InputGradientsOutput

class alphagenome.models.v2.dna_output.InputGradientsOutput(gradients, interval, scalar_score, target_gene, track_metadata, output_type)

Model outputs for an input sequence gradient prediction.

gradients

Mean-centered input gradients times the one-hot input sequence, of shape [sequence_length, 4] (float32). Only the reference base is non-zero, so summing over the last axis gives per-position scores.

interval

Genomic interval corresponding to the sequence, centered on the gene’s median TSS. Positions beyond the chromosome ends are N-padded.

scalar_score

Objective the gradients are taken of: the log of the target gene’s prediction averaged over the tracks in track_metadata (summed over the gene’s exon positions for track outputs such as RNA_SEQ).

target_gene

Metadata of the target gene.

track_metadata

Metadata for the tracks aggregated in the gradient target.

output_type

Output type for which gradients were computed.

gene_id

Ensembl gene identifier from target_gene.

gene_name

Optional gene symbol from target_gene.

biosample_name

Single biosample name shared across track_metadata.

Methods

to_track_data(*[, signed, name])

Converts per-nucleotide input gradients into a 1-bp resolution TrackData.