Background
Flow cytometry, with its high throughput nature, combined with the ability to measure an increasing number of cell parameters at once can surpass the throughput of prevalent genomic and metagenomic approaches in the study of microbiomes. Novel computational approaches to analyze flow cytometry data will result in greater insights and actionability as compared to traditional tools used in the analysis of microbiomes. This paper is a demonstration of the fruitfulness of machine learning in analyzing microbial flow cytometry data generated in anaerobic microbiome perturbation experiments.
Conclusion
Machine learning can benefit the microbial flow cytometry research community by providing rapid screening and characterization tools to discover patterns in the dynamic response of microbiomes to several stimuli.
Results
Autoencoders were found to be powerful in detecting anomalies in flow cytometry data from nanoparticles and carbon sources perturbed anaerobic microbiomes but was marginal in predicting perturbations due to antibiotics. A comparison between different algorithms based on predictive capabilities suggested that gradient boosting (GB) and deep learning, i.e. feed forward artificial neural network with three hidden layers (DL) were marginally better under tested conditions at predicting overall community structure while distributed random forests (DRF) worked better for predicting the most important putative microbial group(s) in the anaerobic digesters viz. methanogens, and it can be optimized with better parameter tuning. Predictive classification patterns with DL (feed forward artificial neural network with three hidden layers) were found to be comparable to previously demonstrated multivariate analysis. The potential applications of this approach have been demonstrated for monitoring the syntrophic resilience of the anaerobic microbiomes perturbed by synthetic nanoparticles as well as antibiotics.