51] cases where the answer was known but the [21:52] model didn't know them. Um what was [21:54] there a moment when it was started [21:56] giving the right answers where you [21:58] thought oh we we can really do this we [22:00] might be able to help crack unsolved [22:02] cases. What was that like? [22:03] >> Yeah I think we we had to iterate a bit [22:06] on how to prompt the model what sort of [22:09] uh mistakes the model needs to avoid and [22:12] it sort of that's why we started with [22:14] the known cases. So we started with the [22:16] known cases, gave the model those cases [22:18] as a initial question and saw how often [22:22] could it get those right and sort of [22:25] sitting together with Alan and Katherine [22:27] we said oh here are some common uh [22:31] patterns of mistakes the model is making [22:33] here I think Alan described the sort of [22:36] complex genetic data very well. uh there [22:38] are so many sort of technical challenges [22:41] in analyzing that data even for human [22:44] analysts that the models also make [22:46] similar mistakes and so you have to [22:48] prompt the model to say oh here be [22:50] careful about this specific error mode [22:53] like you need high sequencing depth to [22:55] believe that the variant is correct and [22:57] not a false positive and similar other [23:00] error modes uh and then I think once we [23:03] started iterating on that we could see [23:04] the sort of the proportion of cases that [23:07] it got correct increasing to like 80 [23:10] 90%. And then we thought, okay, now this [23:12] is worth spending a human analyst time [23:15] on rather than just giving them output [23:18] that might just be incorrect. [23:20] >> Yeah. [23:21] >> Yeah. Interesting. Thank you. Um, [23:24] Katherine, I'd love to hear from you. [23:27] So, you had these 376 cases. Uh, 03 [23:31] identified 18 diagnosises that I think [23:34] were were new or surprising. um and and [23:38] maybe it gave you leads towards those. [23:40] Anyway, what exactly did the model help [23:42] uncover? What like what were you seeing? [23:44] What was like going through that um that [23:47] process with the model where it's [23:48] surfacing something new? [23:51] So we would get output and it would um [23:53] nominate varants um and sometimes it was [23:58] very simple like you know this is a [24:01] known pathogenic variant that in a gene [24:03] that the gene phenotype link just hadn't [24:06] been discovered yet when we first got [24:08] the case. So, it was just a matter of [24:12] like we didn't have that information and [24:15] now it was plain as day and it's super [24:18] in like it was a no-brainer that like [24:21] okay this is the diagnosis um it fit the [24:24] phenotype properly. um other times and [24:27] this is outside of the 18 that is which [24:30] I also think is really cool is the model [24:33] suggested genes gene phenotype [24:36] associations based on the literature and [24:39] the type of variant that they saw and [24:42] one is um it was metase number 151 I [24:46] think or it's a very old old case we're [24:48] at like 3100 cases now so this was one [24:51] that we looked at several times can [24:53] solve it was um vitiligo transposition [24:56] of the great vessels and um pulmonary [24:58] hypertension and the model nominated S1 [25:02] PR1. I'm like what the heck is what? No. [25:05] And then but when we looked at it, it [25:07] took the actual name of the gene and did [25:10] a comprehensive search and found like a [25:12] paper from like 2000 like 26 years ago [25:17] that said um that the S1 PR1 should be [25:22] um considered seriously for the ideology [25:25] of vitiligo and it had solid solid [25:29] reasoning for this. I mean, there were [25:31] several papers that pieced together the [25:34] the pathway and S1 PR1 is completely um [25:38] uh like [25:40] involved. Um and then when I did a quick [25:43] Google search, the person who actually [25:45] cloned S1 PR1 in 1988 is in the building [25:49] next to us [25:52] at Boston Children's. So, we just [25:54] emailed him. We're like, uh, hey, what [25:56] do you think? and he's like, "Yeah, can [25:58] you come by at four o'clock?" And so now [26:01] we're like working on this more. But it [26:03] goes to show that like, you know, if I [26:05] had infinite amount of time, I would [26:07] have gotten there like or a research a [26:10] motivated posttock or researcher would [26:12] have gotten there. But just that model [26:15] is just able to do it so quickly. And [26:19] you know, we get tired looking at papers [26:21] on PubMed. You know, it's like pages and [26:23] pages of articles. go to like page four [26:26] and you think you've done a great job. [26:27] This was going all the way to page like [26:28] 30 or found this, you know. Um, and for [26:34] every one of those there's like that [26:35] this is just one example. We have a ton [26:37] more that we're looking at that um [26:40] hopefully will pan out. [26:41] >> Yeah, that's amaz. That's and the [26:44] advantage of being at Boston Children's [26:46] they're always just a couple minutes [26:47] walk away, right? Those other [26:49] researchers. Um, what did I want to ask [26:52] you the So, it feels like you've cracked [26:55] a hard problem. I don't know how much [26:57] confidence you can assign to like how [26:59] how it generalizes, but like how how do [27:02] we get this out to the world? Like what [27:04] are the other bottlenecks aside from [27:06] your time, right? But the bottlenecks so [27:08] that more families and and clinicians [27:11] can actually get into this same process [27:14] of bringing these diagnoses to the [27:16] families that need them. [27:19] So right now we're working on getting a [27:22] tool that's able to be used by everyone. [27:26] Um so the model is like it was built on [27:29] 03 like a publicly available model but [27:31] like being able to upload your data and [27:33] have the um um have it be secure and and [27:37] to run it like that's you would still [27:40] have to come to us right now but that [27:42] shouldn't be the case. you shouldn't [27:43] have to be at a tertiary medical center [27:45] in order to get like this best-in-class [27:47] like every variant evaluated properly. [27:50] So, um working with Open AI and we got a [27:54] grant at the Manton Center for um Open [27:56] AI, we're building this to make it [27:58] accessible to everyone. Um and that's [28:01] our goal. [28:02] >> Amazing. And and what what should [28:04] families and clinicians know like if if [28:06] they're out there struggling with some [28:07] unnamed disease, like what's the next [28:10] step for them? How do they advocate for [28:11] themselves to get to a solution like [28:14] that? [28:14] >> Well, that's a great thing about the [28:16] Manton Center here is patients can [28:17] self-reer. Uh I know from personal [28:21] experience with my own family members [28:23] that a lot of times um medical care [28:25] teams can be gatekeepers and block um um [28:28] patients from accessing like research or [28:31] getting involved. Um, so the M center is [28:33] great that way that like they can [28:35] directly call us up or send us email and [28:37] be like, "Hey, I think this is um this [28:40] is uh something that's worth someone [28:42] checking out." Um, but uh I think by [28:46] building this tool, we're going to make [28:48] it that uh much I mean the main thing is [28:52] we're cutting down time. So for any [28:55] clinician researcher who knows how to [28:57] analyze a genome but doesn't have the [29:00] time to go back to square one and re [29:03] look at everything when they got a [29:05] negative clinical test or something will [29:06] be able to use it and um periodically [29:09] rerun these cases. Like that's the thing [29:12] here is like the genome may not change [29:14] but the information is changing [29:15] constantly. So being able to rerun [29:18] something um often and inexpensively [29:20] will result in a lot more diagnosis. [29:24] >> Amazing. Um okay, I'd like to ask all [29:26] three of you the same question. We'll [29:27] start with Sash. What do you hope for [29:30] from AI? How how do you hope that AI may [29:33] transform medicine, you know, in in this [29:36] and and adjacent fields? [29:38] >> Yeah, I think the ability I think [29:40] exactly I think Katherine said it [29:41] perfectly is what I want to say. It's [29:43] like the ability to use the models to [29:46] reanalyze your case essentially as every [29:49] new paper is published any new piece of [29:51] knowledge comes in that becomes relevant [29:53] I think should be quite transformative [29:55] and then if the if the model is also [29:58] good enough to just say oh here is a new [30:00] finding that I'm confident in is [30:02] relevant to this case and now I'll [30:04] surface this to a researcher in by an [30:07] email in their inbox or a pop-up [30:09] notification then that will save their [30:11] attention and allow them to focus on the [30:13] really hard cases. Like we were we were [30:15] able to solve 5% of cases but in the [30:18] hard ones there are still like the 95% [30:20] that remain unsolved and maybe the [30:22] researchers can spend time on those ones [30:25] more. [30:26] >> Yeah. [30:27] >> Great. Thank you Alan. I'd love to hear [30:30] >> I mean I would say essentially what it's [30:32] capable of doing now is speeding up and [30:35] making the process of diagnosis much [30:38] more efficient. And so out of 5,000 [30:41] families enrolled with the Manton [30:42] Center, there's maybe 2,000 that still [30:45] lack a diagnosis. And