Introducing a Novel Simulation Tool for Interconnected Differential Expression Signatures and Its Application to Benchmarking

Pharmaceutical research has long used differential gene expression signatures to study external stimuli like pathogenic determinants or small molecule treatments. These signatures measure expression values for multiple tags and are often compared using the concept of connectivity. Despite the scientific community's efforts to produce unbiased datasets for evaluating connectivity-based methods for drug identification and repurposing, the limits of benchmarking data hinder their effectiveness.

To address this, we developed a simulation method to generate pairs of connected differential expression signatures, that is based on a three layers decomposition and relies on a statistical framework with different levels of parametrization. We benchmarked seven connectivity scores methods from the literature using our simulated signatures. We then evaluated the capacity of each method to retrieve the most connected signatures for a specific query, using the area under the precision-recall curves (AUPRC). Moreover, we introduced a novel application perspective by training a Siamese Neural Network with our simulated data to predict the connectivity score.

Overall, our method is a significant advance in pharmaceutical research, providing a reliable way to simulate connected differential expression signatures. It will help develop and evaluate algorithms for comparing signatures to find the most connected or reversed, leading to more effective drug repurposing. An open-source version of the package will be released at the end of 2023.