[{"data":1,"prerenderedAt":154},["ShallowReactive",2],{"content-query-yFc97Clbn7":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"cover":11,"type":12,"body":13,"_type":148,"_id":149,"_source":150,"_file":151,"_stem":152,"_extension":153},"/news/zh/2023-3-28-1","zh",false,"","蛋白质结构预测的解决方案","2021年7月，华为联合昌平实验室、北大BIOPIC与化学学院、深圳湾实验室高毅勤课题组快速复现优化AlphaFold 2，11月开源基于昇思推理工具，效率提升2-3倍。","2023-3-28","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2025/07/25/199b735845bf4106b44b2035dc97bd39.png","news",{"type":14,"children":15,"toc":135},"root",[16,24,29,35,40,45,50,55,62,67,72,77,83,88,93,98,104,109,114,119,125,130],{"type":17,"tag":18,"props":19,"children":21},"element","h2",{"id":20},"项目说明",[22],{"type":23,"value":20},"text",{"type":17,"tag":18,"props":25,"children":27},{"id":26},"案例背景",[28],{"type":23,"value":26},{"type":17,"tag":30,"props":31,"children":32},"p",{},[33],{"type":23,"value":34},"蛋白质结构预测是获得蛋白质功能结构和构象的过程，近半个世纪以来，这一问题一直被誉为“21 世纪的生物物理学”最重要的课题之一。",{"type":17,"tag":30,"props":36,"children":37},{},[38],{"type":23,"value":39},"在过去，因蛋白质构象数量巨大，计算过程复杂，通过 AI 来对蛋白质结构进行预测一直未能取得实质性突破，获取蛋白质空间结构的方法仍然以冷冻电镜、X-ray 等实验技术为主，单个蛋白质的观测成本高达数月及数百万人民币。\n直至 AlphaFold 2 的出现，使得这一问题迎来新的曙光。AlphaFold 2 凭借其接近实验精度的成绩取得 CASP14 蛋白质空间结构预测比赛的榜首，这一成就也被 Nature 誉为“前所未有的进步”。",{"type":17,"tag":30,"props":41,"children":42},{},[43],{"type":23,"value":44},"2021 年 7 月 DeepMind 团队宣布对 AlphaFold 2 的推理源代码进行开源，华为与北京昌平实验室、北京大学生物医学前沿创新中心（BIOPIC）和化学与分子工程学院、深圳湾实验室高毅勤课题组第一时间对其进行了复现及优化，并于同年 11 月开源了基于昇思 MindSpore 的推理工具，效率同比提升 2-3 倍。",{"type":17,"tag":30,"props":46,"children":47},{},[48],{"type":23,"value":49},"日前，华为与北京昌平实验室、北京大学生物医学前沿创新中心（BIOPIC）和化学与分子工程学院、深圳湾实验室高毅勤教授课题组，在全场景 AI 框架昇思 MindSpore 上推出全流程蛋白质结构预测工具 MEGA-Protein。",{"type":17,"tag":18,"props":51,"children":53},{"id":52},"案例简介",[54],{"type":23,"value":52},{"type":17,"tag":56,"props":57,"children":59},"h3",{"id":58},"一alphafold-2-的缺陷",[60],{"type":23,"value":61},"一、AlphaFold 2 的缺陷",{"type":17,"tag":30,"props":63,"children":64},{},[65],{"type":23,"value":66},"为了满足预测精度，AlphaFold 2 必须引入多序列比对的信息。