Tag SNP¶

A tag SNP is a single-nucleotide polymorphism selected from a declared reference population and variant set because its linkage-disequilibrium correlation predicts one or more untyped variants above a chosen threshold, preserving common-variant association coverage while reducing genotyping burden.

Version: v2 · 2026-08-30 · History
Domain-specific #: 2928
Origin domain: statistical genetics
Subdomain: linkage-disequilibrium-based marker selection
Aliases: TagSNP, Tagging SNP, Haplotype-tagging SNP

Core Idea¶

A tag SNP is a single-nucleotide polymorphism deliberately selected as a genotyped representative for other variants whose allelic states it predicts through linkage disequilibrium (LD). Dense common-variant data contain substantial local redundancy: nearby alleles are often correlated because historical chromosomes have not been broken apart by recombination at every generation. A study can therefore genotype a smaller tag set while retaining much of the power it would have had if every common target SNP were directly assayed.^[1]^[2]

“Tag” is a role, not an intrinsic molecular type. The same genomic SNP can be a good tag for a specified target set in one ancestry-matched reference panel and a poor tag in another population, at a stricter threshold, or against a denser variant catalog. Several correlated SNPs may be interchangeable tags; assay compatibility, minor-allele frequency, genotyping quality, or functional priorities may determine which is chosen. A tag SNP need not alter biology, lie in a gene, be the strongest association in a study, or uniquely identify a haplotype.

The defining abstraction is LD-calibrated representative selection: declare a population and target-variant universe, estimate statistical predictability, choose a small assayable subset subject to a coverage criterion, and use the typed subset as proxies in association analysis or as inputs to richer genotype inference. Early haplotype-tagging work emphasized distinguishing common haplotypes; later methods selected pairwise or multi-marker proxies according to statistical power and prediction accuracy.^[1]^[2]^[3] A formal haplotype-block partition and experimentally phased chromosomes can help, but neither is required for the general tag-SNP identity.

Structural Signature¶

The recurring structure is:

reference population + observed genotype or haplotype panel + eligible target SNP universe + candidate assay set + LD/prediction measure + coverage threshold and cost criterion → selected genotyped SNP subset whose states preserve declared information about untyped common variants

Eight roles are load-bearing:

Reference population or sample. LD is estimated in chromosomes intended to represent the study population. Population history, ancestry, admixture, and sample size condition the estimate.
Target variant universe. The study declares which known variants it wants to cover, usually filtered by region, minor-allele frequency, quality, or reference-panel membership.
Candidate tag universe. Only SNPs that can actually be assayed, pass quality constraints, and satisfy design restrictions are eligible representatives.
Predictive relation. Pairwise (r^2), multi-marker haplotype prediction, or another explicitly validated criterion quantifies how well typed information recovers an untyped target.
Coverage threshold. A rule such as \(r^2\geq0.8\) states when a target counts as represented. The number is a design choice, not a universal biological boundary.
Selection objective. The procedure seeks few or low-cost tags while meeting coverage, power, assay, and sometimes forced-inclusion constraints.
Selected representative SNP. A locus becomes a tag only through this selection relation. It may tag itself and one or many other variants.
Downstream use. Genotypes at tags support single-marker tests, haplotype tests, genotype imputation, array coverage, or follow-up localization, with uncertainty inherited from imperfect LD.

For biallelic loci (A) and (B), define

\[ D=p_{AB}-p_Ap_B, \qquad r^2=\frac{D^2}{p_A(1-p_A)p_B(1-p_B)}. \]

Here (r^2) is the squared correlation between allele indicators. In the simplest pairwise design, let (U) be target SNPs, (C) candidate tags, and \(\tau\) the coverage threshold. Select \(S\subseteq C\) to minimize assay cost subject to

\[ \forall u\in U,\quad u\in S\;\;\text{or}\;\; \max_{s\in S} r^2_{su}\geq\tau. \]

This is a coverage or set-selection problem. Multi-marker tagging replaces the single (s) with a combination of typed loci whose haplotype predicts (u). The invariant is not a particular algorithm but reduced assay burden under declared prediction fidelity.

What It Is Not¶

A tag SNP is not simply any SNP marker. A marker may be genotyped for identification, ancestry, linkage, quality control, or direct association without being selected to cover correlated untyped variants. Tag status requires a target set and a prediction/coverage relation.

It is not necessarily a causal or functional variant. An associated tag can be correlated with the causal allele, with several plausible causal alleles, or with a haplotype carrying the effect. Association at the representative identifies a statistical neighborhood; functional experiments, conditional analyses, sequencing, and fine-mapping are needed to resolve causality.

