gsbench

Genomic Selection Model Benchmarking CLI for Plant Breeding

gsbench cross-validates genomic selection models on your genotype/phenotype data and produces a comparison report with prediction accuracy, bias diagnostics, and plots from a single command.

Installation

pip install gsbench

With gradient-boosting models (XGBoost, LightGBM):

pip install gsbench[full]

From source, for development:

git clone https://github.com/josh45-source/gsbench.git
cd gsbench
pip install -e ".[dev]"

Quick Start

gsbench ships with a small simulated example dataset (100 samples x 500 markers, two traits) so you can try it immediately:

# Copy the example genotype/phenotype files to the current directory
gsbench example

# Benchmark models on the example data
gsbench run example_geno.csv example_pheno.csv --trait yield --models GBLUP,BRR,RF --folds 5

This writes gsbench_output/report.html with the full comparison report, gsbench_output/summary.csv, and diagnostic plots under gsbench_output/plots/.

CLI Reference

gsbench run

gsbench run GENO PHENO --trait TRAIT [OPTIONS]

Argument / Option	Default	Description
GENO	—	Path to the genotype file (CSV/TSV, HapMap, or numeric matrix; format auto-detected)
PHENO	—	Path to the phenotype file (CSV/TSV, first column = sample IDs)
--trait	—	Phenotype column to benchmark against (required)
--models	all	all, or a comma-separated list of abbreviations, e.g. GBLUP,BRR,RF
--folds	5	Number of cross-validation folds
--repeats	1	Number of times to repeat k-fold CV (uses RepeatedKFold when > 1)
--maf	0.05	Minimum minor allele frequency; markers below this are dropped
--max-missing	0.2	Maximum per-marker missingness fraction before a marker is dropped
--impute	mean	Missing-genotype imputation: mean or median
--scale	center	Genotype scaling: center, standardize, or none
--output	gsbench_output	Output directory for the report, summary CSV, and plots
--seed	42	Random seed for cross-validation splits
--format	auto	Genotype format override: auto, csv, tsv, hapmap, numeric

gsbench list-models

Prints a table of all registered models and whether their dependencies are installed.

gsbench example

gsbench example [--output DIR]

Copies the bundled example genotype/phenotype CSVs into DIR (defaults to the current directory) and prints the gsbench run command to benchmark them.

Models

Abbreviation	Model	Notes
GBLUP	Genomic BLUP	Kernel ridge regression on the genomic relationship matrix G = ZZ'/p
BRR	Bayesian Ridge Regression	sklearn.linear_model.BayesianRidge on marker dosages
BL	Bayesian LASSO	sklearn.linear_model.ARDRegression, a sparse approximation of BayesB/BayesC
RKHS	RKHS (Gaussian Kernel)	Kernel ridge regression with an RBF kernel; bandwidth chosen by internal CV
RF	Random Forest	sklearn.ensemble.RandomForestRegressor (500 trees)
XGB	XGBoost	Requires pip install gsbench[full]
LGBM	LightGBM	Requires pip install gsbench[full]

Every model implements the same two-method interface (fit / predict), so adding a new one is a matter of subclassing gsbench.models.base.GSModel.

Metrics

Each fold reports r (Pearson correlation), r2, rmse, mae, bias, slope (regression of observed on predicted should be ~1), spearman (rank correlation), and nrmse. Breeders care most about r (prediction accuracy) and spearman (does the model rank genotypes correctly for selection?).

Example Output

Model comparison (prediction accuracy per model, with fold-to-fold error bars):

Predicted vs. observed phenotypes per model:

The full HTML report also includes a boxplot of per-fold accuracy, a bias diagnostic, a runtime comparison, and per-model detail tables.

Companion Tools

gsbench is part of a small plant-breeding data pipeline:

• brapiR2 - pull data from BrAPI servers
• phenoQC - QC for phenotypic trial data
• vcf2dosage - VCF to dosage matrix conversion
• gsbench — benchmark genomic selection models

Pipeline: retrieve → clean → prepare genotypes → benchmark models

License

MIT