要点速览 《科学》（Science）杂志揭晓2025年度十大科学突破榜单，带来一年中最重大的科学发现、可续进展和趋势： 每年，《科学》（Science）杂志都会评选出年度十大科学突破，包括一项科学突破冠军奖以及九项科学突 破入围奖。它们代表了一年中最重大的科学发现、科学进展和趋势。 2025年12月18日， 《科学》揭晓了2025年度十大科学突破榜单 ： 势不可挡的可再生能源 · 势不可挡的可再生能源 · 定制化基因编辑为极罕见疾病带来新希望 · 对抗性传播疾病的新武器 · 神经元对癌症的致命"馈赠" · 天空中的全视之眼 · 揭开"龙人"头骨之谜 · 大语言模型助力科学研究 · 计算的突破助力揭示粒子物理新进展 · 异种移植不断创造历史 · 耐高温水稻 自工业革命以来，人类社会一直依赖于来自"远古的太阳能"——这些能量由数亿年前的植物捕获，储存在化 石燃料中，并从地下挖掘和钻取而来。但今年，势头已明确转向来自"今天的太阳"的能源。可再生能源—— 其中大部分来自阳光本身或最终由太阳驱动的风能——已在多个方面超越了传统能源。 《科学》杂志指出，今年以来， 全球可再生能源以太阳能和风能为主快速增长，其新增发电量已 ...

编辑丨王多鱼 排版丨水成文 每年，《 科学 》 （Science） 杂志都会评选出 年度十大科学突破 ，包括一项 科学突破冠军奖 以及九项 科学突破入围奖 。它们代表了一年中最重大的科学发 现、科学进展和趋势。 撰文丨王聪 2025 年 12 月 18 日， Science 揭晓了 2025 年度的十大科学突破榜单 。其中有 4 项科学突破来自中国—— 可再生能源的迅猛发展 、 揭开 " 龙人"头骨之谜 、 异种器官移植 、 耐高温水稻 。尤其是中国在可再生能源领域的贡献，成为 2025 年度十大科学突破之首。 中国推动可再生能源迅猛发展 自 工业革命 以来，人类社会一直依靠着数亿年前植物捕获并储存在化石燃料中的"古老的太阳能"，通过挖掘和钻探从地下获取这些能源。但今年，势头明显转向 了"今天的太阳能"。 可再生能源 ，其中大部分来自阳光本身或由太阳驱动的风能，在多个方面超越了传统能源。 如今，中国繁忙港口的集装箱里装满了新商品：电动汽车、太阳能电池板、风力涡轮机叶片。在打造自身绿色能源体系的过程中，中国还催生了一个价值近 1800 亿美元的出口产业，让世界其他大部分地区都能用上低成本的可再生能源。 值得一提 ...

文 | 魏文胜 日前通过的《中共中央关于制定国民经济和社会发展第十五个五年规划的建议》中明确提到，要支持创 新药和医疗器械发展。 近期，中国创新药持续成为资本市场关注的焦点：多家企业和科研团队在前沿领域展现出令人瞩目的自 主研发能力，并通过专利授权（License-out）向全球输出创新成果。 事实上，中国医药创新的能力并非一蹴而就，而是在长期积淀中逐渐生长。如今的星星之火，折射出我 国正从"快速追随者"（Fast Follower）向"首创创新者"（First-in-Class）稳步迈进。 生物科技和AI技术叠加助推 以我长期从事的基因编辑研究为例，基因编辑不仅是探索疾病治疗的重要途径，更是一项具有颠覆性意 义的底层技术，正在重新定义我们理解疾病、干预疾病的方式。 所有生物体，包括人类，都是由一套底层"代码" 驱动的，4种碱基在基因组中的排列组合决定了我们的 生理状态和病理特征。每个细胞都携带同样的"代码本"， 但不同细胞根据环境和调控机制差异，会产 生不同的表达模式。一旦这套底层代码出现问题，人体系统便像软件出现"BUG"，轻则影响功能，重则 危及生命，很多遗传性疾病的病因都可以这样简单理解。 基因编辑技 ...

