API reference

The shorkie package is a thin helper layer over baskerville-yeast. It exists so that model loading, sequence preparation, and variant scoring have exactly one implementation — the same one used by the examples, the CLI, and the paper’s own eQTL scorers.

shorkie.config

Path resolution. Every filesystem path in the repository goes through here.

from shorkie import config

config.load(path=None)      # -> Config  (LRU-cached)
config.get("dotted.key", default=None)
config.path("dotted.key")   # -> pathlib.Path
config.repo_root()          # -> pathlib.Path

Resolution order: $SHORKIE_CONFIG → config/paths.yaml → config/paths.example.yaml. ${token} interpolation resolves against environment variables first, then other dotted keys.

shorkie.models.ensemble

The model API.

N_DNA = 4; N_SPECIES = 166; NUM_FEATURES = 170; SCEREVISIAE_COL = 114
fetch_1hot (fasta, chrom, start, end, seq_len)

Fetch a window and one-hot encode it; left-pads with N for negative starts.

make_input (fasta, chrom, start, end, seq_len=16384, species_col=114, mask_pos=None)

Build the (seq_len, 170) model input: DNA one-hot in channels 0–3, species identity in 4–169. mask_pos zeroes the DNA channels at one position — the LM-style masking used for masked-base prediction.

load_ensemble (model_dir, params_file, target_index, num_folds=8, fold_subdir=”f{fold}c0”)

Load an N-fold ensemble from {model_dir}/train/{fold_subdir}/train/model_best.h5, sliced to target_index and wrapped with reverse-complement averaging.

ensemble_predict (models, x) → (1, 1, bins, tracks)

Average the fold predictions. Aliased as predict.

logSED (y_ref, y_alt, gene_slice) → float

log2(Σ_alt + 1) − log2(Σ_ref + 1) over the gene-body bins, averaged across tracks.

logSED_per_track (y_ref, y_alt, gene_slice) → (tracks,)

The same, per track — use this to ask which assays a variant affects.

shorkie.helpers.yeast_helpers

Sequence, attribution, and plotting utilities used by the figure notebooks.

  • make_seq_1hot — plain 4-channel one-hot (no species channel)

  • process_sequence — build model input plus a gene-annotation slice

  • predict_tracks — like ensemble_predict, but concatenates folds instead of averaging

  • get_ism, get_ism_shuffle — in-silico mutagenesis

  • get_prediction_gradient (and _w_rc / _noisy variants) — gradient attribution

  • compute_scores — the broader SAD-style variant statistic family (SUM / logSUM / sqrtSUM / SAX / D1 / D2 / JS / logJS)

  • dna_letter_at, plot_seq_scores, plot_coverage_tracks — sequence logos and coverage plots

shorkie.viz.load_cov

Ground-truth coverage I/O, for comparing predictions against observed data.

  • CovFace(cov_file) — reads bigWig, BED, or HDF5; .read(chrom, start, end)

  • read_coverage(file, chrom, start, end)

  • seq_norm(...) — bin per-nucleotide coverage into the model’s 16 bp output bins, mirroring hound_data’s target processing

shorkie.data

  • bed_helper — extract_intervals, get_exon_mask, generate_beds

  • util — job-submission and parallel-execution helpers

Note

Submodules are not auto-imported. Import them explicitly:

from shorkie.models.ensemble import load_ensemble, make_input, logSED
from shorkie.helpers.yeast_helpers import get_ism
from shorkie.viz.load_cov import read_coverage