AlphaGenome predictions

AlphaGenome reads a megabase of DNA and predicts where it is transcribed, where the chromatin is open, how it splices, and how it folds. We ask it about TAL1 in two cell lines, stack the answers that share units on one axis, and then score an oncogenic insertion against the reference to see the enhancer it creates.

Prerequisites#

• the AlphaGenome plugin, for the adapters these tracks use, the query panel and the variant right-click item. In your own JBrowse, load it by URL; the sessions below already load it

Where the data comes from#

We ran two predictions over the TAL1 locus against AlphaGenome and stored them under stable tokens, following the AlphaGenome team's worked example.

Nothing to download: the plugin reads both stored predictions by URL.

Predicted assays from sequence#

Given a window of reference sequence and a cell type, AlphaGenome predicts the RNA-seq, DNase, ATAC, CAGE, PRO-cap and ChIP-seq signal an experiment would produce there, plus splice junctions and a contact map. Changing one base and asking again gives a second prediction, and the difference is the variant's predicted effect.

The window is a megabase centered on TAL1, which is on in K562, an erythroleukemia line, and off in GM12878, a lymphoblastoid line, the control.

The AlphaGenome plugin#

The plugin is not in the plugin store yet, so it loads by URL from the plugins array in config.json (see configuring plugins):

{
  "plugins": [
    {
      "name": "AlphaGenome",
      "umdUrl": "https://jbrowse.org/demos/alphagenome-plugin/jbrowse-plugin-alphagenome.umd.js"
    }
  ]
}

On JBrowse Desktop, add that URL as a custom plugin named AlphaGenome.

The plugin sends requests to a small service that holds the API key, runs the prediction and stores the arrays. The public instance is the default. To use your own, the page or app that embeds JBrowse calls setApiRoot, a function the plugin module exports, with the service's root URL; config.json has no slot for it.¹

Asking AlphaGenome for a prediction#

AlphaGenome predicts from hg38 sequence, so we load that assembly, the RefSeq genes and the oncogenic TAL1 variants:

{
  "name": "hg38",
  "uri": "https://hgdownload.soe.ucsc.edu/goldenPath/hg38/bigZips/hg38.2bit",
  "refNameAliases": {
    "uri": "https://hgdownload.soe.ucsc.edu/goldenPath/hg38/bigZips/hg38.chromAlias.txt"
  }
}

jbrowse add-assembly https://hgdownload.soe.ucsc.edu/goldenPath/hg38/bigZips/hg38.2bit \
  --name hg38 \
  --refNameAliases https://hgdownload.soe.ucsc.edu/goldenPath/hg38/bigZips/hg38.chromAlias.txt

https://hgdownload.soe.ucsc.edu/goldenPath/hg38/bigZips/hg38.2bit

{
  "type": "FeatureTrack",
  "trackId": "genes",
  "name": "NCBI RefSeq genes",
  "assemblyNames": ["hg38"],
  "adapter": {
    "type": "Gff3TabixAdapter",
    "uri": "https://jbrowse.org/genomes/GRCh38/ncbi_refseq/GCA_000001405.15_GRCh38_full_analysis_set.refseq_annotation.sorted.gff.gz"
  }
}

jbrowse add-track https://jbrowse.org/genomes/GRCh38/ncbi_refseq/GCA_000001405.15_GRCh38_full_analysis_set.refseq_annotation.sorted.gff.gz \
  --trackId genes \
  --name "NCBI RefSeq genes" \
  --assemblyNames hg38

{
  "type": "FeatureTrack",
  "trackId": "genes",
  "name": "NCBI RefSeq genes",
  "assemblyNames": ["hg38"],
  "adapter": {
    "type": "Gff3TabixAdapter",
    "uri": "https://jbrowse.org/genomes/GRCh38/ncbi_refseq/GCA_000001405.15_GRCh38_full_analysis_set.refseq_annotation.sorted.gff.gz"
  }
}

The variants are a BED whose header names REF and ALT columns after the usual four, since the right-click prediction item needs those alleles on each feature. For your own variants, swap uri for a VCF or a BED with the same columns:

{
  "type": "VariantTrack",
  "trackId": "tal1_variants",
  "name": "Oncogenic TAL1 variants",
  "assemblyNames": ["hg38"],
  "adapter": {
    "type": "BedAdapter",
    "uri": "https://jbrowse.org/demos/alphagenome_test.bed"
  }
}

jbrowse add-track https://jbrowse.org/demos/alphagenome_test.bed \
  --trackId tal1_variants \
  --name "Oncogenic TAL1 variants" \
  --assemblyNames hg38 \
  --trackType VariantTrack

{
  "type": "VariantTrack",
  "trackId": "tal1_variants",
  "name": "Oncogenic TAL1 variants",
  "assemblyNames": ["hg38"],
  "adapter": {
    "type": "BedAdapter",
    "uri": "https://jbrowse.org/demos/alphagenome_test.bed"
  }
}

The session below loads the hosted config with the plugin and these tracks and opens the locus with no predictions yet.

