[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$fceqknsgqaESMKJBMt3XgFzoBGZPifXh8f8CxcPQ8AS0":3},{"item":4,"storyV2":93},{"slug":5,"type":6,"track":7,"season":8,"date":9,"featured":10,"teamName":11,"institution":13,"projectTitle":15,"summary":18,"body":20,"tags":24,"aiTags":27,"wetlabTags":30,"awards":33,"logoUrl":36,"materials":37,"gallery":38,"editorial":68,"cover":91},"perturbcell","image","T5","2026","2026-01-01",false,{"zh":12,"en":12},"PerturbCell 造物者小队",{"zh":14,"en":14},"天津大学网络安全学院",{"zh":16,"en":17},"PerturbCell —— 多药扰动预测","PerturbCell — multi-drug perturbation prediction",{"zh":19,"en":19},"融合多模态信息与基因调控网络，预测药物条件下细胞转录组响应",{"zh":21,"en":23},[22],"把处理前后的差异拆开看，才发现模型很可能只是把原来的细胞背景复制了一遍 —— 并没有真正学会药物响应。天津大学 PerturbCell 造物者小队把这次「高分幻觉」当成了转折点：重做评估、增加诊断、加入错误条件对照。",[22],{"zh":25,"en":26},[],[],{"zh":28,"en":29},[],[],{"zh":31,"en":32},[],[],{"zh":34,"en":35},[],[],"\u002Fapi\u002Fteam-showcase\u002Fperturbcell\u002Fmedia\u002Fcdede6f4c16e3f73c9a49b37a",{},[39,47,54,61],{"imageUrl":40,"alt":41,"title":43,"description":45,"position":46},"\u002Fapi\u002Fteam-showcase\u002Fperturbcell\u002Fmedia\u002Fc985f77be39f2109a8465ec55",{"zh":42,"en":42},"六位成员在校园内合影，团队共同推进扰动数据建模、评估与实验接口设计。",{"zh":44,"en":44},"PerturbCell 团队",{"zh":42,"en":42},"center center",{"imageUrl":48,"alt":49,"title":51,"description":53,"position":46},"\u002Fapi\u002Fteam-showcase\u002Fperturbcell\u002Fmedia\u002Fc59b47a8aefc25e550cb81f2e",{"zh":50,"en":50},"六位成员对照白板内容核对模型评估、错误对照与实验反馈接口。",{"zh":52,"en":52},"评估流程集中复核",{"zh":50,"en":50},{"imageUrl":55,"alt":56,"title":58,"description":60,"position":46},"\u002Fapi\u002Fteam-showcase\u002Fperturbcell\u002Fmedia\u002Fcba14d1836405bc1ebe60275b",{"zh":57,"en":57},"团队在白板前梳理多模态输入、模型结构与评估指标之间的对应关系。",{"zh":59,"en":59},"扰动预测方案讨论",{"zh":57,"en":57},{"imageUrl":62,"alt":63,"title":65,"description":67,"position":46},"\u002Fapi\u002Fteam-showcase\u002Fperturbcell\u002Fmedia\u002Fc09d6f84510dc86428f1de49c",{"zh":64,"en":64},"六位成员在天津大学智能与计算学部门前合影。",{"zh":66,"en":66},"学院楼前的团队合影",{"zh":64,"en":64},{"proposition":69,"listEvidence":70},{"zh":19,"en":19},[71,76,81,86],{"value":72,"label":74},{"zh":73,"en":73},"6 人",{"zh":75,"en":75},"成员人数",{"value":77,"label":79},{"zh":78,"en":78},"秦萍",{"zh":80,"en":80},"队长",{"value":82,"label":84},{"zh":83,"en":83},"多药扰动预测",{"zh":85,"en":85},"项目对象",{"value":87,"label":89},{"zh":88,"en":88},"扰动对照评估",{"zh":90,"en":90},"验证方式",{"tone":92,"heroImageUrl":40,"heroImagePosition":46},"indigo",{"version":94,"publicSlug":5,"identity":95,"proposition":99,"lead":100,"listEvidence":101,"narrativeTendency":114,"hero":115,"presentation":120,"chapters":143},2,{"teamName":96,"projectTitle":97,"institution":98,"track":7},{"zh":12,"en":12},{"zh":16,"en":17},{"zh":14,"en":14},{"zh":19,"en":19},{"zh":22,"en":22},[102,105,108,111],{"value":103,"label":104},{"zh":73,"en":73},{"zh":75,"en":75},{"value":106,"label":107},{"zh":78,"en":78},{"zh":80,"en":80},{"value":109,"label":110},{"zh":83,"en":83},{"zh":85,"en":85},{"value":112,"label":113},{"zh":88,"en":88},{"zh":90,"en":90},"evidence-led",{"role":116,"title":117,"caption":118,"alt":119,"position":46,"imageUrl":40},"TEAM",{"zh":44,"en":44},{"zh":42,"en":42},{"zh":42,"en":42},{"template":121,"variant":122,"logo":123,"openingMedia":130},"documentary-profile","rich",{"role":124,"title":125,"caption":127,"alt":129,"position":46,"imageUrl":36},"LOGO",{"zh":126,"en":126},"PerturbCell 