Predicting molecular initiating events using chemical target annotations and gene expression.

Area of Science:

• Computational biology
• Toxicogenomics
• Bioinformatics

Background:

• High-throughput transcriptomic screening generates vast gene expression data for chemical treatments.
• Publicly available datasets with diverse chemical spaces and reference chemicals aid in predicting molecular initiating events.
• Integrating gene expression data with chemical-protein interactions enables prediction of mechanisms of action.

Purpose of the Study:

• To train binary classifiers for predicting mechanisms of action from transcriptomic responses.
• To identify optimal classification algorithms and parameters using MCF7 cell line data.
• To validate classifier performance using holdout data and training-excluded chemicals.

Main Methods:

• Linked LINCS L1000 gene expression data to RefChemDB chemical-protein associations.
• Trained six binary classification algorithms on MCF7 cell line data.
• Validated classifiers using holdout sets, training-excluded chemicals, and permutation testing.
• Compared classifier performance between MCF7 and PC3 cell lines.

Main Results:

• Identified and validated 9 high-performing classifiers for 51 molecular initiating events.
• Achieved high internal accuracies (0.73-0.94) and holdout accuracies (0.68-0.92).
• Demonstrated predictive accuracy extended to training-excluded chemicals.
• Observed variable classifier performance across different cellular contexts (MCF7 vs. PC3).

Conclusions:

• The developed methodology aids in prioritizing candidate perturbagens for targeted screening.
• The approach guides the selection of relevant cellular contexts based on cell line-specific model performance.