1<!DOCTYPE html> 2<html> 3<head> 4 <title>About & feature scheme â Evo2 Variant Effect Database</title> 5 <link rel="stylesheet" href="/-/static/app.css?d59929"> 6 <meta name="viewport" content="width=device-width, initial-scale=1, shrink-to-fit=no"> 7 8 <link rel="stylesheet" href="/static/custom.css"> 9 10 <link rel="stylesheet" href="/static/app-layout.css"> 11 12 13
13<script src="/static/request-client.js"></script>
13 14</head> 15<body class="evo2-doc-body"> 16<div class="not-footer"> 17<header><nav> 18 19 20 <p class="crumbs"> 21 22 <a href="/">home</a> 23 24 25 </p> 26 27 28 29 30</nav></header> 31 32 33 34 35 36 37 38<section class="content"> 39 40<div class="evo2-wrap evo2-doc"> 41 42 <div class="evo2-masthead"> 43 <div class="evo2-brand"> 44 <a class="evo2-logo" href="/" aria-label="Home"><svg viewBox="0 0 32 40" width="36" height="44" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" aria-hidden="true"><path d="M7 3C7 12 25 13 25 20S7 28 7 37M25 3C25 12 7 13 7 20S25 28 25 37"/><path d="M8 6h16M11 11h10M11 17h10M8 22h16M11 28h10M11 33h10"/></svg></a> 45 <a class="evo2-wordmark" href="/">Evo2 Variant Effect Database</a> 46 <nav class="evo2-nav" aria-label="Site"> 47 <a href="/">Home</a> 48 <a href="/guide">Worked examples</a> 49 <a href="/about" aria-current="page">About & fields</a> 50 <a href="/research">Data sources & research</a> 51 <a href="/model-comparison">Model comparison</a> 52 <a href="/updates">Updates</a> 53 <a href="/evo2/variants">Browse table</a> 54 <a href="/evo2">SQL</a> 55 </nav> 56 </div> 57 <nav class="evo2-modulebar" aria-label="Independent analysis modules"> 58 <span class="evo2-modulebar-label">Independent modules</span> 59 <a href="/variant/rs429358">Variant Analysis</a> 60 <a href="/viewer">Region/Region Set Analysis</a> 61 <a href="/gwas">Trait Analysis</a> 62 </nav> 63 <nav class="evo2-modulebar" aria-label="Guided Workflow"> 64 <span class="evo2-modulebar-label">Guided exploration</span> 65 <a href="/workflow">Guided Workflow</a> 66 </nav> 67 68 69</div> 70 71 <h1>About & feature scheme</h1> 72 <p>The <strong>Evo2 Variant Effect Database</strong> provides zero-shot variant-effect scores from the 73 <a href="https://github.com/ArcInstitute/evo2">Evo2</a> DNA foundation model for 74 <strong>6,475,578 common autosomal human variants</strong> (GRCh38, chromosomes 1–22; 75 gnomAD NFE minor-allele-frequency ≥ 5%). It is overwhelmingly noncoding: 76 ~2.61M intergenic, ~2.18M intronic, ~1.12M ncRNA-intronic, and only ~37.9k protein-coding variants. 77 <strong>Research use only — not validated for clinical interpretation.</strong></p> 78 79 <h2 id="score">How a variant is scored</h2> 80 <p>Evo2 is an autoregressive DNA language model. For each variant we give the model the 81 <strong>reference</strong> and the <strong>alternate</strong> sequence and take the difference of their 82 sequence log-likelihoods within a 2-kb reference context centred on the variant (about 1 kb on each side). 83 The alternate allele can be a substitution, insertion or deletion.</p> 84 <p><code>Î = ALT score − REF score</code>. The sequence scores combine log-likelihoods according to 85 the selected strand strategy. <strong>Î</strong> and the word <strong>delta</strong> denote the same quantity; 86 machine-readable column names retain <code>_Delta</code>.</p> 87 <p>Read the score in two parts.</p> 88 <ul> 89 <li><strong>Magnitude — <code>|Î|</code>.</strong> How much the model's view of the whole window 90 changes between the two alleles. Variant-report percentiles and functional stratifications use this 91 magnitude; the raw table sorts by the selected column's stored values. Trait Analysis separately offers the original signed Î annotation and an absolute |Î| 92 sensitivity annotation. A Î 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.</li> 93 <li><strong>Direction — the sign of <code>Î</code>.</strong> Negative means the model finds the 94 ALT allele <em>less</em> expected than REF; positive means <em>more</em> expected. This is the 95 direction of a likelihood shift and nothing more. It is <strong>not</strong> the direction of a 96 phenotypic effect, and a large positive delta is not evidence that a variant is benign — it is 97 still a large departure from the reference sequence the model has learned.