Assessment of methods for network analysis of single-time-point microbial samples

Harnessing information from a single-time-point microbial sample holds transformative potential for advancing personalized medicine. Traditional techniques, including alpha-diversity (richness-based), beta-diversity (dissimilarity-based) measures and neural network models, effectively pinpoint compositional differences and species abundance variations between the test sample and a reference population. Recently, two novel approaches have been proposed, that, instead of assessing abundance of the individual species, focus on the inter-species relationships in the test sample, the network-impact (NI) approach and the individualized dissimilarity-overlap analysis (IDOA). The NI is a measure of the contribution of the test sample to the calculated inter-species correlations in a bottom-up approach, while the IDOA analyzes the relationships between species assemblages and their abundances in a top-down approach. In this research, we comprehensively evaluate and compare the traditional and the new approaches. Employing the Generalized Lotka-Volterra (GLV) model, we create synthetic samples to rigorously test the classification capabilities of each measure in both supervised and semi-supervised setups. Our findings reveal that when the species of the test sample share similar self-dynamics as the reference cohort but distinct inter-species interactions, they typically can not be classified by conventional dissimilarity-based measures. In contrast, IDOA and NI parameters emerge as successful tools in assessing microbial relationships within a single-time-point sample, while a neural network is highly dependent on the training set size. This study underscores the potential of extracting information from the intricate inter-species relationships based on individual microbial snapshots, paving the way for improving microbiome-based personalized medicine.