Regulatory Prediction

Regulatory Prediction uses long-context genomic models to predict signal across a DNA sequence window. It can also compare reference and alternate windows for selected variants.

Use it for

• exploring predicted regulatory signal from DNA sequence;
• creating BED tracks for genome-viewer inspection;
• comparing variant effects on predicted regulatory output;
• testing Enformer, Basenji2, or Borzoi Mini workflows locally.

Inputs

• Reference FASTA/FA/FNA file, or an inline DNA sequence.
• Optional VCF or VCF.GZ file.
• Reference name when needed.
• Window start.
• Output head, usually Human.
• Target index.
• Max variants.

Compatible models

• Enformer Regulatory Prediction
• Basenji2 Human Regulatory
• Borzoi Mini K562 RNA-seq

Outputs

• Regulatory signal CSV.
• Signal BED track.
• Optional variant scores CSV.
• Optional variant score BED track.
• JSON summary.
• Bin count.
• Variant count.
• Top variant delta.
• Warnings.
• Provenance.

How to read the result

Start with the signal track. It shows predicted model output across bins in the input sequence window.

The targetIndex selects which output track to inspect. This is currently a technical parameter. Index 0 is useful for a first smoke test, but scientific interpretation requires knowing which biological assay or target that index represents.

If a VCF is provided, variant scores show how much the predicted signal changed between reference and alternate windows.

Technical details

Tool ID: ai-regulatory-prediction

Each model has its own isolated TensorFlow runtime box. Liatir does not share these runtimes with Nucleotide Transformer or CellTypist because package versions, model files, and input windows differ.