It is not a haplotype or necessarily a unique label for one. A haplotype is a combination of alleles on one chromosome. One tag may discriminate certain common haplotypes, while multiple tags may be required to predict a target. Nor must tagging be preceded by a formal haplotype-block partition: pairwise LD methods can operate directly on genotype correlation, and LD does not obey a single universal block definition.^[2]^[3]

It is not (D'), (r^2), or LD itself. Those are relations or measures between loci; the tag is a selected locus. High (D') can occur when alleles are rare and still provide weak predictive correlation. For association power and genotype prediction, (r^2) is commonly the more direct pairwise tagging criterion.^[2]^[4]

It is not an imputed SNP. A tag is directly typed in the study by design. Imputation combines observed markers with a reference haplotype panel and a probabilistic model to estimate untyped genotypes and their uncertainty. Tags provide input information, but choosing tags and performing imputation are distinct operations.^[5]

It is not automatically the lead, index, or sentinel SNP reported for a GWAS locus. Those terms usually identify the strongest or representative association signal after analysis. Such a SNP may also be an LD proxy, but post hoc signal naming is not the pre-study coverage-selection role that defines a tag SNP.

Scope of Application¶

Tag SNPs arose in human statistical genetics when exhaustive genotyping or sequencing of large cohorts was impractical. Candidate-gene studies used haplotype-tagging or LD-selected marker sets; the International HapMap Project then enabled genome-wide array design by describing common variation and correlations in reference populations.^[6] Commercial fixed arrays likewise chose subsets intended to capture much larger common-variant catalogs directly or through LD.^[7]

The abstraction applies to association-study design, genotyping-panel design, cross-platform coverage evaluation, population-specific customization, replication-marker selection, and genotype-imputation input design. It also transfers within genomics to nonhuman species whenever a reference population, LD structure, target variant universe, assay technology, and prediction criterion are declared.

Whole-genome sequencing changes the economic frontier but does not erase the abstraction. Large cohorts still use arrays followed by imputation; low-density panels can support genomic prediction or imputation in plants and livestock; association reports still interpret an assayed or imputed SNP as a possible tag for nearby causal variation. The role is weakest for novel, private, or rare variants that have no sufficiently correlated common marker, and for genomic regions poorly represented or technically inaccessible in the reference panel.^[7]

Clarity¶

To decide whether a SNP is genuinely a tag, ask five questions. What target variants is it meant to represent? In what reference and study population was LD estimated? What prediction measure and threshold were used? Was the SNP directly assayed and chosen before or as part of panel design? What fraction of the declared targets does it cover alone or with other markers?

A statement such as “rsX is a tag SNP” is incomplete without at least an implicit context. A precise claim is: “In reference population (P), typed SNP (s) tags target SNP (u) at pairwise (r^2=0.92), exceeding the study threshold \(\tau=0.8\).” If the target is instead associated only by proximity, if (D') is high but (r^2) is low, or if the claim comes solely from being the strongest GWAS hit, the tagging identity has not been established.

The causal diagnostic is separate. If editing the tag allele would not alter the phenotype but editing a correlated variant would, the tag remains a valid statistical tag and an invalid causal explanation. That is not a contradiction; representation and mechanism answer different questions.

Manages Complexity¶

Tagging converts a dense marker-assay problem into a coverage problem. Instead of treating millions of common variants as independent genotyping obligations, it groups targets by empirical predictability and selects representatives. Carlson and colleagues made the design criterion explicit: every known common polymorphism should either be directly assayed or exceed a specified (r^2) with a selected tag.^[2]

This compression reduces assay cost, DNA consumption, computational burden, and correlated testing redundancy while retaining much of the power for common-variant association. It also provides an auditable failure surface: targets below threshold, populations unlike the reference, recombination hotspots, rare variants, and unavailable assay probes can be listed rather than hidden inside a vague claim of genome coverage.

