Evo 2：DNAを「書く」ことができるAI——それが合成生物学を変える理由

Evo 2: The AI That Can “Write” DNA—And Why It Changes Synthetic Biology

AIがDNAを「書く」時代へ——9兆塩基対で訓練されたゲノム基盤モデル「Evo 2」が、遺伝子生成や変異予測、さらには人工バクテリオファージの設計まで実現し、合成生物学の転換点を示している。

分からないところをタップすると
↓日本語訳が表示されます↓

Can AI write DNA? In a limited but astonishing sense, yes. On March 4, 2026, Nature published the Evo 2 study, describing a genome foundation model trained on about 9 trillion DNA base pairs from more than 128,000 genomes across the tree of life. Built by Arc Institute, NVIDIA, and academic collaborators, Evo 2 can analyze sequence patterns at single-nucleotide resolution across contexts as long as 1 million base pairs. (nature.com)

What makes Evo 2 remarkable is that it does not merely read DNA; it can also generate it. According to the Nature paper, the model can complete genes and produce genome-scale sequences resembling human mitochondrial DNA, the minimal bacterium Mycoplasma genitalium, and even long yeast chromosomal regions. It also predicts the effects of mutations, including clinically important BRCA1 variants, without task-specific fine-tuning. In other words, Evo 2 treats genomes less like static code and more like a language with grammar, style, and long-range structure. (nature.com)

Still, “AI creates life” would be an exaggeration. Evo 2 does not conjure living organisms out of thin air. In practice, the frontier looks more like a read-write-test loop: AI proposes sequences, researchers synthesize them, and experiments reveal whether biology accepts the design. Arc reports that Evo 2-guided regulatory DNA was synthesized and tested in cells, with some designs producing the predicted chromatin-accessibility patterns. The group also reported the first AI-designed, experimentally validated bacteriophages; after supervised fine-tuning and lab screening, 16 of 285 tested designs successfully propagated in bacteria. (arcinstitute.org)

That is why Evo 2 feels like a turning point. It suggests that synthetic biology may shift from slow trial-and-error toward guided exploration of genetic possibility. At the same time, the safety question is impossible to ignore. Arc says eukaryotic viruses were excluded from Evo 2’s training data, and red-team tests found that attempts to generate pathogenic human viral proteins produced effectively random results. So yes, AI can now “write” DNA—but only under constraints, and only with human judgment, laboratory validation, and serious biosafety guardrails. (arcinstitute.org)

会員登録して
読んだ語数を記録する

AIはDNAを書けるのか？　限定的ではあるが驚くべきことに、答えはイエスだ。2026年3月4日、Natureはゲノム基盤モデルについて記述したEvo 2の研究を発表した。このモデルは、生命の系統樹全体にわたる128,000以上のゲノムから得られた約9兆のDNA塩基対で訓練されている。Arc Institute、NVIDIA、および大学の共同研究者らによって構築されたEvo 2は、100万塩基対にも及ぶ長い文脈にわたって、1塩基単位の解像度で配列パターンを解析することができる。(nature.com)

Evo 2が注目に値するのは、DNAを単に読み取るだけでなく、生成することもできるという点だ。Natureの論文によると、このモデルは遺伝子を補完し、ヒトミトコンドリアDNA、最小の細菌であるマイコプラズマ・ジェニタリウム（Mycoplasma genitalium）、さらには酵母の長い染色体領域に類似した、ゲノム規模の配列を生成することができる。また、タスク特化型のファインチューニングなしに、臨床的に重要なBRCA1変異体を含む変異の影響を予測することもできる。言い換えれば、Evo 2はゲノムを静的なコードとしてではなく、文法やスタイル、長距離構造を持つ言語のように扱っているのだ。(nature.com)

とはいえ、「AIが生命を創造する」というのは誇張であろう。Evo 2は何もないところから生物を生み出すわけではない。実際の最前線は、読み取り・書き込み・テストのループに近い。AIが配列を提案し、研究者がそれを合成し、実験によって生物学的にその設計が受け入れられるかどうかが明らかになる。Arcの報告によると、Evo 2が設計した制御DNAが合成されて細胞内でテストされ、一部の設計では予測どおりのクロマチンアクセシビリティパターンが生じた。同グループはまた、AIが設計し実験的に検証された初のバクテリオファージについても報告した。教師ありファインチューニングとラボでのスクリーニングを経て、テストされた285の設計のうち16が細菌内で増殖に成功した。(arcinstitute.org)

