Model output metadata

Output types, units, resolutions, and track metadata columns for human predictions across AlphaGenome 2 and AlphaGenome 1.

AlphaGenome 2

AlphaGenome 2 defines 15 output types in OutputType (see Migrating from AlphaGenome 1 for output types added or changed relative to AlphaGenome 1).

Modalities, channels, and tracks

A modality is a class of molecular readout (such as chromatin accessibility, gene expression, histone modification, TF binding, splicing, CpG methylation, or 3D chromatin contacts) predicted by one or more output types (for example, chromatin accessibility spans ATAC, DNASE, and SC_ATAC). Each output type predicts a fixed set of channels shared across conditions. For most output types, channels represent technical variants of the same assay (such as sequencing protocol, polyA+ versus total RNA, or forward + and reverse - strands), whereas for ChIP-seq (CHIP_HISTONE and CHIP_TF) each channel corresponds to a specific histone mark or transcription factor target. A track is one channel in one condition. For example:

• Histone ChIP-seq (CHIP_HISTONE) has 43 channels covering 32 histone marks (H3K27ac, H3K4me3, H3K27me3, and so on) across single- and paired-end assays.

• TF ChIP-seq (CHIP_TF) has 911 channels covering 778 transcription factors.

• scATAC-seq (SC_ATAC) has 5 channels, one per sequencing protocol.

• Stranded assays such as RNA-seq and CAGE have separate forward (+) and reverse (-) strand channels.

Each track carries an imputation_mode label (see Navigating imputed tracks).

Output types

Use the tabs below to browse output types by modality. Counts reflect observed human training tracks (dna_model.observed_metadata()).

Expression

Output type	Measures	Units	Resolution	Biosamples	Tracks
RNA_SEQ	Bulk RNA-seq coverage (polyA+ and total RNA, stranded and unstranded)	Normalized read signal	1 bp	306	738
CAGE	Cap Analysis of Gene Expression (CAGE) 5′-cap signal, concentrated at transcription start sites (also detected along exons)	Normalized read signal	1 bp	268	550
SC_GENE_EXPRESSION	Protocol-specific single-cell RNA-seq expression	Counts per million (CPM)	Gene	1943	5028
SC_GENE_EXPRESSION_DENOISED	Denoised single-cell RNA-seq expression (1 channel per strand)	CPM	Gene	Per call	2

• SC_GENE_EXPRESSION has a separate track for each sequencing chemistry, suspension type (cell or nucleus), and strand, and is supported with both condition= and cell_conditioning=.

• SC_GENE_EXPRESSION_DENOISED gives one denoised expression estimate per gene and strand, without the per-protocol tracks of SC_GENE_EXPRESSION. It is human-only and supported exclusively under expression conditioning (cell_conditioning=).

Accessibility

Output type	Measures	Units	Resolution	Biosamples	Tracks
DNASE	DNase I hypersensitive site cleavage (DNase-seq)	Normalized insertion signal	1 bp	330	385
ATAC	Bulk transposase-accessible chromatin (ATAC-seq)	Normalized insertion signal	1 bp	178	179
SC_ATAC	Single-cell ATAC-seq pseudobulks (5 assay protocols per condition)	Normalized insertion signal	1 bp	704	731

• Single-cell datasets (SC_ATAC and SC_GENE_EXPRESSION) were processed with standard pipelines (such as CellRanger or the ENCODE single-cell pipeline) and pseudobulked by author cell-type labels aligned to ontology terms.

Histone, TF, and methylation

Output type	Measures	Units	Resolution	Biosamples	Tracks
CHIP_HISTONE	Histone modification ChIP-seq	Fold-change over control	128 bp	221	1183
CHIP_TF	Transcription factor binding ChIP-seq	Fold-change over control	128 bp	181	1709
METHYLATION	CpG methylation across 2 channels (bulk WGBS and single-cell methylome)	Methylated fraction	1 bp	185	190

• CHIP_HISTONE covers 32 histone marks (30 in human and 11 in mouse) across 43 channels per condition (accounting for both single- and paired-end sequencing), summed per 128 bp bin.

• CHIP_TF covers 778 transcription factor targets (772 in human and 47 in mouse) across 911 channels per condition (accounting for both single- and paired-end sequencing), summed per 128 bp bin.

Splicing

Output type	Measures	Units	Resolution	Biosamples	Tracks
SPLICE_SITES	Condition-specific splice site classification (5 channels per condition)	Predicted probability	1 bp	227	1135
SPLICE_SITE_USAGE	Fraction of transcripts using each donor or acceptor site	Fraction	1 bp	227	560
SPLICE_JUNCTIONS	Split-read counts across candidate donor-acceptor pairings	Normalized junction signal	1 bp	227	280

• SPLICE_SITES predicts 5 channels per condition (donor +, acceptor +, donor -, acceptor -, and padding .) derived from all splicing assays in that condition.

