Skip to contents

assigner

assigner helps answer a practical question: how well can genomic data assign individuals to candidate source populations?

Use assigner to:

  • assign individuals with a native population-genetic likelihood method written entirely in R, using called genotypes (GT) or genotype likelihoods (GL and PL);
  • retain the likelihood of every individual in every candidate population, not only the best and second-best assignments;
  • compare likelihood assignment with elastic net, XGBoost, experimental TabPFN, adegenet, or Eric C. Anderson’s gsi_sim workflows;
  • evaluate assignment honestly with holdout samples, leave-one-out estimation, and repeated marker panels;
  • avoid the high-grading bias described by Anderson (2010) by keeping marker selection inside the training data and holdout individuals untouched;
  • compare random marker panels with panels ranked by population differentiation;
  • calculate and visualize Dlr as a measure of assignment resolution between population pairs;
  • quantify effective reference sample size for GL/PL data and identify potentially important information imbalances among source populations;
  • calculate an optional coverage-aware diagnostic for individuals that may originate from an unsampled source;
  • estimate global and pairwise FST with confidence intervals;
  • run hierarchical AMOVA on incomplete genomic data with auditable locus-wise sample support, selectable molecular distances, Euclidean diagnostics, hierarchy-aware permutations, marker confidence intervals, and population sensitivity analysis; and
  • keep analyses reproducible with recorded arguments, random seeds, intermediate results, and dated output folders.

Ordinary leave-one-out assignment does not remove the high-grading bias created when the same individuals were already used to select apparently informative loci. See Anderson (2010) and the get-started vignette for the Training-Holdout-Leave-one-out design.

assigner accepts genomic data prepared with genometranslator and expects genomic quality control and filtering to have been completed first, for example with radr.

The package distinguishes assignment to predefined reference populations from clustering, mixture estimation, and statistical migrant detection. The get-started vignette explains these distinctions, leave-one-out assignment, unsampled alleles, missing data, validation, and the foundational references behind the methods.

AMOVA for incomplete genomic data

amova_genomic() implements a hierarchical analysis of molecular variance for long genomic datasets. Its default missing-data strategy works locus by locus: individuals without a genotype at the current locus are omitted from that locus only, and the represented populations, sample sizes, degrees of freedom, and unequal-sample-size coefficients are recalculated from the observations that actually contribute.

This follows the genomic, locus-wise philosophy used by Stacks while keeping the analysis accessible from R. It complements rather than replaces established AMOVA implementations in ade4, pegas, poppr, and GenoDive.

amova_result <- assigner::amova_genomic(
  data = "genomic_data.gds",
  hierarchy = c("REGION", "POP_ID"),
  distance = "euclidean",
  missing = "locuswise",
  min.groups = 2,
  min.individuals = 2,
  standardized = FALSE
)

amova_result$global
amova_result$per_locus
amova_result$marker_audit

Available missing-data strategies are "locuswise", "filter", and "complete". The function does not impute genotypes. Identity, nucleotide, Euclidean, Manhattan, and custom squared distances are supported. A Meirmans-style standardized statistic is currently limited to identity distance, where the maximum distance has a defined meaning; distances are never standardized by their observed sample maximum.

The data argument follows the rest of assigner: use a GDS file path or an open GDS object supported by genometranslator::read_genome(). Standard diploid dosage is selected automatically. Use strata when hierarchy columns are kept in a separate strata file, or set value explicitly for haplotype and custom distance analyses.

The AMOVA for incomplete genomic data vignette explains the variance-component theory, distance geometry, standardized differentiation, missing-data assumptions, existing implementations, validation strategy, and recommended literature.

Installation 

install.packages("remotes")
remotes::install_github("thierrygosselin/assigner")

gsi_sim, developed by assigner co-author Eric C. Anderson, is optional and is not installed by the R package. Complete installation instructions are in the get-started vignette. After installing it, verify the setup with:

assigner::check_gsi_sim()

Start here

Record the package version with packageVersion("assigner") and obtain the recommended citation with citation("assigner").