これこそが、Evo 2がターニングポイントのように感じられる理由だ。合成生物学が、遅い試行錯誤から、遺伝的可能性のガイド付き探索へと移行する可能性を示唆している。同時に、安全性の問題は無視できない。Arcによれば、真核生物ウイルスはEvo 2の訓練データから除外されており、レッドチームテストでは、病原性のあるヒトウイルスタンパク質を生成しようとする試みは実質的にランダムな結果しか生まなかったという。つまり、AIは今やDNAを「書く」ことができる——ただし、制約の下でのみ、そして人間の判断、実験室での検証、そして厳格なバイオセーフティのガードレールがあってこそ、である。(arcinstitute.org)

文法

●Cleft / pseudo-cleft for emphasis: “What makes Evo 2 remarkable is that …” → highlights the key point; common in academic writing to foreground evaluation.
●Hedging + stance with modals/conditionals: “would be an exaggeration,” “may shift,” “is impossible to ignore” → cautious, evidence-based claims typical of scientific English.
●Contrastive comparison patterns: “does not merely A; it can also B,” and “less like X and more like Y” → used to reframe an idea and emphasize a new perspective.

語彙

●foundation model: a large, general-purpose model trained on massive data, adaptable to many tasks
●genome / genomic: the full set of genetic material; “genome-scale” = on the scale of an entire genome
●(DNA) base pair: the paired building blocks (A–T, C–G) of DNA; used to measure sequence length
●single-nucleotide resolution: detail at the level of one nucleotide (one “letter” of DNA)
●context (as long as 1 million base pairs): the surrounding sequence window the model can consider at once
●(task-specific) fine-tuning: additional training for a particular task/domain; “without … fine-tuning” implies strong generalization
●conjure (up): to produce as if by magic; often used to deny unrealistic expectations (“not conjure … out of thin air”)
●frontier: the cutting edge / newest area of research or development
●regulatory DNA: DNA that controls gene expression rather than coding for proteins
●guardrails: practical limits/safeguards to reduce risk (common in AI policy/engineering contexts)

表現・慣用句

●in a limited but astonishing sense: a nuanced concession—yes, but only within specific boundaries
●(out) of thin air: from nothing, without any material basis (“does not conjure living organisms out of thin air”)
●trial-and-error: repeated attempts with adjustments based on failures; contrasts with “guided exploration”
●a turning point: a moment when a trend changes direction; implies historical significance
●read-write-test loop: an engineering-style cycle (analyze → generate → experimentally validate), used here as a metaphor for modern biology workflows

by EigoBoxAI
作成:2026/03/18 03:04
レベル:上級 (語彙目安:6000〜8000語)

# Evo 2：DNAを「書く」ことができるAI——それが合成生物学を変える理由
## Evo 2: The AI That Can “Write” DNA—And Why It Changes Synthetic Biology

![thumbnail](https://eigobox.s3.ap-northeast-1.amazonaws.com/g/060d3b7245bbcd3028d687c6b961a17c2f22ba12.png)

