编辑丨王多鱼 排版丨水成文 如果说之前的 基因组大模型 是在逐字认识 ATCG 的排列组合，那么 Genos 团队此次推出的 Gengram 则相当于为其配备了一本" 基因字典 "。数据显示， 搭载 Gengram 后的大模型， 不仅刷新了多项基因组任务的 SOTA 记录，让模型在剪接位点识别等任务上 AUC 提升 16.1% ，还能自己悟出 DNA 双螺旋的物理规 律。 瓶颈：只会 "读字母"的基因组模型 在生物学中，很多 DNA 的功能元件 （例如启动子、剪接位点） 往往由特定的碱基组合 （Motif） 决定，但目前主流的基因组大模型都采用 单碱基分词 的形 式，也就是把 DNA 序列拆成一个个碱基来处理，这种方式虽然精准，但效率极低。 就像 我们看 "刻舟求剑"，是一眼看懂成语的含义，而不是先分析"刻"字有几笔，"舟"字怎么写，然后一个字一个字拼在一起，这不仅浪费算力，还容易让模型在 长达数亿的碱基序列中迷路。 近日， DeepSeek 新发布的 " 外挂大脑 "模式在大模型圈内 爆火 ，仅 16 天后， 国内 一个名叫 Genos 的 团队 将这种模式引入了 生命科学 领域， 提出了 Gengram ...

Core Insights - The launch of Genos, a universal foundational model for the human genome with 100 billion parameters, represents a significant advancement in genomic research, enabling precise single-base resolution identification and supporting ultra-long context analysis of up to one million base pairs [1][2] Group 1: Model Features and Capabilities - Genos integrates multiple public resources, including the Human Pan-Genome Reference Consortium and the Human Genome Structural Variation Map Project, utilizing 636 high-quality human genomes to reduce data bias and better represent human genetic diversity [2] - The model employs a mixture of experts (MoE) architecture, allowing it to activate only the most relevant experts for specific tasks, thus optimizing resource consumption while maintaining a vast knowledge base [2] - Genos has demonstrated superior performance in over half of the classic evaluation tasks, including genome element recognition and mutation pathogenicity prediction, showcasing its strong contextual analysis capabilities [3] Group 2: Clinical and Research Applications - Genos provides a new efficient tool for clinical diagnosis, achieving high accuracy in interpreting pathogenic mutations, especially when combined with the 021 scientific foundational model [3] - The model is designed for easy deployment and use, with both 12 billion and 100 billion parameter versions available, making it accessible for various applications [4] - Genos has been integrated with the DCS Cloud platform, allowing users to predict RNA expression profiles in seconds based solely on DNA sequences, significantly accelerating biological data analysis [5]

Core Insights - The completion of the Human Genome Project in 2003 has provided a foundational understanding of life, yet challenges remain in interpreting this vast amount of genetic data. The launch of the Genos model by BGI Life Sciences and Zhijiang Laboratory aims to address these challenges by offering a deployable genomic universal foundational model with 100 billion parameters [1]. Group 1: Genos Model Advantages - Genos enhances both interpretative and deployment capabilities compared to existing models, primarily due to the expansion of training data. It integrates multiple authoritative public resources, including the Human Pan-Genome Reference Consortium (HPRC) and the Human Genome Structural Variation Consortium (HGSVC), utilizing 636 high-quality human genomes to better reflect human genetic diversity [2]. - The model's deployment capability is improved through a "mixed expert" architecture, which efficiently coordinates relevant algorithms to reduce inference costs and resource consumption, addressing the common issue of large models being difficult to deploy. Genos is also open-source, available in both 1.2 billion and 100 billion parameter versions on platforms like HuggingFace and Modao [2]. Group 2: Clinical and Research Implications - Clinical testing has demonstrated Genos's effectiveness, achieving a 92% accuracy rate in interpreting pathogenic mutations, which increases to 98.3% when combined with scientific foundational models. This performance surpasses existing top-level models [2]. - In research, AI models like Genos can transform the process of identifying pathogenic loci from a "needle in a haystack" approach to "precise navigation," significantly reducing the time required for rare disease and complex mechanism studies [3]. Group 3: Future Prospects and Challenges - The advancement of AI models in genomics is expected to shift drug development from a "trial-and-error" approach to "design-based" innovation, thereby reducing experimental iterations and costs [4]. - Future development faces three main challenges: expanding training databases to include more disease samples, establishing comprehensive ethical and safety standards, and enhancing interdisciplinary collaboration to integrate AI with clinical data systems and biological experimental platforms [4]. - The successful implementation of AI in genomics is anticipated to accelerate the arrival of precision medicine, with the Genos model being a significant step towards unlocking the potential of the life economy [4].

