Science重磅：AI首次生成完整噬菌体基因组，可存活、能抑菌
Ranking
Observed public metrics from 1 member.
Merged summary
TL;DR — 斯坦福团队（King et al., Science, 2026）用基因组语言模型 Evo 1/Evo 2 从头生成完整噬菌体基因组，获得 16 株可存活、能裂解大肠杆菌 C 的合成噬菌体，首次证明生成式 AI 能设计基因组尺度的功能性生命系统，而不止于单基因、单蛋白或基因线路。
- 模型与生成流程：Evo 1/Evo 2 在超 200 万条噬菌体基因组上预训练，再用约 15,000 条微小噬菌体科（Microviridae）序列微调，以 ΦX174（5.4 kb 单链 DNA、11 个基因）起始区共识序列作提示（约 4–9 nt，采样温度 0.7–0.9），Evo 2 的基础生成能力更强。
- 多层过滤与产出：按长度、GC 含量、编码密度、ΦX174 式基因架构、宿主嗜性（主刺突蛋白相似度）及与野生型的进化距离筛选；302 条设计合成后 285 条组装成功，16 株（Evo-Φ）可抑制/裂解 E. coli C，且不抑制另外 6 株菌，宿主特异性保留；存活率与同天然基因组的相似度正相关。
- 新颖序列组合：设计中出现基因增删、基因顺序改变和非编码调控区变化，13 株可存活基因组带有无法由任何单一天然序列解释的突变；复制起点却高度保守，说明模型隐式学到了区域性约束。Evo-Φ36 将 DNA 包装蛋白 J 换成远缘噬菌体 G4 的较短同源物（此前文献报道该组合不可行），冷冻电镜确认衣壳相互作用兼容、颗粒稳定。
- 功能与抗性突破：生成噬菌体的裂解动力学跨度大于天然 ΦX174 类噬菌体；其鸡尾酒组合在 1–2 轮传代内攻克 ΦX174 抗性菌株 CR1/CR2，而天然噬菌体鸡尾酒 5 轮后仍失败，突破株源自 2–3 个生成基因组之间的重组——提示噬菌体疗法价值。
- 局限与治理：目前仅限小基因组，受训练数据偏倚限制；作者明确提示生物安全与生物安保风险，呼吁安全专家介入设计全生命周期并加强审查。
不同来源侧重：学术头条更强调抗性突破细节与生物安全治理呼吁，DrugAI 更强调模型机制、过滤条件与"基因组尺度生成"的里程碑定位。
Sources (2)
Science重磅：AI首次生成完整噬菌体基因组，可存活、能抑菌
TL;DR - Stanford researchers published in Science the first AI-generated complete phage genomes, using the Evo 1/Evo 2 DNA foundation models to design viable ΦX174-like bacteriophages that infect and lyse E. coli C. It marks a shift from single-gene/circuit design to generative whole-genome design, with direct implications for phage therapy and biosecurity governance.
- Pipeline: Evo 1/Evo 2 (pretrained on >2M phage genomes) were fine-tuned on ~15,000 Microviridae sequences, prompted with ΦX174-like start consensus (~4–9 nt prompts, sampling temperature 0.7–0.9), then filtered by length, GC content, ΦX174-like gene architecture, host tropism, and distance from wild-type; Evo 2 showed stronger base generation.
- Yield and specificity: ~300 designed genomes tested → 16 viable phages; viability correlated with similarity to known natural genomes, and all infected E. coli C without inhibiting 6 other tested strains.
- Novel sequence combinations: 13 of the viable genomes carried mutations unexplained by any single natural sequence; Evo-Φ36 swapped DNA packaging protein J for a shorter homolog from distant phage G4 — previously reported as non-viable — and cryo-EM confirmed compatible capsid interactions.
- Function and resistance: generated phages spanned a wider range of lysis kinetics than natural ΦX174-like phages, and cocktails overcame ΦX174-resistant strains CR1/CR2 in 1–2 passages (natural cocktails failed after 5), with breakthrough phages arising from recombination of 2–3 generated genomes; authors flag biosafety/biosecurity risks and call for safety experts across the design lifecycle.
Science | 利用基因组语言模型进行噬菌体从头生成设计
TL;DR - A Science paper (King et al., 2026) uses genome language models (Evo 1/Evo 2) to generate complete bacteriophage genomes de novo, yielding 16 experimentally viable synthetic phages — the first demonstration that generative AI can design functional genome-scale living systems, not just single proteins or circuits.
- Evo models were fine-tuned on ~15,000 Microviridae genomes and prompted with the ΦX174 start region (5.4 kb ssDNA, 11 genes); multi-layer filters enforced host tropism (major spike protein similarity), genome length, GC content, coding density, and evolutionary divergence.
- Of 302 designs synthesized, 285 assembled and 16 ("Evo-Φ") suppressed E. coli C growth, retaining target-host specificity while showing varied infection kinetics.
- Designs contained gene gains/losses, altered gene order, and non-coding regulatory changes; the replication origin stayed highly conserved, implying the model implicitly learned regional constraint. One Evo-Φ borrowed a DNA-packaging protein from distant phage G4, confirmed as a stable particle by cryo-EM.
- Cocktails of designed phages rapidly overcame ΦX174-resistant E. coli, while natural ΦX174-like cocktails did not — suggesting therapeutic value. Limits noted: small genomes only, training-data bias, and biosafety review needs.