imaging-genetics-models

Create an NPZ bundle with genotype [subjects, variants] and phenotype arrays, optionally including variant_id and covariates. Then run train_reference.py with --model association and an output directory to get association_results.csv with effect estimates and P values.

Prepare an NPZ file containing a genotype matrix and a weights array of matching variant length, optionally with subject_id. Run the skill's training script with --model prs to produce predictions.csv containing a polygenic_score column per subject.

Use the lmm mode when your sample contains related individuals and you can supply a kinship matrix in the NPZ bundle. Kinship-aware linear mixed models correct for relatedness and population structure that standard association tests ignore.

No, PLINK2 remains an external executable with independent installation terms. The skill only builds audited PLINK2 command lines, which you execute separately through an approved shell workflow after installing PLINK2 yourself.

All modes consume NPZ bundles with named arrays: genotype and phenotype for association, plus kinship for LMM, weights for PRS, or X and Y matrices for PLS/CCA. Outputs include metrics.json, config.json, and run_manifest.json for provenance.

Never split related individuals across train and test folds, and fit dimensionality reduction and covariate residualization inside training data only. Record genome build, allele orientation, QC thresholds, ancestry definition, and phenotype construction with every analysis.