Core Insights - The article reports the launch of the world's first deployable genomic universal foundational model, Genos, by BGI Life Sciences Research Institute and Zhijiang Laboratory, featuring over 100 billion parameters [1] - Genos supports ultra-long context analysis of up to one million base pairs and achieves precise identification at single-base resolution [1] Group 1: Model Features - Genos integrates multiple authoritative public resources, including the Human Pan-Genome Reference Consortium (HPRC) and the Human Genome Structural Variation Consortium (HGSVC) [1] - The model utilizes high-quality human genomes from 636 "telomere-to-telomere" level datasets, covering diverse global populations, enhancing the understanding of human genetic diversity [1] Group 2: Availability - The Genos model is fully open-sourced on platforms such as HuggingFace and Modao, offering two versions with 1.2 billion and 10 billion parameters to meet varying needs [1]

Group 1 - ByteDance's Seed team launched a 3D generative model called Seed3D 1.0 [1] - Kuaishou's StreamLake officially introduced a "tool + model + platform" AI programming product matrix [1] - Inspur released China's first one-stop development platform for scientific large models named OneScience [1] Group 2 - The general-purpose model of the human genome with 100 billion parameters, named "Genos," was officially released [1] - Alibaba's Quark AI glasses started pre-sale on the 24th, with prices starting at 3,699 yuan [1] - Elon Musk stated that the third generation of Optimus may be released in the first quarter of next year [1] Group 3 - The National People's Congress Standing Committee's Legislative Affairs Commission proposed to add content promoting AI safety and development to the Cybersecurity Law [1] - The UK government reached an agreement with OpenAI to expand data hosting capabilities [1]

Core Viewpoint - The launch of Genos, the world's first hundred billion parameter universal basic model for the human genome, represents a significant advancement in genomic research, transitioning from merely reading base sequences to understanding the underlying logic of life [1] Group 1: Product Development - Genos is a genome foundational model optimized for the human genome, capable of supporting ultra-long context analysis of up to one million base pairs [1] - The model achieves precise identification at single-base resolution, enhancing the accuracy of genomic analysis [1] Group 2: Impact on Industry - The introduction of Genos is expected to bring transformative changes to clinical disease diagnosis, personal genome interpretation, and cutting-edge scientific research [1] - The related findings have been published in the international academic journal GigaScience, indicating the model's credibility and relevance in the scientific community [1]

Core Insights - The launch of Genos, the world's first deployable genomic universal foundational model with over 100 billion parameters, was announced by BGI Life Sciences Research Institute and Zhijiang Laboratory [1][2] - Genos supports ultra-long context analysis of up to one million base pairs and achieves precise identification at single-base resolution [1] - The model integrates high-quality human genomes from various authoritative public resources, covering 636 "telomere-to-telomere" level genomes, enhancing the understanding of human genetic diversity [1] Model Performance - Genos demonstrates a 92% accuracy rate in interpreting pathogenic mutations directly related to clinical applications, which increases to 98.3% when combined with scientific foundational models [1] - The model surpasses existing best-level models in multiple comprehensive evaluations [1] Accessibility - Genos is fully open-sourced on platforms like HuggingFace and Modao, offering two versions with 1.2 billion and 100 billion parameters to meet diverse needs [2]