Genomic Variant Effect

Genomic Variant Effect scores variants by comparing model embeddings for reference and alternate sequence windows.

Use it for

• local variant prioritization experiments;
• quick checks on small VCF/VCF.GZ files;
• generating BED tracks from model-derived variant scores;
• building exploratory genomics pipelines.

Inputs

• Reference FASTA/FA/FNA file, or an inline reference sequence.
• VCF or VCF.GZ variant file.
• Reference name when needed.
• Window start.
• Flank size.
• Max variants.
• Max tokens.
• Nucleotide Transformer AI Model.

Compatible models

• Nucleotide Transformer v2 50M
• Nucleotide Transformer v2 500M

Outputs

• Variant scores CSV.
• BED genome track.
• JSON summary.
• Variant count.
• Top embedding delta.
• Warnings.
• Provenance.

How to read the result

The main score is based on how much the alternate sequence changes the model embedding compared with the reference sequence.

A higher score means a bigger model-representation change. It does not mean the variant is automatically pathogenic, causal, or clinically important.

Always check warnings. The most important warning is a REF mismatch: it means the VCF reference allele does not match the FASTA sequence Liatir used.

Good first pipeline

1. Reference FASTA input.
2. VCF input.
3. Genomic Variant Effect with Nucleotide Transformer 50M.
4. BED output into a genome viewer.
5. Inspect top scores and provenance.

Technical details

Tool ID: ai-genomic-variant-effect

Scoring method: 1 - cosine_similarity(reference_embedding, alternate_embedding).

The current scorer scans VCF/VCF.GZ records sequentially and stops at Max variants. It does not require a .tbi index for the current sequential workflow.