Incorporating microbial community data with machine learning techniques to predict feed substrates in microbial fuel cells

Abstract

The complicated interactions that occur in mixed-species biotechnologies, including biosensors, hinder chemical detection specificity. This lack of specificity limits applications in which biosensors may be deployed, such as those where an unknown feed substrate must be determined. The application of genomic data and well-developed data mining technologies can overcome these limitations and advance engineering development. In the present study, 69 samples with three different substrate types (acetate, carbohydrates and wastewater) collected from various laboratory environments were evaluated to determine the ability to identify feed substrates from the resultant microbial communities. Six machine learning algorithms with four different input variables were trained and evaluated on their ability to predict feed substrate from genomic datasets. The highest accuracies of 93 ± 6% and 92 ± 5% were obtained using NNET trained on datasets classified at the phylum and family taxonomic level, respectively. These accuracies corresponded to kappa values of 0.87 ± 0.10, 0.86 ± 0.09, respectively. Four out of six of the algorithms used maintained accuracies above 80% and kappa values higher than 0.66. Different sequencing method (Roche 454 or Illumina sequencing) did not affect the accuracies of all algorithms, except SVM at the phylum level. All algorithms trained on NMDS-compressed datasets obtained accuracies over 80%, while models trained on PCoA-compressed datasets presented a 10–30% reduction in accuracy. These results suggest that incorporating microbial community data with machine learning algorithms can be used for the prediction of feed substrate and for the potential improvement of MFC-based biosensor signal specificity, providing a new use of machine learning techniques that has substantial practical applications in biotechnological fields.