# BEND 2.0 native HF protocol map

Eight scientific endpoints are preserved. Seven downstream endpoints have separate embeddings and native-output subleaderboards so a shared checkpoint never collapses two readouts into one score. The sequence-prior endpoint is independent of downstream readout and is published once per checkpoint.

| Scientific task | Native task ID | Unit | Better |
|---|---|---|---|
| Prior loss | `prior_loss_bits_per_base` | bits/base | Lower |
| Splicing effects | `opensplice_embeddings_spearman_rho` | ρ | Higher |
| Splicing effects | `opensplice_native_outputs_spearman_rho` | ρ | Higher |
| Cell contrast | `context_embeddings_spearman_rho` | ρ | Higher |
| Cell contrast | `context_native_outputs_spearman_rho` | ρ | Higher |
| Tissue identification | `tissue_identification_embeddings_accuracy_percent` | % | Higher |
| Tissue identification | `tissue_identification_native_outputs_accuracy_percent` | % | Higher |
| Background interaction | `background_embeddings_spearman_rho` | ρ | Higher |
| Background interaction | `background_native_outputs_spearman_rho` | ρ | Higher |
| Single-edit rank | `repair_percentile_embeddings_residual_midrank_percentile` | percentile | Lower |
| Single-edit rank | `repair_percentile_native_outputs_residual_midrank_percentile` | percentile | Lower |
| Cell selectivity | `cell_selective_embeddings_utility_log2` | log₂ utility | Higher |
| Cell selectivity | `cell_selective_native_outputs_utility_log2` | log₂ utility | Higher |
| Multi-cell gain | `atlas_profile_embeddings_gain_over_random_log2` | log₂ gain | Higher |
| Multi-cell gain | `atlas_profile_native_outputs_gain_over_random_log2` | log₂ gain | Higher |

## Prior loss

Predict the hidden central base of 512 human reference-genome windows. Lower loss means better calibrated conditional probabilities.

Metric: **Hidden-base conditional log loss**, in **bits/base**. 512 windows of 1,200 bp from separate 1-Mb genomic blocks; central base hidden.

- Interface-sensitive: NTv3 uses 12% additional masking and averages the individual losses of two queries; GENERATORv2 uses temperature-scaled probabilities.
- Only rows with a comparable nucleotide-probability interface report this task. Missing prior scores are omitted from the combined score, never imputed.

## Splicing effects

Predict mutation-induced changes in alternative splicing; correlate within each held-out gene, then average genes equally.

Metric: **Macro-gene Spearman correlation**, in **ρ**. 590,104 interventions in 466 genes; five folds hold out complete genes for fitted readouts.

- Native splice heads retain historical supervision; fitted readouts are trained on the benchmark’s source folds.
- OpenSplice is the full gene-held-out endpoint, not a cross-assay transfer score.

## Cell contrast

Predict the HepG2-minus-K562 difference in a mutation’s reporter effect, then correlate within regulatory elements and average elements equally.

Metric: **Macro-element Spearman correlation**, in **ρ**. 170,292 substitutions across 128 elements; complete elements and related allelic backgrounds remain together during fitting and evaluation.

- The experimental reference has physical counts for 127 of 128 elements and uses repeat subsets; model values use the pooled posterior effects.

## Tissue identification

Match a predicted expression profile across held-out genes to the correct measured tissue among 51 candidates.

Metric: **Top-1 tissue identification accuracy**, in **%**. 5,782 held-out genes and 51 represented tissues; 12,146 training and 1,350 validation genes.

- This holds out genes within represented tissues; it does not test transfer to unseen tissues.
- The 99.0% experimental reference averages both donor-half directions; no one-direction interval is transferred to it.

## Background interaction

Predict how the same mutation’s effect differs across two parental genetic backgrounds, isolating background dependence.

Metric: **Macro-element Spearman correlation**, in **ρ**. 68,026 paired mutations in 114 elements; paired backgrounds remain together.

- The current experimental reference is HepG2–K562 transfer (0.421), not the superseded replicate-subset background value.
- Transfer across cells is an experimental comparison, not a noise ceiling.

## Single-edit rank

Choose an alternative single-base edit that matches the effect of a reference edit. Score the first nomination by its measured residual’s rank among eligible candidates.

Metric: **Mean measured residual midrank percentile**, in **percentile**. 331 response targets in 110 elements; 578–597 eligible substitutions per element; direct reversion is excluded.

- All candidate measurements are withheld until evaluation; predictions rank proximity to the model’s own predicted reference-edit effect.

## Cell selectivity

Select one substitution or no edit to increase activity in the target cell while penalizing the absolute effect in the other cell: U = target − |other|.

Metric: **Mean cell-selective utility**, in **log₂ utility**. 167,958 substitutions, 280 parental backgrounds and 127 loci; both cells serve as targets; loci receive equal weight.

- Native outputs use held-locus, nonnegative per-cell calibration slopes; they are not zero-shot in this task.
- Model utilities use all five RNA repeats; experimental selection uses reciprocal disjoint repeat subsets.
- Original confirmatory tests concern within-model controls; a sorted point estimate is not a significance claim.

## Multi-cell gain

Select ten enhancers for each multi-cell ON/OFF program. Utility rewards the weakest requested activity and penalizes the strongest unwanted activity; report gain over intent-matched random selection.

Metric: **Mean utility gain over intent-matched random selection**, in **log₂ gain**. 2,411 measured 145-nt candidates, 16 overlapping programs and 12 cell types; select 10 per request. Fit on 5,246 separate source constructs after excluding candidate families.

- The native rows and Enformer fit linear heads to native functional outputs on source constructs.
- The two RNA-repeat markers are descriptive repeat measurements, not confidence intervals or independent program replications.
- Positive gain does not imply positive absolute utility; AlphaGenome fitted has 0.833 gain but −0.096 absolute utility.