The abstraction does not eliminate multiple testing, population stratification, genotyping error, phenotype misclassification, or the need for replication. It manages the marker-coverage dimension of study design. De Bakker and colleagues showed that tag-selection and analysis strategies trade genotyping investment, test burden, and power; multi-marker tests may improve coverage or rare-haplotype sensitivity but can change the multiple-testing and common-variant power balance.^[8]

Abstract Reasoning¶

The pairwise model supports several useful deductions. If a causal allele is tested indirectly through a marker correlated with it at (r^2), the simplest asymptotic association models treat the marker as carrying roughly an (r^2) fraction of the information available from direct genotyping. A design may therefore need on the order of (1/r^2) times as many samples to recover comparable power, all else equal. This is why (r^2), rather than physical distance alone or high (D'), governs many tag-selection rules.^[2]^[8]

The optimization also exposes nontransitivity. If \(r^2(A,B)\geq0.8\) and \(r^2(B,C)\geq0.8\), one cannot infer \(r^2(A,C)\geq0.8\). Coverage must be checked target by target. A greedy algorithm that selects the candidate covering the most uncovered targets can work well, but it need not find a unique or globally optimal tag set under every constraint.

Lowering \(\tau\) produces fewer tags and cheaper assays but more information loss. Raising it increases marker count and expected fidelity. Adding forced tags for functional candidates can improve interpretability while reducing pure cost optimality. Allowing multi-marker predictors can cover targets no single tag reaches, but prediction becomes dependent on phasing or genotype-pattern estimation and a more complex analysis model.^[3]^[8]

Population transfer is an empirical question, not an automatic failure. Some tag sets transfer well between related populations, but recombination history, drift, bottlenecks, admixture, allele-frequency changes, and reference-panel sampling can alter (r^2). Coverage should be recalculated in an ancestry-relevant panel rather than inferred from continental labels alone.^[4]^[9]

Knowledge Transfer¶

Within genetics, the abstraction transfers exactly from candidate-gene panels to genome-wide arrays, from human studies to agricultural genomics, and from direct proxy testing to the observed backbone used by genotype imputation. Each application preserves the same roles: a reference panel, candidate assays, targets, predictability measure, threshold, selected representatives, and a downstream inferential task.

The reasoning also transfers from design to interpretation. When an association is reported at a tag, the relevant unit is the correlated set, not the printed rs identifier alone. When a panel is moved to another population, recalibrate coverage. When sequencing reveals additional targets, recompute which are tagged. When an assay fails, determine whether another member of the same LD bin can substitute without crossing the fidelity threshold.

Outside genetics, “choose correlated representatives to cover redundant variables” instantiates Compression and resembles feature selection. Calling a survey item, sensor, or financial instrument a “tag SNP,” however, is metaphorical because SNP identity, haplotypes, recombination, allele frequency, and population LD are absent. The portable reasoning belongs to the parent abstractions, not to the domain-specific term.

Examples¶

Pairwise toy selection. Suppose target SNPs are (A,B,C,D), and the threshold is \(r^2\geq0.8\). Candidate (B) correlates with (A) at 0.92 and (C) at 0.87; (D) has no partner above 0.8. Selecting (B) and (D) directly covers all four targets: (B,D) by assay and (A,C) through (B). Selecting only (A,D) fails if (r^2(A,C)=0.78), even though both (A) and (C) correlate strongly with (B). The example illustrates thresholded coverage and nontransitivity.

Candidate-gene design. Carlson and colleagues selected tagSNPs across 100 candidate genes so that common variants were directly assayed or exceeded a chosen pairwise (r^2). At the relatively stringent (r^2>0.8) threshold, their selected tags resolved more than 80 percent of observed haplotypes across the studied genes, and they concluded that tags should be selected separately for populations with different ancestry when LD differs.^[2] The example shows an explicit target universe, threshold, compact marker set, and population qualification.

HapMap-enabled genome coverage. Phase II HapMap characterized more than 3.1 million SNPs in four population samples. It estimated that untyped common variation had average maximum (r^2) of roughly 0.90–0.96 depending on population, while contemporary commercial arrays captured common HapMap SNPs less uniformly, especially in the African sample; some common variants remained untaggable, often around recombination hotspots.^[7] A “genome-wide” array was therefore a calibrated proxy panel, not an exhaustive census.

Population-transfer check. A panel chosen in one HapMap sample is proposed for a new study. De Bakker and colleagues compared dense data across HapMap and eleven other populations and found little simulated power loss in many transfers, while still framing transferability as a question of effective coverage.^[9] The correct lesson is to measure portability, not to assume either universal equivalence or universal failure.

Association without causation. A typed tag (T) is associated with disease, and two untyped variants (U_1,U_2) both have high (r^2) with (T). The result establishes a locus-level signal but cannot identify which variant is functional. Denser imputation, conditional tests, fine-mapping, sequencing, and functional assays may discriminate them. Reporting (T) as “the disease mutation” would confuse proxy evidence with mechanism.

