henhaobio
Sign in

bioRxivGenomicsPreprint

Genome modeling and design across all domains of life with Evo 2

Garyk BrixiiD, Matthew G. DurrantiD, Jerome Ku, Michael PoliiD, Greg Brockman, Daniel ChangiD, Gabriel A. GonzaleziD, S. B. KingiD, David B. LiiD, Aditi T. MerchantiD, Mohsen NaghipourfariD, Eric NguyeniD, Chiara Ricci-TamiD, David W. RomeroiD, Gwanggyu SuniD, Ali Taghibakshi, Anton Vorontsov, Brandon Yang, Myra Deng, Liv GortoniD, Nam Nguyen, Nicholas K. WangiD, Etowah AdamsiD, Stephen A. BaccusiD, Steven DillmanniD, , Brian HieiD

PDFDOI
In-depth full-text digest available · Jump to digest ↓

Abstract

All of life encodes information with DNA. While tools for sequencing, synthesis, and editing of genomic code have transformed biological research, intelligently composing new biological systems would also require a deep understanding of the immense complexity encoded by genomes. We introduce Evo 2, a biological foundation model trained on 9.3 trillion DNA base pairs from a highly curated genomic atlas spanning all domains of life. We train Evo 2 with 7B and 40B parameters to have an unprecedented 1 million token context window with single-nucleotide resolution. Evo 2 learns from DNA sequence alone to accurately predict the functional impacts of genetic variation--from noncoding pathogenic mutations to clinically significant BRCA1 variants--without task-specific finetuning. Applying mechanistic interpretability analyses, we reveal that Evo 2 autonomously learns a breadth of biological features, including exon-intron boundaries, transcription factor binding sites, protein structural elements, and prophage genomic regions. Beyond its predictive capabilities, Evo 2 generates mitochondrial, prokaryotic, and eukaryotic sequences at genome scale with greater naturalness and coherence than previous methods. Guiding Evo 2 via inference-time search enables controllable generation of epigenomic structure, for which we demonstrate the first inference-time scaling results in biology. We make Evo 2 fully open, including model parameters, training code, inference code, and the OpenGenome2 dataset, to accelerate the exploration and design of biological complexity.

In-depth summary

Claude Opus 5.5 · xhigh effort · 17 Oct 2026 · from the full text of version 1 (21 Feb 2025) · CC BY-ND 4.0

A 40B-parameter DNA model trained on 9.3 trillion tokens scores coding and noncoding variants zero-shot, exposes interpretable sequence features, and writes chromosome-scale sequences.

Earlier genome language models were limited to prokaryotes or short contexts, so they could not cover the noncoding, intron-rich architecture of eukaryotic genomes or score the variant classes clinicians care about. Evo 2 pairs a curated multi-domain corpus with a convolutional multi-hybrid architecture (StripedHyena 2) to reach a 1 million token context at single-nucleotide resolution, and is released with parameters, training and inference code, and the OpenGenome2 corpus.

Key results

  1. ShownWithout variant-specific training or alignments, model likelihood changes classify human pathogenic variants, leading on noncoding, non-SNV and splice variants while trailing specialised predictors on coding SNVs.

    On ClinVar (release 2024.02.28, two-star or better) the 40B and 7B models ranked fourth and fifth on coding SNVs behind AlphaMissense, ESM-1b and GPN-MSA, but beat all compared models on coding non-SNVs and on noncoding SNVs and non-SNVs; they also led on SpliceVarDB and on BRCA1 and BRCA2 with coding and noncoding variants combined. AUROC values sit in figures absent here. · 2.3. Evo 2 enables accurate human clinical…; 4.3.13. ClinVar variant effect prediction; 4.3.14. SpliceVarDB variant effect prediction; 4.3.15. BRCA1/2 zero-shot classification

  2. ShownA supervised classifier on 40B-model embeddings separates loss-of-function BRCA1 variants from the rest better than any zero-shot benchmark compared.

    Test-set AUROC 0.94 and AUPRC 0.84 on BRCA1 coding SNVs, AUROC 0.95 and AUPRC 0.86 across all BRCA1 SNVs, after choosing block 20 of the 40B model (AUROC 0.92) and averaging embeddings over a 128nt window; labels came from the source study (823 LOF vs 3,070 functional or intermediate SNVs), with 20% held out for testing. · 2.3. Evo 2 enables accurate human clinical…; 4.3.15. BRCA1/2 zero-shot classification; 4.3.16. BRCA1 supervised classification

  3. ShownSparse autoencoders on the model's internal activations recover latent features that line up with unannotated biological structure, from prophages and CRISPR spacers to exon boundaries, transcription factor motifs and protein secondary structure.

    A Batch-TopK SAE (32,768 features, k = 64) on 1 billion layer-26 activations gave f/19746, firing on annotated prophages in E. coli including CPZ-55 and on scrambled CRISPR spacers, plus features for coding regions, introns, exon starts and ends, frameshifts, alpha-helices and beta-sheets; exon features transferred to the woolly mammoth genome, absent from training. · 2.4. Feature interpretation in Evo 2 from…; 4.4.1. SAE training and dataset composition; 4.4.3. Prophage feature; 4.4.8. Exon-intron features

  4. ShownPrompted generation produces genome-scale sequences whose annotated content resembles natural organelle, bacterial and yeast genomes, though below native feature density.

