Revumenib

Co-Scientist：基于 Gemini 2.0 的多 Agent 结构化科学思维引擎 本文介绍了 Google DeepMind 联合 Google Research、斯坦福大学等机构开发的Co-Scientist，这是全球首款基于Gemini 大模型构建的多 Agent 结构化科学思维引擎，核心通过异步多 Agent 协作架构、Elo 锦标赛式假设进化机制和测试时计算扩展范式，实现科学假设的生成、批判、迭代与优化。研究在急性髓系白血病（AML）药物重定位、肝纤维化新型治疗靶点发现、抗菌耐药性（AMR）基因转移机制解析三个复杂度递增的生物医药领域完成端到端湿实验验证，发现了具有临床潜力的单药与协同联合疗法，鉴定出有效抗纤维化表观遗传靶点，且仅用 2 天就独立复现了当时未发表的突破性 AMR 机制。系统评估显示其假设质量全面超越OpenAI o1/o3-mini-high、DeepSeek R1等前沿推理模型及人类领域专家，标志着 AI 从科研辅助工具向平等科学合作者的范式转变。 一、研究背景与核心问题 1.现代科学的核心挑战：学科深度专业化与跨学科突破需求的矛盾日益突出；每年新增数百万篇科学文献，人类研究者难以全面掌握领域进展，导致知识整合效率低下。 2.现有 AI 科研工具的不足：传统工具（如文献检索、数据可视化、Deep Research）仅能完成信息整合与辅助分析，无法生成原创性、可验证的科学假设，更不能指导实验设计与临床转化。 3.技术突破契机：大语言模型的长上下文推理能力、多 Agent 系统的协作能力，以及 "测试时计算扩展" 范式（通过增加推理时间提升输出质量）的发展，为构建真正的 AI 科学合作者提供了可能。 二、Co-Scientist 系统架构与工作原理 Co-Scientist是一款模拟科学方法核心流程的多 Agent AI 系统，基于Gemini 2.0构建，采用 "科学家在环" 的协同设计。 1.核心工作流程科学家输入自然语言研究目标→Supervisor Agent 解析并生成研究计划→多 Agent 协同生成、评审、排序假设→迭代进化优化→输出排名靠前的假设与完整研究概览。 2.六大专业 Agent 的分工与协作 Agent 名称 核心职能 关键技术 Generation 生成初始假设与研究方向 文献搜索、模拟科学辩论、迭代假设识别 Reflection 批判性评审假设 初始筛选、全文献核查、深度验证、实验模拟 Ranking 假设优先级排序 Elo 评分系统、多轮科学辩论、位置偏差消除 Proximity 假设空间管理 语义相似度计算、聚类去重、多样性保障 Evolution 假设迭代优化 缺陷修复、证据增强、组合创新、简化验证 Meta-review 全局反馈与总结 提炼评审共性问题、优化后续推理、生成研究报告 3.核心创新点 ￮测试时计算扩展：假设质量随推理时间持续提升，无明显饱和迹象 ￮锦标赛式进化：通过两两辩论竞争筛选优质假设，模拟科学共同体的评审机制 ￮无微调自改进：通过 Meta-review 的全局反馈优化 Agent 提示词，无需重新训练模型 三、系统性能全面评估 1. 消融分析：验证各组件的不可替代性 组件与配置 核心贡献 量化验证结果 Reflection Agent + 外部搜索 防止幻觉，准确评估新颖性 非新颖假设评分从 6.14/10 降至 2.38/10；正确性评分从 7.4 升至 8.46/10 Ranking Agent + 科学辩论 提升排序准确性 显著减少位置偏差，高价值假设排名提升 Evolution Agent 迭代优化 提升假设整体质量 GPQA 基准精度从 70.9% 升至 75.4%；假设质量评分从 4.7 升至 5.6 Meta-review Agent 全局反馈 持续改进推理能力 正确性预测 AUC 从 0.521 升至 0.597 2. 测试时计算扩展效果 •测试集：203 个跨学科研究目标（以生物医药为主，涵盖数学、物理） •实验设计：将生成过程分为 10 个等时间桶，计算每个桶的最高 Elo 评分和前 10 名平均 Elo 评分 •核心结论：两个指标均随时间呈持续线性上升趋势，表明增加计算资源可稳定提升假设质量 3. 与前沿模型及人类专家的对比 •测试集：7 位生物医学专家精心设计的15 个高难度开放科学问题 •对比基线：Gemini 2.0 Pro Experimental、Gemini 2.0 Flash Thinking Experimental、OpenAI o1、OpenAI o3-mini-high、DeepSeek R1、人类专家 "最佳猜测" •结果：Co-Scientist 通过迭代推理，最终 Elo 评分全面超越所有对比模型和人类专家；且能进一步优化人类专家提出的假设，使其评分超过自身独立生成的结果 4. 人类专家盲评 •测试集：11 个研究问题 •评估维度：新颖性（5 分制）、影响力（5 分制）、整体偏好排名（1 = 最优） •结果： ￮Co-Scientist 平均偏好排名：2.36（所有模型中最低） ￮新颖性评分：3.64/5 ￮影响力评分：3.09/5 ￮所有指标均显著优于其他对比模型 四、真实世界端到端实验验证 研究选择了三个复杂度递增的生物医药领域，所有实验均由领域专家全程参与并独立验证： 应用领域 研究目标 核心方法 关键实验结果 临床 / 科学价值 AML（急性髓系白血病） 药物重定位 从 2300 种已批准药物中筛选对 AML 有效的单药与联合疗法 专家筛选 + 体外细胞活力实验（4 种 AML 细胞系 + 1 种正常细胞系） 1. 5 种候选药物中 3 种有效：Binimetinib（IC50=2nM）、Pacritinib、Cerivastatin2. 3 种全新单药中 KIRA6 效果最佳：对 KG-1a 细胞 IC50=10nM，与 TK6 正常细胞有 18 倍治疗窗口3. 7 种联合疗法中，5 种在 MOLM-13 细胞中呈协同作用 为 AML 提供了多个具有临床潜力的候选药物，特别是针对干细胞样 AML 细胞的新靶点 肝纤维化新靶点发现 鉴定参与肌成纤维细胞形成的新型表观遗传靶点 人类肝脏类器官模型 + 活细胞成像 1. 提出 3 个排名最高的表观遗传靶点2. 2 种靶向药物显示显著抗纤维化活性且无细胞毒性3. 其中 Vorinostat 为 FDA 已批准的皮肤 T 细胞淋巴瘤药物 实现了跨疾病领域的知识迁移，为肝纤维化治疗提供了快速转化的候选方案 AMR（抗菌药物耐药性） 机制解析 解释 cf-PICIs 为何能跨多种细菌物种传播耐药基因 仅提供基础背景信息，让 Co-Scientist 独立提出机制 仅用 2 天就提出核心机制：cf-PICIs 通过结合多种噬菌体的尾部蛋白，形成嵌合感染颗粒，从而扩展宿主范围 与同期未发表的 Cell 论文结论完全一致，证明 AI 可独立做出与人类顶尖团队相当的基础科学发现 五、局限性与未来发展方向 1.当前关键局限性 ￮知识边界受限：仅能访问开放获取文献，无法获取付费内容和未发表的阴性实验结果 ￮事实性风险：仍存在大语言模型的幻觉问题，可能传播来源文献中的错误结论 ￮通用性不足：目前仅在生物医药领域验证，其他学科的表现有待测试 ￮伦理与生态风险：可能导致研究方向同质化，加剧科学可重复性危机 2.