谷歌母公司Alphabet的科学家连续两年斩获诺贝尔奖，再次凸显了这家科技巨头在人工智能和量子计算 等前沿基础研究领域的深厚实力，这些技术被认为将对未来商业和市场格局产生颠覆性影响。 最新动态是，2025年诺贝尔物理学奖授予了三位在量子力学领域做出开创性贡献的物理学家，其中包括 谷歌量子AI实验室的现任硬件首席科学家Michel Devoret，以及曾领导其硬件团队多年的John Martinis。 这一荣誉不仅是对他们个人成就的认可，也为谷歌在下一代计算技术竞赛中的领先地位提供了有力背 书。 谷歌首席执行官Sundar Pichai迅速对此表示祝贺，并强调获奖者的研究为公司在量子计算领域取得的最 新突破奠定了基础。他在社交媒体上表示，他们的工作为未来纠错量子计算机的实现铺平了道路。这一 表态向市场传递了明确信号：谷歌对基础科学的长期投资正在转化为其在未来关键技术领域的核心竞争 力。 聚焦量子突破 本年度的诺贝尔物理学奖授予了Michel Devoret、John Martinis和John Clarke，以表彰他们"在电路中发 现宏观量子力学隧穿和能量量子化"。据瑞典皇家科学院称，获奖者通过一系列实验证明， ...

起猛了，苹果怎么搞起跨界AI模型了？？ 发布了一个基于流匹配的蛋白质折叠模型SimpleFold，被网友戏称为"iFold"。 SimpleFold没有花里胡哨的专属模块设计，就靠通用的Transformer模块，搭配流匹配生成范式，3B参数版本追平了该领域顶流模型谷歌AlphaFold2的性 能。 苹果这波跨界看来玩的是化繁为简。 MacBook Pro跑起来不费力 首先来说说蛋白质折叠是怎么一回事。 核心是将"一串"氨基酸折成特定的3D形状，这样蛋白质才能发挥作用。 而蛋白质折叠模型就是从氨基酸的一级序列预测它的三维空间构象。 之前最厉害的模型，比如谷歌的AlphaFold2，虽然实现了突破，但用了很多复杂的专属设计。 比如要分析大量相似蛋白质的序列，依赖多序列对比（MSA）构建进化信息、靠三角注意力优化空间约束、推理时需调用超算级算力，普通实验室不太 能用得起。 但这款"iFold"用通用AI框架解决了这个问题。 不同于扩散模型的逐步去噪，流匹配通过学习从随机噪声分布到蛋白质构象分布的光滑映射，实现一步式生成原子坐标。 在训练阶段，团队构建了包含900万条数据的混合数据集，训练出了100M到3B参数的多 ...

闻乐 发自 凹非寺 量子位 | 公众号 QbitAI 苹果这波跨界看来玩的是 化繁为简 。 起猛了，苹果怎么搞起跨界AI模型了？？ 发布了一个基于 流匹配 的蛋白质折叠模型 SimpleFold ，被网友戏称为"iFold"。 SimpleFold没有花里胡哨的专属模块设计，就靠通用的Transformer模块，搭配流匹配生成范式，3B参数版本追平了该领域顶流模型谷歌 AlphaFold2的性能。 MacBook Pro跑起来不费力 首先来说说蛋白质折叠是怎么一回事。 核心是将"一串"氨基酸折成特定的3D形状，这样蛋白质才能发挥作用。 而蛋白质折叠模型就是从氨基酸的一级序列预测它的三维空间构象。 之前最厉害的模型，比如谷歌的AlphaFold2，虽然实现了突破，但用了很多复杂的专属设计。 比如要分析大量相似蛋白质的序列，依赖多序列对比（MSA）构建进化信息、靠三角注意力优化空间约束、推理时需调用超算级算力，普通 实验室不太能用得起。 但这款"iFold"用通用AI框架解决了这个问题。 SimpleFold在架构上采用多层Transformer编码器作为核心骨干，仅通过自适应层归一化适配蛋白质序列特征，相当于用 ...

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].

