该研究首先提出了一种新颖的蛋白质表征方法——将蛋白质一维序列与三维结构相结合形成"结构感知"词 汇表并据此训练出了 蛋白质语言大模型 —— Saprot 。在此基础上，团队进一步推出了 SaprotHub 开源 平台 。 该平台旨在将 Saprot 等一系列先进蛋白质语言模型的能力开放给生命科学领域研究者，它也是 开放蛋白 质模型联盟 （Open Protein Modeling Consortium，OPMC） 为推动全球科研协作、共建开源社区而迈 出的关键第一步。 编辑丨王多鱼 排版丨水成文 如同人类拥有语言，生命世界也有一套由氨基酸序列构成的"分子语言"—— 蛋白质 。近年来，人工智能 （AI） 领域的 蛋白质语言模型 （PLM） 展现出解码这套语言的强大能力，能够精准预测蛋白质的结构与 功能 。 然而，这些尖端模型的训练与使用，往往需要深厚的机器学习专业知识和编程能力，这在 AI 开发者与广大 生物学家之间形成了一道鸿沟。 为了打破这一壁垒， 2025 年 10 月 24 日， 西湖大学 原发杰 团队 在 Nature Biotechnology 期刊 发表 了题为： Democratizing Pr ...

Core Viewpoint - The rapid development of antibiotic resistance outpaces the discovery of new antibiotics, highlighting the potential of antimicrobial peptides (AMPs) as promising alternatives due to their broad-spectrum antimicrobial activity and unique mechanisms of action [2][6]. Group 1: Antimicrobial Peptides (AMPs) - AMPs are small molecules (10-50 amino acids) that play a crucial role in the host immune defense system, targeting bacteria, fungi, viruses, and parasites [2]. - The mechanisms of AMPs differ from traditional antibiotics, primarily disrupting pathogen cell membranes or interfering with metabolic processes [2][6]. - Despite their potential, the discovery of AMPs remains challenging, necessitating advanced tools like machine learning and deep learning to accelerate research [6][8]. Group 2: Generative Artificial Intelligence in AMP Design - Generative artificial intelligence, particularly through models like AMP-Diffusion, offers a powerful approach for designing AMPs by exploring sequence space systematically [3][7]. - AMP-Diffusion utilizes a pre-trained latent diffusion model to generate potent AMP sequences, ensuring integration with established protein language models like ESM-2 [7][9]. - The model has successfully generated 50,000 candidate AMP sequences, with 76% demonstrating low toxicity and effective bacterial killing capabilities [8][9]. Group 3: Research Findings and Implications - The research team synthesized and validated 46 top-ranking AMP candidates, which exhibited broad-spectrum antimicrobial activity, including against multidrug-resistant strains, with low cytotoxicity [8][9]. - In preclinical mouse models, lead AMPs significantly reduced bacterial load, showing efficacy comparable to polymyxin B and levofloxacin without adverse effects [8][9]. - AMP-Diffusion represents a robust platform for antibiotic design, addressing the urgent need for new antimicrobial agents in the face of rising antibiotic resistance [8][9].