未来发展路线图 ￮短期（1-2 年）：增强文献溯源能力、扩展数据库访问、优化事实核查机制 ￮中期（3-5 年）：整合多模态数据推理、引入实验反馈的强化学习、支持更复杂的实验设计 ￮长期（5 年以上）：对接实验室自动化平台，实现 "假设生成 - 实验执行 - 结果分析 - 迭代优化" 的全闭环自主科学发现 六、研究核心贡献 1.技术创新：首次提出并实现了基于测试时计算扩展的多 Agent 科学推理框架，证明了 AI 系统可以生成原创、可验证的科学假设。 2.范式突破：建立了 "科学家 - AI" 协同的新型科研范式，AI 不再是简单的辅助工具，而是能够平等参与科学发现过程的合作者。 3.实际价值：在三个关键生物医药领域取得了实质性进展，为加速药物研发、应对抗菌耐药性等全球健康挑战提供了新的技术手段。 关键问题与答案 问题 1（技术原理侧重）：Co-Scientist 的 "测试时计算扩展" 范式与传统大模型的 "参数规模扩展" 有何本质区别？为什么说前者更适合科学发现任务？ 答案：两者的核心区别在于提升能力的维度不同： •参数规模扩展：通过增加模型参数量和训练数据量来提升通用能力，是 "训练时一次性投入"，模型能力在部署后固定不变；其优势是通用能力强，但在需要深度推理和原创性的任务中容易遇到瓶颈。 •测试时计算扩展：模型参数固定，通过增加推理阶段的计算资源（更多的 Agent 调用、更多轮的迭代、更复杂的辩论）来提升输出质量，是 "测试时按需投入"；其优势是可以针对具体问题进行深度探索，输出质量随计算时间持续提升。 对于科学发现任务，测试时计算扩展更具优势： 1.科学问题往往需要深度、迭代式的推理，而非一次性的快速回答 2.可以根据问题的复杂度灵活分配计算资源，平衡效率与质量 3.无需重新训练模型即可利用最新的基础模型进展，具有更好的可扩展性 4.能够模拟人类科学家 "反复思考、不断修正" 的研究过程，更符合科学发现的规律 问题 2（实际应用侧重）：Co-Scientist 在 AML 药物重定位中发现的 KIRA6 为什么具有重要的临床转化价值？ 答案：KIRA6 作为一种新型 IRE1α 抑制剂，其临床转化价值主要体现在三个方面： 1.高度的细胞选择性：对原始干细胞样 AML 细胞（KG-1a）的 IC50 低至 10nM，而对正常淋巴母细胞 TK6 的 IC50 为 180nM，具有18 倍的治疗窗口，这意味着在有效杀伤癌细胞的同时，对正常细胞的毒性较低，安全性更好。 2.针对难治性 AML 亚型：KG-1a 细胞代表了 AML 中最具侵袭性、对传统化疗最耐药的干细胞群体，目前临床上缺乏针对这一群体的有效治疗手段；KIRA6 的发现为靶向 AML 干细胞提供了新的方向。 3.明确的作用机制：Co-Scientist 不仅预测了 KIRA6 的疗效，还阐明了其通过抑制 IRE1α-XBP1 内质网应激通路发挥作用的分子机制，为后续的临床前研究和临床试验设计提供了坚实的理论基础。 此外，KIRA6 是一种已经经过临床前安全性评估的化合物，这将大大缩短其从实验室到临床的转化时间。 问题 3（行业影响侧重）：Co-Scientist 的出现会对未来的科学研究模式产生哪些深远影响？ 答案：Co-Scientist 代表了 AI 赋能科学发现的一个重要里程碑，将从三个层面重塑未来的科研模式： 1.科研效率层面：将科学家从文献检索、数据整理、初步假设筛选等重复性劳动中解放出来，使其能够专注于更具创造性的科学问题设计和实验验证；预计可将单个研究项目的周期缩短数倍。 2.科研协作层面：形成 "人类科学家提出问题、AI 生成假设、人类验证假设" 的新型协作模式；AI 将成为每个科研团队的 "虚拟博士后"，能够 24 小时不间断地进行知识整合和推理。 3.科研生态层面： ￮降低科研门槛：资源有限的研究机构也能通过 AI 获得与顶尖机构相当的知识整合能力，促进科学民主化 ￮催生新的研究方向：AI 能够发现人类难以察觉的跨学科联系，开辟全新的研究领域 ￮推动科研评价体系变革：需要建立针对 AI 生成科研成果的评审标准和验证机制，确保科学研究的严谨性和可重复性 论文图表内容 一、表格翻译 表 1 | 用于 Co-Scientist 端到端验证的三项生物医药实际应用 应用方向 药物重定位 新型治疗靶点发现 解析基因转移进化机制 核心挑战 复杂检索 鉴定新型靶点 理解复杂系统 复杂度 中等 高 极高 研究规模 中等，数据受限 中等，实验受限 大规模，数据与计算受限 未知因素 有限 广泛 庞大且动态 补充数据表 1 | 急性髓系白血病（AML）细胞系（MOLM-13、KG-1a）联合用药实验的药物半抑制浓度（IC50） 药物名称 IC50（微摩尔 /μM）- MOLM-13 细胞 IC50（微摩尔 /μM）- KG-1a 细胞 MSA-2 无响应 无响应 柳氮磺吡啶 无响应 无响应 CB-839 16.98 8.79 JNJ-64619178（奥那美司他） 0.0046 0.024 匹诺塞他 无响应 无响应 塞利尼索 0.039 0.055 JQ1 0.011 0.12 维奈克拉 0.009 0.006 奥拉帕利 2.12 7.1 SNDX 0.010 无响应 帕博西尼 0.026 0.07 补充数据表 2 | 药物联合协同实验总结 药物组合 MOLM-13 细胞 KG-1a 细胞 JNJ-64619178 + 塞利尼索 协同作用 混合响应 帕博西尼 + 塞利尼索 协同作用 协同作用 JNJ-64619178 + SNDX-5613 协同作用 混合响应 CB-839 + 柳氮磺吡啶 无协同作用 无协同作用 维奈克拉 + 匹诺美司他 混合响应 协同作用 JNJ-64619178 + 塞利尼索 + SNDX-5613 协同作用 协同作用 JQ1 + 奥拉帕利 + MSA-2 协同作用 无协同作用 二、图表图例翻译 图 1 | Co-Scientist 设计、多智能体架构及实验验证总结 (a) 系统概述：展示 Co-Scientist 结构化科学推理引擎的各组件、多智能体系统，以及与科学家的交互模式。科学家输入自然语言研究目标后，Co-Scientist 生成新型研究假设；系统基于 Gemini 调用生成、反思、排序、进化、相似度分析、元评审等专用智能体，在锦标赛框架内持续生成、辩论、优化假设，通过锦标赛反馈实现迭代升级，形成自优化闭环，以产出可解决复杂科学问题的高质量新假设。Co-Scientist 借助网页检索、专用 AI 模型等工具提升假设的依据性与可靠性；科学家可通过自然语言与系统对话，明确研究目标、设定约束、提供反馈、引导研究方向。(b) 多智能体架构：监督智能体解析用户自然语言研究目标，在异步任务队列中为专用工作智能体动态分配算力资源；红色框为 Co-Scientist 专用智能体（各有独立逻辑与功能），蓝色框为科学家在环的输入与反馈，深灰色箭头为系统内信息流，红色箭头为专用智能体间的信息反馈闭环。(c) 三大生物医药场景端到端验证：Co-Scientist 提出急性髓系白血病（AML）新型重定位药物（上）、鉴定肝纤维化新型表观遗传靶点（中）、独立复现一项未发表的抗菌耐药（AMR）相关细菌基因转移新机制（下）；所有 AI 生成的假设均通过独立体外实验验证。 