Machine learning based prioritisation of genes associated with osteoarthritis joint damage in animals.

Abstract

OBJECTIVES: To systematically prioritise unstudied genes for their potential to modulate joint damage phenotypes in animal models of osteoarthritis (OA). METHODS: We curated 567 known protein-coding OA-associated genes from the literature and integrated them with 1139 uniformly processed skeletal transcriptomic response datasets and a human protein-protein interaction network to prioritise unstudied candidate OA genes. Three machine learning models, XGBoost, SVM, and Random Forest, were trained using repeated 5-fold cross-validation and evaluated on an independent test set of the latest literature-reported associations. Model interpretability was assessed using SHAP values analysis. RESULTS: The XGBoost model achieved the highest performance on the held-out test dataset, with an AUROC of 0.84 (95% CI: 0.81-0.87). As expected, key predictive features included transcriptomic responses from both animal models and human OA tissues, as well as protein-protein interaction network features. Notably, SHAP value analysis highlighted specific biological pathways including oxidative phosphorylation and complement factor pathways as influential in the model. The top-prioritised genes included regulators of the NF-κB pathway, such as IER3, SOCS3, and NFKBIA. Incorporation of OpenTargets druggability data highlighted putative clinically tractable genes for further investigation, including MMP2 and MAP3K8. CONCLUSIONS: Our trained machine learning model effectively prioritised newly reported OA-associated genes, demonstrating its potential as a systematic gene prioritisation tool. An accompanying SkeletalVis R package enables researchers to explore over 1000 transcriptomic responses and the trained model predictions for their own studies.