| Microorganisms inhabit almost every imaginable environment in the biosphere,play integral and unique roles in ecosystems,and are involved in the biogeochemical cycling of essential elements,such as carbon,oxygen,nitrogen,sulfur,phosphorus,and various metals.Their structure,function,and dynamics are critical to our existen ce.Nevertheless,microorganisms do not exist as individual population in the environment.Rather,they form complex assemblages that perform essential ecosystem functions and maintain ecosystem stability.In nature,microorganisms usually exist in diverse communities.Many microorganisms coexist through ecological interactions and jointly play roles in the geobiochemical circulation.As a powerful tool to study the interactions among microbial groups,microbial association network has attracted the attention of many microbiologists and ecologists,who use the form of association network to decipher the potential interactions among microbial groups.Much effort has been made toward the methodological development for constructing microbial association networks.However,microbial profiles suffer dramatically from zero values,which hamper accurate association network construction.In this study,we investigated the effects of zero-value issues associated with microbial association network construction.Using the TARA Oceans microbial profile as an example,different zero-valuetreatment approaches(treating zero values(i.e.,NA values)as 0,using nearest neighbor algorithm fill zero values,using 0.01 instead of zero values,excluding pairwise zero values and the unpaired zero values are replaced with 0.01,excluding pairwise zero values but unpaired zero values are kept as is,exclude all samples as long as paired or unpaired zero values are observed)were comparatively investigated using different correlation methods(Spearman,Pearson,Kendall).The results suggested dramatic variations of correlation coefficient values for differently treated microbial profiles,and the second method,which uses k-nearest neighbor algorithm instead of zero values,is significantly different from the association networks constructed by other methods,so it is not recommended to use this method for the construction of association network.Most specifically,correlation coefficients among less frequent microbial taxa were more affected,whichever method was used,so rare and abundant taxa shall be differently treated prior to network construction.Negative correlation coefficients were more problematic and sensitive to network construction,as many of them were inferred from lowoverlapped microbial taxa.The results also provided potential clues explaining why negative associations were rarely found in many studies.Since microbial data are usually compositional,we performed centered log transformed(clr).which showed that the centrality of the correlation coefficient distribution was improved for microbial abundance data.Further,we constructed and compared the consensus networks by extracting the overlapped nodes and edges in the networks constructed using different data filtering methods(first,third,fifth and sixth)and correlation methods(Spearman,Pearson and Kendall).As a results,the influence of different data processing methods on microbial consensus network is no less than that of correlation calculation methods.In addition,SparCC was indirectly evaluated in this study,and the results suggest that SparCC is also likely to be affected by data filtering methods.As microbial association network analyses have become a widely used technique in the field of microbial ecology and environmental science,we urge cautions be made to critically consider the zerovalue issues in microbial data.Consequently,microbial association networks were greatly differed.Among various approaches,we recommend sequential calculation of correlation coefficients for microbial taxa pairs by excluding paired zero values.Filling missing values with pseudo-values is not recommended. |
PDF Full Text Request