图 2 | 测试时算力扩展提升 Co-Scientist 科学推理与假设质量 (a) 基于 Elo 自动评分的测试时算力扩展效果：将 203 个跨学科科研目标生成的假设分为 10 个连续时间桶，最优假设 Elo 评分（左）、前 10 假设平均 Elo 评分（右）均持续上升，证明系统可自我优化，生成更高质量的科学假设。(b) 与主流大模型及人类专家的自动评分对比：在 15 个专家定制的高难度生物医药目标中，Co-Scientist 通过测试时算力扩展，超越人类领域专家与前沿大模型（OpenAI o1、o3-mini-high、DeepSeek-R1 等），生成更高质量的新假设。(c) 人类专家盲评结果：11 个生物医药开放问题中，专家对各模型假设的新颖性与影响力评分（左，5 分制）、整体偏好排名（右，数值越低越优）；Co-Scientist 在新颖性、影响力得分及专家偏好度上均为最高。（注：a、b 为均值 + 95% 置信区间；c 为均值 ± 标准误，样本量 n=203/15/11） 图 3 | Co-Scientist 生成的急性髓系白血病（AML）单药重定位候选药物体外生物学验证 Co-Scientist 筛选出兼具临床前依据与全新治疗潜力的 AML 活性化合物。(a-c) MOLM-13 细胞经比美替尼、帕克替尼、西立伐他汀处理的剂量反应曲线：AI 推荐的已有临床依据药物，展现强效抗白血病活性。(d-h) 全新候选药物 KIRA6（IRE1α 抑制剂）在 4 种 AML 细胞系及正常 TK6 细胞中的剂量反应曲线：KIRA6 对 KG-1a AML 细胞具有高度选择性细胞毒性，与正常细胞相比存在 18 倍治疗窗口，证明 AI 的检索、推理与活性化合物筛选能力。（X 轴：药物浓度 /μM，对数标尺；Y 轴：生长抑制率；数据为 n=3 次独立实验均值 ± 标准差） 图 4 | Co-Scientist 预测的急性髓系白血病（AML）协同多药组合验证 Co-Scientist 在高维组合空间中高效筛选有效联合用药方案，在 MOLM-13、KG-1a 细胞中完成验证。(a-b) 双联组合 JNJ-64619178 + 塞利尼索的定量协同分析：基于 Chou-Talalay 法绘制联合指数（CI）- 受影响分数（Fa）曲线，红色虚线为叠加效应（CI=1.0），下方为协同区（CI<1）、上方为拮抗区（CI>1），确认 AI 推荐组合存在强协同作用。(c-d) 三联组合 JQ1 + 奥拉帕利 + MSA2 的超额效应热图：基于最高单药（HSA）、Bliss 独立模型量化不同浓度组合的协同效应，红色为协同、蓝色为拮抗；证明 AI 可无需大规模湿实验筛选，直接发现高活性复杂联合方案。（所有实验 n=3 次生物学重复） 补充数据图表图例 补充数据图 1 | 基于 Elo 自动评分的 Co-Scientist 辅助专家优化效果 通过自优化机制，Co-Scientist 可逐步迭代优化人类专家的 “最优猜想” 假设（15 个定制科研目标）；Elo 评分为自动评分，非独立客观真值；误差线为均值标准误（SEM），数据为均值 ±SEM。 补充数据图 2 | Co-Scientist 与基线模型的大模型偏好排名自动评估 15 个专家定制目标中，4 个大模型（OpenAI o3-mini、o1-preview、Gemini 2.0 Pro/Flash）作为评审的平均偏好排名，数值越低排名越优；箱线图中线为中位数，箱体为上下四分位，须线为 1.5 倍四分位距内极值，菱形为均值。 补充数据图 3 | 重定位药物比美替尼在其他细胞系中的剂量反应曲线 比美替尼对 KG-1a、HL-60、TK6 细胞均有生长抑制活性；其作用靶点为 RAS-RAF-MEK-ERK 通路，该通路在正常 TK6 细胞中本非高表达 / 必需，因此 TK6 细胞的 IC50 显著高于 AML 细胞，与作用机制一致；数据为 n=3 次独立实验均值 ± 标准差。 补充数据图 4 | 对 MOLM-13 细胞无明显效果的重定位药物剂量反应曲线 专家筛选药物中，普伐他汀、富马酸二甲酯（DMF）在测试浓度内对 MOLM-13 细胞无明显效果；全新候选药物中，来氟米特、纳武拉坦在测试浓度内对 MOLM-13 细胞无明显效果；数据为 n=3 次独立实验均值 ± 标准差。 补充数据图 5 | Co-Scientist 推荐的 AML 细胞双联药物组合全面协同分析 延续图 4 (a-b) 的验证逻辑，补充其他 AI 预测双联组合在 MOLM-13、KG-1a 细胞中的药理互作；基于 Chou-Talalay 法定量协同效应，红色虚线为叠加效应（CI=1.0），下方为协同、上方为拮抗；结果显示帕博西尼 + 塞利尼索等组合在不同遗传背景下均稳定协同，其余组合呈细胞背景依赖性响应。 补充数据图 6 | Co-Scientist 预测的三联药物组合超额效应热图 延续图 4 (c-d) 的验证逻辑，补充三联组合 JNJ-64619178+SNDX-5613 + 塞利尼索在 AML 细胞中的互作矩阵；基于 HSA、Bliss 模型量化协同 / 拮抗，红色为协同、白色为叠加、蓝色为拮抗；热图展示第三种药物浓度范围内的最大正效应。 三个案例简要说明 AML 药物重定位AML：急性髓系白血病 药物重定位：把已上市 / 临床成熟老药，跳出原有适应症，发掘其抗 AML 新疗效，用于白血病治疗。核心要点核心优势 省去全新药物从头研发周期、成本、高风险，安全性已知，快速推进 AML 临床应用。作用逻辑 老药通过靶向白血病干细胞、抑制增殖、诱导凋亡、逆转耐药、调控免疫等机制，干预 AML 发病通路。常用重定位方向 抗炎药、降脂药、降糖药、抗生素、抗寄生虫药等跨界用于 AML 联合化疗 / 靶向药，减毒增效、克服 AML 化疗耐药应用价值 适配难治复发 AML、老年不耐受强化疗患者，填补临床缺药短板，是低成本抗白血病新药研发捷径。现存局限 多数停留在细胞 / 动物实验，临床证据不足；剂量、联用方案、适应症精准度需进一步验证。 肝纤维化新靶点发现肝纤维化新靶点发现 简要解析 核心本质肝纤维化是肝脏慢性损伤后，肝星状细胞过度活化、大量胶原异常沉积形成疤痕组织，持续进展会走向肝硬化、肝癌。传统治疗多抗炎、保肝，缺乏精准阻断纤维化通路的手段。 新靶点研究逻辑从细胞信号通路、表观遗传、代谢重编程、炎症互作、肠道 - 肝轴等维度，找到调控肝星状细胞活化 / 凋亡、抑制胶原合成、促进纤维降解的关键分子、受体、基因或非编码 RNA，即为新型靶点。 主流热门新靶点方向 信号通路靶点：阻断 TGF-β/Smad、Wnt/β-catenin 等促纤维化核心通路关键节点 细胞因子靶点：靶向 PDGF、CTGF 等促纤维化因子 代谢靶点：调控肝细胞 / 星状细胞脂代谢、糖代谢紊乱 免疫靶点：调控肝内巨噬细胞、炎症微环境，减轻促纤维化免疫反应 表观靶点：调控甲基化、组蛋白修饰，逆转纤维化相关基因异常表达临床价值 突破传统对症治疗，实现精准抗纤维化 有望开发小分子抑制剂、单抗、核酸药物等新型抗肝纤维化药 可用于早中期纤维化逆转，延缓肝病终末期进程 为病毒性肝炎、脂肪肝、自身免疫性肝病等所致纤维化提供通用治疗新思路现存短板 多数靶点仍处于基础实验 / 动物模型阶段，肝脏靶向性、全身副作用、临床转化难度大，距离成熟上市药物还有较长研发周期。 AMR（ 抗菌药物耐药性 ） 机制解析AMR 抗菌药物耐药性 核心机制（极简解析） 灭活破坏：细菌产生灭活酶，直接水解 / 修饰药物，让药失效（如 β- 内酰胺酶破青霉素）。 