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]

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

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