Core Insights - The article discusses the significance of AlphaFold2, an AI tool developed by DeepMind, in predicting protein structures, particularly the giant protein titin, which has eluded complete structural analysis for over 70 years [1][3][4] Group 1: AlphaFold2 and Protein Structure Prediction - AlphaFold2 has revolutionized the field of protein structure prediction, achieving over 90% accuracy in predicting protein structures from amino acid sequences during the global protein structure prediction competition (CASP) in 2020 [6][4] - The database created by AlphaFold now contains over 200 million predicted protein structures, covering 98.5% of the human proteome, enabling researchers worldwide to explore protein functions more efficiently [6][4] - AlphaFold2 was utilized during the early stages of the COVID-19 pandemic to predict the structures of viral proteins, aiding in understanding the virus's mechanisms and potential treatments [8][10] Group 2: Applications in Disease Research - Researchers are using AlphaFold to study the impact of genetic mutations on diseases, such as osteoporosis, by comparing the structures of normal and mutated proteins [11][13] - The introduction of AlphaMissense allows scientists to assess the pathogenic potential of missense mutations, successfully categorizing 89% of human missense mutations and creating a directory for further research [13][11] Group 3: Environmental and Pharmaceutical Innovations - AlphaFold2 is also being applied to address environmental issues, such as plastic pollution, by helping scientists design enzymes that can efficiently degrade single-use plastics [14][17] - The integration of AlphaFold2 into drug discovery platforms, like Insilico Medicine's Pharma.AI, has led to the identification of a candidate drug for idiopathic pulmonary fibrosis, Rentosertib, which is currently in Phase II clinical trials [18][20] Group 4: Future Developments - The article highlights ongoing advancements in protein research, including the discovery of a new protein larger than titin and the release of AlphaFold3 and AlphaProteo, which enhance predictions of protein interactions and custom protein design [23][21] - Other AI models, such as RoseTTAFold and I-TASSER, are also contributing to solving long-standing challenges in protein folding, indicating a collaborative effort in the field [23]

Core Insights - AlphaFold2, developed by Google DeepMind, has revolutionized scientific research by enabling accurate predictions of protein structures based solely on amino acid sequences since its launch in November 2020 [1][4][7]. Group 1: Impact on Scientific Research - Over the past five years, AlphaFold2 has assisted researchers worldwide in predicting millions of protein structures, marking a second renaissance in structural biology [7]. - The tool has significantly accelerated discovery processes, with researchers like Andrea Pauli stating that every project now utilizes AlphaFold [12]. - The Nature paper describing AlphaFold2 has garnered nearly 40,000 citations, indicating sustained interest from the scientific community [12]. Group 2: Applications and Discoveries - AlphaFold-Multimer, an extension of AlphaFold2, has enabled the discovery of three critical proteins involved in fertilization, challenging previous assumptions about the simplicity of sperm-egg interactions [8][10]. - The TMEM81-IZUMO1-SPACA6 protein complex plays a vital role in mediating sperm-egg binding, highlighting the complexity of fertilization mechanisms [10]. Group 3: User Engagement and Accessibility - AlphaFold has been accessed by approximately 3.3 million users across over 190 countries, with more than 1 million users from low- and middle-income countries, showcasing its global reach and accessibility [15]. - The AlphaFold database (AFDB) contains over 240 million predicted protein structures, covering nearly all known proteins on Earth [15]. Group 4: Influence on Structural Biology and Computational Biology - Researchers using AlphaFold have submitted about 50% more protein structures to the Protein Data Bank (PDB) compared to those who did not use the tool [18]. - AlphaFold has opened new research directions in computational biology, including AI-assisted drug discovery and protein design, leading to increased funding and interest in these areas [21]. Group 5: Future Prospects - AlphaFold2 is expected to aid in understanding disease mechanisms and potentially lead to new therapies, with AlphaFold3 anticipated to enhance drug discovery capabilities [24].

