Making new predictions

Once the models are trained and the best parameter combinations have been identified, they can be used to make predictions about the function of new sequence variants. Once you have in silico mutated structures of theoretical variants, simply compute a DynaSigDF object from them, as was show in computing_signatures.

You can then use your saved DynaSigML_Model objects to make predictions about the new variants:

from dynasigml.dynasig_df import load_pickled_dynasig_df
from dynasigml.dynasig_ml_model import load_dynasigml_model_from_file

dsdf_new_variants = load_pickled_dynasig_df("dsdf_new_variants.pickle")
dsml_model = load_dynasigml_model_from_file("dsml_inverted.pickle")
all_file_ids = dsdf_new_variants.get_file_ids()

# get the data with the beta scaling factor used by the best LASSO model
data_lasso = dsdf_double_mutants.get_data_array(all_file_ids, beta=dsml_model.get_best_beta_lasso())
data_lasso = data_lasso[:, 2:] # reject first two columns (beta value and dummy experimental measure)
predictions_lasso = dsml_model.predict_lasso(data_lasso)

# same as above, but uses the best user-supplied ML model instead of LASSO regression
data_ml = dsdf_double_mutants.get_data_array(all_file_ids, beta=dsml_model.get_best_beta_ml())
data_ml = data_ml[:, 2:]
predictions_ml = dsml_model.predict_ml(data_ml)

lasso_zscores = (predictions_lasso - np.mean(predictions_lasso)) / np.std(predictions_lasso)
ml_zscores = (predictions_ml - np.mean(predictions_ml)) / np.std(predictions_ml)
with open("predictions_new_variants.df", "w") as f:
    f.write("file_id pred_lasso pred_ml pred_lasso_zscore pred_ml_zscore sum_zscores\n")
    for file_id, pred_lasso, pred_ml, lasso_z, ml_z in zip(all_file_ids, predictions_lasso, predictions_ml,
                                                             lasso_zscores, mlp_zscores):
        f.write("{} {} {} {} {} {}\n".format(file_id, pred_lasso, pred_ml, lasso_z, ml_z, lasso_z + ml_z))

The predicted fitness values are now saved in the predictions_double_mutants.df file, along with the Z-scores for the LASSO and ML predictions, and the sum of both Z-scores. The highest Z-score sum is indicative of both models agreeing on this variant leading to high maturation efficiency.