靶点改变：细菌修改药物结合靶点，药物无法结合抑菌，结构变异避药效。 外排泵亢进：细菌主动把进入体内的药物快速泵出，胞内药浓度不足。 通透性下降：细胞膜孔道减少 / 关闭，阻止药物进入菌体内部。 代谢旁路替代：绕开药物抑制的代谢通路，换途径正常生长繁殖。 基因水平转移：通过质粒、转座子、噬菌体跨菌传耐药基因，快速扩散耐药。 生物智能：在生物先进产业场景中构建“状态感知-实时认知-自主决策-精准执行-学习提升”的生物科学智能（Biology_and_AI）；实现生物产业转型升级、DT驱动业务、价值创新创造的产业互联生态链。 生物产业+物理AI=生物智能 产业智能官：Science_and_AI 加入知识星球“生物智能研究院”：生物产业OT技术（自动化+机器人+工艺+精益）和新一代IT技术（云计算+物联网+区块链+大数据+人工智能）深度融合，在场景中构建“状态感知-实时认知-自主决策-精准执行-学习提升”的生物科学智能；实现生物产业转型升级、DT驱动业务、价值创新创造的产业互联生态链。 生物科学智能作为第四次工业革命的核心驱动力，将进一步释放历次科技革命和产业变革积蓄的巨大能量，并创造新的强大引擎；重构设计、生产、供应链和服务等经济活动各环节，形成从宏观到微观各领域的智能化新需求，催生新技术、新产品、新产业、新业态和新模式；引发经济结构重大变革，深刻改变人类生产生活方式和思维模式，实现社会生产力的整体跃升。 生物产业智能化技术分支用来的今天，从业者必须了解如何将生物科学智能全面渗入整个公司、产品、业务等商业场景中，利用生物科学智能形成数字化、网络化和智能化力量，实现行业的重新布局、企业的重新构建和焕然新生。 版权声明：产业智能官（ID：Science_and_AI）发表的文章，除非确实无法确认，我们都会注明作者和来源，涉权请联系协商解决，联系、投稿邮箱：wolongzy@qq.com

– Oral presentation will highlight favorable outcomes observed among 21 adults and children with KMT2Ar, NPM1m, or NUP98r acute leukemia who received revumenib post-transplant – – Data underscoring unique aspects of revumenib’s PK profile, including the ability to administer it with gastric acid reducing agents and low-fat meals, will be presented – – Design of ongoing pivotal trials of revumenib plus low or high-intensity chemotherapy in newly diagnosed NPM1m or KMT2Ar acute leukemia will be highlighted – NEW YORK , May 21, 2026 (GLOBE NEWSWIRE) -- Syndax Pharmaceuticals (Nasdaq: SNDX), a commercial-stage biopharmaceutical company advancing innovative cancer therapies, today announced the acceptance of four Revuforj® (revumenib) abstracts for presentation at the American Society of Clinical Oncology (ASCO) Annual Meeting, taking place May 29 – June 2, 2026 , in Chicago . “Our strong presence at ASCO highlights our scientific leadership in menin inhibition and our deep commitment to advancing cancer care. We and our collaborators will present new evidence that moves the field forward, including an oral presentation of data showing favorable outcomes among patients who received revumenib in the post-transplant setting. This presentation will provide additional evidence in an area of high physician interest and inform further clinical research,” said Nick Botwood , MBBS, Head of Research & Development and Chief Medical Officer at Syndax. Dr. Botwood continued, “Building on the strong body of efficacy data that distinguishes revumenib, we also look forward to presenting PK data which highlight other important aspects of its profile, including the ability to administer revumenib with commonly prescribed gastric acid reducing agents without the risk of reducing efficacy.” Key revumenib presentations at ASCO 2026: An oral presentation of safety and efficacy data from 21 adults and children with KMT2Ar, NPM1m, or NUP98r acute leukemia who received revumenib as maintenance following hematopoietic stem cell transplantation (HSCT). The presentation will include the observed overall survival and relapse rate, along with a comparison to a historical cohort of patients with the same genetic subtypes of acute leukemia treated prior to the advent