[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$fcGpIh_btVkVjYYUNU6f91XsQ9kGPWXLLyu723W_ho-c":3},{"item":4,"storyV2":72},{"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":47,"cover":70},"prometa","image","T4","2026","2026-01-01",false,{"zh":12,"en":12},"ProMeta",{"zh":14,"en":14},"大连理工大学生物工程学院",{"zh":16,"en":17},"酵母启动子设计 —— AI 驱动的闭环优化","Yeast promoter design — AI-driven closed-loop optimization",{"zh":19,"en":19},"用人工智能预测并设计解脂耶氏酵母启动子，并在两个外部数据集上验证预测器",{"zh":21,"en":23},[22],"ProMeta 训练四路启动子强度预测器，并用内部交叉验证和两个外部数据集比较单模型、stacking 与简单平均的泛化表现。",[22],{"zh":25,"en":26},[],[],{"zh":28,"en":29},[],[],{"zh":31,"en":32},[],[],{"zh":34,"en":35},[],[],"\u002Fapi\u002Fteam-showcase\u002Fprometa\u002Fmedia\u002Fc740fc40d3fbb3936d4133f04",{},[39],{"imageUrl":40,"alt":41,"title":43,"description":45,"position":46},"\u002Fapi\u002Fteam-showcase\u002Fprometa\u002Fmedia\u002Fc8ee67acdade209ede9e71ee4",{"zh":42,"en":42},"原始生成 3360 条 → 通过格式与打分输入 3222 条 → 冻结模型打分 3222 条 → 每个目标取 Top100 共 700 条 → Top30 共 210 条 → Top6 共 42 条。",{"zh":44,"en":44},"WGAN 候选生成与筛选漏斗",{"zh":42,"en":42},"center center",{"proposition":48,"listEvidence":49},{"zh":19,"en":19},[50,55,60,65],{"value":51,"label":53},{"zh":52,"en":52},"4 人",{"zh":54,"en":54},"成员人数",{"value":56,"label":58},{"zh":57,"en":57},"惠晨一",{"zh":59,"en":59},"队长",{"value":61,"label":63},{"zh":62,"en":62},"解脂耶氏酵母",{"zh":64,"en":64},"项目底盘",{"value":66,"label":68},{"zh":67,"en":67},"荧光与交叉验证",{"zh":69,"en":69},"验证方式",{"tone":71,"heroImageUrl":40,"heroImagePosition":46},"ai",{"version":73,"publicSlug":5,"identity":74,"proposition":78,"lead":79,"listEvidence":80,"narrativeTendency":93,"hero":94,"presentation":99,"chapters":107},2,{"teamName":75,"projectTitle":76,"institution":77,"track":7},{"zh":12,"en":12},{"zh":16,"en":17},{"zh":14,"en":14},{"zh":19,"en":19},{"zh":22,"en":22},[81,84,87,90],{"value":82,"label":83},{"zh":52,"en":52},{"zh":54,"en":54},{"value":85,"label":86},{"zh":57,"en":57},{"zh":59,"en":59},{"value":88,"label":89},{"zh":62,"en":62},{"zh":64,"en":64},{"value":91,"label":92},{"zh":67,"en":67},{"zh":69,"en":69},"evidence-led",{"role":95,"title":96,"caption":97,"alt":98,"position":46,"imageUrl":40},"HERO",{"zh":44,"en":44},{"zh":42,"en":42},{"zh":42,"en":42},{"template":100,"variant":101,"logo":102,"openingMedia":106},"documentary-profile","rich",{"role":103,"alt":104,"position":46,"imageUrl":36},"LOGO",{"zh":105,"en":105},"ProMeta队徽",[],[108,119,128,161,195,219