遗传

• 前沿聚焦 •    

从“读写基因组”到“生成基因组”:人工智能驱动的基因组设计与合成

黄煜,许鸿翔,谢建平   

  1. 西南大学生命科学学院现代生物医药研究所,重庆 400715
  • 收稿日期:2026-08-12 修回日期:2026-10-16 发布日期:2026-11-03
  • 基金资助:
    国家自然科学基金项目(编号:82472325)[Supported by the National Natural Science Foundation of China (No. 82472325)

From genome reading and writing to genome generation: artificial intelligence in genome design and synthesis

Yu Huang, Hongxiang Xu, Jianping Xie   

  1. Institute of Modern Biopharmaceuticals, College of Life Sciences, Southwest University, Chongqing 400715, China
  • Received:2026-08-12 Revised:2026-10-16 Online:2026-11-03

摘要: DNA测序、编辑和化学合成技术的发展使生命科学逐渐具备在基因组尺度“读”和“写”遗传信息的能力,但如何从功能目标出发理性设计完整基因组,长期受到基因间上位性、调控元件耦联、基因组空间结构以及宿主适应等多尺度复杂性的限制。近年来,以基因组语言模型和生物学基础模型为代表的人工智能技术开始直接从海量天然DNA序列中学习进化形成的序列约束,并由基因功能预测逐步拓展至调控元件、复杂分子系统乃至基因组尺度的生成设计。2026年,King等利用Evo和Evo 2基因组语言模型,以噬菌体ΦX174为模板生成完整噬菌体基因组,并通过DNA合成和实验筛选获得16种具有感染活性的人工设计噬菌体,首次建立了从生成式人工智能设计到完整基因组实验验证的闭环。本文以该工作为切入点,概述了人工智能在DNA序列建模、功能元件生成、多基因系统设计及完整基因组设计方面的主要进展,分析了人工智能与DNA合成、组装及设计—构建—测试—学习循环融合的发展趋势,并讨论了当前模型架构区别、训练目标、实验成功率、合成成本以及生物安全等关键问题。随着基因组基础模型、高通量合成和自

关键词: 人工智能, 基因组语言模型, 生成式基因组学, 基因组设计, DNA合成, 合成生物学

Abstract:

Advances in DNA sequencing, genome editing, chemical synthesis and large-fragment assembly have provided increasingly powerful capabilities to read and write genetic information at the genome scale. However, compared with the rapidly improving ability to synthesize DNA, the rational design of complete genomes from desired biological functions remains a major challenge in synthetic biology. Genome function emerges from multiscale interactions among coding sequences, regulatory elements, noncoding regions, genome architecture, epistasis and host environments, rather than from a simple combination of individual genes. Recent advances in artificial intelligence (AI), particularly DNA foundation models and genomic language models, offer a new paradigm in which statistical and evolutionary constraints can be learned directly from massive collections of natural genomic sequences. AI-enabled genome engineering has consequently progressed from sequence-function prediction toward generative design of regulatory elements, multigene systems and genome-scale sequences. In 2026, King et al. used the genomic language models Evo 1 and Evo 2 to generate complete bacteriophage genomes guided by the ΦX174 genome. Computational selection followed by DNA synthesis and experimental screening yielded 16 viable designed phages, establishing an integrated loop from generative AI and genome design to DNA synthesis and functional validation. This review takes that study as a starting point and summarizes major advances in artificial intelligence for DNA sequence modeling, functional element generation, multigene system design, and complete genome design. It analyzes the trend of integrating AI with DNA synthesis, assembly, and the design-build-test-learn cycle, and discusses key issues including current model architectures, training objectives, experimental success rates, synthesis costs, and biosafety. With the further integration of genomic foundation models, high-throughput synthesis, and automated experimental platforms, genome engineering is expected to gradually move from "modifying natural genomes" toward "generating genomes by function."

Key words: artificial intelligence, genomic language model, generative genomics, genome design, DNA synthesis, synthetic biology