of revumenib. A poster presentation characterizing the pharmacokinetics (PK) of revumenib, with an emphasis on differentiating aspects of its PK profile, including the ability to 1) administer revumenib with gastric acid reducing agents without the risk of reduced efficacy, 2) ensure optimal exposure in the presence of strong CYP3A4 inhibitors using a clear revumenib dose adjustment strategy, and 3) administer revumenib with low-fat meals. The accepted abstracts listed below are now available online on the ASCO conference website. Copies of the oral and poster presentations will be made available in the ‘Publications & Meeting Presentations’ section of the Syndax website after the data are presented. Full list of revumenib abstracts accepted for presentation at ASCO 2026 (all times in CDT): About Revuforj® (revumenib) Revuforj (revumenib) is an oral, first-in-class menin inhibitor that is FDA approved for the treatment of relapsed or refractory (R/R) acute leukemia with a lysine methyltransferase 2A gene (KMT2A) translocation as determined by an FDA-authorized test in adult and pediatric patients one year and older. Revuforj is also indicated for the treatment of R/R acute myeloid leukemia (AML) with a susceptible nucleophosmin 1 (NPM1) mutation in adult and pediatric patients one year and older who have no satisfactory alternative treatment options. Multiple trials of revumenib are ongoing or planned across the treatment landscape, including in combination with standard of care therapies in newly diagnosed patients with NPM1m or KMT2Ar AML. Revuforj (revumenib) IMPORTANT SAFETY INFORMATION WARNING: DIFFERENTIATION SYNDROME, QTc PROLONGATION, and TORSADES DE POINTES Differentiation syndrome, which can be fatal, has occurred with Revuforj. Signs and symptoms may include fever, dyspnea, hypoxia, pulmonary infiltrates, pleural or pericardial effusions, rapid weight gain or peripheral edema, hypotension, and renal dysfunction. If differentiation syndrome is suspected, immediately initiate corticosteroid therapy and hemodynamic monitoring until symptom resolution. QTc prolongation and Torsades de Pointes have occurred in patients receiving Revuforj. Correct hypokalemia and hypomagnesemia prior to and during treatment. Do not initiate Revuforj in patients with QTcF > 450 msec. If QTc interval prolongation occurs, interrupt, reduce, or permanently discontinue Revuforj. WARNINGS AND PRECAUTIONS Differentiation Syndrome: Revuforj can cause fatal or life-threatening differentiation syndrome (DS). Symptoms of DS, including those seen in patients treated with Revuforj, include fever, dyspnea, hypoxia, peripheral edema, pleuropericardial effusion, acute renal failure, rash, and/or hypotension. In clinical trials, DS occurred in 60 (25%) of 241 patients treated with Revuforj at the recommended dosage for relapsed or refractory acute leukemia. Among those with a KMT2A translocation, DS occurred in 33% of patients with acute myeloid leukemia (AML), 33% of patients with mixed-phenotype acute leukemia (MPAL), and 9% of patients with acute lymphoblastic leukemia (ALL); DS