Applications of weighted association networks applied to compositional data in biology

Abstract

Summary Next‐generation sequencing technologies have generated, and continue to produce, an increasingly large corpus of biological data. The data generated are inherently compositional as they convey only relative information dependent upon the capacity of the instrument, experimental design and technical bias. There is considerable information to be gained through network analysis by studying the interactions between components within a system. Network theory methods using compositional data are powerful approaches for quantifying relationships between biological components and their relevance to phenotype, environmental conditions or other external variables. However, many of the statistical assumptions used for network analysis are not designed for compositional data and can bias downstream results. In this mini‐review, we illustrate the utility of network theory in biological systems and investigate modern techniques while introducing researchers to frameworks for implementation. We overview (1) compositional data analysis, (2) data transformations and (3) network theory along with insight on a battery of network types including static‐, temporal‐, sample‐specific‐ and differential‐networks. The intention of this mini‐review is not to provide a comprehensive overview of network methods, rather to introduce microbiology researchers to (semi)‐unsupervised data‐driven approaches for inferring latent structures that may give insight into biological phenomena or abstract mechanics of complex systems.

Interactions between fecal gut microbiome, enteric pathogens, and energy regulating hormones among acutely malnourished rural Gambian children

“…Networks are graphical structures used to represent relationships between discrete objects where these discrete objects are referred to as nodes and connections between nodes as edges [51] . The connections are weighted by a numeric value.…”

VEBA: a modular end-to-end suite for in silico recovery, clustering, and analysis of prokaryotic, microeukaryotic, and viral genomes from metagenomes

“…Another alternative would be aligning reads to a concatenated assembly but—due to the likelihood of similar but distinct strains of the same species occurring in multiple samples—reads will either be randomly assigned or multi-mapped. The former would result in another sparse matrix and latter in a multi-mapped counts table both of which violate assumptions of compositional data analysis [ 87 ] with the latter known to introduce downstream analytical complications [ 88 – 92 ]. Further, sample-specific and consensus metagenomics is analogous to amplicon-sequence variants [ 21 , 93 ] and operational taxonomic units [ 22 ] in that MAGs yielded by the former can be added to existing databases as their construction is not dependent on multiple samples.…”

“…Networks were implemented using the following approach: (1) split feature matrix into (1a) mature plastic biofilm samples and (1b) early plastic biofilm samples; (2) ρ proportionality for ensemble co-occurrence of Network Mature and Network Early separately [ 87 , 88 , 91 , 92 ] using the EnsembleNetworkX Python package [ 96 ] with 1000 iterations; (3) compute differential connectivity via Network Mature —Network Early ; (4) consider only edges that have positive associations in both conditions (negative ρ associations are non-trivial to interpret) and have a differential connectivity of at least 0.1; and (5) hive plot of differential connectivity edges implemented via Hive NetworkX [ 97 ] . Network analysis was performed only on the Plastisphere dataset as this had several taxa for each domain which was not the case in MarineAerosol or Netherton datasets.…”

Rapid reductions in population size drive evolutionary divergence in diatoms

“…Taking a compositional approach, we used the centered log-ratio (CLR) transformation on raw transcript counts by taking the log of each count and dividing by the geometric mean using the compositional Python package (Espinoza et al, 2020). Hierarchical clustering and PCoA ordinations was performed using SciPy (Virtanen et al ., 2020) and Soothsayer (Espinoza et al ., 2021) Python packages.…”