],{"id":109,"eyebrow":110,"title":112,"theme":114,"type":115,"quote":116,"attribution":118},"team-motto",{"zh":111,"en":111},"队伍口号",{"zh":113,"en":113},"队伍声音","people","quote",{"zh":117,"en":117},"读懂序列，设计精准基因开关",{"zh":12,"en":12},{"id":120,"eyebrow":121,"title":123,"body":125,"theme":127,"type":127},"project-background",{"zh":122,"en":122},"项目背景",{"zh":124,"en":124},"开关，和调速器",{"zh":126,"en":126},"启动子就像基因表达的「开关和调速器」，决定细胞中某个基因何时启动、表达多强。两个词都要。开关管有无，调速器管强弱 —— 而代谢工程真正需要的是后者：一条通路上不同酶的表达量得配比得当，多一分堆积、少一分断流。所以这个项目的目标不是「找到一个强启动子」，而是队伍写在最后的那句：获得一批强度梯度明确、序列多样、便于应用的调控元件。强度梯度明确，是为了能挑；序列多样，是为了同时用多个时不会因同源重组而失稳；便于应用，是为了真的能装进载体。","problem",{"id":129,"eyebrow":130,"title":132,"body":134,"theme":136,"type":137,"facts":138,"image":153},"research-context",{"zh":131,"en":131},"研究路径",{"zh":133,"en":133},"四路模型，先在内部比一轮",{"zh":135,"en":135},"团队在同一核心数据集上比较四路启动子强度预测器，统一采用 5 折折外预测（OOF）。CNN 集成、LSTM\u002FGRU 集成、DNABERT2-W29 与 Stacking 的 Pearson R 分别为 0.5290、0.5370、0.5030 和 0.5658。Stacking 的二级学习器仅使用基础模型的折外预测结果，避免将训练集拟合值混入集成评估；这些内部比较结果仍需通过独立外部数据检验泛化表现。","source","source-context",[139,143,148],{"label":140,"value":142},{"zh":141,"en":141},"底盘",{"zh":62,"en":62},{"label":144,"value":146},{"zh":145,"en":145},"外部验证集",{"zh":147,"en":147},"n=82、n=85",{"label":149,"value":151},{"zh":150,"en":150},"生成模型",{"zh":152,"en":152},"WGAN",{"role":154,"title":155,"caption":157,"alt":159,"position":46,"imageUrl":160},"RESULT",{"zh":156,"en":156},"内部 5 折 OOF 的回归性能",{"zh":158,"en":158},"CNN 集成 0.5290、LSTM\u002FGRU 集成 0.5370、DNABERT2-W29 0.5030，Stacking 0.5658（Pearson R）。注里写明：所有结果基于相同核心数据集的 5 折 OOF 预测，Stacking 二级学习器仅使用基础模型的 OOF 结果。",{"zh":158,"en":158},"\u002Fapi\u002Fteam-showcase\u002Fprometa\u002Fmedia\u002Fc65aff7836d7ba460be227256",{"id":162,"eyebrow":163,"title":165,"theme":167,"type":168,"semantics":169,"steps":170},"project-process",{"zh":164,"en":164},"项目过程",{"zh":166,"en":166},"项目推进与验证路径","process","process-flow","hybrid",[171,180,187],{"order":172,"phase":173,"title":175,"body":177,"status":179},1,{"zh":174,"en":174},"阶段 1",{"zh":176,"en":176},"外部数据集上的模型比较",{"zh":178,"en":178},"团队在两个外部数据集上比较内部最强的 Final stacking 与三个基础模型简单平均（Base3 平均）。Literature-82 上，Base3 平均与 CNN、DNABERT 的 Spearman ρ 均为 0.60，Final stacking 为 0.57；Kendall τ 分别为 0.43 和 0.41，不一致对比例分别为 0.29 和 0.30。Wang2023-rescued85 上两者 Spearman ρ 分别为 0.46 和 0.45。两个外部数据集都未显示 stacking 优于简单平均，因此团队把外部验证与内部排序分开报告。","described",{"order":73,"phase":181,"title":183,"body":185,"status":179},{"zh":182,"en":182},"阶段 2",{"zh":184,"en":184},"生成这一侧：3360 条，筛到 42 条",{"zh":186,"en":186},"预测器之外，另一半是生成。