Migrating from AlphaGenome 1

Both models are available through the alphagenome package:

from alphagenome.models.v2 import dna_client  # AlphaGenome 2 (recommended)
from alphagenome.models.v1 import dna_client as dna_client_v1  # AlphaGenome 1

When to use each model

• Use AlphaGenome 2 for most analyses: predicting tracks or scoring variants in a reference condition (condition=) or under expression conditioning on your own human scRNA-seq count matrix (cell_conditioning=), predicting single-cell accessibility (SC_ATAC), denoised single-cell gene expression (SC_GENE_EXPRESSION_DENOISED), protocol-specific single-cell gene expression (SC_GENE_EXPRESSION), or CpG methylation (METHYLATION), computing input gradients for enhancer-to-gene (E2G) attribution (predict_interval_gradients), encoding single-cell profiles (encode_cells), or using cross-modality track imputation.

• Use AlphaGenome 1 to reproduce results from the AlphaGenome paper [Avsec et al., 2026], score variants across all tissues and cell types at once without choosing a condition, or predict PRO-cap tracks.

AlphaGenome 2 architecture. Takes a DNA sequence of up to 1 Mb, the species, and a condition (a reference condition, or a cell embedding from the cell encoder under expression conditioning) to predict observed and imputed coverage tracks, gene expression, splicing, and contact maps.

AlphaGenome 1 architecture. Takes a DNA sequence of up to 1 Mb and predicts observed multi-task coverage tracks, splicing, and pairwise contact maps across human and mouse assays.

API migration reference

Both models share the same core prediction and scoring methods; the differences are in how you choose conditions, discover metadata, and read scores.

Conditions

	AlphaGenome 1	AlphaGenome 2
Predict or score one variant or interval	predict_*: optional ontology_terms list (pass None for all tissues and cell types); score_*: no ontology_terms argument (all tissues and cell types are always scored)	Required: one reference condition (condition=) or expression conditioning on your own human scRNA-seq counts (cell_conditioning= plus a cell_condition_name= label; up to 256 cells per call, raw unique molecular identifier (UMI) counts with Ensembl IDs)
Batch methods (predict_sequences, predict_intervals, predict_variants, score_intervals, score_variants, and in AlphaGenome 2 predict_interval_gradients_batch)	predict_*: optional ontology_terms list (pass None for all tissues and cell types) applied across all items; score_*: no ontology_terms argument (predict_interval_gradients_batch not available)	One reference condition (conditions=) or expression-conditioning input (cell_conditionings= plus cell_condition_names=) per item. Pass conditions= and variant_scorers= as keywords when porting AlphaGenome 1 score_variants calls.
Condition format	Single ontology CURIE	Composite condition (ontology CURIEs joined by underscores), single CURIE, or a ModelCondition

Metadata

	AlphaGenome 1	AlphaGenome 2
What can I predict?	dna_model.output_metadata()	dna_model.supported_conditions(), one row per condition
Which tracks were measured?	Same table	dna_model.observed_metadata()
Biosample column	ontology_curie	biosample_curie, plus an imputation_mode for every track

Output types

Added in AlphaGenome 2 (see Model output metadata for descriptions):

• SC_GENE_EXPRESSION_DENOISED: denoised single-cell gene expression under cell_conditioning=, one channel per strand in CPM

• SC_GENE_EXPRESSION

• SC_ATAC

• METHYLATION

• CELL_EMBEDDINGS

Changed in AlphaGenome 2

• SPLICE_SITES: the 4 condition-agnostic tracks of AlphaGenome 1 are now condition-specific, with 5 channels per condition (donor +, acceptor +, donor -, acceptor -, padding .).

• CONTACT_MAPS: targets are not coarse-grained, so values are not directly comparable to AlphaGenome 1 contact maps.

Removed in AlphaGenome 2

• PROCAP: use CAGE for transcription start site activity.

Variant scores

	AlphaGenome 1	AlphaGenome 2
Score columns	raw_score and quantile_score	raw_score only (no quantile_score column)

How empirical quantile scores work in AlphaGenome 1

Different output types yield raw variant scores on different scales (for example, splice site usage scores lie in \([0, 1]\) whereas RNA-seq log-fold change scores are signed and not restricted to that range). AlphaGenome 1 (alphagenome.models.v1) estimates a background distribution for each recommended scorer and track across ~300,000 common variants (MAF > 0.01 in any gnomAD v3 population; see [Avsec et al., 2026]) and converts each raw score into a percentile rank bounded by ±0.99999.

For signed variant scorers, \([0, 1]\) rank probabilities are linearly mapped to \([-1, 1]\) so that negative scores map near \(-1\), scores near zero map to \(0\), and strongly positive scores map near \(+1\).

Key behavioral differences

• One condition per call: With AlphaGenome 2, each call to predict_sequence, predict_interval, predict_variant, or score_variant predicts for one condition; loop over conditions when comparing multiple tissues or cell types (or pass one reference condition or expression-conditioning input per item to the batch methods).

• Complete channel output with per-track imputation labels: For every condition, AlphaGenome 2 predicts all channels across requested output types and labels each track in .metadata['imputation_mode'] as observed, imputed_supported, or imputed_exploratory (see Navigating imputed tracks).