</li> 98 </ul> 99 <p>Delta is a relative model score on the log-likelihood scale: <strong>not</strong> a calibrated 100 probability, <strong>not</strong> a pathogenicity call, and not tied to any gene, tissue or mechanism. The 101 scale is also not comparable across model sizes — see the next section. The 102 <a href="/variant/rs429358">variant report</a> converts a raw delta into a percentile against all 6.48 M 103 scored variants and against variants of the same functional class, which is the form in which the magnitude 104 can actually be judged.</p> 105 106 <h2 id="which">Which configuration should I use?</h2> 107 <p><strong>Default: <code>Evo2_40B_AvgRC_Delta</code></strong>, matching the configuration used for the 108 manuscript's main case studies. Explicitly selected configurations are preserved when following analysis links.</p> 109 <p>There is no single best configuration for every task. The <a href="/model-comparison">model-comparison 110 section</a> presents the frozen comparisons by data source, outcome and analysis sample. Use the corresponding 111 captions and supporting tables, and report the model size, strand strategy and score transformation.</p> 112 <p>The two model sizes differ in both parameter count and training exposure: the Evo 2 publication reports 113 2.4 trillion training tokens for 7B and 9.3 trillion for 40B from OpenGenome2 114 (<a href="https://doi.org/10.1038/s41586-026-10176-5">Brixi et al., Nature, 2026</a>). 115 This training corpus is distinct from the GRCh38 reference used to build each variant's scoring window. 116 The strand strategies below also change how position-wise likelihoods are combined. Scores should be 117 compared within a stated configuration; all six are shown in the variant report.</p> 118 119 <h2>Two models × three strand strategies</h2> 120 <p>Every variant carries scores from <strong>both</strong> Evo2 model sizes, each computed under 121 <strong>three</strong> strategies for handling DNA's two strands — 6 score sets in total. Each set 122 has three columns: the ALT log-likelihood, the REF log-likelihood, and their delta.</p> 123 <table> 124 <thead><tr><th>
124Term</th><th>Meaning</th></tr></thead> 125 <tbody> 126 <tr><td>Evo2-7B</td><td>The 7-billion-parameter Evo2 model.</td></tr> 127 <tr><td>Evo2-40B</td><td>The 40-billion-parameter Evo2 model. Performance depends on the evaluated task and sample.</td></tr> 128 <tr><td>RC</td><td><strong>Reverse complement.</strong> DNA is double-stranded; the same variant can be read on the forward strand or on the complementary (reverse-complement) strand.</td></tr> 129 <tr><td>noRC</td><td>Scored on the forward strand only (no reverse complement).</td></tr> 130 <tr><td>avgRC</td><td>The average of the forward-strand and reverse-complement scores.</td></tr> 131 <tr><td>wtRC / weightRC</td><td>Weighted reverse-complement scoring (<code>WeightedRC</code> 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.</td></tr> 132 </tbody> 133 </table> 134 <p>This produces the 18 model-score columns below (6 sets × alt / ref / delta):</p> 135 <table> 136 <thead><tr><th>Model</th><th>noRC</th><th>avgRC</th><th>weightRC</th></tr></thead> 137 <tbody> 138 <tr><td>Evo2-7B</td><td>Evo2_7B_NoRC_{Ref,Alt,Delta}</td><td>Evo2_7B_AvgRC_{Ref,Alt,Delta}</td><td>Evo2_7B_WeightedRC_{Ref,Alt,Delta}</td></tr> 139 <tr><td>Evo2-40B</td><td>Evo2_40B_NoRC_{Ref,Alt,Delta}</td><td>Evo2_40B_AvgRC_{Ref,Alt,Delta}</td><td>Evo2_40B_WeightedRC_{Ref,Alt,Delta}</td></tr> 140 </tbody> 141 </table> 142 143 <h2 id="fields">Database fields</h2> 144 145 <h3>Identity & position</h3> 146 <table> 147 <thead><tr><th>Column</th><th>Type</th><th>Description</th></tr></thead> 148 <tbody> 149 <tr><td>Variant_ID</td><td>text</td><td>Unique key, Chromosome:Position:Ref:Alt (GRCh38)</td></tr> 150 <tr><td>Chromosome</td><td>text</td><td>Chromosome (1–22; autosomes only)</td></tr> 151 <tr><td>Position</td><td>int</td><td>Base-pair position (GRCh38)</td></tr> 152 <tr><td>RSID</td><td>text</td><td>dbSNP Reference SNP ID (build 156)</td></tr> 153 <tr><td>Reference_Allele</td><td>text</td><td>Reference allele</td></tr> 154 <tr><td>Alternate_Allele</td><td>text</td><td>Alternate allele</td></tr> 155 </tbody> 156 </table> 157 158 <h3>Allele frequency — gnomAD v4.1 <em>joint</em> (9 columns)</h3> 159 <p>Allele frequencies are from the gnomAD v4.1 <strong>joint</strong> release (exomes + genomes combined). 