Resolving microbial membership using Abundance and Variability In Taxonomy ('AVIT )

Abstract

Development of NGS has revolutionized the analysis in microbial ecology contributing to our deeper understanding of microbiota in health and disease. However, the quality, quantity and confidence of summarized taxonomic abundances are in need of further scrutiny due to sample dependent and independent effects. In this article we introduce 'AVIT (Abundance and Variability In Taxonomy), an unbiased method to enrich for assigned members of microbial communities. As opposed to using a priori thresholds, 'AVIT uses inherent abundance and variability of taxa in a dataset to determine the inclusion or rejection of each taxa for further downstream analysis. Using in-vitro and in-vivo studies, we benchmarked performance and parameterized 'AVIT to establish a framework for investigating the dynamic range of microbial community membership in clinically relevant scenarios.

Figure 1. Abundance and Variability In Taxonomy (‘AVIT) methodology.

(A) In ‘AVIT we look into both the column (individual sample) and row (across samples) and using metrics like average abundance within a sample, maximum abundance within a sample and average abundance over the whole data-set, we take into account not just abundance but also variability into removal of noise (potentially erroneous species) from 16S taxonomic data. Individual characteristics of each arm can be seen in Supplementary materials S1-S2. (B) Different levels/modes of ‘AVIT used in the current study; normal, strong and extra-strong and its corresponding parameters. Primarily, the three levels used for demonstration in the current study differed in the ranges of parameters for the raw count cut-off (RC_co). (C) Current study plan. We used a combination of in-vitro (20 strains mock community) and in-vivo (monoinoculated GF mice and Altered Schaedlers Flora mice) to develop, identify parameters and validate ‘AVIT methodology in house. Subsequently, we applied ‘AVIT to clinical samples from Lauber et al. study (including lean healthy, obese diabetic and obese non-diabetic patients) to assess the applicability of the methodology and highlight the implications of our findings.

Figure 2. Application of ‘AVIT at different levels normal, strong and extra strong, on separate in-vitro Equimolar, Single Strain and staggered pool samples of 20 strain mock community.

(A) Strains used in the study and corresponding matches at the genus level after using the 16S pipeline. (B) Composition of the staggered pools with the corresponding species and their concentrations. (C) Number of species retained in different samples upon using abundance only noise reduction and using ‘AVIT at different stringency levels. (D) Species correctly/incorrectly retained or removed in 0.005 relative abundance based cutoff and different levels of ‘AVIT application to single strain, equimolar and staggered pool samples.

Figure 3. In-vivo application of ‘AVIT for a monoinoculation study and analysis of Altered Schaedlers Flora (ASF) Mice fecal samples.

(A) 15 Germ-free (GF) mice were used for the monoinoculation study. 7 GF mice were monoinoculated with E. coli (subsequently refered to as MI mice) and fecal samples were collected on Day 1 and Day 14 after monoinoculation. Additionally fecal samples were collected from all the 15 GF mice before monoinoculation (at Day 0) and from 8 remaining GF mice on Day 14. Additionally, 5 pure E. coli samples were used as controls. 13 fecal samples were collected from ASF bred mice. (B) Variation of sequence depths observed for different groups of samples. (C) Number of members retained after different forms of noise reduction for different samples. (D) Members correctly/incorrectly retained or removed in 0.005 relative abundance based cutoff and different levels of ‘AVIT application to the GF + MI, MI* (including one erroneously sequenced sample) samples, MI (excluding the erroneously sequenced sample) samples, E. coli in-vitro samples and samples from ASF mice respectively.

Figure 4. Application of ‘AVIT for fecal samples from Lauber et al.¹⁶ including samples from lean healthy, obese diabetic and obese non-diabetic individuals before and after fiber supplementation.

As compared to abundance only method of noise reduction, using ‘AVIT we firstly retain a different set of members. Additionally, based on parameter choices and stringency levels in ‘AVIT, we retain different sub-groups of members. Based on the presence/absence of members across different methodologies, we categorized the members in terms of high to low resolution. Blue: high resolution – Category 1, Purple: medium-high resolution – Category 2, Orange: medium resolution – Category 3, Green: medium-low resolution – Category 4 and Brown: low resolution – Category 5. This categorization of members allowed us to implicitly consider this scoring downstream for analysis and for rank ordering hypothesis generation.