Genomic Region
A genomic region is a contiguous interval on a genome assembly. Region-level modeling treats the interval as the input object rather than a whole genome.
A region can be represented as:
where and are start and end coordinates.
Coordinate conventions are part of the data contract:
Mixing these conventions can shift every extracted sequence by one base.
Modeling Uses
- Sequence representation learning over fixed-length windows.
- Regulatory or functional annotation prediction.
- Variant context extraction.
- Retrieval of sequence neighborhoods.
Region Contract
| Field | Why It Matters |
|---|---|
| Assembly | coordinates are only meaningful relative to a reference version |
| Chromosome / contig | naming conventions differ across sources |
| Start / end convention | off-by-one errors change the sequence |
| Strand | reverse complement may be required |
| Window rule | centered, padded, clipped, or variable-length extraction changes context |
| Overlap policy | overlapping windows can leak between splits |
Checks
- Is the coordinate system 0-based or 1-based?
- Is the end coordinate inclusive or exclusive?
- Is the genome assembly version fixed?
- Are overlapping windows split carefully to avoid leakage?
- Does strand orientation matter for the task?
- Are regions lifted over, filtered, or padded before modeling?
- Is the split unit region, gene, chromosome, sample, family, or time?