←News detail

Shanghai ObserverUpdated Oct 25 20:57

Westlake University team publishes Pinal model in Nature: a single sentence can generate candidate proteins

English text is being prepared.

AI summary

A research team from Westlake University, in collaboration with the Hong Kong University of Science and Technology (Guangzhou) and the Shenzhen Institute of Advanced Technology of the Chinese Academy of Sciences, built an AI model called Pinal that, with just a single sentence of text description, can generate candidate proteins whose sequences differ from known natural proteins and which may possess actual functions. The related results were published online in the journal Nature. Using protein three-dimensional structures as an intermediate bridge, the team constructed a synthetic training dataset covering approximately 135 million protein sequences and approximately 1.7 billion protein-text pairs. The model first generates a structural blueprint based on the text, then arranges amino acids and scores the candidate sequences; in four types of experiments—fluorescent proteins, enzymes that degrade PET plastic, alcohol dehydrogenases, and shuttle H proteins—it designed proteins that performed the target functions. The researchers stated that this direction is still at the proof-of-concept stage, and fine-grained functional design such as specifying emission wavelengths remains a challenge.

Why read this

AI score 58

The Westlake University team used three-dimensional structures as an intermediate bridge to enable natural language to directly generate candidate proteins, and presented validation results from four types of experiments.

Topics

Tags

Body

记者10月22日从西湖大学获悉,该校科研团队与香港科技大学(广州)、中国科学院深圳先进技术研究院团队合作,搭建了名为Pinal的人工智能(AI)模型。只需给该模型提供一句文字描述,它就能生成与已知天然蛋白序列不同、并可能具备实际功能的候选蛋白质。相关研究成果日前在线发表于《自然》杂志。 蛋白质是生物大分子,其借助层层折叠出的精确空间结构,发挥各种特定功能。“在冷冻电镜重构出的三维结构中,蛋白质呈现复杂而精密的折叠形态,成千上万个原子在纳米尺度内有序排布。”西湖大学工学院特聘研究员原发杰介绍,蛋白质设计的难点正在于其结构高度精密,局部变化可能影响整体折叠和功能。 近年来,随着技术的快速发展,AI已能够预测蛋白质结构与序列之间的关系。西湖大学科研团队希望更进一步,直接用自然语言描述蛋白质的功能,让AI生成对应的蛋白质序列。 Pinal概念图。(图片由 GPT-Image-2生成)西湖大学供图 在研究过程中,科研团队发现,这一尝试需跨越巨大的数据鸿沟:已知蛋白序列数据规模庞大,但经实验验证功能的数据相对稀缺,AI模型难以建立功能文字和蛋白序列之间的可靠对应关系。 为破解难题,科研团队提出以蛋白质三维结构作为中间桥梁的技术路线,以此来开发和训练Pinal模型。他们整合了多种来源的蛋白质数据,构建了覆盖约1.35亿条不同蛋白质序列的训练数据集。通过为同一蛋白质配备多种语言表述,团队最终构建了包含约17亿条“蛋白质-文本”配对的AI合成训练数据集,为Pinal模型学习语言描述与蛋白质之间的关联提供了基础。 该模型运行大致分为三步:先根据文字生成蛋白质的结构蓝图,再依据结构排布氨基酸,最后对众多候选氨基酸序列打分,挑出最符合目标的候选序列。 科研团队选取荧光蛋白、降解PET塑料的酶、醇脱氢酶以及穿梭H蛋白四类功能迥异的蛋白质开展实验验证。四类实验共同证实,Pinal模型可以设计出区别于已知天然蛋白序列且在实验中发挥目标功能的蛋白质。 原发杰表示,目前基于自然语言描述的蛋白质设计模型仍处于概念验证阶段,更多精细化的蛋白质功能设计需求,例如通过语言指令指定荧光蛋白的发光波段,是团队以及科技界当下面临的科研挑战。

Source

Shanghai Observer