MSA 的数量决定了 AlphaFold 2 的预测精度，因此 AlphaFold 2 有着自己的缺陷：",{"type":17,"tag":30,"props":68,"children":69},{},[70],{"type":23,"value":71},"自然界中的「孤儿序列」以及人造序列存在着缺少 MSA 或没有 MSA 的情况，导致 AlphaFold 2 等模型对相关的结构预测准确度大幅下降。",{"type":17,"tag":30,"props":73,"children":74},{},[75],{"type":23,"value":76},"AlphaFold 2 等标准检索 MSA 流程，数据库规模大，搜索时间长，不利于展开研究。",{"type":17,"tag":56,"props":78,"children":80},{"id":79},"二ai-msa-引擎",[81],{"type":23,"value":82},"二、AI MSA 引擎",{"type":17,"tag":30,"props":84,"children":85},{},[86],{"type":23,"value":87},"昇思 MindSpore 社区与昌平实验室、北京大学生物医学前沿创新中心（BIOPIC）和化学与分子工程学院、深圳湾实验室高毅勤教授课题组新提出的 AI MSA 引擎具有如下几个优势：",{"type":17,"tag":30,"props":89,"children":90},{},[91],{"type":23,"value":92},"于原始 MSA 质量不高或者数量少的蛋白，将 AI MSA 引擎接入 AlphaFold 2 后可以明显改善结构预测的质量。\n端到端推理性能大幅提升，训练完成的 AI MSA 引擎不需要额外配置数据库。",{"type":17,"tag":30,"props":94,"children":95},{},[96],{"type":23,"value":97},"AI MSA 引擎是一种对下游结构模型普适的预训练方案，可直接接入下游结构预测模型。",{"type":17,"tag":56,"props":99,"children":101},{"id":100},"三昇思-mindspore-ai-框架",[102],{"type":23,"value":103},"三、昇思 MindSpore AI 框架",{"type":17,"tag":30,"props":105,"children":106},{},[107],{"type":23,"value":108},"AI MSA 引擎训练参数量，数据量巨大，训练要求高，昇思 MindSpore AI 框架充分发挥以下优势，支撑完成 AI MSA 引擎的训练。",{"type":17,"tag":30,"props":110,"children":111},{},[112],{"type":23,"value":113},"昇思 MindSpore 在昇腾 AI 基础软硬件平台上与昇腾 CANN 深度结合，通过深度协同优化的高性能算子库，充分释放硬件的算力。",{"type":17,"tag":30,"props":115,"children":116},{},[117],{"type":23,"value":118},"昇思 MindSpore 采用了多段并行流水线的方式来构建数据处理 pipeline，大幅提高了数据处理过程的吞吐量。\n昇思 MindSpore 支持大集群高效训练，实现了优质的计算通信比，通过三层 AI 分布式编程范式，大幅提升分布式并行程序开发效率。",{"type":17,"tag":56,"props":120,"children":122},{"id":121},"四全流程蛋白质结构预测工具-mega-protein",[123],{"type":23,"value":124},"四、全流程蛋白质结构预测工具 MEGA-Protein",{"type":17,"tag":30,"props":126,"children":127},{},[128],{"type":23,"value":129},"MEGA-Protein 包含 AI MSA 引擎、蛋白质折叠训练推理流程、蛋白质结构打分、蛋白质结构预测数据集 PSP 等关键技术，能够高精度高性能地预测蛋白质结构和功能，其中 AI MSA 引擎能够在 MSA 少样本甚至零样本的情况下，帮助 AlphaFold 2 等模型维持甚至提高推理精度，有效突破了 AlphaFold 2 的缺陷。",{"type":17,"tag":30,"props":131,"children":132},{},[133],{"type":23,"value":134},"这是在实现 AlphaFold 2 从训练到推理全流程打通且效率同比提升 2 至 3 倍后，取得的又一次成功。",{"title":7,"searchDepth":136,"depth":136,"links":137},4,[138,140,141],{"id":20,"depth":139,"text":20},2,{"id":26,"depth":139,"text":26},{"id":52,"depth":139,"text":52,"children":142},[143,145,146,147],{"id":58,"depth":144,"text":61},3,{"id":79,"depth":144,"text":82},{"id":100,"depth":144,"text":103},{"id":121,"depth":144,"text":124},"markdown","content:news:zh:2023-3-28-1.md","content","news/zh/2023-3-28-1.md","news/zh/2023-3-28-1","md",1776506058626]