Molecular Structure-Based Large-Scale Prediction of Chemical-Induced Gene Expression Changes

Area of Science:

• Toxicogenomics
• Computational Chemistry
• Bioinformatics

Background:

• Quantitative Structure-Activity Relationship (QSAR) models predict chemical-induced biological responses.
• Modeling chemical-induced genomewide gene expression changes presents challenges due to model complexity.

Purpose of the Study:

• To evaluate the variable nearest neighbor (v-NN) method for modeling chemical-induced genomewide gene expression changes.
• To assess the utility of predicted gene expression signatures in identifying organ injuries.

Main Methods:

• Utilized a dataset of 13,150 compounds from the NIH Library of Integrated Network-based Cellular Signatures program.
• Applied the v-NN method to predict gene expression signatures.
• Compared predicted and experimentally derived signatures for predicting liver, kidney, and heart injuries.

Main Results:

• Achieved 62% prediction accuracy for compound signatures with a correlation coefficient of 0.61 in 10-fold cross-validation.
• Predicted and experimental signatures showed similar performance in identifying organ injuries (AUCs 0.70-0.77).
• Signatures predicted using multiple near neighbors outperformed experimental data, suggesting improved signal-to-noise ratio.

Conclusions:

• The v-NN method is a practical approach for large-scale, genomewide, chemical-induced gene expression modeling.
• Averaging information from near neighbors may enhance the accuracy of gene expression measurements.
• This method holds promise for toxicological assessments and drug safety evaluations.