项目团队用 WGAN 按目标强度生成候选序列，然后一层层筛。漏斗图把每一层的数量都标了出来：原始生成 3360 条 → 通过格式与打分输入 3222 条 → 由冻结模型打分 3222 条 → 每个目标取 Top100，合计 700 条 → Top30，210 条 → Top6，42 条。如果打分器跟着生成器一起更新，生成器很容易学会去骗打分器，而不是真的写出好序列。另外两张图分别检查生成器听不听话：一张看设定的目标强度与生成候选的平均预测强度是否跟得上，另一张看 GC 含量与预测强度的关系，并按两种过滤口径（balanced_only 与 strict_safe）分色。",{"order":188,"phase":189,"title":191,"body":193,"status":179},3,{"zh":190,"en":190},"阶段 3",{"zh":192,"en":192},"模型、生成与实验如何衔接",{"zh":194,"en":194},"工作分为模型训练与集成、候选启动子生成、实验推进和数据回流。干实验侧负责文献、模型和候选筛选，湿实验侧负责 DNA 合成、荧光验证及实测强度与预测值对照，总体统筹负责项目设计和进度衔接。",{"id":196,"eyebrow":197,"title":199,"theme":201,"type":202,"items":203},"project-figure",{"zh":198,"en":198},"项目证据",{"zh":200,"en":200},"项目图表与方法记录","evidence","figures",[204,211],{"role":154,"title":205,"caption":207,"alt":209,"position":46,"imageUrl":210},{"zh":206,"en":206},"启动子设计闭环",{"zh":208,"en":208},"流程从启动子文库和序列—强度数据出发，经特征提取、模型训练与条件生成形成候选，再由 DNA 合成和荧光测定回传实测结果。",{"zh":208,"en":208},"\u002Fapi\u002Fteam-showcase\u002Fprometa\u002Fmedia\u002Fcba5ad0fcab40c6873b42364d",{"role":212,"title":213,"caption":215,"alt":217,"position":46,"imageUrl":218},"DRY_LAB",{"zh":214,"en":214},"启动子候选处理脚本",{"zh":216,"en":216},"代码整理启动子序列特征与候选结果，为模型训练、生成候选过滤和后续实验清单提供统一输入。",{"zh":216,"en":216},"\u002Fapi\u002Fteam-showcase\u002Fprometa\u002Fmedia\u002Fc351eaf7073a244f07d17c4d3",{"id":220,"eyebrow":221,"title":223,"theme":114,"type":225,"items":226},"team-gallery",{"zh":222,"en":222},"团队协作",{"zh":224,"en":224},"计算、实验与候选复核记录","gallery",[227,235,242],{"role":228,"title":229,"caption":231,"alt":233,"position":46,"imageUrl":234},"TEAM",{"zh":230,"en":230},"团队成员与项目分工",{"zh":232,"en":232},"成员介绍图列出四名团队成员及其项目分工，用于说明项目协作安排。",{"zh":232,"en":232},"\u002Fapi\u002Fteam-showcase\u002Fprometa\u002Fmedia\u002Fcd86e5f67eca6c18e5f99d551",{"role":154,"title":236,"caption":238,"alt":240,"position":46,"imageUrl":241},{"zh":237,"en":237},"目标强度与预测结果对照",{"zh":239,"en":239},"横轴为设定的目标强度档位，两条曲线分别记录目标值与生成候选的平均预测强度，用于检查条件生成是否贴近设定目标。",{"zh":239,"en":239},"\u002Fapi\u002Fteam-showcase\u002Fprometa\u002Fmedia\u002Fc1eb173be50652b908c999d86",{"role":154,"title":243,"caption":245,"alt":247,"position":46,"imageUrl":248},{"zh":244,"en":244},"ProMeta项目图表",{"zh":246,"en":246},"GC 含量与预测强度的关系，按两种过滤口径分色（balanced_only 与 strict_safe）。",{"zh":246,"en":246},"\u002Fapi\u002Fteam-showcase\u002Fprometa\u002Fmedia\u002Fc5d384806c7641c29ec141f21"]