occurred in 18% of patients with NPM1m AML. DS was Grade 3 or 4 in 12% of patients and fatal in 2 patients. The median time to initial onset was 9 days (range 3-41 days). Some patients experienced more than 1 DS event. Treatment interruption was required for 7% of patients, and treatment was withdrawn for 1%. Reduce the white blood cell count to less than 25 Gi/L prior to starting Revuforj. If DS is suspected, immediately initiate treatment with systemic corticosteroids (e.g., dexamethasone 10 mg IV every 12 hours in adults or dexamethasone 0.25 mg/kg/dose IV every 12 hours in pediatric patients weighing less than 40 kg) for a minimum of 3 days and until resolution of signs and symptoms. Institute supportive measures and hemodynamic monitoring until improvement. Interrupt Revuforj if severe signs and/or symptoms persist for more than 48 hours after initiation of systemic corticosteroids, or earlier if life-threatening symptoms occur such as pulmonary symptoms requiring ventilator support. Restart steroids promptly if DS recurs after tapering corticosteroids. QTc Interval Prolongation and Torsades de Pointes: Revuforj can cause QT (QTc) interval prolongation and Torsades de Pointes. Of the 241 patients treated with Revuforj at the recommended dosage for relapsed or refractory acute leukemia in clinical trials, QTc interval prolongation was reported as an adverse reaction in 86 (36%) patients. QTc interval prolongation was Grade 3 in 15% and Grade 4 in 2%. The heart-rate corrected QT interval (using Fridericia’s method) (QTcF) was greater than 500 msec in 10%, and the increase from baseline QTcF was greater than 60 msec in 24%. Revuforj dose reduction was required for 7% due to QTc interval prolongation. QTc prolongation occurred in 21% of the 34 patients less than 17 years old, 35% of the 146 patients 17 years to less than 65 years old, and 46% of the 61 patients 65 years or older. One patient had a fatal outcome of cardiac arrest, and one patient had non-sustained Torsades de Pointes. Correct electrolyte abnormalities, including hypokalemia and hypomagnesemia, prior to and throughout treatment with Revuforj. Perform an electrocardiogram (ECG) prior to initiation of Revuforj, and do not initiate Revuforj in patients with QTcF >450 msec. Perform an ECG at least once weekly for the first 4 weeks and at least monthly thereafter. In patients with congenital long QTc syndrome, congestive heart failure, electrolyte abnormalities, or those who are taking medications known to prolong the QTc interval, more frequent ECG monitoring may be necessary. Concomitant use with drugs known to prolong the QTc interval may increase the risk of QTc interval prolongation. Interrupt Revuforj if QTcF increases >480 msec and <500 msec, and restart Revuforj at the same dose twice daily after the QTcF interval returns to ≤480 msec Interrupt Revuforj if QTcF increases >500 msec or by >60 msec from baseline, and restart Revuforj twice daily at the lower-dose level after the QTcF interval returns to ≤480 msec Permanently discontinue Revuforj in patients with ventricular arrhythmias and in