Abstract Accurate identification of pain-related genes remains challenging due to the complex nature of pain pathophysiology and the subjective nature of pain reporting in humans, or inferring pain states in animals on the basis of behaviour. Here, we use a machine learning approach to identify possible “pain genes”. Labelling was based on a gold-standard list of genes with validated involvement across pain conditions, and was trained on a selection of -omics (eg. transcriptomics, proteomics, etc.), protein-protein interaction (PPI) network features, and biological function readouts for each gene. Multiple classifiers were trained, and the top-performing model was selected to predict a “pain score” per gene. The top ranked genes were then validated against pain-related human SNPs to validate against human genetics studies. Functional analysis revealed JAK2/STAT3 signal, ErbB, and Rap1 signalling pathways as promising targets for further exploration, while network topological features contribute significantly to the identification of “pain” genes. As such, a PPI network based on top-ranked genes was constructed to reveal previously uncharacterised pain-related genes including CHRFAM7A and UNC79. These analyses can be further explored using the linked open-source database at https://livedataoxford.shinyapps.io/drg-directory/ , which is accompanied by a freely accessible code template and user guide for wider adoption across disciplines. Together, the novel insights into pain pathogenesis can indicate promising directions for future experimental research.

Predicting pain genes: multi-modal data integration using probabilistic classifiers and interaction networks