Core Viewpoint - The article discusses the breakthrough technology BindCraft, which allows for the one-shot design of functional protein binders with a success rate of 10%-100%, significantly improving the efficiency of protein design compared to traditional methods [2][3][5]. Summary by Sections BindCraft Technology - BindCraft is an open-source, automated platform for de novo design of protein binders, achieving high-affinity binders without the need for high-throughput screening or experimental optimization [3][5]. - The technology leverages AlphaFold2's weights to generate protein binders with nanomolar affinity, even in the absence of known binding sites [3][5]. Applications and Results - The research team successfully designed binders targeting challenging targets such as cell surface receptors, common allergens, de novo proteins, and multi-domain nucleases like CRISPR-Cas9 [3][7]. - Specific applications include: 1. Designing antibody drugs targeting therapeutic cell surface receptors like PD-1 and PD-L1, achieving nanomolar affinity without extensive design and screening [7]. 2. Blocking allergens, with a designed binder for birch pollen allergen Bet v1 showing a 50% reduction in IgE binding in patient serum tests [7][8]. 3. Regulating CRISPR gene editing by designing a new inhibitory protein that significantly reduces Cas9's gene editing activity in HEK293 cells [8]. 4. Neutralizing deadly bacterial toxins, with a designed protein completely eliminating cell death caused by the toxin from Clostridium perfringens [8]. 5. Modifying AAV for targeted gene delivery by integrating mini-binders that specifically target HER2 and PD-L1 expressing cancer cells [8]. Impact and Future Potential - BindCraft addresses long-standing success rate bottlenecks in protein design and offers direct solutions for allergy treatment, gene editing safety, toxin neutralization, and targeted gene therapy [9]. - The open-source nature of the technology allows ordinary laboratories to design custom proteins, potentially reshaping drug development, disease diagnosis, and biotechnology [9].

Core Viewpoint - Latent Labs has developed a groundbreaking generative AI model, Latent-X, which enables the design of functional protein binders with atomic-level precision, significantly improving the drug discovery process [6][7][26]. Group 1: Company Background - Simon Kohl, a former researcher at DeepMind, founded Latent Labs after leaving the company in late 2022, focusing on advanced protein design models to aid biopharmaceutical companies [2]. - Latent Labs secured $50 million in funding in February 2025 to further its mission in drug development [2]. Group 2: Technology and Innovation - Latent-X can design functional protein binders, including macrocyclic peptides and small protein binders, with unprecedented efficiency and accuracy [6][7]. - The model generates reliable protein binders by solving geometric challenges at the atomic level, producing high-affinity and specificity binders [7][20]. - Latent-X demonstrated a significant improvement in efficiency, requiring only 30-100 candidates per target to achieve results that typically need millions of candidates [7][18]. Group 3: Performance Validation - The research team tested Latent-X on seven benchmark target proteins related to viral infections, tumor regulation, and neurodegeneration [11][12]. - Latent-X achieved a hit rate of 91%-100% for macrocyclic peptides and 10%-64% for small protein binders across the target proteins [18]. - The best-designed macrocyclic peptides reached micromolar affinity, while small protein binders achieved picomolar affinity, surpassing other design models [19]. Group 4: Features and Usability - Latent-X allows users to generate both protein sequences and structures simultaneously, outperforming previous methods that generated them sequentially [23]. - The platform is user-friendly, requiring no coding skills, and provides a complete workflow for laboratory validation [21][29]. - Latent-X is scalable and has successfully generated various therapeutic binders, with plans for further expansion [22]. Group 5: Competitive Advantage - Latent-X excels in generating binders for previously unseen targets, achieving higher simulation hit rates with fewer samples compared to other models [24][28]. - The model adheres to atomic-level biochemical rules, creating structures that are chemically viable and suitable for drug development [28].

Core Viewpoint - The article discusses the imminent transformation brought by artificial intelligence (AI) and robotics, emphasizing that 8 billion people are already involved in this change, with predictions that robots will outnumber humans in the next decade [1][10]. Group 1: AGI Development - Predictions suggest that Artificial General Intelligence (AGI) could be achieved within five years, requiring the integration of three intelligence waves: generative AI, robotics, and AI for Science [5]. - Current focus remains on information intelligence, with advancements in generative AI like ChatGPT demonstrating conversational capabilities, while natural image and video generation still require 4-5 years of development [5]. Group 2: Challenges in AGI - AGI development faces challenges such as "boundary cognition loss" in large language models, leading to the generation of "hallucinations" or fabricated information [6]. - While the hallucination rate has decreased, new model issues have emerged, necessitating context-specific responses in applications like art creation and information retrieval [6]. Group 3: AI for Science - AI for Science is viewed as a transformative force in research, with initiatives like the Tsinghua University Intelligent Industry Research Institute focusing on creating cross-disciplinary foundational models to enhance drug discovery and molecular screening [8]. - AI can significantly narrow down drug candidate searches from billions to millions, improving efficiency in addressing the over 90% of diseases lacking effective treatments [8]. Group 4: Robotics and Future Predictions - Human-like robots are identified as a breakthrough in AI physical intelligence, with predictions that their numbers will surpass humans in ten years [10]. - China is expected to lead the global humanoid robot industry, leveraging its complete supply chain, a young engineering talent pool, and a large unified market to replicate the success of the mobile internet era [10].

