[{"data":1,"prerenderedAt":296},["ShallowReactive",2],{"content-query-kOslaq9vwQ":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"cover":11,"type":12,"body":13,"_type":290,"_id":291,"_source":292,"_file":293,"_stem":294,"_extension":295},"/technology-blogs/zh/1071","zh",false,"","昇思MindSpore再突破！蛋白质结构预测训练推理全流程开源，助力生物医药发展","国产AI框架具备了强大的AI for Science底层软件能力，同时也为相关科研工作者提供新的选择。","2022-03-14","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2022/03/17/11a146984e17451681e74adf94747b01.png","technology-blogs",{"type":14,"children":15,"toc":287},"root",[16,24,33,78,90,95,100,105,110,115,120,130,135,142,147,154,159,166,171,176,181,186,194,205,212,220,230,246,261,271],{"type":17,"tag":18,"props":19,"children":21},"element","h1",{"id":20},"昇思mindspore再突破蛋白质结构预测训练推理全流程开源助力生物医药发展",[22],{"type":23,"value":8},"text",{"type":17,"tag":25,"props":26,"children":27},"p",{},[28],{"type":17,"tag":29,"props":30,"children":32},"img",{"alt":7,"src":31},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2022/03/17/34e342bb02714a1abeafe90e3cd77a75.gif",[],{"type":17,"tag":25,"props":34,"children":35},{},[36,38,44,46,51,53,65,66,71,73],{"type":23,"value":37},"近日，",{"type":17,"tag":39,"props":40,"children":41},"strong",{},[42],{"type":23,"value":43},"昇思MindSpore团队",{"type":23,"value":45},"与",{"type":17,"tag":39,"props":47,"children":48},{},[49],{"type":23,"value":50},"昌平实验室",{"type":23,"value":52},"、",{"type":17,"tag":39,"props":54,"children":55},{},[56,58,63],{"type":23,"value":57},"北京大学生物医学前沿创新中心（BIOPIC）",{"type":17,"tag":39,"props":59,"children":60},{},[61],{"type":23,"value":62},"和",{"type":23,"value":64},"化学与分子工程学院",{"type":23,"value":52},{"type":17,"tag":39,"props":67,"children":68},{},[69],{"type":23,"value":70},"深圳湾实验室高毅勤教授课题组",{"type":23,"value":72},"及",{"type":17,"tag":39,"props":74,"children":75},{},[76],{"type":23,"value":77},"鹏城实验室陈杰团队****基于全场景AI框架昇思MindSpore实现AlphaFold2蛋白质结构训练。",{"type":17,"tag":25,"props":79,"children":80},{},[81,83,88],{"type":23,"value":82},"这是继2021年11月发布推理工具之后又一次突破，",{"type":17,"tag":39,"props":84,"children":85},{},[86],{"type":23,"value":87},"本次训练意味着国产AI框架具备了强大的AI for Science底层软件能力，同时也为相关科研工作者提供新的选择",{"type":23,"value":89},"。",{"type":17,"tag":25,"props":91,"children":92},{},[93],{"type":23,"value":94},"该联合工作依托鹏城云脑II、昇腾AI Atlas集群进行，单步迭代性能提升40%，TM-score达85分（国际权威评测数据集CASP14）。相关训练代码将在昇思MindSpore社区进行开源并定期扩展与维护。",{"type":17,"tag":25,"props":96,"children":97},{},[98],{"type":23,"value":99},"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":25,"props":101,"children":102},{},[103],{"type":23,"value":104},"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。",{"type":17,"tag":25,"props":106,"children":107},{},[108],{"type":23,"value":109},"直至AlphaFold2的出现，使得这一问题迎来新的曙光。AlphaFold2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":25,"props":111,"children":112},{},[113],{"type":23,"value":114},"2021年7月DeepMind团队宣布对AlphaFold2的推理源代码进行开源，昇思团队与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":25,"props":116,"children":117},{},[118],{"type":23,"value":119},"由于开源范围仅限推理，相关从业者无法基于此进行优化，因此许多团队积极地投入训练过程的复现（Pytorch、JAX等）。AlphaFold2模型本身存在内存需求大，数据处理繁琐，控制编译复杂等特点，对基础AI框架存在着巨大挑战。",{"type":17,"tag":25,"props":121,"children":122},{},[123,125],{"type":23,"value":124},"近期，昇思MindSpore研发团队联合高毅勤课题组（刘思睿，张骏等）、鹏城实验室全面打通AlphaFold2的训练。",{"type":17,"tag":39,"props":126,"children":127},{},[128],{"type":23,"value":129},"相同配置下采用混合精度，单步迭代时间由20秒（V100）优化为12秒（昇腾910），性能提升40%。依托昇思MindSpore内存复用能力, 训练序列长度由384提升至512。",{"type":17,"tag":25,"props":131,"children":132},{},[133],{"type":23,"value":134},"为了尽可能客观地评估训练结果，我们选取了AlphaFold2论文附录中提到的87条验证集进行验证。