160 Variants were included on <strong>Non-Finnish European MAF ≥ 5%</strong>; the per-ancestry AFs 161 below show how frequency varies across populations (a variant common in one ancestry can be rare in another).</p> 162 <table> 163 <thead><tr><th>Column</th><th>Type</th><th>Description</th></tr></thead> 164 <tbody> 165 <tr><td>Joint_AF</td><td>real</td><td>Joint alternate-allele frequency, all ancestries combined</td></tr> 166 <tr><td>African_AF</td><td>real</td><td>African / African-American</td></tr> 167 <tr><td>Admixed_American_AF</td><td>real</td><td>Admixed American</td></tr> 168 <tr><td>Ashkenazi_Jewish_AF</td><td>real</td><td>Ashkenazi Jewish</td></tr> 169 <tr><td>East_Asian_AF</td><td>real</td><td>East Asian</td></tr> 170 <tr><td>Finnish_AF</td><td>real</td><td>Finnish</td></tr> 171 <tr><td>Middle_Eastern_AF</td><td>real</td><td>Middle Eastern</td></tr> 172 <tr><td>Non_Finnish_European_AF</td><td>real</td><td>Non-Finnish European (the MAF ≥ 5% inclusion cohort)</td></tr> 173 <tr><td>South_Asian_AF</td><td>real</td><td>South Asian</td></tr> 174 </tbody> 175 </table> 176 177 <h3>Evo2 model scores (18 columns)</h3> 178 <table> 179 <thead><tr><th>Column pattern</th><th>Type</th><th>Description</th></tr></thead> 180 <tbody> 181 <tr><td>Evo2_{7B,40B}_{NoRC,AvgRC,WeightedRC}_Alt</td><td>real</td><td>Log-likelihood of the ALT sequence</td></tr> 182 <tr><td>Evo2_{7B,40B}_{NoRC,AvgRC,WeightedRC}_Ref</td><td>real</td><td>Log-likelihood of the REF sequence</td></tr> 183 <tr><td>Evo2_{7B,40B}_{NoRC,AvgRC,WeightedRC}_Delta</td><td>real</td><td>Î = ALT score − REF score under the selected strategy. Strength is |Î|; the sign is the direction of the likelihood shift, not of a phenotypic effect</td></tr> 184 </tbody> 185 </table> 186 187 <h3>SCREEN Registry V4 cCRE annotation</h3> 188 <p>These fields use the eight GRCh38 BED classes supplied in SCREEN Registry V4. A variant overlaps an 189 element when its reference-allele interval <code>[Position â 1, Position â 1 + length(REF))</code> 190 intersects that element. The result describes coordinate overlap, not a validated regulatory effect. 191 Multiple classes and element accessions are retained, separated by <code>|</code>.</p> 192 <table> 193 <thead><tr><th>
193Column</th><th>Type</th><th>Description</th></tr></thead> 194 <tbody> 195 <tr><td>SCREEN_V4_Classes</td><td>text</td><td>Distinct overlapping SCREEN V4 cCRE classes. Separate from the legacy <code>ENCODE_cCRE_Annotation</code> field.</td></tr> 196 <tr><td>SCREEN_V4_Accessions</td><td>text</td><td>Distinct overlapping SCREEN V4 element identifiers (EH38E accessions).</td></tr> 197 <tr><td>SCREEN_V4_Overlap_Count</td><td>int</td><td>Number of distinct overlapping elements; zero means no overlap in this registry.</td></tr> 198 </tbody> 199 </table> 200 201 <h3>Gene & functional annotation (ANNOVAR / Ensembl)</h3> 202 <table> 203 <thead><tr><th>Column</th><th>Type</th><th>Description</th></tr></thead> 204 <tbody> 205 <tr><td>Functional_Annotation</td><td>text</td><td>Region class (intergenic, intronic, exonic, UTR…)</td></tr> 206 <tr><td>Gene</td><td>text</td><td>Ensembl gene symbol(s)</td></tr> 207 <tr><td>Gene_Detail</td><td>text</td><td>Transcript / distance detail</td></tr> 208 <tr><td>Exonic_Function</td><td>text</td><td>Coding consequence (nonsynonymous / synonymous / stopgain…)</td></tr> 209 <tr><td>Amino_Acid_Change</td><td>text</td><td>Amino-acid change & transcript (coding variants)</td></tr> 210 <tr><td>ENCODE_cCRE_Annotation</td><td>text</td><td>Legacy functional/regulatory label, retained for the original strata. Values such as Coding and Intronic are not SCREEN Registry V4 cCRE classes.