Imputation boundary. A genotyping array directly measures selected markers. An imputation program compares the observed pattern with reference haplotypes and assigns genotype probabilities to additional variants. The typed markers play tagging roles; the estimated variants are imputed targets. Imputation adds a model and uncertainty rather than converting every predicted variant into a directly observed tag.^[5]

Structural Tensions¶

Efficiency versus coverage. Fewer tags lower cost and correlated-test burden; stricter coverage requires more assays. A threshold makes the tradeoff visible but cannot choose it for the study.

Common-variant economy versus rare-variant blindness. LD tagging works best for variants represented and sufficiently correlated in the reference panel. Rare, recent, private, or technically difficult variants are more likely to escape, so excellent common-variant coverage is not genome completeness.

Population specificity versus portability. Tag identity is population-conditioned, yet many tags transfer acceptably among related samples. Treating tags as universal is unsafe; treating every ancestry difference as total nontransferability wastes empirical information.

Pairwise simplicity versus multi-marker efficiency. Pairwise (r^2) produces transparent bins and tests. Multi-marker combinations can improve coverage, but add phase/prediction dependence, analysis choices, and a larger hypothesis space.^[8]

Prediction versus causal interpretation. Strong LD is exactly what makes a tag efficient and what makes the associated causal variant ambiguous. The design succeeds by collapsing distinctions that fine-mapping later must reopen.

Frozen panel versus changing variant catalog. A marker set can meet its threshold against one reference release and lose coverage when new variants, populations, or sequencing data are added. Tag status must be versioned with its target universe and reference.

Structural–Framed Character¶

Tag SNP is strongly structural and strongly domain-framed. Its structural form is precise: correlated variables, a target universe, a fidelity threshold, a cost objective, and representative subset selection. Remove the LD-calibrated proxy relation or the reduced-assay purpose and the identity disappears.

The framing is irreducibly genetic. SNP alleles, chromosomes, haplotypes, recombination, population ancestry, minor-allele frequency, genotyping, and association power determine what the roles mean. A correlated representative variable in another field may share the structure but is not literally a tag SNP.

Structural Core vs. Domain Accent¶

The portable core is remove redundant measurements by retaining representatives that predict the omitted variables above a declared fidelity threshold. Compression captures the economy; Correlation supplies the empirical dependence; Proxy–Target Fidelity asks how much information survives; feature selection describes retaining original variables rather than constructing latent combinations.

The domain accent fixes the variables as SNP genotypes, the dependence as population LD, the fidelity measure commonly as pairwise or multi-marker (r^2), and the downstream purpose as genetic association, imputation, or genomic prediction. It also creates domain-specific failure modes: recombination hotspots, ancestry mismatch, phasing uncertainty, rare variants, assay design, and causal ambiguity.

Tag SNP instantiates Compression, the minimal prospective DAG parent. The dense genotype representation contains redundancy; tag selection retains a shorter set of directly assayed coordinates subject to a reconstruction, coverage, or power criterion. This is lossy compression specialized to population-genetic variables. Compression does not determine SNPs, LD, allele-frequency thresholds, population matching, or association-study use, so it strictly subsumes rather than covers the child.

Correlation is constitutive but not a second minimal parent. Pairwise (r^2) measures the association that licenses proxying, yet correlation alone does not select representatives, set a coverage threshold, or minimize genotyping cost. Proxy–Target Fidelity is a related diagnostic for whether the typed locus faithfully tracks the untyped target in the intended population.

Dimensionality Reduction is related in the broad goal of fewer variables, but the live prime explicitly distinguishes constructed low-dimensional features from feature selection. Tagging retains original SNP coordinates. It is therefore not used as the parent even though informal genomics prose may call the result dimensionality reduction.

Sampling (Representativeness) concerns selecting statistical units to represent a population. Tag-SNP selection instead selects variables/loci to represent correlated variables, so it is not the correct genus.

Relationships to Other Abstractions¶

Current abstraction Tag SNP Domain-specific

Parents (1) — more general patterns this builds on

Tag SNP is a kind of Compression Prime

Tag SNP instantiates Compression, the minimal prospective DAG parent.

