About & feature scheme
The Evo2 Variant Effect Database provides zero-shot variant-effect scores from the Evo2 DNA foundation model for 6,475,578 common autosomal human variants (GRCh38, chromosomes 1–22; gnomAD NFE minor-allele-frequency ≥ 5%). It is overwhelmingly noncoding: ~2.61M intergenic, ~2.18M intronic, ~1.12M ncRNA-intronic, and only ~37.9k protein-coding variants. Research use only — not validated for clinical interpretation.
How a variant is scored
Evo2 is an autoregressive DNA language model. For each variant we give the model the reference and the alternate sequence and take the difference of their sequence log-likelihoods within a 2-kb reference context centred on the variant (about 1 kb on each side). The alternate allele can be a substitution, insertion or deletion.
Δ = ALT score − REF score. The sequence scores combine log-likelihoods according to
the selected strand strategy. Δ and the word delta denote the same quantity;
machine-readable column names retain _Delta.
Read the score in two parts.
- Magnitude —
|Δ|. How much the model's view of the whole window changes between the two alleles. Variant-report percentiles and functional stratifications use this magnitude; the raw table sorts by the selected column's stored values. Trait Analysis separately offers the original signed Δ annotation and an absolute |Δ| sensitivity annotation. A Δ near zero indicates a limited likelihood shift for this model and context; it does not by itself establish that a variant is tolerated or benign. - Direction — the sign of
Δ. Negative means the model finds the ALT allele more surprising than REF; positive means less surprising. This is the direction of a likelihood shift and nothing more. It is not the direction of a phenotypic effect, and a large positive delta is not evidence that a variant is benign — it is still a large departure from the reference sequence the model has learned.
Delta is a relative model score on the log-likelihood scale: not a calibrated probability, not a pathogenicity call, and not tied to any gene, tissue or mechanism. The scale is also not comparable across model sizes — see the next section. The variant report converts a raw delta into a percentile against all 6.48 M scored variants and against variants of the same functional class, which is the form in which the magnitude can actually be judged.
Which configuration should I use?
Default: Evo2_40B_AvgRC_Delta, matching the configuration used for the
manuscript's main case studies. Explicitly selected configurations are preserved when following analysis links.
There is no single best configuration for every task. The model-comparison section presents the frozen comparisons by data source, outcome and analysis sample. Use the corresponding captions and supporting tables, and report the model size, strand strategy and score transformation.
The two model sizes differ in both parameter count and training exposure: the Evo 2 publication reports 2.4 trillion training tokens for 7B and 9.3 trillion for 40B from OpenGenome2 (Brixi et al., Nature, 2026). This training corpus is distinct from the GRCh38 reference used to build each variant's scoring window. The strand strategies below also change how position-wise likelihoods are combined. Scores should be compared within a stated configuration; all six are shown in the variant report.
Two models × three strand strategies
Every variant carries scores from both Evo2 model sizes, each computed under three strategies for handling DNA's two strands — 6 score sets in total. Each set has three columns: the ALT log-likelihood, the REF log-likelihood, and their delta.
| Term | Meaning |
|---|---|
| Evo2-7B | The 7-billion-parameter Evo2 model. |
| Evo2-40B | The 40-billion-parameter Evo2 model. Performance depends on the evaluated task and sample. |
| RC | Reverse complement. DNA is double-stranded; the same variant can be read on the forward strand or on the complementary (reverse-complement) strand. |
| noRC | Scored on the forward strand only (no reverse complement). |
| avgRC | The average of the forward-strand and reverse-complement scores. |
| wtRC / weightRC | Weighted reverse-complement scoring (WeightedRC in field names). The manuscript describes position weights that decrease with distance from the variant, applied on both strands. Frozen comparison captions retain any unresolved formula or weight-documentation limitations; the site does not recompute these scores. |
This produces the 18 model-score columns below (6 sets × alt / ref / delta):
| Model | noRC | avgRC | weightRC |
|---|---|---|---|
| Evo2-7B | Evo2_7B_NoRC_{Ref,Alt,Delta} | Evo2_7B_AvgRC_{Ref,Alt,Delta} | Evo2_7B_WeightedRC_{Ref,Alt,Delta} |
| Evo2-40B | Evo2_40B_NoRC_{Ref,Alt,Delta} | Evo2_40B_AvgRC_{Ref,Alt,Delta} | Evo2_40B_WeightedRC_{Ref,Alt,Delta} |
Database fields
Identity & position
| Column | Type | Description |
|---|---|---|
| Variant_ID | text | Unique key, Chromosome:Position:Ref:Alt (GRCh38) |
| Chromosome | text | Chromosome (1–22; autosomes only) |
| Position | int | Base-pair position (GRCh38) |
| RSID | text | dbSNP Reference SNP ID (build 156) |
| Reference_Allele | text | Reference allele |
| Alternate_Allele | text | Alternate allele |
Allele frequency — gnomAD v4.1 joint (9 columns)
Allele frequencies are from the gnomAD v4.1 joint release (exomes + genomes combined). Variants were included on Non-Finnish European MAF ≥ 5%; the per-ancestry AFs below show how frequency varies across populations (a variant common in one ancestry can be rare in another).