Group 1 - The 2024 Nobel Prize in Physics was awarded to Geoffrey Hinton, while the Chemistry Prize went to Demis Hassabis and John Jumper for their work on AlphaFold2, marking a significant year for AI in the Nobel context [1][2] - Michel Devoret received the Nobel Prize in Physics for his contributions to quantum hardware, which is less related to AI [3][2] - The 2023 Nobel Prize in Economic Sciences was awarded to Joel Mokyr, Philippe Aghion, and Peter Howitt for their insights on how innovation drives sustainable development [2][7] Group 2 - Philippe Aghion and Peter Howitt's work on "creative destruction" highlights the dual nature of innovation, which can lead to both the creation of new products and the obsolescence of older ones [10][11] - Their research emphasizes the need to maintain the mechanisms of creative destruction to avoid economic stagnation [16][10] - The Nobel laureates' definitions of AI touch on its potential impact on economic growth and the challenges it poses to traditional labor roles [18][19] Group 3 - Aghion and Howitt argue that AI represents the latest form of automation, which has historically been a key driver of economic growth [20][22] - They discuss the "Baumol's cost disease," which suggests that productivity gains in certain sectors do not necessarily translate to overall economic growth due to rising costs in labor-intensive industries [23][26] - The potential for AI to enhance productivity is tempered by the limitations posed by sectors that are difficult to automate, which could hinder overall economic progress [27][29] Group 4 - The discussion on post-AGI economics suggests that even with advanced AI, economic growth may still be constrained by the slow progress in certain critical tasks [31][32] - Contrasting views suggest that AI-augmented R&D could significantly boost economic growth rates, potentially doubling them if AI technologies are widely adopted [33][34] - The notion that AI could permanently enhance productivity across various fields indicates a transformative potential for future economic growth [35]

Core Insights - Alphabet's scientists have won the Nobel Prize for two consecutive years, highlighting the company's strong capabilities in cutting-edge research areas like artificial intelligence and quantum computing, which are expected to disrupt future business and market landscapes [1][2] - The 2025 Nobel Prize in Physics was awarded to three physicists, including Google Quantum AI Lab's current hardware chief scientist Michel Devoret, recognizing their groundbreaking contributions to quantum mechanics [1][3] - Google CEO Sundar Pichai emphasized that the awardees' research lays the foundation for the company's latest breakthroughs in quantum computing, signaling that long-term investments in fundamental science are translating into core competitive advantages in key technology areas [1][5] Quantum Breakthrough Focus - The Nobel Prize in Physics was awarded to Michel Devoret, John Martinis, and John Clarke for their discoveries in macroscopic quantum tunneling and energy quantization in circuits, demonstrating the peculiar properties of the quantum world in tangible systems [3] - Devoret's role as hardware chief scientist at Google Quantum AI Lab is crucial in building scalable, fault-tolerant quantum computers, while Martinis previously led the team that achieved "quantum supremacy" in 2019 [3] AI Nobel Moment - The recent Physics award follows last year's Nobel Prize in Chemistry awarded to DeepMind's CEO Sir Demis Hassabis and senior research scientist John Jumper for their contributions to protein structure prediction through the AI model AlphaFold2, which has significant applications in pharmaceuticals and materials science [4] R&D Strength of Tech Giants - Winning Nobel Prizes in consecutive years illustrates Google's R&D prowess and its long-term commercial moat, with AI and quantum computing being critical variables in future market competition [5] - Sundar Pichai expressed pride in working at a company with five Nobel laureates, highlighting Google's culture of attracting and retaining top scientific talent, which is seen as a driver of continuous innovation [5] - For investors, these awards serve as important indicators of the company's innovative capacity and future potential, suggesting that Google's sustained investment in fundamental science is positioning it favorably for the next technological revolution [5]