those who develop QTc interval prolongation with signs or symptoms of life-threatening arrhythmia Embryo-Fetal Toxicity: Revuforj can cause fetal harm when administered to a pregnant woman. Advise pregnant women of the potential risk to a fetus. Advise females of reproductive potential and males with female partners of reproductive potential to use effective contraception during treatment with Revuforj and for 4 months after the last dose of Revuforj. ADVERSE REACTIONS Fatal adverse reactions occurred in 9 (4%) patients who received Revuforj, including 4 with sudden death, 2 with differentiation syndrome, 2 with hemorrhage, and 1 with cardiac arrest. Serious adverse reactions were reported in 184 (76%) patients. The most frequent serious adverse reactions (≥10%) were infection (29%), febrile neutropenia (20%), bacterial infection (15%), differentiation syndrome (13%), and hemorrhage (11%). The most common adverse reactions (≥20%) including laboratory abnormalities, were phosphate increased (51%), hemorrhage (48%), nausea (48%), infection without identified pathogen (46%), aspartate aminotransferase increased (44%), alanine aminotransferase increased (40%), creatinine increased (38%), musculoskeletal pain (37%), febrile neutropenia (37%), electrocardiogram QT prolonged (36%), potassium decreased (34%), parathyroid hormone intact increased (34%), alkaline phosphatase increased (33%), diarrhea (29%), bacterial infection (27%), triglycerides increased (27%), phosphate decreased (25%), differentiation syndrome (25%), fatigue (24%), edema (24%), viral infection (23%), decreased appetite (20%), and constipation (20%). DRUG INTERACTIONS Drug interactions can occur when Revuforj is concomitantly used with: Strong CYP3A4 inhibitors: reduce Revuforj dose Strong or moderate CYP3A4 inducers: avoid concomitant use with Revuforj QTc-prolonging drugs: avoid concomitant use with Revuforj. If concomitant use is unavoidable, obtain ECGs when initiating, during concomitant use, and as clinically indicated. Withhold Revuforj if the QTc interval is >480 msec. Restart Revuforj after the QTc interval returns to ≤480 msec SPECIFIC POPULATIONS Lactation: advise lactating women not to breastfeed during treatment with Revuforj and for 1 week after the last dose. Pregnancy and testing: Revuforj can cause fetal harm when administered to a pregnant woman. Verify pregnancy status in females of reproductive potential within 7 days prior to initiating Revuforj. Infertility: based on findings in animals, Revuforj may impair fertility. The effects on fertility were reversible. Pediatric: monitor bone growth and development in pediatric patients. Geriatric: no overall differences were observed in the effectiveness of Revuforj between patients who were 65 years and older, and younger patients. Compared to younger patients, the incidences of QTc prolongation and edema were higher in patients 65 years and older. To report SUSPECTED ADVERSE REACTIONS, contact Syndax Pharmaceuticals at 1-888-539-3REV or FDA at 1-800-FDA-1088 or www.fda.gov/medwatch. Please see Full Prescribing Information, including