</td></tr> 211 </tbody> 212 </table> 213 214 <h3>Conservation</h3> 215 <table> 216 <thead><tr><th>Column</th><th>Type</th><th>Description</th></tr></thead> 217 <tbody> 218 <tr><td>PhastCons_100way</td><td>real</td><td>phastCons conservation, 100-way vertebrate (0–1)</td></tr> 219 <tr><td>PhyloP_100way</td><td>real</td><td>phyloP conservation, 100-way vertebrate (signed)</td></tr> 220 </tbody> 221 </table> 222 223 <h3>Repeats</h3> 224 <table> 225 <thead><tr><th>Column</th><th>Type</th><th>Description</th></tr></thead> 226 <tbody> 227 <tr><td>Repeat_Name</td><td>text</td><td>RepeatMasker repeat name</td></tr> 228 <tr><td>Repeat_Class</td><td>text</td><td>RepeatMasker repeat class (LINE/SINE/LTR/DNA…)</td></tr> 229 <tr><td>Repeat_Family</td><td>text</td><td>RepeatMasker repeat family</td></tr> 230 </tbody> 231 </table> 232 <p style="color:var(--c-muted);font-size:.88rem">Missing values are stored as NULL (source tokens <code>.</code> / <code>NA</code>).</p> 233 234 <h2>Quick start</h2> 235 <ul> 236 <li>Search by <a href="/evo2/lookup_by_rsid?rsid=rs429358">rsID</a>, <a href="/viewer?gene=APOE">gene</a>, or <a href="/viewer?chr=19&start=43900000&end=45900000">region</a> from the home page.</li> 237 <li>The <a href="/viewer">region viewer</a> plots an interactive Î landscape for any gene or window.</li> 238 <li><a href="/variant/rs429358">Variant Analysis</a>, <a href="/regionset">Region/Region Set Analysis</a> and <a href="/gwas">Trait Analysis</a> are independent modules. The <a href="/workflow">Guided Workflow</a> starts from one explicitly selected study; it does not change the independent gene examples or recompute S-LDSC.</li> 239 <li>Filter and facet the whole table at <a href="/evo2/variants">/evo2/variants</a>.</li> 240 <li>Run SQL at <a href="/evo2">/evo2</a>; add <code>.json</code> to any page for the API.</li> 241 </ul> 242 243 <h2>Reading evidence by source</h2> 244 <p>Evidence is organized by its database or original study, with the source version, record identity and 245 matching status retained. A source that has no matching record is not evidence of no biological effect. 246 Original association P-values, PIP, experimental readouts, clinical classifications and third-party model 247 scores describe different quantities and should be read on their own scales.</p> 248 <p>The Variant view contains individual variant records. Existing gene, intron, assay or signal models 249 belong to Region/Gene, Region set or study views, where their original sample size and the current 250 selection's coverage can be distinguished. Those models are Evo2VED analyses of the named source, 251 rather than statistics supplied by that source or models refitted to the current window.</p> 252 253 <h2 id="cite">Cite</h2> 254 <ul> 255 <li>Zhu et al. <em>Evo2VED: a database of DNA large language modelâderived scores for genome-wide variant interpretation.</em> 256 Manuscript under revision, 2026. <a href="/updates">Website release history</a>. Publication details will be added when available.</li> 257 <li>Brixi, G., Durrant, M. G., Ku, J., et al. <em>Genome modelling and design across all domains of life with Evo 2.</em> 258 <strong>Nature</strong> (2026). <a href="https://doi.org/10.1038/s41586-026-10176-5">doi:10.1038/s41586-026-10176-5</a>.</li> 259 </ul> 260 <p>License: <a href="https://www.apache.org/licenses/LICENSE-2.0">Apache-2.0</a>.</p> 261 262 <a class="evo2-back" href="/">← Back to search</a> 263</div> 264 265</section> 266</div> 267<footer class="ft">Powered by <a href="https://datasette.io/" title="Datasette v">Datasette</a> 268 269</footer> 270
271<script> 272document.body.addEventListener('click', (ev) => { 273 /* Close any open details elements that this click is outside of */ 274 var target = ev.target; 275 var detailsClickedWithin = null; 276 while (target && target.tagName != 'DETAILS') { 277 target = target.parentNode; 278 } 279 if (target && target.tagName == 'DETAILS') { 280 detailsClickedWithin = target; 281 } 282 Array.from(document.getElementsByTagName('details')).filter( 283 (details) => details.open && details != detailsClickedWithin 284 ).forEach(details => details.open = false); 285}); 286</script>
286 287 288 289 290 291</body> 292</html>
Line numbers count LF bytes from the start of the resource, as the search results do. Vendor segments are library code the classifier recognised; they are stored but not indexed. Bytes are shown as Latin1 characters, one per byte.