ADAPT: Analysis of Microbiome Differential Abundance by Pooling Tobit Models

Abstract

Motivation: Microbiome differential abundance analysis (DAA) remains a challenging problem despite multiple methods proposed in the literature. The excessive zeros and compositionality of metagenomics data are two main challenges for DAA.

Results: We propose a novel method called "Analysis of Microbiome Differential Abundance by Pooling Tobit Models" (ADAPT) to overcome these two challenges. ADAPT interprets zero counts as left-censored observations to avoid unfounded assumptions and complex models. ADAPT also encompasses a theoretically justified way of selecting non-differentially abundant microbiome taxa as a reference to reveal differentially abundant taxa while avoiding false discoveries. We generate synthetic data using independent simulation frameworks to show that ADAPT has more consistent false discovery rate control and higher statistical power than competitors. We use ADAPT to analyze 16S rRNA sequencing of saliva samples and shotgun metagenomics sequencing of plaque samples collected from infants in the COHRA2 study. The results provide novel insights into the association between the oral microbiome and early childhood dental caries.

Availability and implementation: The R package ADAPT can be installed from Bioconductor at https://bioconductor.org/packages/release/bioc/html/ADAPT.html or from Github at https://github.com/mkbwang/ADAPT. The source codes for simulation studies and real data analysis are available at https://github.com/mkbwang/ADAPT_example.

Figure 1.

Illustration of ADAPT with a toy example. (a) Three microbiome taxa (taxa 1, 4, and 7) are DA between two ecosystems. Neither the observed counts nor the relative abundances can be directly compared for DAA. (b) ADAPT treats zero counts as left-censored at the detection limit (one in this case). ADAPT first calculates the fold change of relative abundances. It then selects a subset of taxa (taxa 2, 3, and 5) whose fold changes equal the median as reference taxa. After scaling the counts by the sum of three reference taxa, ADAPT can recover the DA taxa without false positives by comparing the normalized counts.

Figure 2.

Simulation studies for comparing ADAPT with nine other DAA methods. We simulate synthetic metagenomics sequencing count data under two contrasting conditions using the SparseDOSSA framework. The number of samples is the same between the two conditions. We generate 500 replicates for each simulation setting and report the mean of performance metrics. (a) False positive rates (type I errors) of all methods except for ANCOM under simulation settings with no DA taxa. The total number of taxa is 500. The total sample size is 50 or 100. The average library size is the same (balanced) for two conditions at ${10}^4$ or different (unbalanced) between two conditions (${10}^4$ for one condition and ${10}^5$ for the other). (b) False discovery rates and power under simulation settings with different proportions of DA taxa. The sample size is 100. The total number of taxa is 500. The proportion of DA taxa is 5%, 10%, 20%, or 30%. The average library size is $2\times {10}^4$ for both conditions. The average absolute abundance fold change of DA taxa is 5. The directions of absolute abundance changes of DA taxa may be balanced or unbalanced.

Figure 3.

Microbiome DAA between children who developed ECC and those who did not. (a) Thirty-eight out of 155 total ASVs in the saliva samples collected at 12 months old are DA based on at least one method. ADAPT detects 27 DA ASVs. (b) Fourteen out of 590 taxa in the plaque samples collected between 36 and 60 months old are DA based on at least one method. ADAPT detects 12 DA taxa. (c) Volcano plot for DAA of saliva samples. (d) Volcano plot for DAA of plaque samples.