Variant Effect Prediction

Variant effect prediction estimates how a sequence variant changes a biological or functional target. In this wiki, it is treated as a genome-level sequence modeling task, not as a clinical interpretation workflow.

A variant can be represented as:

$$v=(r,b_{\mathrm{ref}},b_{\mathrm{alt}})$$

where $r$ is the genomic position or region, $b_{\mathrm{ref}}$ is the reference allele, and $b_{\mathrm{alt}}$ is the alternate allele.

A model often compares reference and alternate sequence contexts:

$${\Delta }_{\theta }(v)=f_{\theta }(x_{\mathrm{alt}})-f_{\theta }(x_{\mathrm{ref}})$$

The prediction target must say what “effect” means:

$$\text{effect}=\text{change in a specified observable under a specified context}$$

Without that context, a variant effect score is only a generic ranking signal.

Task Contract

Field	Question
Variant unit	SNV, insertion/deletion, haplotype, edited region, or window
Reference context	assembly, strand, window length, flanking sequence
Alternate context	how the alternate sequence is constructed
Target observable	annotation, assay readout, expression proxy, constraint score, or benchmark label
Label source	measured, curated, simulated, predicted, or derived
Split unit	region, locus, chromosome, source, variant family, or species
Metric	classification, regression, ranking, calibration, or retrieval metric

Reference and Alternate Windows

For a fixed window operator $W$, a common construction is:

$$x_{\mathrm{ref}}=W(G,r),x_{\mathrm{alt}}=\mathrm{edit}(x_{\mathrm{ref}},v)$$

where $G$ is the reference genome and $r$ is the region around the variant. The edit operation should be deterministic and documented.

Choice	Risk
window length	too short loses regulatory or protein-context information
strand handling	reverse-complement mismatch
indel handling	coordinate shift and alignment ambiguity
multiple variants	single-variant assumption breaks
assembly version	ref/alt sequence mismatch

Label Types

Label Type	Meaning	Caution
functional assay readout	measured change in a defined experiment	assay/source-specific
computational annotation	label produced by another model or rule	teacher leakage and circular evidence
conservation or constraint proxy	evolutionary signal	not the same as experimental function
public benchmark label	task-specific evaluation label	benchmark scope may be narrow
clinical category	interpretation label	out of scope here unless explicitly public and non-sensitive

Evaluation Boundary

Variant effect prediction can be evaluated as classification, regression, ranking, or prioritization. The metric should match the output claim.

Claim	Evidence Needed
variant classification	label semantics, threshold, calibration or confusion matrix
effect-size regression	unit, transform, uncertainty, error scale
prioritization	ranked-list metric and candidate set definition
cross-region generalization	region/locus split and overlap checks
cross-source robustness	source-held-out or annotation-version-held-out evaluation

For a score difference ${\Delta }_{\theta }(v)$, calibration is not automatic. A larger score may rank variants without being a well-calibrated probability or measured effect size.

Leakage Risks

Risk	Check
nearby variants in train and test	group by region or locus
overlapping windows	split before window extraction
same source evidence in labels	record source and version
annotation-derived labels used as features	separate input features from targets
homologous or duplicated regions	group duplicated sequence contexts
post-hoc threshold tuning	tune threshold on validation only

Checks

• What target does “effect” mean?
• Are reference and alternate contexts generated from the same assembly?
• Are variants near each other split carefully to avoid regional leakage?
• Are labels measured, curated, simulated, or inferred?
• Is the model evaluated on held-out regions, variant types, or sources?
• Is the task classification, regression, ranking, or prioritization?
• Does the metric match the claim rather than only the model output format?
• Are threshold choices, candidate sets, and failed variants included in the protocol?
• Is the note staying within sequence/region/variant modeling rather than clinical interpretation?

Related

• Genomic region
• Genome annotation
• K-mer
• Genome
• Sequence
• Dataset
• Classification
• Regression
• Ranking
• Calibration
• Example unit
• Split unit
• Leakage
• OOD generalization