A benchmark of sequence-level perturbation tasks for evaluating DNA foundation models. Each task presents pairs of genomic sequences — one real (unperturbed) and one structurally altered — and asks whether the model assigns higher log-likelihood to the original.

Metric: pairwise discrimination accuracy = mean(LL(original) > LL(perturbed))

Tasks

`syn_human` · `syn_mouse` — Synonymous codon substitution (20,000 pairs each)

Codons within a real CDS are replaced with the highest-frequency synonym for the target species, while the upstream and downstream flanking sequence is left unchanged. Amino acid identity is preserved by construction. The model should prefer the natural codon usage over the artificially optimised variant.

Human (syn_human): GENCODE v45 annotations on GRCh38/hg38; codon frequencies from CoCoPUTs (H. sapiens RefSeq CDS); 9,616 unique genes.
Mouse (syn_mouse): GENCODE vM34 annotations on GRCm39/mm39; codon frequencies from CoCoPUTs (M. musculus RefSeq CDS); 10,253 unique genes.
Window size: 8,192 bp centred on the CDS (cds_start_in_seq / cds_end_in_seq).
Mean fraction of codons changed: ~55% (range 10–95%).

`motif_human` — CAG repeat insertion (20,000 pairs) / Triplet Expansion

A 30 bp codon-aligned region beginning 60 bp downstream of the first complete codon of the CDS exon is replaced with 10 consecutive CAG triplets (CAGCAGCAGCAGCAGCAGCAGCAGCAGCAG), mimicking the pathological trinucleotide repeat expansions underlying polyglutamine disorders (Huntington's disease, SCAs, DRPLA). The substitution is length- and reading-frame-preserving (30 bp in, 30 bp out), so the total window remains 8,192 bp. All sequence outside the patch is identical between original and perturbed.

Annotations: GENCODE v45 / GRCh38; all CDS exons (all ranks); 9,705 unique genes.
Window layout: first complete CDS codon always at position 8,102; patch always at positions 8,162–8,192 (patch_start_in_seq–patch_end_in_seq). For exons with phase 1 or 2, the 1–2 split-codon bases from the preceding exon are placed immediately before position 8,102, so the upstream genomic context is 8,102 bp (phase 0), 8,101 bp (phase 1), or 8,100 bp (phase 2).

Schema

All configs share the key columns:

Column	Description
`original_sequence`	Real, unperturbed genomic sequence — positive
`sequence`	Structurally altered sequence — negative
`cds_start_in_seq` / `cds_end_in_seq`	CDS boundaries within the window (bp offset)
`chr`, `strand`	Genomic locus
`gene_name`, `transcript_id`	GENCODE annotation
`benchmark_id`	Unique row identifier

Config-specific columns (e.g. patch_start_in_seq, n_codons_changed, fraction_codons_changed, codon_usage_source) are described in the dataset info above.

Usage

from datasets import load_dataset

# Synonymous codon substitution — human
syn_human = load_dataset(
    "HuggingFaceBio/perturbation-bench", "syn_human", split="test"
)

# Synonymous codon substitution — mouse
syn_mouse = load_dataset(
    "HuggingFaceBio/perturbation-bench", "syn_mouse", split="test"
)

# CAG motif insertion
motif = load_dataset(
    "HuggingFaceBio/perturbation-bench", "motif_human", split="test"
)

Evaluation

A ready-to-run scorer for Carbon, GENERator, and Evo2 is available at evaluation/perturbation_tasks.py in the Carbon release repository.

python evaluation/perturbation_tasks.py \
    --task syn_human \
    --model HuggingFaceBio/Carbon-3B \
    --bf16

python evaluation/perturbation_tasks.py \
    --task motif_human \
    --model arcinstitute/evo2_7b --backend evo2 \
    --bf16

Citation

@article{allal2026carbon,
  title={Carbon: Decoding the Language of Life},
  author={Allal, Loubna Ben and Li, Qiuyi and Fiusco, Maurizio and Tunstall, Lewis and Rasul, Kashif and Beeching, Ed and Aubakirova, Dana and Pati{\~n}o, Carlos and Frere, Thibaud and Lozhkov, Anton and others},
  journal={bioRxiv},
  pages={2026--05},
  year={2026},
  publisher={Cold Spring Harbor Laboratory}
}