Machine learning applied to higher order functional representations of omics data reveals biological pathways associated with Parkinson‘s Disease

Background: Despite the increasing prevalence of Parkinson’s Disease (PD) and research efforts to understand its underlying molecular pathogenesis, early diagnosis of PD remains a challenge. Machine learning analysis of blood-based omics data is a promising non-invasive approach to finding molecular fingerprints associated with PD that may enable an early and accurate diagnosis.

Description: We applied several machine learning classification methods to public omics data from PD case/control studies. We used aggregation statistics and Pathifier’s pathway deregulation scores to generate higher order functional representations of the data such as pathway-level features. The models’ performance and most relevant predictive features were compared with individual feature level predictors.
The resulting diagnostic models from individual features and Pathifier’s pathway deregulation scores achieve significant Area Under the Curve (AUC, a receiver operating characteristic curve) scores for both cross-validation and external testing. Furthermore, we identify plausible biological pathways associated with PD diagnosis.

Conclusions: We have successfully built machine learning models at pathway-level and single-feature level to study blood-based omics data for PD diagnosis. Plausible biological pathway associations were identified. Furthermore, we show that pathway deregulation scores can serve as robust and biologically interpretable predictors for PD.