Core Insights - Apple has launched a protein folding model called SimpleFold, which has been informally referred to as "iFold" by users [1][2]. Group 1: Model Overview - SimpleFold utilizes a general Transformer module combined with flow matching generation paradigm, achieving performance comparable to Google's AlphaFold2 with a 3 billion parameter version [2][4]. - The model simplifies the complex processes involved in protein folding, which traditionally required extensive computational resources and specialized designs [4][6]. Group 2: Technical Innovations - The core innovation of SimpleFold lies in the introduction of flow matching generation technology, which allows for a smooth mapping from random noise distribution to protein conformation distribution, enabling one-step generation of atomic coordinates [7]. - The training dataset for SimpleFold consisted of 9 million entries, resulting in multi-scale models ranging from 100 million to 3 billion parameters, with the SimpleFold-3B model achieving 95% of AlphaFold2's performance on the CAMEO22 benchmark [9]. Group 3: Performance Metrics - In the CAMEO22 benchmark, SimpleFold-3B achieved a TM-score of 0.837 and a GDT-TS score of 0.916, indicating strong performance relative to other models [10]. - The model also demonstrated efficiency, with inference times of only two to three minutes for processing 512 residue sequences on a MacBook Pro equipped with the M2 Max chip, significantly faster than traditional models that require hours [11]. Group 4: Research Team Background - The lead author of the research, Yuyang Wang, has a strong academic background in mechanical engineering and machine learning, with experience at Apple focusing on diffusion models [12]. - The corresponding author, Jiarui Lu, has a background in machine learning and has contributed to Apple's open-source project ToolSandbox, showcasing expertise in large model benchmarking [15].

Core Viewpoint - Apple has launched a cross-domain AI model named SimpleFold for protein folding prediction, which has been informally referred to as "iFold" by users [1]. Group 1: Model Overview - SimpleFold utilizes a straightforward design based on general Transformer modules, achieving performance comparable to Google's AlphaFold2 with a 3 billion parameter version [2][8]. - The model simplifies the complex processes involved in protein folding prediction, making it more accessible for ordinary laboratories [3][7]. Group 2: Technical Details - The core of protein folding involves predicting the three-dimensional structure of a protein from its amino acid sequence [5][6]. - SimpleFold employs a multi-layer Transformer encoder as its backbone, adapting protein sequence features through adaptive layer normalization [10]. - The key innovation lies in the introduction of flow matching generation technology, which allows for smooth mapping from random noise distribution to protein conformation distribution, enabling one-step generation of atomic coordinates [11][12]. Group 3: Performance Metrics - The training dataset for SimpleFold consisted of 9 million entries, resulting in multi-scale models ranging from 100 million to 3 billion parameters. The 3 billion parameter model achieved 95% of AlphaFold2's performance on the CAMEO22 benchmark [14]. - In the CASP14 high-difficulty test set, SimpleFold outperformed similar flow matching models like ESMFold [15]. Group 4: Efficiency - On a MacBook Pro equipped with the M2 Max chip, SimpleFold can process a sequence of 512 residues in just two to three minutes, significantly faster than traditional models that require hours [18]. Group 5: Research Team - The lead author of the research, Yuyang Wang, has a strong academic background with degrees from Tongji University and Carnegie Mellon University, focusing on mechanical engineering and machine learning [18]. - The corresponding author, Jiarui Lu, also has a solid educational foundation from Tsinghua University and Carnegie Mellon University, and has contributed to Apple's open-source project ToolSandbox [21][22].

Core Viewpoint - Apple has released a lightweight protein folding prediction AI model called SimpleFold, which utilizes flow matching methods to reduce computational costs while maintaining predictive performance, potentially advancing drug development and new material exploration [1][4]. Group 1: Technology and Innovation - SimpleFold replaces traditional complex modules like multiple sequence alignment with flow matching methods, significantly lowering computational costs and making protein-related research more accessible to various research teams [1][4]. - The flow matching technique, derived from diffusion models, allows for direct generation of protein structures from random noise, bypassing multiple denoising steps, thus enhancing generation speed and reducing computational load [4]. Group 2: Performance Evaluation - Multiple model versions of SimpleFold, ranging from 100 million to 3 billion parameters, were evaluated against the CAMEO22 and CASP14 benchmarks, focusing on generalization, robustness, and atomic-level accuracy [4]. - SimpleFold outperformed similar flow matching models like ESMFold and demonstrated performance comparable to leading protein folding prediction models [4][5]. Group 3: Comparative Performance Metrics - In the CAMEO22 test, SimpleFold achieved approximately 95% of the performance of AlphaFold2 and RoseTTAFold2, while the smaller SimpleFold-100M version exceeded 90% of ESMFold's performance, validating its competitive edge in the protein prediction field [5].