    250 generated 16 kb mitochondrial sequences carried the expected counts of coding, tRNA and rRNA genes under MitoZ, with AlphaFold 3 complexes matching expected folds; for M. genitalium (about 580 kb), nearly 70% of predicted genes had significant Pfam hits versus 18% for Evo 1 131k; 330 kb generated from a yeast chromosome III prompt contained tRNAs, promoters and introns but fewer features per kb than native yeast. · 2.5 Genome-scale generation across the domains…

  5. ShownBeam search guided by chromatin accessibility predictors steers generation to put open-chromatin regions where specified, and design quality improves predictably with inference-time compute.

    Scoring each new 128-bp chunk with an Enformer plus Borzoi ensemble (DNase tracks in 129 ES-E14 cells, sequence placed in a mouse chrX locus), sampling 30 or more chunks and keeping the top two reached AUROC around 0.9 for matching the target open/closed pattern, with a log-linear gain as beam width grew; patterns spelled Morse messages, and dinucleotide frequencies matched mm39, unlike a uniform-proposal control. · 2.6. Generative epigenomics via inference-time…

Methods, data and reproducibility

Two autoregressive DNA models (7B and 40B parameters) on the StripedHyena 2 convolutional multi-hybrid were trained by next-token prediction over byte-level sequence, first at 8,192-token context weighted toward genic windows, then midtrained to 1 million tokens. They were evaluated by likelihood differences, embedding classifiers, sparse autoencoders, and guided generation.

  • OpenGenome2 training corpus spanning bacteria, archaea, eukarya and bacteriophage, excluding viruses infecting eukaryotes · all domains of life · n = 8.84 trillion nucleotides; 9.3 trillion tokens for 40B, 2.4 trillion for 7B
  • Prokaryotic reference genomes from GTDB releases 214.1 and 220.0 · bacteria and archaea · n = 113,379 genomes (357 billion nucleotides)
  • Filtered eukaryotic reference genomes downloaded from NCBI on May 31, 2024, clustered by Mash distance · eukaryotes · n = 15,032 genomes, 6.98 trillion nucleotides
  • Organelle genomes, metagenomes, transcripts, ncRNAs and EPDnew promoters · multiple · n = 32,240 organelle genomes; 41,253 metagenomes
  • Clinical and experimental variant benchmarks: ClinVar, SpliceVarDB, BRCA1 and BRCA2 functional assays, DART-Eval regulatory tasks · human · n = BRCA2: 1,156 LOF vs 5,681 functional or intermediate SNVs
  • Deep mutational scanning sets for proteins and ncRNAs, plus human viral proteins as a negative control · bacteria, human, viruses · n = 9 prokaryotic, 6 human and 18 viral protein DMS sets; 9 ncRNA DMS sets
  • Gene essentiality: DEG bacterial genomes, two phage genomes, and a Cas13 lncRNA knockdown screen in 5 cell lines · bacteria, phage, human · n = 56 bacterial genomes; 778 essential lncRNAs · NC_001416; NC_005856
  • Mutation scans around start and stop codons and across genomic region types · 20 prokaryotic and 16 eukaryotic species · n = 1,024 genes sampled per species; 8,192 bp context windows
Methods & tools
Evo 2 (7B, 40B), StripedHyena 2, Savanna training stack, Vortex inference stack, Batch-TopK sparse autoencoder, Enformer, Borzoi, AlphaFold 3, MitoZ, Prodigal and HHpred/Pfam, geNomad, mmseqs
Reagents & environment
Training used Savanna, built from DeepSpeed, GPT-NeoX and Transformer Engine, with mixed precision, FP8 linear layers and RMSNorms, and 3D data/tensor/context parallelism with ZeRO-3; inference runs on Vortex. No version pins or container are given in this text.
Assay conditions & compute
Scale is described rather than itemised: 40B pretraining ran 6.6T tokens at 1024 context then 1.1T at 8192, with 600B tokens of midtraining; StripedHyena 2 is reported up to 1.3× faster than optimised Transformers at 16 thousand context and 3× at 1 million. Exon classifier work used one NVIDIA H100.
Reproducibility gaps
The data and code sections list OpenGenome2, repositories, Hugging Face weights, NVIDIA NIM endpoints and two web tools, but every URL is missing from this text, so none can be reproduced here. Several results point to supplementary tables and figures that are also absent.
Limitations (6)
  • Sequences from viruses infecting eukaryotes were deliberately excluded, so the model performs near randomly on human viral proteins; the authors note task-specific post-training could undo this safeguard. · stated by the authors
  • On coding SNVs the models stay behind specialised predictors such as AlphaMissense, ESM-1b and GPN-MSA. · stated by the authors
  • Generated yeast sequence has lower tRNA and gene density than the native chromosome, and generation here is unconstrained autoregressive sampling. · stated by the authors
  • Performance depends heavily on data engineering: training on raw whole eukaryotic genomes degraded downstream results, with zero-shot BRCA1 AUROC 0.793 for the ablation versus 0.891 for the released composition. · stated by the authors
  • Design success for chromatin accessibility is judged by the same Enformer and Borzoi ensemble that guides the search, so no independent measurement supports the designs. · noted in this summary
  • None of the generated mitochondrial, bacterial or yeast sequences were synthesised or tested in cells; quality is assessed by annotation tools and structure prediction. · noted in this summary

Read this if you score genetic variants, annotate poorly characterised genomes, or want a generative DNA model you can steer with your own sequence-to-function predictor.

Citations
218

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISNguyen '24Mathur '26Pearce '26Merchant '25Amgarten '26Black '26
74% semantic matchSynthetic Biology
Eric Nguyen, Michael Poli, Matthew G. Durrant+17 more·Science·Nov 14, 2024· 593 citations