MiCoNE: turning 16S amplicon data into co-occurrence networks

MiCoNE is a Nextflow pipeline with a Python command-line interface. It converts raw 16S sequencing reads into microbial co-occurrence networks. Each step accepts several alternative methods, so the same dataset can be run through many combinations of tools and the resulting networks compared. It was joint work with collaborators at Boston University and the Segrè lab.

Each step in the workflow has more than one accepted method

Going from 16S reads to a network of microbial associations takes five or six steps, and each step has several defensible answers. Reads can be denoised with DADA2 or Deblur, or clustered at 97% identity against Greengenes. Taxonomy can be assigned with a naive Bayes classifier trained on Greengenes, or by BLAST against NCBI RefSeq. The count matrix can be filtered or left alone. Edges can be inferred with SparCC, SpiecEasi or FlashWeave. Each of these choices has been used in a published study.

How much the final network depends on those choices had not been measured. Papers reported a network and the method that produced it, and readers could not tell whether a different but equally reasonable pipeline would have produced the same edges. MiCoNE makes the comparison mechanical: hold the data fixed, vary one step, and compare the networks that come out.

Five stages, with interchangeable methods at each

The MiCoNE pipeline

Stage	What it does
Sequence processing	Demultiplexes and quality-trims reads
Denoising and clustering	Builds sequence features and removes chimeras
Taxonomy assignment	Maps representative sequences to reference taxonomies
Feature-table processing	Groups, normalises and filters abundance tables
Network inference	Compares association methods and builds a consensus network

Methods written in R, Julia and Python run in isolated environments, and Nextflow passes alternative choices through the same downstream analysis. A pipeline decision therefore becomes an experimental variable rather than an undocumented assumption.

The taxonomy database accounts for two thirds of the variance

The evaluation used a faecal microbiome transplant dataset with samples from children with autism spectrum disorder and healthy controls, sweeping the available tools at each step. The variance in the inferred networks was distributed very unevenly across the steps. The taxonomy database accounted for 65.4% of it and filtering of the count matrix during OTU processing for 26.8%. Network inference accounted for 6.553%, denoising and clustering for 0.648%, and chimera checking for 0.003%.

The reference databases disagree substantially. Below the order level, any pair of Greengenes, SILVA and NCBI RefSeq agrees on less than 45% of assignments, and even at the order level more than 51% of assignments mismatch. Denoising methods behaved differently: they produced nearly identical abundance profiles and diverged only on rare sequences.

Network inference methods disagree on individual edges

The association methods disagreed sharply about which edges to call, so the pipeline computes a consensus network rather than relying on any single method. The default is the scaled-sum method at a threshold of 0.333, chosen by benchmarking against synthetic data generated with NorTA and seqtime. The other defaults come from the same benchmarks: DADA2 for denoising, remove_bimera for chimera checking, and the Greengenes 13_8 classifier with naive Bayes for taxonomy.

Citation

Kishore, D., Birzu, G., Hu, Z., DeLisi, C., Korolev, K. S., & Segrè, D. (2023). Inferring microbial co-occurrence networks from amplicon data: a systematic evaluation. mSystems, 8(4), e00961-22. doi:10.1128/msystems.00961-22

Each step in the workflow has more than one accepted method

Five stages, with interchangeable methods at each

The taxonomy database accounts for two thirds of the variance

Network inference methods disagree on individual edges

Links

Citation