Core Viewpoint - The article discusses the development of a peptide-encoded organ-selective targeting (POST) method that enhances the delivery of mRNA to extrahepatic organs using lipid nanoparticles (LNP) [4][11]. Group 1: mRNA Delivery and LNP Technology - mRNA-based gene and protein replacement technologies present significant opportunities for vaccine, cancer treatment, and regenerative therapy development [2]. - LNPs have been widely adopted as delivery vehicles for mRNA COVID-19 vaccines, demonstrating their safety and efficacy [2]. - Achieving organ-selective delivery of LNPs containing mRNA remains challenging, particularly for extrahepatic organs [2][4]. Group 2: Advances in Organ-Selective Delivery - Recent studies have made progress in organ-selective delivery through simple binary charge modulation and lipid chemical modifications, but these strategies are limited by the rational design of the LNP-environment interface [2][4]. - The POST method utilizes specific amino acid sequences to engineer the surface of LNPs, allowing for efficient mRNA delivery to extrahepatic organs after systemic administration [4][7]. Group 3: Mechanism and Applications - The targeting mechanism of the POST system is based on the optimization of the mechanical affinity between peptide sequences and plasma proteins, forming a specific protein corona around the LNPs [4][9]. - The POST code does not rely on the charge of LNPs for organ selectivity, but rather on the unique protein corona formed, which is influenced by the amino acid sequence [9]. - The POST code is applicable to various LNP formulations and can facilitate the selective delivery of mRNA to organs such as the placenta, bone marrow, adipose tissue, and testes [9][11]. Group 4: AI and Computational Design - The research team developed an AI-based framework using a Transformer-based protein language model to generate peptide sequences with high mechanical affinity for specific proteins, demonstrating the potential of computational design in guiding LNP organ targeting [9][11]. - The peptide sequence RRRYRR was shown to enable selective delivery of mRNA to the lungs, supporting the feasibility of using computer-aided rational design for POST-LNP organ-selective delivery [9][11].

Core Viewpoint - The article discusses the significant environmental challenges posed by plastic waste, particularly PET, and introduces a novel enzyme discovery model, VenusMine, which utilizes protein language models for efficient identification of highly effective PET hydrolases [2][6][13]. Group 1: Enzyme Discovery Model - VenusMine is a protein language model that integrates structural analysis to efficiently mine for PET hydrolases from vast protein databases [6][7]. - The model identifies and clusters target proteins based on the crystal structure of IsPETase, followed by screening for solubility and thermal stability [7][8]. Group 2: Findings and Results - The research team successfully discovered a series of PET hydrolases, with KbPETase from Kibdelosporangium banguiense exhibiting a catalytic efficiency 97 times higher than that of IsPETase [3][8]. - Among the 34 candidate proteins, 14 demonstrated PET degradation activity within the temperature range of 30-60 °C, with KbPETase showing a melting temperature 32°C higher than IsPETase [8][12]. Group 3: Structural Insights - X-ray crystallography and molecular dynamics simulations revealed that KbPETase possesses a conserved catalytic domain and enhanced intramolecular interactions, supporting its improved functionality and thermal stability [12].

Core Viewpoint - The research introduces a machine-learning-assisted strategy called CAGE-Prox vivo for precise protein activation in living organisms, providing a universal platform for time-resolved biological studies and on-demand therapeutic interventions [1][13]. Group 1: Research Background - The study emphasizes the importance of gain-of-function research in understanding biological processes and disease pathology, highlighting various protein engineering techniques that have been developed to manipulate proteins [4]. - Current techniques, while effective, often rely on complex protein constructs that may alter the natural function of target proteins [4][5]. Group 2: CAGE-Prox Strategy - CAGE-Prox is a more universal strategy for controlled activation of a wide range of protein targets, independent of the amino acid residue type at the active site [5]. - The strategy utilizes a light-degradable tyrosine residue (ONBY) to temporarily mask protein activity, allowing for high temporal resolution in studying stimulated cellular processes [5][6]. Group 3: CAGE-Prox vivo Development - The CAGE-Prox vivo strategy incorporates a non-natural amino acid, trans-cyclooctene-tyrosine (TCOY), which can be introduced near the active site of target proteins to temporarily deactivate their function [7][9]. - The research team developed an integrated machine learning process to evolve an aminoacyl-tRNA synthetase (aaRS) that can efficiently incorporate TCOY into proteins [10][11]. Group 4: Applications of CAGE-Prox vivo - The CAGE-Prox vivo system enables precise killing of tumor cells by temporarily inactivating the anthrax lethal factor (LF) and then restoring its activity through a small molecule-triggered bioorthogonal reaction [9][10]. - The strategy also allows for the construction of safer bispecific antibodies that only regain their tumor-targeting function upon specific chemical activation, reducing the risk of cytokine storms and related toxicities [11][12].