---

[["Can AI write DNA?","AIはDNAを書けるのか？"],["In a limited","限定的ではあるが"],["but astonishing sense, yes.","驚くべき意味で、イエスだ。"],["On March 4, 2026,","2026年3月4日、】【"],["Nature published","Nature誌は発表した"],["the Evo 2 study,","Evo 2の研究を、"],["describing a genome","ゲノム基盤モデルを"],["foundation model","記述したものであり、"],["trained on about","約9兆の"],["9 trillion DNA base pairs","DNA塩基対で訓練された"],["from more than 128,000 genomes","128,000以上のゲノムから得た"],["across the tree of life.","生命の系統樹全体にわたる。"],["Built by Arc Institute,","Arc Institute、"],["NVIDIA,","NVIDIA、"],["and academic collaborators,","および学術協力者によって構築された"],["Evo 2 can analyze","Evo 2は解析できる"],["sequence patterns","配列パターンを"],["at single-nucleotide resolution","一塩基レベルの解像度で"],["across contexts as long as","最大で"],["1 million base pairs.","100万塩基対の文脈にわたって。"],["([nature.com](https://www.nature.com/articles/s41586-026-10176-5))","（[nature.com](https://www.nature.com/articles/s41586-026-10176-5)）"],["What makes Evo 2 remarkable","Evo 2が注目に値する理由は"],["is that it does not","それが単に"],["merely read DNA;","DNAを読むだけでなく、"],["it can also generate it.","生成もできるということだ。"],["According to the Nature paper,","Nature論文によると、"],["the model can complete genes","このモデルは遺伝子を補完し"],["and produce","そして生成できる"],["genome-scale sequences","ゲノム規模の配列を"],["resembling human","ヒトの"],["mitochondrial DNA,","ミトコンドリアDNA、"],["the minimal bacterium","最小細菌"],["Mycoplasma genitalium,","マイコプラズマ・ジェニタリウム、"],["and even long","さらには長い"],["yeast chromosomal regions.","酵母の染色体領域に類似した。"],["It also predicts","また予測する"],["the effects of mutations,","変異の影響を、"],["including clinically important","臨床的に重要な"],["BRCA1 variants,","BRCA1変異体を含む、"],["without task-specific fine-tuning.","タスク特化のファインチューニングなしで。"],["In other words,","言い換えれば、"],["Evo 2 treats genomes","Evo 2はゲノムを"],["less like static code","静的なコードとしてではなく"],["and more like a language","むしろ言語として扱う"],["with grammar, style,","文法、文体、"],["and long-range structure.","そして長距離構造を持つ。"],["([nature.com](https://www.nature.com/articles/s41586-026-10176-5))","（[nature.com](https://www.nature.com/articles/s41586-026-10176-5)）"],["Still,","しかし、"],["\"AI creates life\"","「AIが生命を創造する」は"],["would be an exaggeration.","誇張であろう。"],["Evo 2 does not","Evo 2は"],["conjure living organisms","生物を生み出すわけではない"],["out of thin air.","無から。"],["In practice,","実際には、"],["the frontier looks more like","最前線はむしろ"],["a read-write-test loop:","読み書きテストの循環に近い："],["AI proposes sequences,","AIが配列を提案し、"],["researchers synthesize them,","研究者がそれを合成し、"],["and experiments reveal","実験が明らかにする"],["whether biology","生物学がその設計を"],["accepts the design.","受け入れるかどうかを。"],["Arc reports that","Arcの報告によると"],["Evo 2-guided regulatory DNA","Evo 2が導いた制御DNA は"],["was synthesized and","合成され"],["tested in cells,","細胞内でテストされ、"],["with some designs producing","一部の設計は生み出した"],["the predicted","予測された"],["chromatin-accessibility patterns.","クロマチンアクセシビリティパターンを。"],["The group also reported","同グループはまた報告した"],["the first AI-designed,","初のAI設計による"],["experimentally validated","実験的に検証された"],["bacteriophages;","バクテリオファージを。"],["after supervised fine-tuning","教師ありファインチューニングと"],["and lab screening,","実験室でのスクリーニングの後、"],["16 of 285 tested designs","テストされた285の設計のうち16が"],["successfully propagated","うまく増殖した"],["in bacteria.","細菌内で。"],["([arcinstitute.org](https://arcinstitute.org/news/evo-2-one-year-later))","（[arcinstitute.org](https://arcinstitute.org/news/evo-2-one-year-later)）"],["That is why","そういうわけで"],["Evo 2 feels like","Evo 2は"],["a turning point.","転換点のように感じられる。"],["It suggests that","それは示唆している"],["synthetic biology may shift","合成生物学が移行しうると"],["from slow trial-and-error","遅い試行錯誤から"],["toward guided exploration","導かれた探索へと"],["of genetic possibility.","遺伝的可能性の。"],["At the same time,","同時に、"],["the safety question","安全性の問題は"],["is impossible to ignore.","無視できない。"],["Arc says","Arcによると"],["eukaryotic viruses were excluded","真核生物ウイルスは除外された"],["from Evo 2's training data,","Evo 2の訓練データから、"],["and red-team tests found","そしてレッドチームテストで判明した"],["that attempts to generate","生成しようとする試みは"],["pathogenic human viral proteins","病原性のあるヒトウイルスタンパク質を"],["produced effectively","実質的に"],["random results.","ランダムな結果を生んだと。"],["So yes,","つまり確かに、"],["AI can now \"write\" DNA—but","AIは今やDNAを「書ける」が—"],["but only under constraints,","ただし制約の下でのみ、"],["and only with human judgment,","そして人間の判断、"],["laboratory validation,","実験室での検証、"],["and serious","そして厳格な"],["biosafety guardrails.","バイオセーフティの安全策があってこそだ。"],["([arcinstitute.org](https://arcinstitute.org/news/evo-2-one-year-later))","（[arcinstitute.org](https://arcinstitute.org/news/evo-2-one-year-later)）"]]