--- license: cc-by-4.0 pretty_name: IMAGINE-AI contig-level Evo 2 embeddings tags: [biology, genomics, microbiology, embeddings, dna-language-model, evo2] size_categories: [10K/.npz` | one file per genome (11,862 files, ~465 GB in total; 1 MB to 2.6 GB, median 15 MB) | | `manifests/genomes.csv` | one row per listed genome (11,868; 6 have `status = rejected` and no `.npz`) | | `manifests/pilot_genomes.csv` | selection checks for the 12 pilot genomes | | `manifests/labels.csv` | phenotype labels for the pilot genomes only (Madin, BacDive, Bergey's) | `` is digits 4–6 of the accession number: `GCF_000007365.1` → `contigs/evo2/007/` (in Python, `acc[7:10]`). This keeps every folder well under 10,000 files. ### Each `.npz` | Key | dtype, shape | Meaning | |---|---|---| | `embedding` | float32 [32, 4096] | genome vector, all layers (identical to the row in `genome-embeddings/evo2/`) | | `contig_means` | float32 [n_contigs, 32, 4096] | one vector per contig, all layers | | `contig_ids` | str [n_contigs] | FASTA record IDs, same order as `contig_means` | | `meta` | JSON string | per-genome provenance: model revision, layers, pooling policy, `fasta_sha256`, `n_contigs`, `n_windows`, `total_bp`, IUPAC cleaning per contig, runtime and software environment | Layer *i* is `blocks.i`: the residual stream after Evo 2 block *i*, before the final norm. ```python from huggingface_hub import hf_hub_download import numpy as np, json acc = "GCF_000007365.1" p = hf_hub_download("ImagineG2P/Contig_embeddings", f"contigs/evo2/{acc[7:10]}/{acc}.npz", repo_type="dataset") z = np.load(p) contigs = z["contig_means"][:, 24] # [n_contigs, 4096], layer blocks.24 ids = z["contig_ids"] meta = json.loads(str(z["meta"])) ``` To fetch a subset without downloading everything, use `snapshot_download(..., allow_patterns=[...])` or `hf download ImagineG2P/Contig_embeddings --repo-type dataset --include "contigs/evo2/007/*"`. ## Model and pooling | | | |---|---| | Model | [`arcinstitute/evo2_7b_base`](https://huggingface.co/arcinstitute/evo2_7b_base) @ `074097e9` | | Tokens | one byte per base (CharLevelTokenizer) | | Window | 8,192 bp, forward strand only | | Precision | bf16 inference, float32 pooling | **Pooling.** Mean over tokens within each 8,192-bp window, then an equal-weight mean over the windows of a contig (the short last window counts as much as a full one) → `contig_means`. The genome vector `embedding` is an **equal-weight mean over contigs**, so a 1-kb contig counts as much as a 5-Mb chromosome. For a length-weighted genome vector, weight `contig_means` by contig length (from the assembly FASTA; the total is `total_bp` in `meta`). Non-ACGT IUPAC codes were set to N before embedding (counts in `meta`). ## Genomes All public NCBI assemblies (GCF_/GCA_), each with a `priority` from the IMAGINE-AI master list: 0 pilot genomes (11), 1 reference genomes for the WVU isolates' species (80), 2 QC-passed genomes from the growth-rate master table (283), 3 other genomes in that table (554), 4 [`ImagineG2P/fasta`](https://huggingface.co/datasets/ImagineG2P/fasta) (10,940). `fna_sha256` in `manifests/genomes.csv` is the hash of the exact FASTA embedded. ## Notes and limits - Frozen-model features, not fine-tuned. MicroGlot contig vectors are not included. - Draft assemblies vary widely in contig count (`n_contigs` in `manifests/genomes.csv`); see the pooling note. - Code: `Hackathon/pilot/` in the IMAGINE-AI repository (`embed_evo2.py`, `export_genome_embeddings.py`). ## License and citation CC-BY-4.0. Please cite IMAGINE-AI (DOE DE-FOA-0003612), Evo 2 (Brixi et al., *Nature* 2026; weights Apache-2.0, Arc Institute) and NCBI for the assemblies.