Core Viewpoint - The article discusses the introduction of AlphaGenome, a new AI tool by DeepMind that predicts the effects of single nucleotide mutations in human DNA sequences, enhancing the understanding of gene regulation and disease biology [2][3]. Group 1: AlphaGenome Overview - AlphaGenome is a DNA sequence model that can process up to 1 million base pairs and predict various molecular characteristics related to gene regulation [2][9]. - The model builds on previous DeepMind models like Enformer and complements AlphaMissense, focusing on the 98% of the genome that is non-coding and crucial for gene regulation [10][12]. Group 2: Unique Features of AlphaGenome - AlphaGenome offers high-resolution predictions in the context of long DNA sequences, allowing for detailed biological insights without compromising on sequence length or resolution [12]. - It provides comprehensive multi-modal predictions, enabling scientists to gain a deeper understanding of complex gene regulation processes [13]. - The model can efficiently score mutations, assessing their impact on various molecular characteristics in just one second [14]. - AlphaGenome can directly model splicing sites, which is significant for understanding rare genetic diseases [15]. - It achieves state-of-the-art performance across various genomic prediction benchmarks, outperforming or matching existing models in multiple evaluations [16][18]. Group 3: Applications and Research Directions - AlphaGenome can aid in disease understanding by accurately predicting the effects of gene disruptions, potentially identifying new therapeutic targets [23]. - Its predictions can guide the design of synthetic DNA with specific regulatory functions [24]. - The model accelerates basic research by helping to map key functional elements of the genome [25]. - DeepMind researchers have utilized AlphaGenome to explore mechanisms related to cancer mutations, demonstrating its capability to link non-coding mutations to disease genes [26][27]. Group 4: Limitations and Future Directions - Despite its advancements, AlphaGenome faces challenges in capturing the effects of regulatory elements that are far apart in the genome [32]. - The model has not been specifically designed or validated for individual genome predictions, limiting its application in complex traits or diseases influenced by broader biological processes [32]. - DeepMind is continuously improving the model and collecting feedback to address these limitations [32]. - Currently, the API is open for non-commercial use, focusing on scientific research rather than direct clinical applications [32].

Core Viewpoint - The emergence of D-I-TASSER, a new protein structure prediction tool, demonstrates significant advancements in protein folding prediction, outperforming existing models like AlphaFold2 and AlphaFold3 in accuracy and coverage [3][8]. Group 1: D-I-TASSER Development and Performance - D-I-TASSER was developed by a collaborative research team and has shown superior performance in the CASP15 competition, excelling in both single-domain and multi-domain protein structure predictions [3][8]. - The tool successfully predicted structures for 19,512 proteins from the human proteome, achieving 81% domain coverage and 73% full-length sequence coverage, which is a notable improvement over AlphaFold2 [3][12][14]. - D-I-TASSER integrates deep learning with physical simulations, utilizing multiple sources of information to enhance prediction accuracy [8][14]. Group 2: Technical Innovations - The core innovation of D-I-TASSER lies in its hybrid approach, combining deep learning with physical modeling to refine protein structure predictions [8][17]. - The tool employs an upgraded DeepMSA2 for multi-sequence alignment, increasing information retrieval from metagenomic databases by 6.75 times [11]. - D-I-TASSER's modeling process includes a unique workflow of automatic domain cutting, independent prediction, and dynamic assembly, resulting in improved accuracy and reduced orientation errors [8][11]. Group 3: Challenges and Future Directions - Despite its impressive performance, D-I-TASSER faces challenges such as reduced prediction accuracy for orphan proteins and higher computational time compared to pure deep learning models [20]. - The research indicates that the ultimate solution to protein folding may lie in the deep synergy between data-driven methods and physical simulations [17][20]. - The D-I-TASSER model and its human protein structure prediction database have been made open-source, promoting further research and collaboration in the field [17].

整理｜冬梅、核子可乐 去年成功斩获诺贝尔奖之后，Demis Hassabis 决定与一位国际象棋世界冠军打场扑克以示庆 祝。Hassabis 一直痴迷于游戏，这股热情也成为他 AI 先驱之路上的契机与驱力。 近日，做客一档名为《60 分钟》的访谈栏目，讲述了他如何带领众多研究者追逐新的技术"圣 杯"——通用人工智能（AGI），一种兼具人类灵活性与超人般速度与知识储备的硅基智能形态。 除此之外，他也在访谈中透露了 DeepMind 未来的研究方向以及有可能亮相的产品和技术。 "天才少年"Hassabis AI 之旅始于国际象棋 Hassabis 于 2010 年与他人共同创立了了 AI 公司 DeepMind，2014 年该公司被谷歌以 5 亿多 美元收购。2017 年，他发明了 AI 算法 AlphaZero，它只需要国际象棋规则和四个小时的自对 弈，就能成为有史以来最强的国际象棋选手，击败人类国际象棋大师。 2024 年，Hassabis 与同为诺贝尔化学奖得主的 DeepMind 总监约翰·江珀 (John Jumper) 共同 获得了诺贝尔化学奖，获奖原因是他创建了一个 AI 模型 AlphaFold2 ...