Interpretation of the fitted ML models using genomic and environmental data

Cathy Westhues

2022-12-16

Model training and prediction with the maize toy dataset (Genomes to Fields data)

data(geno_G2F)
data(pheno_G2F)
data(map_G2F)
data(info_environments_G2F)
data(soil_G2F)

METdata_G2F_training <-
  create_METData(
    geno = geno_G2F,
    pheno = pheno_G2F[pheno_G2F$year%in%c(2014,2015,2016),],
    map = map_G2F,
    climate_variables = NULL,
    compute_climatic_ECs = TRUE,
    et0=T,
    info_environments =  info_environments_G2F[info_environments_G2F$year%in%c(2014,2015,2016),],
    soil_variables = soil_G2F[soil_G2F$year%in%c(2014,2015,2016),],
    path_to_save = "~/Data/PackageMLpredictions/learnmet_plus/benchmarking_g2f/results_g2f_forward_3"
  )

METdata_G2F_new <-
  create_METData(
    geno = geno_G2F,
    pheno = as.data.frame(pheno_G2F[pheno_G2F$year%in%2017,] %>% dplyr::select(-pltht,-yld_bu_ac,-earht)),
    map = map_G2F,
    climate_variables = NULL,
    compute_climatic_ECs = TRUE,
    et0=T,
    info_environments = info_environments_G2F[info_environments_G2F$year%in%2017,],
    soil_variables = soil_G2F[soil_G2F$year%in%2017,],
    path_to_save = "~/Data/PackageMLpredictions/learnmet_plus/benchmarking_g2f/results_g2f_forward_3",
    as_test_set = T
  )
met_pred <- predict_trait_MET(
  METData_training = METdata_G2F_training,
  METData_new = METdata_G2F_new,
  trait = 'yld_bu_ac',
  prediction_method = 'xgb_reg_1',
  use_selected_markers = F,
  lat_lon_included = F,
  year_included = F,
  save_model = T,
  num_pcs = 200,
  include_env_predictors = T,
  save_splits = T,
  seed = 100,
  save_processing = T,
  path_folder = '~/g2f/res_xgb/cv0'
)

fitted_split <- met_pred$list_results[[1]]

learnMET::variable_importance_split(object = fitted_split,
                                    path_plot  = '~/g2f/res_xgb/cv0',
                                    type = 'model_specific')

learnMET::variable_importance_split(object = fitted_split,
                                    path_plot  = '~/g2f/res_xgb/cv0',
                                    type = 'model_agnostic')

learnMET::ALE_plot_split(fitted_split, 
                         path_plot = '~/g2f/res_xgb/cv0', 
                         variable ='freq_P_sup10_2',
                         nb_bins = 6)