如图2所示，平均TM-score达到85分，基本持平AlphaFold2。",{"type":17,"tag":25,"props":136,"children":137},{},[138],{"type":17,"tag":29,"props":139,"children":141},{"alt":7,"src":140},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2022/03/17/36c3cfb5b7f5433fb94cc321136470a4.jpg",[],{"type":17,"tag":25,"props":143,"children":144},{},[145],{"type":23,"value":146},"表1. T1052-D1推理精度",{"type":17,"tag":25,"props":148,"children":149},{},[150],{"type":17,"tag":29,"props":151,"children":153},{"alt":7,"src":152},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2022/03/17/dfa9fe51f207423595bc6f9edd4bc1bd.jpg",[],{"type":17,"tag":25,"props":155,"children":156},{},[157],{"type":23,"value":158},"图.1 T1052-D1 预测结构图",{"type":17,"tag":25,"props":160,"children":161},{},[162],{"type":17,"tag":29,"props":163,"children":165},{"alt":7,"src":164},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2022/03/17/7ebb8749c07541409e7802f27d9b7f3c.jpg",[],{"type":17,"tag":25,"props":167,"children":168},{},[169],{"type":23,"value":170},"图2. CASP14 87 targets TM-score 对比",{"type":17,"tag":25,"props":172,"children":173},{},[174],{"type":23,"value":175},"昇思MindSpore对蛋白质结构预测训练推理的支持填补了国产AI软硬件的空白。",{"type":17,"tag":25,"props":177,"children":178},{},[179],{"type":23,"value":180},"在训练精度接近AlphaFold2的基础上，研发团队将在算法、规模和软硬件支持等方向上持续改进，并计划开放共享训练数据集供同仁使用。",{"type":17,"tag":25,"props":182,"children":183},{},[184],{"type":23,"value":185},"同时也期望与更多学术界和工业界伙伴合作，进一步提升模型精度、扩展应用场景。",{"type":17,"tag":25,"props":187,"children":188},{},[189],{"type":17,"tag":39,"props":190,"children":191},{},[192],{"type":23,"value":193},"代码开源路径",{"type":17,"tag":25,"props":195,"children":196},{},[197],{"type":17,"tag":198,"props":199,"children":203},"a",{"href":200,"rel":201},"https://gitee.com/mindspore/mindscience/tree/dev/MindSPONGE/mindsponge/fold",[202],"nofollow",[204],{"type":23,"value":200},{"type":17,"tag":25,"props":206,"children":207},{},[208],{"type":17,"tag":29,"props":209,"children":211},{"alt":7,"src":210},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2022/03/17/b926dafdb4ac48fda3c4d6a189b88279.jpg",[],{"type":17,"tag":25,"props":213,"children":214},{},[215],{"type":17,"tag":39,"props":216,"children":217},{},[218],{"type":23,"value":219},"MindSpore官方资料",{"type":17,"tag":25,"props":221,"children":222},{},[223,228],{"type":17,"tag":39,"props":224,"children":225},{},[226],{"type":23,"value":227},"官方QQ群",{"type":23,"value":229}," : 486831414",{"type":17,"tag":25,"props":231,"children":232},{},[233,238,240],{"type":17,"tag":39,"props":234,"children":235},{},[236],{"type":23,"value":237},"官网",{"type":23,"value":239},"：",{"type":17,"tag":198,"props":241,"children":244},{"href":242,"rel":243},"https://www.mindspore.cn/",[202],[245],{"type":23,"value":242},{"type":17,"tag":25,"props":247,"children":248},{},[249,254,255],{"type":17,"tag":39,"props":250,"children":251},{},[252],{"type":23,"value":253},"论坛",{"type":23,"value":239},{"type":17,"tag":198,"props":256,"children":259},{"href":257,"rel":258},"https://bbs.huaweicloud.com/forum/forum-1076-1.html",[202],[260],{"type":23,"value":257},{"type":17,"tag":25,"props":262,"children":263},{},[264,269],{"type":17,"tag":39,"props":265,"children":266},{},[267],{"type":23,"value":268},"Gitee",{"type":23,"value":270}," : https : //gitee.com/mindspore/mindspore",{"type":17,"tag":25,"props":272,"children":273},{},[274,279,281],{"type":17,"tag":39,"props":275,"children":276},{},[277],{"type":23,"value":278},"GitHub",{"type":23,"value":280}," : ",{"type":17,"tag":198,"props":282,"children":285},{"href":283,"rel":284},"https://github.com/mindspore-ai/mindspore",[202],[286],{"type":23,"value":283},{"title":7,"searchDepth":288,"depth":288,"links":289},4,[],"markdown","content:technology-blogs:zh:1071.md","content","technology-blogs/zh/1071.md","technology-blogs/zh/1071","md",1776506112055]