Hierarchy paths (3) — routes to 3 parentless roots

Tag SNP → Compression → Abstraction

Show alternative paths (2)

Neighborhood in Abstraction Space¶

Tag SNP sits in a sparse region of the domain-specific corpus (86^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (1565 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With¶

Single-nucleotide polymorphism: a variant class; only selected SNPs with a declared proxy role are tags.
Genetic marker / marker SNP: any assayed locus used to track inheritance or association; broader than tag SNP.
Haplotype: an allele combination on one chromosome, not a representative locus.
Haplotype block: a region defined under a block convention; useful context but not required by pairwise or multi-marker tagging.
Linkage disequilibrium: population-level nonrandom allelic association, the relation used for tagging rather than the selected marker.
(D'): a normalized LD measure sensitive to historical recombination; high (D') does not guarantee high predictive (r^2).
Causal or functional variant: a variant that changes a biological mechanism; a tag may merely correlate with it.
Lead, index, or sentinel SNP: a top reported association signal; post-analysis prominence is not tag-selection identity.
Imputed SNP: an untyped genotype estimated probabilistically from observed markers and a reference panel.
Fine-mapping variant: a member of a credible causal set or prioritized candidate after denser locus analysis.
Ancestry-informative marker: selected to distinguish ancestry components, not necessarily to cover neighboring SNPs.
Feature selection generally: the cross-domain method; Tag SNP is its LD- and genomics-specific realization.

References¶

[1] Johnson, G. C. L., Esposito, L., Barratt, B. J., et al. (2001). “Haplotype tagging for the identification of common disease genes.” Nature Genetics, 29, 233–237. https://doi.org/10.1038/ng1001-233 The foundational study develops haplotype-tagging markers to capture common variation without assaying every SNP. registry ↩a ↩b

[2] Carlson, C. S., Eberle, M. A., Rieder, M. J., Yi, Q., Kruglyak, L., & Nickerson, D. A. (2004). “Selecting a maximally informative set of single-nucleotide polymorphisms for association analyses using linkage disequilibrium.” American Journal of Human Genetics, 74(1), 106–120. https://doi.org/10.1086/381000 The paper formalizes (r^2)-threshold coverage, presents a tag-selection algorithm, connects (r^2) to association power, and evaluates population differences. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g

[3] Halperin, E., Kimmel, G., & Shamir, R. (2005). “Tag SNP selection in genotype data for maximizing SNP prediction accuracy.” Bioinformatics, 21(Suppl. 1), i195–i203. https://doi.org/10.1093/bioinformatics/bti1021 The paper frames tag selection as choosing a small informative subset and evaluates prediction directly from unphased genotype data. registry ↩a ↩b ↩c

[4] Pritchard, J. K., & Przeworski, M. (2001). “Linkage disequilibrium in humans: models and data.” American Journal of Human Genetics, 69(1), 1–14. https://doi.org/10.1086/321275 The review relates LD to recombination, gene conversion, demography, admixture, and association-mapping density. registry ↩a ↩b

[5] Marchini, J., & Howie, B. (2010). “Genotype imputation for genome-wide association studies.” Nature Reviews Genetics, 11(7), 499–511. https://doi.org/10.1038/nrg2796 The review defines model-based inference of untyped genotypes, performance determinants, uncertainty, fine-mapping, and cross-study harmonization. registry ↩a ↩b

[6] International HapMap Consortium. (2005). “A haplotype map of the human genome.” Nature, 437, 1299–1320. https://doi.org/10.1038/nature04226 The Phase I map documents recombination hotspots, block-like LD, common haplotypes, neighboring-SNP correlations, and tag-based association-study design. registry ↩

[7] International HapMap Consortium. (2007). “A second generation human haplotype map of over 3.1 million SNPs.” Nature, 449, 851–861. https://doi.org/10.1038/nature06258 The Phase II resource quantifies average maximum (r^2), array coverage differences, untaggable common variants, recombination-hotspot effects, and imputation gains. registry ↩a ↩b ↩c

[8] de Bakker, P. I. W., Yelensky, R., Pe'er, I., Gabriel, S. B., Daly, M. J., & Altshuler, D. (2005). “Efficiency and power in genetic association studies.” Nature Genetics, 37(11), 1217–1223. https://doi.org/10.1038/ng1669 The study compares pairwise and multi-marker tagging, genotyping investment, analysis choices, and statistical power using HapMap ENCODE data. registry ↩a ↩b ↩c ↩d

[9] de Bakker, P. I. W., Burtt, N. P., Graham, R. R., et al. (2006). “Transferability of tag SNPs in genetic association studies in multiple populations.” Nature Genetics, 38(11), 1298–1303. https://doi.org/10.1038/ng1899 The study measures effective coverage and simulated association power when tags selected in HapMap samples are deployed across other population samples. registry ↩a ↩b