{
  "defaultSession": {
    "name": "TAL1",
    "views": [
      {
        "id": "alphagenome_lgv",
        "type": "LinearGenomeView",
        "assembly": "hg38",
        "loc": "chr1:47,189,833..47,259,832",
        "tracks": ["genes", "tal1_variants"]
      }
    ]
  }
}

jbrowse set-default-session --session - << 'EOF'
{
  "name": "TAL1",
  "views": [
    {
      "id": "alphagenome_lgv",
      "type": "LinearGenomeView",
      "assembly": "hg38",
      "loc": "chr1:47,189,833..47,259,832",
      "tracks": ["genes", "tal1_variants"]
    }
  ]
}
EOF

AlphaGenome predictions… in the view menu opens the query panel, which takes four inputs:

• what to predict: any of the eleven output types, or a preset group
• which cell types and tissues: a search over the biosample catalog. This page uses K562 and GM12878 for the locus, and CD34+ progenitors for the variant
• how wide a window: 16 kb, 100 kb, 500 kb or 1 Mb, centered on the view
• a variant: left empty, the panel predicts the reference

A wide request takes minutes, and the browser polls until it finishes. The service keys each request by its content, so repeating a request this page already made returns at once.²

Predicted TAL1 expression in two cell lines on one axis#

Open the list of a finished prediction's tracks, often thousands, and stack two on one axis:

• Type polyA plus into the list's filter.
• Tick K562 and GM12878.
• Click Add 2 selected to put both in one multi-wiggle track on a shared y-axis.

The panel stacks any tracks in the same units this way. For one track per pick, untick Stack on a shared scale.

STIL, to the right, is predicted in both lines, which shows the GM12878 row loaded.

Where the chromatin is open#

Add four accessibility rows to the prediction:

• Type DNase into the track filter and tick K562 and GM12878.
• Type ATAC and tick both again; ticks survive the filter change.
• Click Add 4 selected to put all four in one track on one axis.

The GM12878 rows have peaks too, at open chromatin that is not transcribed.

Predicted TAL1 splice junctions in each cell line#

Splice junctions come back as arcs joining exons, as in a sashimi plot. Add the K562 and GM12878 polyA junctions and zoom in to TAL1.³

The predicted GM12878 contact map over the whole window#

AlphaGenome predicts Hi-C-style contact maps at 2 kb bins for about a dozen cell lines, GM12878 among them. Add the GM12878 contact map and zoom out to the whole predicted megabase:

Scoring a TAL1 variant against the reference#

In T-cell acute lymphoblastic leukemia (T-ALL), TAL1 is switched on in a lineage where it should be silent. One route is a small insertion upstream of the gene that creates a binding site. The variant track holds the insertions reported in T-ALL patients and cell lines, all at one site.

Right-click any variant and choose Predict variant effect with AlphaGenome. The query panel opens with the variant loaded, and running it predicts the window twice, once for the reference and once with the variant.

The stored prediction is for the Jurkat insertion, chr1:47239296 C>CCGTTTCCTAACC, in CD34+ common myeloid progenitors, the closest AlphaGenome cell type to the CD34+ hematopoietic cells Mansour et al. studied. To reach it, replace K562 in the biosample box with "common myeloid progenitor, CD34-positive" and type the insertion into the variant box. Any other variant or biosample makes a live API call.

The panel warns when the variant falls outside the window. In variant mode AlphaGenome reports only the junctions the variant could affect, so clear the variant to get the whole locus back.

Predicted effect of the Jurkat insertion on TAL1#

A variant prediction returns each assay as two tracks: the reference and alternate curves together, and a difference row (alternate minus reference) where positive is a gain from the insertion.⁴ To read the differences:

• Add the CD34+ DNase, polyA plus RNA-seq and H3K27ac tracks.
• Zoom to TAL1 and the insertion, and close the reference and alternate tracks to keep the three difference rows.
• Untick Y axis... → Clip extreme outliers on the DNase difference row, because on a row this sparse the default clipping flattens the gain at the insertion.

Accessibility rises at the insertion. H3K27ac, the mark of an active enhancer, rises across the locus, and predicted transcription rises over the TAL1 exons.

Sharing prediction tracks between sessions#

The panel adds predictions as session tracks whose adapters read the stored arrays through presigned URLs that expire within the hour. A track config copied into another session stops loading once they expire, so to share a prediction, re-open the panel there and repeat the query, which returns the same stored arrays with fresh URLs.

See also#

• Tutorial: RNA-seq visualization
• Tutorial: ChromHMM chromatin states
• Tutorial: Structural variants from Hi-C
• Tutorial: TP53 from prediction to crystal
• Config guide: Plugins

Citations#

• Avsec Ž, et al. AlphaGenome: advancing regulatory variant effect prediction with a unified DNA sequence model. bioRxiv (2025). https://doi.org/10.1101/2025.06.25.661532
• Mansour MR, et al. An oncogenic super-enhancer formed through somatic mutation of a noncoding intergenic element. Science 346:1373-1377 (2014). https://doi.org/10.1126/science.1259037

Feedback on this tutorial is welcome: contact us.