it can can rapidly [30:50] point us to the diagnosis of a good [30:53] number of these based on known [30:54] information. Over time, it can uh as new [30:58] information becomes available. um [31:00] agentic approaches can probably um go [31:04] and re-review these data on a regular [31:06] basis in the background and then flag [31:08] new findings to us. Um and then what [31:11] Katherine mentioned, this case with the [31:13] vitiligo represents the next step. It's [31:15] actually synthesizing available [31:17] information to propose a hypothesis [31:19] about something new that we don't [31:20] already know. I'm not at all worried [31:22] about losing my job. We're still going [31:24] to need uh people like us to analyze [31:27] this information, to interpret it, and [31:29] to understand the implications, what the [31:31] next steps should be, but it makes our [31:33] process much faster and much more [31:35] efficient. [31:36] >> Yeah. Amazing. Katherine, um, is there [31:39] anything you'd like to add about your [31:41] hopes for how AI may bring a [31:43] transformation to medicine? [31:46] I just think that we're so on the right [31:48] track and we're just getting started and [31:51] you know I can envision a world where [31:54] like you run your an undiagnosed case is [31:58] run through and then we match it with [32:01] the diagnosis but like you know working [32:04] with a clinician getting the right [32:07] treatment or the right clinical trial or [32:10] just being able to make the world so [32:14] much more accessible. posible. Um I I [32:17] think we have a lot to do, but [32:20] what um we've been able to do already is [32:24] just so incredibly exciting and um I [32:28] feel like the reception has been [32:30] positive and it's because like how can [32:32] you argue with this? like how can you be [32:35] upset with um diagnosing kids and making [32:41] um things just better for families. [32:45] >> Yeah. Yeah. I think there's huge [32:47] potential for AI to bring relief to [32:50] their suffering. Um okay, I need to uh [32:54] interject with a brief comment. So, [32:56] OpenAI supported this research with [32:58] Seash and the Maten Center. But to be [33:00] clear, now that we have the open [33:02] Foundation, the grant you got is from [33:04] the foundation, right? [33:05] >> Okay. So, those were two separate [33:06] things, but just we we're we're in a new [33:09] configuration now and and it's important [33:11] to um make make clear what each of us is [33:13] doing here. Um the [33:17] I think we're going to open this up to [33:18] questions from the community. Um Oh, [33:21] cool. Okay. So, we have Elizabeth [33:23] Stewart, science writer and editor at [33:25] Innovate Bio. She says, "If whole genome [33:28] sequencing and analysis was an analysis [33:30] done early on in a diagnostic journey, [33:33] how much faster would rare diseases be [33:35] diagnosed?" So really like how should [33:37] people do be doing whole genome analyses [33:40] early on and what will that solve for [33:41] them? [33:42] >> Alan, I think you should take this one. [33:43] >> Okay. Well, I mean, yes, yes, and yes. [33:48] Um, there's been a lot of concern in the [33:50] medical community because insurance [33:52] companies have been slow to reimburse [33:54] the cost of this testing. the testing is [33:56] below $1,000, which is a fraction of [33:58] what it costs for a typical MRI, which [34:01] they don't object to spending money on. [34:03] Um, all of us, we've been involved in a [34:06] number of initiatives. Some of it [34:07] involving newborn genome sequencing. Um, [34:11] some of it involving sequencing of [34:13] everybody who's born and sick enough to [34:16] go to a neon neonatal ICU. um and it's [34:20] uh unit and it's possible now to make [34:23] this diagnosis literally within days to [34:25] weeks. In fact, some of our colleagues [34:27] here recently received [34:30] um [34:31] a um what's the name of the pri uh prize [34:35] for um [34:38] for the fastest genome sequencing? Um, [34:41] I've got it. So, uh, anyway, um, [34:45] a prize for sequencing and diagnosing a [34:47] patient within 24 hours. Now, we don't [34:49] do that routinely, but I think everybody [34:52] who has an undiagnosed or uncertain [34:54] aspect to their medical condition should [34:57] have their genome sequenced. Um, they're [34:59] going to be a good number where we learn [35:01] something important. They're going to be [35:02] some where we don't, but uh, the [35:05] information remains available and may [35:06] inform their future medical care. [35:08] >> Yeah. Interesting. Thank you. [35:10] >> Sorry. Give us Book of World Records. [35:13] >> So, it's getting much faster and it it [35:15] can be very fast and it's getting much [35:16] much cheaper like a few