项目标识",{"zh":128,"en":128},"标识将药物、DNA 与网络节点组合在一起，对应单药和多药扰动响应预测任务。",{"zh":128,"en":128},[131,135,139],{"role":116,"title":132,"caption":133,"alt":134,"position":46,"imageUrl":48},{"zh":52,"en":52},{"zh":50,"en":50},{"zh":50,"en":50},{"role":116,"title":136,"caption":137,"alt":138,"position":46,"imageUrl":55},{"zh":59,"en":59},{"zh":57,"en":57},{"zh":57,"en":57},{"role":116,"title":140,"caption":141,"alt":142,"position":46,"imageUrl":62},{"zh":66,"en":66},{"zh":64,"en":64},{"zh":64,"en":64},[144,155,164,198,231,274],{"id":145,"eyebrow":146,"title":148,"theme":150,"type":151,"quote":152,"attribution":154},"team-motto",{"zh":147,"en":147},"队伍口号",{"zh":149,"en":149},"队伍声音","people","quote",{"zh":153,"en":153},"让扰动可预测，让细胞可计算",{"zh":12,"en":12},{"id":156,"eyebrow":157,"title":159,"body":161,"theme":163,"type":163},"project-background",{"zh":158,"en":158},"项目背景",{"zh":160,"en":160},"为什么要在电脑里做扰动实验",{"zh":162,"en":162},"生命系统精准预测，是合成生物学走向理性设计的核心瓶颈。具体到药物扰动：传统湿实验依赖低通量试错，周期长、成本高，难以系统解析单药及多药联合扰动下细胞状态的复杂响应。「多药联合」这四个字是难度的来源。单药的组合数是线性的，两两组合就是平方，再加上剂量、细胞类型，空间大到不可能穷举。这个问题让湿实验面临两个挑战 —— 搜索空间太大，真实的生物系统又太复杂。AI 可以从大规模单细胞数据和药物的结构、靶点、机制等信息中学习规律，快速提出优先验证的预测；细胞实验则负责检验这些预测、发现模型的盲区。AI 用于提高候选筛选效率，实验用于验证预测结果。","problem",{"id":165,"eyebrow":166,"title":168,"theme":170,"type":171,"semantics":172,"steps":173},"project-process",{"zh":167,"en":167},"项目过程",{"zh":169,"en":169},"项目推进与验证路径","process","process-flow","hybrid",[174,183,190],{"order":175,"phase":176,"title":178,"body":180,"status":182},1,{"zh":177,"en":177},"阶段 1",{"zh":179,"en":179},"那个接近满分的分数",{"zh":181,"en":181},"团队曾得到接近满分的总体相关性，但把处理前后的差异单独评估后发现，模型可能主要复制了原有细胞背景，没有学会药物响应。团队因此重做评估、增加诊断，并加入错误条件对照。","described",{"order":94,"phase":184,"title":186,"body":188,"status":182},{"zh":185,"en":185},"阶段 2",{"zh":187,"en":187},"评估不只看总体相关性",{"zh":189,"en":189},"不只看总体相关性，还检查差异表达的方向、关键基因的命中率，以及打乱或移除药物信息之后的表现变化。四项里有三项是专门针对上面那个陷阱设计的 —— 方向对不对、关键基因抓没抓到、把药物信息拿掉之后分数会不会掉。如果拿掉药物信息分数几乎不掉，那就说明模型压根没在用它。Learn 阶段则是根据「幅度低估」或「方向错误」这类失败模式来调整模型 —— 不是笼统地「再训一版」，而是对着具体的失败类型改。",{"order":191,"phase":192,"title":194,"body":196,"status":182},3,{"zh":193,"en":193},"阶段 3",{"zh":195,"en":195},"先冻结测试集",{"zh":197,"en":197},"数据规模大，不同数据体系差异明显，单次划分和单个随机种子都可能误导结论。",{"id":199,"eyebrow":200,"title":202,"theme":204,"type":205,"items":206},"project-figure",{"zh":201,"en":201},"项目证据",{"zh":203,"en":203},"项目图表与方法记录","evidence","figures",[207,215,223],{"role":208,"title":209,"caption":211,"alt":213,"position":46,"imageUrl":214},"RESULT",{"zh":210,"en":210},"多模态扰动模型结构",{"zh":212,"en":212},"结构图展示单细胞表达、药物结构与文本知识进入模型后生成扰动响应预测的路径。",{"zh":212,"en":212},"\u002Fapi\u002Fteam-showcase\u002Fperturbcell\u002Fmedia\u002Fc851852