Background: Heterogeneity of disease is a major concern in medical research, which is commonly recognized as subtypes with different pathogenesis, exhibiting distinct prognosis and treatment effects. The classification of population into homogenous subgroups is challenging, especially for the complex diseases. Recent studies show that gut microbiome compositions play a vital role in disease development, and it is of great interest to cluster the patients according to their microbial profiles. There are a variety of beta diversity measures to quantify the dissimilarity between the compositions of different samples for clustering. However, using different beta diversity measures results in different clusters, and it is difficult to make a choice.
Results: Considering the microbial compositions from 16S rRNA sequencing, which are presented as a high-dimensional vector with large proportion of extremely small or even zero-valued elements, we set up three simulation experiments to mimic the microbial compositional data and evaluate the performance of different beta diversity measures in clustering. Their performance in two real datasets demonstrates the validity of the simulation experiments.
Conclusion: It is shown that J-divergence/Jensen-Shannon divergence/square root of Jensen-Shannon divergence and Bhattacharyya/Hellinger can capture the compositional changes at low abundance elements more efficiently, and Bhattacharyya/Hellinger is suggested.