How variant scoring works

AlphaGenome 2 scores a variant by predicting reference (REF) and alternate (ALT) profiles in one condition, restricting both to a spatial mask, and summarizing their difference per track (and per gene, junction, or CpG site where applicable).

Scoring pipeline

Each variant scorer transforms paired REF and ALT predictions into scalar scores through four steps:

1. Predict REF and ALT

Variant scoring begins by running forward passes on the REF and ALT sequences across the requested genomic interval and output type (such as RNA_SEQ, ATAC, or SPLICE_SITES).

Indel coordinate alignment

For insertions and deletions (indels), the ALT profile is aligned to REF genomic coordinates before comparison. Inserted bases are summarized by taking the maximum predicted signal across the inserted segment, while deleted bases are assigned zero signal in ALT so both profiles share identical positions.

2. Apply spatial mask

A spatial mask restricts scoring to the positions relevant to the assay, discarding predictions outside the masked coordinates:

• Variant-centered windows capture local cis-regulatory changes around the variant: 501 bp for ATAC, DNASE, CAGE, and CHIP_TF; 1201 bp for SC_ATAC and shared CpG sites in METHYLATION; 2001 bp for broad histone marks in CHIP_HISTONE; and 1 Mb for 3D interactions in CONTACT_MAPS.

• Gene annotations from a GTF reference restrict scoring to exons (RNA_SEQ), gene bodies including introns (SPLICE_SITES, SPLICE_SITE_USAGE), top splice sites (SPLICE_JUNCTIONS), or 400-bp windows around 3’ cleavage sites (POLYADENYLATION).

3. Aggregate and compare

Within the spatial mask, each scorer reduces positions along the sequence axis (mean, sum, or max), applies any optional log or \(\ell_2\) scaling, and compares ALT against REF to produce one scalar value per track (and per gene or CpG site when applicable).

How AggregationType names encode order of operations

Options in AggregationType list operations from right to left, with the rightmost operation applied first to the masked predictions:

• DIFF_LOG2_SUM sums predicted coverage across the window first, transforms with \(\log_2(\text{sum} + 1)\), and then takes ALT minus REF.

• DIFF_SUM_LOG2 transforms each position with \(\log_2(\text{value} + 1)\), sums across the mask, and then takes ALT minus REF.

4. Tidy and summarize

Raw scoring outputs are returned as AnnData objects per variant and scorer. Calling tidy_scores() flattens them into a tidy table with raw_score, track metadata (including imputation_mode), optional strand matching (match_gene_strand), and optional filtering to protein-coding genes.

Recommended variant scorers

RECOMMENDED_VARIANT_SCORERS provides 23 preconfigured scorers across expression, chromatin, splicing, methylation, 3D contacts, and active-allele baseline activity. get_recommended_scorers() returns the subset supported for a given organism, and includes SC_GENE_EXPRESSION_DENOISED only when cell_conditioning=True:

Expression and polyadenylation

Gene-expression scorers (GeneMaskLFCScorer) are signed (ALT minus REF); POLYADENYLATION (PolyadenylationScorer) is unsigned:

Key	Mask	Score
RNA_SEQ	Gene exons	\(\log(\text{mean}(\text{ALT}) + 0.001)\) \(- \log(\text{mean}(\text{REF}) + 0.001)\)
SC_GENE_EXPRESSION_DENOISED	Gene row (GeneExpressionData)	\(\log(\text{ALT} + 0.001) - \log(\text{REF} + 0.001)\) on gene-level CPM (one channel per strand; requires cell_conditioning=)
SC_GENE_EXPRESSION	Gene row (GeneExpressionData)	\(\log(\text{ALT} + 0.001) - \log(\text{REF} + 0.001)\) on gene-level CPM (one track per chemistry, suspension type, and strand)
POLYADENYLATION	400-bp windows at 3’ cleavage sites	Max absolute log-fold change of distal/proximal PAS usage ratios [Linder et al., 2025]

Chromatin, TF, and methylation

All chromatin, transcription factor, and methylation scorers are signed (positive when ALT increases signal or methylation relative to REF). Spatial tracks use CenterMaskScorer; methylation uses MethylationScorer:

Key	Mask	Score
ATAC, DNASE, CAGE, CHIP_TF	501 bp centered on variant	\(\log_2((\text{sum}(\text{ALT})+1)/(\text{sum}(\text{REF})+1))\)
SC_ATAC	1201 bp centered on variant	\(\log_2((\text{sum}(\text{ALT})+1)/(\text{sum}(\text{REF})+1))\)
CHIP_HISTONE	2001 bp centered on variant	\(\log_2((\text{sum}(\text{ALT})+1)/(\text{sum}(\text{REF})+1))\)
METHYLATION	1201 bp centered on variant (shared CpGs)	Per-CpG \(\operatorname{logit}(\text{ALT})-\operatorname{logit}(\text{REF})\) of the methylated fraction, for CpG sites present in both alleles (CpGs created or destroyed by the variant are omitted)

RNA splicing and 3D contacts

All splicing and 3D contact scorers (GeneMaskSplicingScorer, SpliceJunctionScorer, and ContactMapScorer) are unsigned magnitudes:

Key	Mask	Score
SPLICE_SITES	Gene body (exons and introns)	\(\max\lvert\text{ALT} - \text{REF}\rvert\) across donor and acceptor probabilities
SPLICE_SITE_USAGE	Gene body (exons and introns)	\(\max\lvert\text{ALT} - \text{REF}\rvert\) across splice site usage fractions
SPLICE_JUNCTIONS	Top-k annotated and predicted splice sites	\(\max\lvert\log(\text{ALT}) - \log(\text{REF})\rvert\) across junction read counts
CONTACT_MAPS	1 Mb centered on variant	Mean \(\lvert\text{ALT} - \text{REF}\rvert\) across interactions with the variant bin

Active-allele scorers

Active-allele scorers (GeneMaskActiveScorer and CenterMaskScorer) quantify baseline signal intensity rather than allelic change by taking the maximum of the aggregated REF and ALT signals within the mask:

Key	Mask	Score
RNA_SEQ_ACTIVE	Gene exons	\(\max(\text{mean}(\text{ALT}), \text{mean}(\text{REF}))\)
SC_GENE_EXPRESSION_ACTIVE	Gene row (GeneExpressionData)	\(\max(\text{sum}(\text{ALT}), \text{sum}(\text{REF}))\)
ATAC_ACTIVE, DNASE_ACTIVE, CAGE_ACTIVE, CHIP_TF_ACTIVE	501 bp centered on variant	\(\max(\text{sum}(\text{ALT}), \text{sum}(\text{REF}))\)
SC_ATAC_ACTIVE	1201 bp centered on variant	\(\max(\text{sum}(\text{ALT}), \text{sum}(\text{REF}))\)
CHIP_HISTONE_ACTIVE	2001 bp centered on variant	\(\max(\text{sum}(\text{ALT}), \text{sum}(\text{REF}))\)

Next steps

Scoring a variant

Run recommended scorers on single variants and batches, tidy outputs, and inspect top cell types and genes.

Scoring a variant

Analyzing a splicing variant

Compare splice site probabilities, splice site usage, and splice junction read counts on pathogenic variants.

Analyzing a splicing variant

Variant scorers API reference

Inspect all scorer classes, supported mask widths, and tidy_scores() options.

api/models_v2.html#variant-scorers