204f20a61ce4b3bc77",{"role":216,"title":217,"caption":219,"alt":221,"position":46,"imageUrl":222},"DRY_LAB",{"zh":218,"en":218},"训练任务运行日志",{"zh":220,"en":220},"终端记录模型训练与评估任务，便于复核数据版本、参数和运行结果。",{"zh":220,"en":220},"\u002Fapi\u002Fteam-showcase\u002Fperturbcell\u002Fmedia\u002Fca8882395f9545a5e6de215ac",{"role":224,"title":225,"caption":227,"alt":229,"position":46,"imageUrl":230},"WET_LAB",{"zh":226,"en":226},"扰动验证设备与耗材",{"zh":228,"en":228},"实验台上的设备、移液器与耗材用于细胞样品处理和标准化验证。",{"zh":228,"en":228},"\u002Fapi\u002Fteam-showcase\u002Fperturbcell\u002Fmedia\u002Fc9692bfdab6a914eb4b88fad1",{"id":232,"eyebrow":233,"title":235,"theme":150,"type":237,"items":238},"team-gallery",{"zh":234,"en":234},"团队协作",{"zh":236,"en":236},"计算、实验与候选复核记录","gallery",[239,246,253,260,267],{"role":216,"title":240,"caption":242,"alt":244,"position":46,"imageUrl":245},{"zh":241,"en":241},"模型版本多指标对照",{"zh":243,"en":243},"表格按 MSE、差异表达相关性和 Top-K 命中等指标比较多个模型版本，不同指标对应的最优版本并不相同。",{"zh":243,"en":243},"\u002Fapi\u002Fteam-showcase\u002Fperturbcell\u002Fmedia\u002Fc360879c14ede9d8f64cabc85",{"role":208,"title":247,"caption":249,"alt":251,"position":46,"imageUrl":252},{"zh":248,"en":248},"项目简介示意图",{"zh":250,"en":250},"多模态输入（单细胞转录组、药物结构与文本知识）汇入模型，输出单药与多药扰动下的转录响应预测。",{"zh":250,"en":250},"\u002Fapi\u002Fteam-showcase\u002Fperturbcell\u002Fmedia\u002Fc54855a63d0a28d6f5efd2e4d",{"role":224,"title":254,"caption":256,"alt":258,"position":46,"imageUrl":259},{"zh":255,"en":255},"细胞实验准备区",{"zh":257,"en":257},"超净工作台内的耗材与样品位置，记录后续关键药物扰动验证的实验准备。",{"zh":257,"en":257},"\u002Fapi\u002Fteam-showcase\u002Fperturbcell\u002Fmedia\u002Fcf898adf7382565de1c461a08",{"role":208,"title":261,"caption":263,"alt":265,"position":46,"imageUrl":266},{"zh":262,"en":262},"基线模型评估表",{"zh":264,"en":264},"屏幕列出多种基线模型及指标结果，用于比较不同方法在扰动预测任务中的表现。",{"zh":264,"en":264},"\u002Fapi\u002Fteam-showcase\u002Fperturbcell\u002Fmedia\u002Fcc68189ca81e8009356847688",{"role":208,"title":268,"caption":270,"alt":272,"position":46,"imageUrl":273},{"zh":269,"en":269},"实验结果记录表",{"zh":271,"en":271},"文档按实验版本整理运行设置与评价结果，为模型复核保留可追溯记录。",{"zh":271,"en":271},"\u002Fapi\u002Fteam-showcase\u002Fperturbcell\u002Fmedia\u002Fc5a9924d4f987345a09d6af36",{"id":275,"eyebrow":276,"title":278,"theme":280,"type":281,"highlight":282,"full":291},"team-interview",{"zh":277,"en":277},"队伍专访",{"zh":279,"en":279},"听 PerturbCell 造物者小队 讲述项目路径","media","interview",{"version":283,"title":284,"description":286,"durationSeconds":288,"videoUrl":289,"posterUrl":290},"highlight",{"zh":285,"en":285},"项目精华",{"zh":287,"en":287},"项目核心内容与队伍表达的精华剪辑。",158.774813,"\u002Fapi\u002Fteam-showcase\u002Fperturbcell\u002Fmedia\u002Fc617fd2ada3103d8cb0a0d099","\u002Fapi\u002Fteam-showcase\u002Fperturbcell\u002Fmedia\u002Fca11150196021d1d2820474a6",{"version":292,"title":293,"description":295,"durationSeconds":297,"videoUrl":298,"posterUrl":290},"full",{"zh":294,"en":294},"完整访谈",{"zh":296,"en":296},"按采访问题完整收录的队伍访谈。",372.4,"\u002Fapi\u002Fteam-showcase\u002Fperturbcell\u002Fmedia\u002Fcb2ecd73df839ff22d48b6ac2"]