Predicting characterization of microbiome taxonomy from imaging using machine learning approaches

For this study, a total of 47 mock human skin microbiome communities were created using microorganisms collected from human donors and grown in vitro for between eight and 32 days. Each mock community sample was split. Ten mL of each sample was used to determine the taxonomy of the community, using metatranscriptomics and Kraken2 to provide population-level taxonomic information; five mL of each sample was used for imaging. The resulting micrographs served as the basis for establishing a new analysis pipeline that sequentially used two different methods for machine learning and one statistical technique: (1) confocal microscopy images were segmented into individual cells using the generalist, deep learning, publicly available machine learning model Cellpose; (2) continuous probability density functions describing the joint distribution of the cell area and eccentricity were found using algorithms expressing the statistical technique of kernel density estimation; (3) these probability density functions were used as input for convolutional neural networks, that were trained to predict both the taxonomic diversity and the most common bacterial class, independently of metatranscriptomics. Specifically, models were made to predict the Shannon index (a quantitative measure of taxonomic diversity) and to predict the most common bacterial class, for each micrograph. Measured Shannon indices (based on metatranscriptomics) ranged from nearly 0 to 1.4. The model predictions of Shannon indices had a mean squared error of 0.0321 +/- 0.0035. The model predictions of the most common taxonomic class of bacteria had an accuracy of 94.0% +/- 0.7%.