上海交通大学洪亮教授课题组与上海人工智能实验室的青年研究员谈攀，在蛋白质突变-性质预测领域取得了重大突破，这一成果已发表在《Nature Communications》期刊上。这一研究采用了一种创新的策略，即基于语言模型的蛋白质功能小样本预测方法（FSFP），在利用极少量的实验室数据的情况下，显著提升了传统蛋白质预训练大模型的性能，极大地加速了蛋白质工程的进程。
蛋白质工程，作为现代生物技术的重要分支，旨在通过设计和改造蛋白质分子来实现特定的功能。传统方法往往依赖于大量的湿实验，通过反复试错来筛选出具有理想性质的蛋白质突变体，这一过程不仅耗时耗力，且筛选的突变库范围有限。而洪亮教授团队与谈攀研究员的合作，提出了一种全新的训练策略，通过“few-shot learning”（小样本学习），使语言模型能够高效地学习和预测蛋白质突变后的性质，从而大幅减少了对湿实验的依赖。
FSFP方法的核心在于其利用了语言模型的高效信息处理能力，通过少量的实验数据训练，模型能够快速理解蛋白质突变与性质之间的复杂关系，进而预测出具有高潜力的突变体。这一突破性成果不仅缩短了从设计到验证蛋白质新功能的时间周期，而且拓宽了可能的突变序列库，为蛋白质工程领域带来了革命性的变化。
《Nature Communications》作为国际顶级学术期刊，该论文的发表标志着FSFP方法在蛋白质突变-性质预测领域的广泛应用前景，有望推动蛋白质工程、酶工程等多个生物技术领域的发展，加速新药物、生物催化剂等高价值生物产品的研发进程。这一研究成果不仅展示了人工智能与生物科学的深度结合，也为未来的科学研究提供了新的思路和工具。
英语如下：
News Title: “Shanghai Jiao Tong University- Shanghai AI Lab Innovates Method, Efficiently Predicts Protein Mutation Properties”
Keywords: Prof. Hong Liang’s Research Team, Shanghai AI Lab, Protein Function Prediction
News Content: Prof. Hong Liang’s research team at Shanghai Jiao Tong University, in collaboration with young researchers at the Shanghai Artificial Intelligence Lab, has made a significant breakthrough in the field of predicting protein mutation properties. This achievement has been published in the prestigious journal ‘Nature Communications’. The study employs an innovative approach, a protein function small sample prediction method (FSFP) based on language models, which significantly enhances the performance of traditional protein pre-training large models by utilizing a minimal amount of laboratory data, greatly accelerating the process of protein engineering.
As a crucial branch of modern biotechnology, protein engineering aims to design and modify protein molecules to achieve specific functions. Traditional methods often rely on extensive wet experiments, through repeated trial and error to screen for protein mutants with ideal properties, a process that is both time-consuming and labor-intensive, with a limited scope of screened mutation libraries. The collaboration between Prof. Hong Liang’s team and researcher Tan Pan proposes a novel training strategy, utilizing “few-shot learning” (small sample learning), enabling language models to efficiently learn and predict the properties of protein mutations, thereby significantly reducing reliance on wet experiments.
The core of the FSFP method lies in its utilization of the high-efficiency information processing capabilities of language models. Through a minimal amount of experimental data training, the model can rapidly understand the complex relationship between protein mutations and properties, and predict high-potential mutants. This breakthrough not only shortens the cycle from design to validation of protein new functions but also broadens the possible mutation sequence libraries, bringing revolutionary changes to the field of protein engineering.
The publication of this paper in ‘Nature Communications’, an international top academic journal, marks the FSFP method’s broad application prospects in the field of predicting protein mutation properties. It is expected to propel the development of various biotechnology fields such as protein engineering and enzyme engineering, accelerating the development of high-value bioproducts such as new drugs and biocatalysts. This research result not only demonstrates the deep integration of artificial intelligence and biological science, but also provides new ideas and tools for future scientific research.
【来源】https://www.jiqizhixin.com/articles/2024-07-08-23
Views: 3