Transcriptomic data and biomedical literature synergize in finding pharmacologic gene regulators

Most disorders caused by a deficiency or excess of one gene product lack targeted therapies. Since these disorders can be modeled with a gene overexpression, knockout, or knockdown, drugs that oppose the transcriptomic effects of such perturbations may be promising therapeutic candidates. RNA-Sequencing (RNA-Seq) studies can fuel this drug-prioritization, but their labels, written in plain language, must be annotated manually. Hence, we introduce Signature-based Networks from Automatically Curated Knockout, Knockdown, and Small-molecule Studies (SNACKKSS), which automatically curates gene-disruption and drug-treatment studies from the Gene Expression Omnibus and, in partnership with uniformly computed read count datasets, feeds the labels and RNA-Seq data directly into regulatory relationship predictions. Through cross-validation, we show that SNACKKSS' predictions (specifically, from a variant called "SA4") make a unique contribution to finding protein-inhibiting compounds, even alongside existing predictors. We demonstrate the benefit of aggregating multiple predictive tools, and provide this powerful ensemble alongside SNACKKSS. Importantly, we advise researchers to test complex machine learning models on multiple devices, since the hardware architecture can affect their output. Nonetheless, our findings support the value of integrating automatically curated RNA-Seq signatures with literature- and co-expression-based predictors for drug-repurposing prioritization.