hundred bucks [35:18] whereas it used to be thousands of [35:19] dollars. So, so it's more and more [35:21] available. Um, okay. So, next question. [35:23] Jason Duca. He's at Crossing Point IT [35:26] Solutions. Do you see potential for AI [35:28] to connect unrelated aspects of a [35:30] patient's health and reveal diagnostic [35:33] patterns that most clinicians might [35:35] otherwise miss? [35:37] I totally think that's possible. Um I I [35:41] think there's some exciting research [35:42] going on right now which is like looking [35:45] well I just heard about the other day I [35:46] was looking at changes in an eye and how [35:49] that could indicate other organ [35:50] problems. Um and you can see how how AI [35:54] would be really well suited to doing [35:56] that kind of work or assisting with that [35:58] kind of work. Um we actually we knew [36:01] that the model when we were working [36:03] together um developing this tool was [36:05] ready to go. When we had given it um uh [36:08] 20 solved cases and it got 19 correct [36:11] but when we looked at the one it hadn't [36:13] gotten correct we were wrong and that [36:14] there were two things going on at the [36:17] individual and it had caught both of [36:19] them when we had only been really [36:21] looking at one variant. And I think [36:23] that's going to happen a lot more where [36:25] we're able to tease out what's due to [36:28] like what what phenotypes are due to [36:32] which gene variance. Um where right now [36:34] perhaps we're not as exact as we could [36:38] be because you know you find one thing [36:40] and you're like oh yes this must be it [36:42] but like and it could be an expansion a [36:44] phenotypic expansion but maybe there's [36:46] actually a second thing going on. [36:47] >> Yeah. Interesting. So, Sat Lana Romanova [36:50] is asking a question that's a segue [36:52] here. She's asking, "How much of the [36:54] gain do you think came from re in the [36:57] diagnosis that got surfaced came from [36:59] reasoning versus data integration [37:02] just exposure to the right data?" [37:04] >> That's a great question. But [37:05] >> yeah, I don't remember if we tested any [37:09] other models on it. We tested sort of [37:10] the state-of-the-art model at the time. [37:14] I think [37:15] >> data integration was definitely [37:17] important. And I think we spent a bunch [37:19] of effort on giving the model somewhat [37:22] curated data not not the three billion [37:25] baseware set but a reduced like a [37:27] thousand 10,000 variant set and the [37:30] symptom list was also curated. I think [37:33] partly the vision the future is that you [37:37] don't need to do this curation. You give [37:38] it both the person's full medical record [37:41] with all their uh lab tests and imaging [37:45] and anything else they have along with [37:47] their full genome file and the model [37:49] should be good enough to do all of this [37:52] already. We haven't benchmarked it but [37:54] maybe as the models are getting better [37:56] this is something that could happen. [37:58] >> Okay. Interesting. Um well folks [38:03] there's some music going on. [38:05] >> How's my phone? Sorry. [38:07] >> All right. Believe it or not, people [38:08] still make phone calls occasionally. [38:11] >> I I thought that was how they were [38:12] telling us to wrap this up. So, thank [38:15] you. This has been amazing. Um, I've [38:18] learned so much today. I'll bet our [38:19] community has. Thank you to the [38:21] community, too, for such excellent [38:23] questions and for giving us your time. [38:25] Um, and I just want to say thank you so [38:27] deeply for your work. Like, there are [38:29] many families, including people in my [38:30] extended family, who whose health will [38:32] depend on this research getting better. [38:35] And I'm I'm so excited for uh these [38:38] advances and and what they'll do for [38:40] humanity. Um so just as a note, this [38:43] study used opening eye 03 deep research, [38:45] but we've got newer general purpose [38:47] models that search and synthesize [38:49] through scientific material that are [38:52] designed for deeper life sciences work [38:54] like GPT Rosland. Um so those advances I [38:57] think that that we're reporting today, [38:59] we'll continue to see advances over the [39:01] coming years. Um so thank you for [39:03] working with us. Uh shout out to the [39:05] grant from the OpenAI Foundation uh [39:07] which will support your your further [39:09] work and I hope to see everyone in the [39:13] community back here for our next [39:15] exciting events. We're going to keep on [39:17] hammering on AI and healthcare and how [39:20] it affects rare diseases and rare [39:22] cancers over the course of a year. So [39:24] thanks again everybody. We'll see you [39:25] soon.