Differential Abundance Analysis with Bayes Shrinkage Estimation of Variance (DASEV) for Zero-Inflated Proteomic and Metabolomic Data

Abstract

Mass spectrometry (MS) is frequently used for proteomic and metabolomic profiling of biological samples. Data obtained by MS are often zero-inflated. Those zero values are called point mass values (PMVs). Zero values can be further grouped into biological PMVs and technical PMVs. The former type is caused by true absence of a compound and the later type is caused by a technical detection limit. Methods based on a mixture model have been developed to separate the two types of zeros and to perform differential abundance analysis comparing proteomic/metabolomic profiles between different groups of subjects. However, we notice that those methods may give unstable estimate of the model variance, and thus lead to false positive and false negative results when the number of non-zero values is small. In this paper, we propose a new differential abundance analysis method, DASEV, which uses an empirical Bayes shrinkage method to more robustly estimate the variance and enhance the accuracy of differential abundance analysis. Simulation studies and real data analysis show that DASEV substantially improves parameter estimation of the mixture model and outperforms current methods in identifying differentially abundant features.

A Review on Differential Abundance Analysis Methods for Mass Spectrometry-Based Metabolomic Data

“…Simulation studies that compared the performance of different methods were conducted in the literature. Gleiss [22,28,50]. In summary, if the proportion of BPMVs are similar between the two groups, one-part tests generate acceptable results.…”

“…SDA is able to handle both normally and non-normally distributed features simultaneously, and outperforms two-part t-test and two-part Wilcoxon test for non-normally distributed features [28]. Mixture models, especially DASEV, can provide less biased estimates on both proportions of TPMVs and BPMVs when the TPMV proportion is high [22,50]. LIM-LRT, DASEV, and SDA all yield good true positive rates when the PMV proportion is not very high [22,28,50].…”

“…Mixture models, especially DASEV, can provide less biased estimates on both proportions of TPMVs and BPMVs when the TPMV proportion is high [22,50]. LIM-LRT, DASEV, and SDA all yield good true positive rates when the PMV proportion is not very high [22,28,50].…”

“…Huang et al ( 2020) noticed that the variance estimation from LIM could be unstable in presence of a large proportion of zero values, which affected the DA analysis results [50]. To address this issue, they adapted the variance shrinkage approach proposed by Smyth (2004) for microarray data to the mixture model setting, where data from the ensemble of metabolites were borrowed to achieve a more robust variance estimation of each individual metabolite [30].…”

“…After all model estimates are obtained until convergence, a likelihood ratio test is applied for DA analysis. Huang et al (2020) also extended LIM to allow covariate adjustment, where a logistic regression model was used to characterize the association between covariates and the proportion of BPMVs and a linear model was used to characterize the association between covariates and the mean of non-BPMVs [50].…”