binchenlab/InsilicoCell

Two different scenarios were considered for prediction performance evaluation: sample-level holdout validation and entity-level holdout validation.

Entity-level holdout validation:

This scenario was used for evaluating model prediction performance on input samples containing previously unseen entities, such as unseen cell lines and unseen compounds, which did not appear in the training set and were viewed by the model as new cell lines and new compounds in the test set to predict on. To construct the entity-level holdout test sets, we first pooled data from all seven tasks and compiled a list of all unique drugs, genes, proteins and cell lines. We then randomly selected 5% of entities from each category. All input sample containing the selected entity were excluded from model training and reserved exclusively for testing. These set-aside samples solely served as test sets for model prediction and performance evaluation. Depending on the specific type of set-aside entities, the entity-level holdout can be specified throughout this paper as drug-level holdout, gene-level holdout, protein-level holdout or cell-level holdout.
Data files include:
"drug-induced_gene_expression_change_entity-level_holdout_test_set.csv"
"drug-protein_binding_entity-level_holdout_test_set.csv"
"TF-gene_association_entity-level_holdout_test_set.csv"
"drug_sensitivity_entity-level_holdout_test_set.csv"
"gene_effect_score_entity-level_holdout_test_set.csv"
"gene_mutation_entity-level_holdout_test_set.csv"
"CNV_entity-level_holdout_test_set.csv"

Sample-level holdout validation:

After excluding input samples for entity-level holdout validation as aforementioned, we also created another scenario to evaluate model performance based on sample-level holdout validation using the rest of the data. This was achieved by randomly selecting 10% of all input samples from the left data, serving as a test set that did not participate in model training. In this scenario, all the input samples in the test set were never seen during training, though the entities in the test set could have appeared in the training set. Accordingly, model performance is generally better under sample-level validation than under the more stringent entity-level validation setting.
Data files include:
"drug-induced_gene_expression_change_sample-level_holdout_test_set.csv"
"drug-protein_binding_sample-level_holdout_test_set.csv"
"TF-gene_association_sample-level_holdout_test_set.csv"
"drug_sensitivity_sample-level_holdout_test_set.csv"
"gene_effect_score_sample-level_holdout_test_set.csv"
"gene_mutation_sample-level_holdout_test_set.csv"
"CNV_sample-level_holdout_test_set.csv"

After setting aside input samples for all the aforementioned test sets based on sample-level holdout validation and entity-level holdout validation, the rest of the data served as the training set for InsilicoCell. Check our github repo on how to run InsilicoCell for prediction. We will release the training set after the acceptance of the paper.