Allele Caller¶

The dnanet.evaluation.allele_caller module implements allele calling strategies that translate pixel-level model outputs into discrete allele calls.

Design Pattern: Strategy¶

AlleleCaller defines the interface; concrete implementations provide different allele-calling algorithms.

Classes¶

from dnanet.evaluation.allele_caller import (
    AlleleCaller,
    FromSegmentationImageCaller,
    NearestBasePairCaller,
    ExactBasePairCaller,
)

NearestBasePairCaller¶

Assigns each predicted region to the nearest allele by base-pair distance. The allele with the smallest absolute difference to the predicted base-pair position is selected.

ExactBasePairCaller¶

Assigns each predicted region to the allele that has no more than 0.3 base-pair distance from the predicted position. Peaks that fall outside any bin by more than 0.3 bp are marked as "Out of Bin".

Both classes inherit from FromSegmentationImageCaller which handles: - Normalizing prediction images (binary or multiclass mode) - Finding connected components of positive predictions - Translating pixel positions to allele calls via the reference panel

Common constructor args: - threshold — Probability threshold for positive predictions (default: 0.5) - exclude_non_autosomal — Filter out non-autosomal markers (default: False) - prediction_mode — 'auto', 'binary', or 'multiclass_labels' (default: 'auto')