AI for Bio · Benchmark 38

BEND

Evaluates DNA language models' frozen representations on human-genome functional tasks.

bench-38Type BEvidence L2Genomics & RNA● Active Board中文

Overview

NameBEND · aliases: BEND, BEND 2.0
Full nameBenchmarking DNA Language Models on Biologically Meaningful Tasks
Ambiguous names (not auto-linked)bend
FamilyNot a family entry. BEND 2.0 is treated as a new version of the same benchmark, not a separate family member.
DomainGenomics & RNA
InclusionSupplementary entry
TypeB · Fixed datasets with frozen representations and a light downstream head.
Evidence levelL2 · Retrospective; the variant-effect tasks are zero-shot and nothing is blind.
PublisherTechnical University of Denmark, University of Copenhagen (Academic lab)
First release · paper v1
Basis: arXiv 2311.12570 v1 source
Other milestones
  • Preprint v1: · arXiv 2311.12570 v1 source
  • Official announcement: † · ICLR 2024 conference paper (only the year is sourced) source
Latest update · new version
Basis: BEND 2.0 preprint (arXiv 2611.03806 v1)
BEND 2.0 preprint released, reorganized around capabilities needed for DNA design: 8 tasks and 5 experimental datasets; the original repository's last commit was 2024-10-31. source
The most recent sourced event, not a claim that nothing changed since.
Next expectedThe authors have announced no date for a next version.
Status● Active · Version 2.0 released 2026-11-02; the original and 2.0 use different scopes and must be cited separately.
ScaleOriginal: 7 tasks, 285 to 2,005,617 samples, 512 to 100,096 bp. BEND 2.0: 8 tasks over 5 experimental datasets.
  • Tasks: 7 (BEND v4)
  • Sample-count range: 285-2,005,617 (BEND v4)
  • Sequence-length range: 512-100,096 bp (BEND v4)
  • Tasks: 8 (BEND 2.0)
  • Experimental datasets: 5 (BEND 2.0)
Wet-lab refreshStatic data · Reuses published genome annotations and functional genomics data; no new experiments.
Own wet lab needed to evaluate a new modelNo, run it as a download
Models reported0 counted
Unverified fields† release_milestones.official_announcement, ranking.entries
As of2026-11-16

What it measures

  • Task: Freeze pretrained embeddings and train a light head on human-genome functional tasks: gene finding, enhancer annotation, chromatin accessibility, histone modification, methylation and two variant-effect tasks.
  • Truth: Published genome annotations and functional genomics measurements, with public labels for the variant-effect tasks.
  • Primary metric: Per task: MCC, AUPRC or AUROC.
  • Secondary metrics: BEND 2.0 adds correlation and design-task metrics.
  • Version traps: The original and 2.0 define tasks differently and their numbers cannot be mixed; state the embedding chunk length, whether the official head configuration is used, and whether a readout is native or fitted.

Current ranking

Board

MetricPer-task metrics with no single official ranking metric: conditional log loss (bits/base, lower better), macro Spearman correlation, top-1 tissue identification accuracy
Eval setBEND 2.0 Table 1 snapshot: 8 endpoints across 5 capability categories, 17 model-interface rows
Snapshot
Boardhttps://machinelearninglifescience.github.io/BEND2
Retrieved2026-11-16

Caveat. Values belong to BEND 2.0 (a 2026-11 preprint) and come from its site data file's Table 1 snapshot; there is no official aggregate, ranks hold per endpoint only, and native and fitted readouts are separate rows.

Shown in the overview table:

SpliceAI
0.674
Per-task metrics with no single official ranking metric: conditional log loss (bits/base, lower better), macro Spearman correlation, top-1 tissue identification accuracy · BEND 2.0 Table 1 snapshot: 8 endpoints across 5 capability categories, 17 model-interface rows
Board ·
Best site model: AlphaGenome (#2)

RankModelScore (verbatim)Metric · eval set / protocolSource type · dateSource and note
1SpliceAI0.674Per-task metrics with no single official ranking metric: conditional log loss (bits/base, lower better), macro Spearman correlation, top-1 tissue identification accuracy
BEND 2.0 Table 1 snapshot: 8 endpoints across 5 capability categories, 17 model-interface rows
Board
source Not a site model; first on this endpoint, as a specialist model's native output.
2AlphaGenome0.650Per-task metrics with no single official ranking metric: conditional log loss (bits/base, lower better), macro Spearman correlation, top-1 tissue identification accuracy
BEND 2.0 Table 1 snapshot: 8 endpoints across 5 capability categories, 17 model-interface rows
Board
source Second on this endpoint with a fitted readout and 0.621 native; first on five of the eight endpoints.
4NTv3 650M post0.564Per-task metrics with no single official ranking metric: conditional log loss (bits/base, lower better), macro Spearman correlation, top-1 tissue identification accuracy
BEND 2.0 Table 1 snapshot: 8 endpoints across 5 capability categories, 17 model-interface rows
Board
source Fourth on this endpoint; first on the sequence-prior endpoint at 1.375 bits/base, where lower is better.
9Evo2 20B0.367Per-task metrics with no single official ranking metric: conditional log loss (bits/base, lower better), macro Spearman correlation, top-1 tissue identification accuracy
BEND 2.0 Table 1 snapshot: 8 endpoints across 5 capability categories, 17 model-interface rows
Board
source Ninth on this endpoint; prior loss 1.388 bits/base (rank 3); last on the design endpoints.

Models reported

Grouped by relation type; each line shows its source type. Counted relations add up to 0. A pairing whose evidence rests only on this survey's own page is shown as unverified.

None. No model among the 112 has a recorded pairing with this entry.

Caveats

  1. Human genome only, and frozen-embedding results are not a fine-tuning ceiling.
  2. The enhancer task is small and sparse and all models do poorly; text and tables disagree on histone modification.
  3. The original baselines are mostly models from around 2023, so newer DNA models are absent from that board.
  4. The two versions differ in scope: 2.0 is organized around design capabilities and its numbers cannot sit beside the original's.
  5. The site models on the 2.0 board are not yet recorded as pairings; their relation type awaits the next round.

How to use it

Take data and configs from the official repository, extract frozen embeddings per task and train a small head, which costs far less than fine-tuning. State the embedding chunk length and whether the official head configuration is used; report both mean rank and per-task values rather than a single task; keep the original and 2.0 apart; and report the zero-shot variant tasks alongside TraitGym.

Pairs well with: GUE + Nucleotide Transformer benchmark · Basenji/Enformer test set · TraitGym · ClinVar hold-out protocol

Sources

  1. BEND paper (arXiv 2311.12570 v1) (preprint) ·
  2. BEND paper full text v4 (preprint) ·
  3. ICLR 2024 conference paper page (paper)
  4. Official repository frederikkemarin/BEND (repository)
  5. BEND 2.0 preprint (arXiv 2611.03806) (preprint) ·
  6. BEND 2.0 leaderboard page (leaderboard) ·

Corrections

1 correction
FieldWasNowBasisFound
ranking.eval_setoriginal BEND standingBEND 2.0 Table 1 snapshot (2026-10-22), per endpointsource2026-11-16

Part of the AI for Bio benchmarks survey · all 44 benchmarks · open in the explorer. Record as of 2026-11-16.