Genome Annotation

Genome annotation assigns labels or metadata to genomic regions. In machine learning, annotation can be a training label, auxiliary feature, evaluation target, or provenance field.

An annotated region can be represented as:

$$(r_i,y_i,m_i)$$

where $r_i$ is the region, $y_i$ is the annotation label, and $m_i$ records source metadata.

The annotation is meaningful only relative to a reference system:

$$a_i=(\text{assembly},\text{coordinate},\text{region},\text{label},\text{source},\text{version})$$

Changing the assembly, annotation source, or version can change the task even if the sequence window looks similar.

Annotation Sources

Source Type	Use	Risk
curated database	stable labels and metadata	version drift and curation bias
public reference annotation	genes, transcripts, regions, functional labels	assembly and coordinate mismatch
experimental assay	measured signal for regions or variants	assay/source-specific labels
computational prediction	weak label or auxiliary feature	teacher-model leakage
rule-based derived label	deterministic preprocessing target	hidden rule encodes target information

Label Roles

The same annotation can play different roles in a model pipeline.

Role	Example Question
target label	can the model predict the annotation from sequence?
input feature	can annotation context improve another task?
filter	which regions are included or excluded?
grouping key	which regions must stay together across splits?
provenance field	where did this label come from?

Do not mix these roles silently. If an annotation is used for filtering or grouping, it may affect the evaluation claim even if it is not a model input.

Coordinate Contract

Genome annotations need a coordinate contract:

$$r=(\text{assembly},\text{chromosome},s,e,\text{strand})$$

where $s$ and $e$ are start and end coordinates. A public note should state whether intervals are 0-based or 1-based when that matters.

Field	Why
assembly	maps sequence to coordinates
chromosome or contig	defines the coordinate namespace
start/end convention	prevents off-by-one label shifts
strand	affects sequence extraction and reverse complement
annotation version	prevents label drift across releases

Split and Leakage

Annotation tasks can leak through overlapping windows, near-duplicate regions, shared evidence sources, and annotation-derived filters.

$$\mathrm{overlap}(r_{\mathrm{train}},r_{\mathrm{test}})=0$$

is a minimum condition for region-level generalization, not a complete guarantee. For broader claims, group by region family, chromosome, source, species, or annotation evidence as needed.

Leakage Path	Check
overlapping sequence windows	split regions before window extraction
same gene or locus across splits	group by locus or gene when relevant
labels derived from same source	separate source or version if claiming source generalization
annotation used as filter	record filtering before interpreting the metric
reverse-complement duplicates	group both orientations

Checks

• Which annotation source and version are used?
• Is the label measured, curated, or predicted by another model?
• Are annotation versions stable across train and test?
• Does metadata leak the target label?
• Are ambiguous or conflicting annotations handled explicitly?
• What genome assembly and coordinate convention are used?
• Is the annotation a target, feature, filter, grouping key, or provenance field?
• Are overlapping windows, adjacent regions, and reverse complements split safely?
• Is the public claim limited to sequence/region/variant modeling rather than broad omics?

Related

• Annotation and labeling
• Metadata and provenance
• Genomic region
• Variant effect prediction
• Example unit
• Split unit
• Dataset
• Genome
• Leakage