BOXED WARNINGS. About Syndax Syndax Pharmaceuticals is a commercial-stage biopharmaceutical company advancing innovative cancer therapies. Highlights of the Company's pipeline include Revuforj® (revumenib), an FDA-approved menin inhibitor, and Niktimvo™ (axatilimab-csfr), an FDA-approved monoclonal antibody that blocks the colony stimulating factor 1 (CSF-1) receptor. Fueled by our commitment to reimagining cancer care, Syndax is working to unlock the full potential of its pipeline and is conducting several clinical trials across the continuum of treatment. For more information, please visit www.syndax.com/ or follow the Company on X and LinkedIn. Forward-Looking Statements This press release contains forward-looking statements within the meaning of the Private Securities Litigation Reform Act of 1995. Words such as "anticipate," "believe," "could," "estimate," "expects," "intend," "may," "plan," "potential," "predict," "project," "should," "will," "would" or the negative or plural of those terms, and similar expressions (as well as other words or expressions referencing future events, conditions or circumstances) are intended to identify forward-looking statements. These forward-looking statements are based on Syndax's expectations and assumptions as of the date of this press release. Each of these forward-looking statements involves risks and uncertainties. Actual results may differ materially from these forward-looking statements. Forward-looking statements contained in this press release include, but are not limited to, statements about the progress, timing, clinical development and scope of clinical trials, the reporting of clinical data for Syndax's product candidates, the acceptance of Syndax and its partners' products in the marketplace, sales, marketing, manufacturing and distribution requirements, and the potential use of its product candidates to treat various cancer indications and fibrotic diseases. Many factors may cause differences between current expectations and actual results, including: unexpected safety or efficacy data observed during preclinical or clinical trials; clinical trial site activation or enrollment rates that are lower than expected; changes to Revuforj's or Niktimvo’s commercial availability; changes in expected or existing competition; changes in the regulatory environment; failure of Syndax's collaborators to support or advance collaborations or product candidates; and unexpected litigation or other disputes. Other factors that may cause Syndax's actual results to differ from those expressed or implied in the forward-looking statements in this press release are discussed in Syndax's filings with the U.S. Securities and Exchange Commission, including the "Risk Factors" sections contained therein. Except as required by law, Syndax assumes no obligation to update any forward-looking statements contained herein to reflect any change in expectations, even as new information becomes available. Syndax Contact Sharon Klahre Syndax Pharmaceuticals, Inc. sklahre@syndax.com Tel 781.684.9827 SNDX-G Source: Syndax Pharmaceuticals, Inc.