Evo 2: Genome Modeling and Design Across All Domains of Life

Evo 2: A Foundation Model for DNA

• 💡 Evo 2 is a novel biological foundation model designed for genome modeling and design across all domains of life.
• 🧠 It learns from DNA sequence alone, treating DNA as a language, similar to how large language models learn human language through generative pre-training.
• 🎯 This approach allows Evo 2 to capture evolutionary pressures and the functional importance encoded within DNA.

Architecture and Training Data

• 🛠️ Evo 2 utilizes a new machine learning architecture combining transformer layers with short, medium, and long hyena convolutional layers to enhance throughput and perplexity.
• 📊 The model was trained on an unprecedented 9.3 trillion DNA base pairs from a highly curated genomic atlas (OpenGenome2), spanning bacteria, eukarya, and archaea.
• 🌱 Training involved a two-phase strategy: an initial short-context phase (8,000 base pairs) focusing on genic regions, followed by a context extension phase (up to 1 million base pairs) emphasizing whole eukaryotic genomes.

Zero-Shot Predictive Power

• ✅ Evo 2 accurately predicts the functional impacts of genetic variation (e.g., pathogenic mutations, BRCA1 variants) without task-specific fine-tuning.
• 📈 It correlates likelihoods with fitness for proteins (bacterial and human) and non-coding RNAs, often outperforming or matching state-of-the-art protein and RNA language models.
• 🔬 The model effectively predicts gene essentiality in prokaryotes and long non-coding RNA essentiality in human cell lines, and variant effects (substitutions and indels) across coding and non-coding regions.

Uncovering Biological Features

• 🔍 Through mechanistic interpretability, Evo 2 autonomously learns a breadth of biological features directly from DNA sequences.
• 🧬 These include exon-intron boundaries, transcription factor binding sites, protein structural elements (alpha helix, beta sheets), tRNAs, and prophage genomic regions.
• ⚠️ The model can even identify specific features that activate for "bugs in DNA code," such as frame shifts in coding sequences.

Genome Generation and Design

• 🚀 Evo 2 demonstrates robust generative capabilities, producing mitochondrial, prokaryotic, and eukaryotic sequences at genome scale with high naturalness and coherence.
• 💡 Generated sequences maintain synteny and realistic gene statistics, while also exhibiting diversity from the training data.
• 🧪 It enables controllable generation of epigenomic structure, such as chromatin accessibility, by integrating with prediction models and using inference-time search for targeted design.
• 🌐 The model, including its parameters, code, and dataset, is fully open-source to foster further research and application in biological complexity.