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

Area of Science:

• Biotechnology
• Microbiology
• Bioinformatics

Background:

• Mixed-species biotechnologies, like biosensors, face challenges in chemical detection specificity due to complex microbial interactions.
• This lack of specificity limits their use in applications such as determining unknown feed substrates.

Purpose of the Study:

• To evaluate the ability of machine learning algorithms to identify feed substrates from microbial genomic data.
• To assess the impact of different taxonomic levels and data compression methods on prediction accuracy.

Main Methods:

• Six machine learning algorithms were trained and evaluated using genomic datasets from 69 samples with three substrate types (acetate, carbohydrates, wastewater).
• Datasets were classified at phylum and family taxonomic levels, and data compression techniques (NMDS, PCoA) were applied.
• Performance was measured using accuracy and kappa values, considering different sequencing methods (Roche 454, Illumina).

Main Results:

• The Neural Network (NNET) algorithm achieved the highest accuracies (93% ± 6% at phylum level, 92% ± 5% at family level).
• Four out of six algorithms maintained accuracies above 80% and kappa values higher than 0.66.
• NMDS-compressed datasets yielded accuracies over 80%, while PCoA-compressed datasets showed a 10-30% reduction.

Conclusions:

• Microbial community genomic data, combined with machine learning, can effectively predict feed substrates.
• This approach offers potential improvements in microbial fuel cell (MFC)-based biosensor signal specificity.
• Machine learning techniques provide practical applications for advancing biotechnological fields.