| Column | Type | Description |
|---|---|---|
| Joint_AF | real | Joint alternate-allele frequency, all ancestries combined |
| African_AF | real | African / African-American |
| Admixed_American_AF | real | Admixed American |
| Ashkenazi_Jewish_AF | real | Ashkenazi Jewish |
| East_Asian_AF | real | East Asian |
| Finnish_AF | real | Finnish |
| Middle_Eastern_AF | real | Middle Eastern |
| Non_Finnish_European_AF | real | Non-Finnish European (the MAF ≥ 5% inclusion cohort) |
| South_Asian_AF | real | South Asian |
Evo2 model scores (18 columns)
| Column pattern | Type | Description |
|---|---|---|
| Evo2_{7B,40B}_{NoRC,AvgRC,WeightedRC}_Alt | real | Log-likelihood of the ALT sequence |
| Evo2_{7B,40B}_{NoRC,AvgRC,WeightedRC}_Ref | real | Log-likelihood of the REF sequence |
| Evo2_{7B,40B}_{NoRC,AvgRC,WeightedRC}_Delta | real | Δ = ALT score − REF score under the selected strategy. Strength is |Δ|; the sign is the direction of the likelihood shift, not of a phenotypic effect |
SCREEN Registry V4 cCRE annotation
These fields use the eight GRCh38 BED classes supplied in SCREEN Registry V4. A variant overlaps an
element when its reference-allele interval [Position − 1, Position − 1 + length(REF))
intersects that element. The result describes coordinate overlap, not a validated regulatory effect.
Multiple classes and element accessions are retained, separated by |.
| Column | Type | Description |
|---|---|---|
| SCREEN_V4_Classes | text | Distinct overlapping SCREEN V4 cCRE classes. Separate from the legacy ENCODE_cCRE_Annotation field. |
| SCREEN_V4_Accessions | text | Distinct overlapping SCREEN V4 element identifiers (EH38E accessions). |
| SCREEN_V4_Overlap_Count | int | Number of distinct overlapping elements; zero means no overlap in this registry. |
Gene & functional annotation (ANNOVAR / Ensembl)
| Column | Type | Description |
|---|---|---|
| Functional_Annotation | text | Region class (intergenic, intronic, exonic, UTR…) |
| Gene | text | Ensembl gene symbol(s) |
| Gene_Detail | text | Transcript / distance detail |
| Exonic_Function | text | Coding consequence (nonsynonymous / synonymous / stopgain…) |
| Amino_Acid_Change | text | Amino-acid change & transcript (coding variants) |
| ENCODE_cCRE_Annotation | text | Legacy functional/regulatory label, retained for the original strata. Values such as Coding and Intronic are not SCREEN Registry V4 cCRE classes. |
Conservation
| Column | Type | Description |
|---|---|---|
| PhastCons_100way | real | phastCons conservation, 100-way vertebrate (0–1) |
| PhyloP_100way | real | phyloP conservation, 100-way vertebrate (signed) |
Repeats
| Column | Type | Description |
|---|---|---|
| Repeat_Name | text | RepeatMasker repeat name |
| Repeat_Class | text | RepeatMasker repeat class (LINE/SINE/LTR/DNA…) |
| Repeat_Family | text | RepeatMasker repeat family |
Missing values are stored as NULL (source tokens . / NA).
Quick start
- Search by rsID, gene, or region from the home page.
- The region viewer plots an interactive Δ landscape for any gene or window.
- Variant Analysis, Region/Region Set Analysis and Trait Analysis are independent modules. The Guided Workflow starts from one explicitly selected study; it does not change the independent gene examples or recompute S-LDSC.
- Filter and facet the whole table at /evo2/variants.
- Run SQL at /evo2; add
.jsonto any page for the API.
Reading evidence by source
Evidence is organized by its database or original study, with the source version, record identity and matching status retained. A source that has no matching record is not evidence of no biological effect. Original association P-values, PIP, experimental readouts, clinical classifications and third-party model scores describe different quantities and should be read on their own scales.
The Variant view contains individual variant records. Existing gene, intron, assay or signal models belong to Region/Gene, Region set or study views, where their original sample size and the current selection's coverage can be distinguished. Those models are Evo2VED analyses of the named source, rather than statistics supplied by that source or models refitted to the current window.
Cite
- Zhu et al. Evo2VED: a database of DNA large language model–derived scores for genome-wide variant interpretation. Manuscript under revision, 2026. Website release history. Publication details will be added when available.
- Brixi, G., Durrant, M. G., Ku, J., et al. Genome modelling and design across all domains of life with Evo 2. Nature (2026). doi:10.1038/s41586-026-10176-5.
License: Apache-2.0.
← Back to search