3D contacts and cells

Output type	Measures	Units	Resolution	Biosamples	Tracks
CONTACT_MAPS	Pairwise 3D chromatin contact frequency (Micro-C and Hi-C), normalized for distance decay	Log-fold over distance expectation	2048 bp	73	73
CELL_EMBEDDINGS	256-dimensional cell embeddings from the cell encoder (cell_conditioning=)	Embedding vector	Cell	Per call	NA

Track metadata columns

To inspect the metadata table for all observed human training tracks:

observed = dna_model.observed_metadata(
    organism=dna_client.Organism.HOMO_SAPIENS
)

Each attribute on observed (such as rna_seq) is a DataFrame with one row per track. The columns are:

Column	Applies to	Description
name	All track outputs	Track identifier (for example, CL:0000047 polyA plus RNA-seq)
strand	All track outputs	+ (forward), - (reverse), or . (unstranded)
biosample_curie	All track outputs	Ontology CURIE or composite condition (for example, UBERON:0000178_CL:0000907)
biosample_name	All track outputs	Plain-text biosample description (for example, motor neuron)
imputation_mode	Predicted outputs	observed, imputed_supported, or imputed_exploratory
data_pipeline, data_source	Most bulk outputs	Preprocessing pipeline and source consortium
paper_doi	Single-cell outputs	Publication DOI for the source single-cell dataset
histone_mark	chip_histone outputs	Profiled histone mark (for example, H3K27ac)
transcription_factor	chip_tf outputs	Assayed transcription factor target
splice_site_type	splice_sites outputs	donor, acceptor, or padding

Every row of observed_metadata() is an observed training track. On predicted outputs (TrackData.metadata and GeneExpressionData.track_metadata) and in variant scores, imputation_mode takes all three values. tidy_scores() renames name and strand to track_name and track_strand to distinguish them from gene_name and gene_strand.

How stranded (+, -) and unstranded (.) assays differ

Strand-specific transcriptional assays (stranded RNA-seq, CAGE, SC_GENE_EXPRESSION_DENOISED, and SC_GENE_EXPRESSION) have separate forward (+) and reverse (-) strand channels to distinguish overlapping antisense transcripts. Chromatin accessibility and unstranded assays (ATAC-seq, DNase-seq, and unstranded GTEx RNA-seq) report a single unstranded (.) track per channel.

For SPLICE_JUNCTIONS, strand is stored per junction rather than per track row, so track metadata contains half as many rows as the total junction tracks.

Observed and imputed tracks

AlphaGenome 2 predicts all channels of the requested output types for the chosen condition and labels each track in metadata['imputation_mode']. To browse all reference conditions and see which output types have observed training data in each condition:

conditions = dna_model.supported_conditions(
    organism=dna_client.Organism.HOMO_SAPIENS
)

Each row is a reference condition, and each output-type column records the best imputation mode across that output type’s channels ('observed' means at least one channel was measured). See Navigating imputed tracks for details on imputation modes and the Choosing a condition tutorial for interactive filtering.

AlphaGenome 1

AlphaGenome 1 (alphagenome.models.v1) predicts 11 output types across human and mouse assays. See the Methods section of the AlphaGenome paper [Avsec et al., 2026] for dataset processing and normalization details.

Output types

Output type	Measures	Units	Resolution	Biosamples	Tracks
RNA_SEQ	Bulk RNA-seq coverage (polyA+ and total RNA, stranded and unstranded)	Normalized read signal	1 bp	285	667
CAGE	Cap Analysis of Gene Expression (CAGE) 5′-cap signal, concentrated at transcription start sites (also detected along exons)	Normalized read signal	1 bp	264	546
PROCAP	Precision run-on sequencing and capping (PRO-cap) at start sites	Normalized read signal	1 bp	6	12
DNASE	DNase I hypersensitive site cleavage (DNase-seq)	Normalized insertion signal	1 bp	305	305
ATAC	Bulk transposase-accessible chromatin (ATAC-seq)	Normalized insertion signal	1 bp	167	167
CHIP_HISTONE	Histone modification ChIP-seq across 24 marks	Fold-change over control	128 bp	219	1116
CHIP_TF	Transcription factor ChIP-seq across 43 targets	Fold-change over control	128 bp	163	1617
SPLICE_SITES	Condition-agnostic donor and acceptor probabilities (4 tracks total)	Predicted probability	1 bp	NA	4
SPLICE_SITE_USAGE	Fraction of transcripts using each donor or acceptor site	Fraction	1 bp	282	734
SPLICE_JUNCTIONS	Split-read counts across candidate donor-acceptor pairings	Normalized junction signal	1 bp	282	734
CONTACT_MAPS	Coarse-grained pairwise 3D contact frequency [Zhou, 2022]	Log-fold over distance expectation	2048 bp	12	28

Track metadata columns

To access track metadata for AlphaGenome 1 human outputs:

output_metadata = dna_model.output_metadata(
    organism=dna_client.Organism.HOMO_SAPIENS
)

Each output type attribute (for example, output_metadata.rna_seq) is a DataFrame with name, strand, ontology_curie, and biosample_name columns (plus type-specific columns such as gtex_tissue on rna_seq for GTEx Consortium [2020] tissues; see the AlphaGenome 1 quick start).

Note

For GTEx tissue 'Brain - Cerebellar hemisphere', metadata uses UBERON:0002245 (cerebellar hemisphere) instead of the UBERON ID listed in the GTEx documentation (UBERON:0002037).