| name | borzoi-regulatory-prediction |
| description | Run pinned Borzoi models to predict RNA-seq and regulatory tracks from long DNA sequence and compare reference versus alternate alleles. Use for regulatory variant hypotheses with explicit genome assembly, sequence-window, tissue, and track provenance. |
| license | MIT |
Borzoi Regulatory Prediction
Gate
Ask once before cloning/installing a pinned Borzoi revision, downloading model
weights, or using GPU compute. Keep genomic sequence local unless remote use is
approved. Follow $cx-compute-environment.
Workflow
- Pin genome assembly, coordinates, strand, reference/alternate alleles,
transcript/annotation release, tissue/track IDs, input length, and boundary
handling. Verify reference alleles against the assembly.
- Pin code commit, ensemble/checkpoint checksums, preprocessing, target metadata,
shifts/reverse-complement ensemble, precision, hardware, and seeds.
- Smoke-test an official example. Predict into
artifacts/<run-id>/borzoi/; retain input windows, coordinate maps, target
labels, raw tracks, reference/alternate differences, plots, logs, and failures.
- Summarize effect direction/magnitude by declared tissues and transcripts;
report ensemble variability and sensitivity to window placement. Compare
simple conservation/motif and nearby-QTL baselines where relevant.
- Record with
$science-provenance; review with $science-review.
Boundaries
- Predicted expression effects are not causal, pathogenic, or clinical evidence.
- Prevent overlapping/LD-correlated windows and the same locus from crossing
evaluation splits; audit training-data and annotation-release overlap.