suzuki-2001/evo2-sae-handson

Evo 2 sparse autoencoder activations for the Escherichia coli K-12 MG1655 genome

SAE activations for every base of the chromosome (RefSeq GCF_000005845.2, NC_000913.3, 4,641,652 bp), with the embeddings, annotation and structures used by the hands-on notebook.

How the activations were made

Evo 2 7B (arcinstitute/evo2_7b) was run over the circular chromosome in 16,384 bp windows at a stride of 15,360 bp, discarding the first 1,024 bp of each window as context. Each base's blocks.26 embedding, 4,096 values, was encoded with the Goodfire BatchTopK SAE (Goodfire/Evo-2-Layer-26-Mixed, sae-layer26-mixed-expansion_8-k_64.pt, sha256 prefix 598f131a2e71f1c4) the way the Evo 2 notebook does it:

a = relu(x @ W + b_enc)
z = BatchTopK_64(a)

BatchTopK keeps the largest 64 x L activations over the L bases it is given, so the result depends on the window. Rows here hold the largest 128 activations per base, twice what BatchTopK keeps on average, so the cut-off can be reapplied for any window. A base that exceeds the cut-off more than 128 times loses the surplus — 0.16% of the kept activations over the 30,720 bp of region_emb.pt, 0.57% over a 5 kb window, always the smallest values just above the cut-off.

Files

File	Bytes	Description
ecoli_values.f16	1,188,262,912	the largest 128 activations per base, float16
ecoli_indices.u16	1,188,262,912	their feature ids, uint16, 0–32767
ecoli_indptr.i64	37,133,224	row pointers, int64, 4,641,653 entries, 128 apart
ecoli_meta.json	1,000	provenance of the three files above
region_emb.pt	251,659,966	blocks.26 embeddings for NC_000913.3:4,162,560-4,193,280; 30,720 × 4,096 bfloat16
context.pt	83,888,037	five window overlaps, each 1,024 bases embedded twice, with 0–1,023 and with ≥ 15,360 bp of left context
scramble.pt	183,506,633	the CRISPR array embedded sixteen ways: natural, and five draws of each of three edits; 16 × 1,400 × 4,096 bfloat16
class_sums.npz	656,598	per-class totals of the per-region mean activation, over 7,279 regions and 32,768 features, with the genome-wide BatchTopK cut-off applied
ecoli.bed	182,381	4,340 CDS, 86 tRNA, 22 rRNA, 20 CRISPR repeats, 18 spacers, 6 prophages
proteins.json	5,468	CDS coordinates, UniProt ids and DSSP strings for seven proteins; residue i is codon i
eftu_trna.pdb	379,071	EF-Tu (AlphaFold P0CE48) on 1B23, keeping the 1B23 tRNA as chain B
rpoBC.pdb	1,701,413	RNA polymerase β and β′ chains from 6C9Y
pspA.pdb	149,849	AlphaFold P0AFM6
ompG.pdb	205,982	AlphaFold P76045
crp.pdb	139,643	AlphaFold P0ACJ8
groL.pdb	331,451	AlphaFold P0A6F5
sha256sums.txt	—	checksums for all of the above

Reading one region

Rows lo:hi occupy [2*indptr[lo], 2*indptr[hi]) of the values and indices files, and the host serves HTTP Range requests, so a few-kb region costs a MB or two. Apply BatchTopK over the window before reading a feature.

import numpy as np, requests

BASE = "https://huggingface.co/datasets/suzuki-2001/evo2-sae-handson/resolve/main"

def part(name, first, last):
    r = requests.get(f"{BASE}/{name}",
                     headers={"Range": f"bytes={first}-{last - 1}"})
    r.raise_for_status()
    return r.content

lo, hi = 4_175_250, 4_177_250  # tRNA array and tufB
ptr = np.frombuffer(part("ecoli_indptr.i64", 8 * lo, 8 * (hi + 1)), np.int64)
first, last = int(ptr[0]), int(ptr[-1])
values = np.frombuffer(part("ecoli_values.f16", 2 * first, 2 * last), np.float16)
indices = np.frombuffer(part("ecoli_indices.u16", 2 * first, 2 * last), np.uint16)
rows = np.repeat(np.arange(hi - lo), np.diff(ptr))

value = values.astype(np.float32)                       # BatchTopK over this window
want = 64 * (hi - lo)
floor = np.partition(value, len(value) - want)[len(value) - want]
keep = value >= floor

hit = keep & (indices == 30262)  # the paper's tRNA feature
track = np.zeros(hi - lo, np.float32)
track[rows[hit]] = value[hit]

Sources and licences

• Genome NC_000913.3 — NCBI imposes no restrictions on use or redistribution; rights remain with the original submitters
• Evo 2 7B weights, arcinstitute/evo2_7b — Apache 2.0
• SAE checkpoint, Goodfire/Evo-2-Layer-26-Mixed — MIT
• AlphaFold models P0CE48, P0A8V2, P0A8T7, P0AFM6, P76045, P0ACJ8, P0A6F5 — CC-BY 4.0, EMBL-EBI
• Experimental structures 1B23 and 6C9Y — CC0 1.0, RCSB PDB
• DSSP strings — mkdssp 4.6.1 (BSD-2-Clause), run on the AlphaFold models above
• Prophage intervals — geNomad (Lawrence Berkeley National Laboratory academic / non-commercial licence); CPZ-55, doi:10.3389/fgene.2019.00065
• CRISPR repeats and spacers — derived from the genome sequence

Released under CC-BY 4.0.

Citation

Brixi G, Durrant MG, Ku J, Naghipourfar M, Poli M, Sun G, et al. Genome modelling and design across all domains of life with Evo 2. Nature 2026. doi:10.1038/s41586-026-10176-5

Archived at https://doi.org/10.5281/zenodo.21856795