Deconvolving specific from non-specific effects in differential gene expression experiments

Understanding the mechanisms of action upon cellular perturbations is a fundamental endeavor in molecular and chemical biology. Differential expression analysis is a widely used approach for probing these mechanisms, yet it presents substantial interpretational challenges due to the presence of secondary effects and the complex impact of experimental treatments on gene expression. To address this, we introduce orthos, an approach that employs Deep Generative Networks to disentangle specific and non-specific effects of perturbations on gene expression. Trained on large collections of human and mouse gene expression contrasts compiled for this work, orthos isolates non-specific effects by learning the patterns of expression changes that manifest time and again in unrelated experiments. We demonstrate, in diverse experimental settings, that the specific component obtained from the decomposition is a more informative and robust experimental signature and a better proxy for the direct molecular effects of a treatment compared to the original contrast, thereby drastically enhancing the interpretability of differential expression results. In addition, orthos allows identification of experiments with similar specific effects, aiding in the mapping of new treatments to their mechanisms of action. In summary, orthos constitutes a novel strategy in the analysis and interpretation of gene expression data and offers a powerful platform for the study of genetic, physiological, and pharmacological treatments in basic and applied research.