Machine Learning Analysis of Microbiome Data to Predict Cocaine Addiction.

Cocaine use disorder is a significant public health problem in the US and worldwide. The microbiome may influence behavioral response to cocaine via gut-brain interactions. Microbial and behavioral variables are measured on mice to test whether differences in the microbiome account for the inter-subject variation in cocaine use. However, complex data requires sophisticated data science methods to analyze effectively. This study aimed to take a novel machine learning approach to first reduce data dimension, then cluster mice according to their cocaine use patterns, and finally classify mice into the resultant clusters based on microbial features. Both linear and nonlinear dimensionality reduction methods were employed and compared to generate new coordinates of cocaine use behavior. Based on the best coordinates, K-means identified three clusters of mice: mice that became addicted quicker and administered more cocaine; mice that used fewer doses in response to different dosages; mice that required much more infusion sessions to acquire addiction eventually. An artificial neural network was used to differentiate these behavioral groups based on the abundance of 57 bacterial genera produced micro- and macro-averaged AUC of 0.73 and 0.67, respectively. Model interpretation of this network helped identify risk and protective factors. An increased abundance of Escherichia shigella and Enterococcus was associated with high-risk mice, whereas an increased abundance of Akkemensia and Erysipelotrichaceae incertae sedis was shown to be protective factors. [ABSTRACT FROM AUTHOR]