Problem

Going from metagenomes to community metabolic models meant chaining a dozen separate tools by hand, and models built from reference genomes miss the strain-level differences in a real sample.

Approach

I designed and released metaGEM, a Snakemake workflow that runs quality control, assembly, binning, bin refinement, taxonomy, metabolic model reconstruction and community simulation in one workflow, on a single machine or a Slurm cluster.

My contribution

First author and lead developer. The paper credits me with co-conceiving the project, developing the pipeline with Filip Buric, running all the analyses and co-writing the paper. I wrote the documentation, maintain the Bioconda package, published six releases (v1.0.0 to v1.0.5, 2021 to 2023) and still answer users' GitHub issues.

Tools

Snakemake, Python, CarveMe, SMETANA, MEGAHIT, CONCOCT, MaxBin2, MetaBAT2, metaWRAP, GTDB-Tk, MEMOTE, Slurm HPC

Links

• Paper (NAR)
• GitHub
• Docs
• Snakemake catalog

Problem

Ecological theory predicts that many soil bacteria cannot make some essential nutrients and rely on neighbours for them, which would also help explain why most soil bacteria will not grow alone in the lab. This had not been tested systematically in real soil communities. For agtech, a strain that relies on neighbours for nutrients may not establish on its own.

Approach

The Kost lab isolated 6,931 bacteria from 27 soil communities in Germany and tested which could grow without added nutrients. We sequenced 62 strains and built metabolic models for them to see whether their genomes predict these dependencies. The lab then grew auxotrophs with strains from the same soil, in pairs and small groups, and the models predicted which neighbours could supply the missing nutrients.

My contribution

Co-first author, shared with Ghada Yousif. I co-designed the genome sequencing and modelling and did most of the genomic analysis and metabolic modelling: sequence analysis with metaGEM, genome-scale models of the 62 sequenced strains and their comparison with the Kost lab's growth tests, the insertion-sequence and gene-loss analysis, pangenomics, the phylogenetic regression and the community simulations. I also co-wrote the first draft. The auxotrophy screen was the Kost lab's experimental work. The models missed most auxotrophies, and the genome analysis linked many missed cases to insertion sequences and gene loss in pathway genes, which became a result of the paper. The code repository maps every figure and table to the code and data that produce it.

What I used

metaGEM, CarveMe, SMETANA, ReFramed, Pangenomics, R (phyloglm), Python

Press coverage

• Phys.org
• Osnabrück University (German)

Links

• Paper (Nat. Microbiol.)
• bioRxiv preprint
• Code (GitHub)

Problem

Cheddar starter cultures mix Streptococcus thermophilus with several Lactococcus strains. How these strains interact during ripening, and how that shapes flavour, was largely unknown beyond pairwise tests in simplified lab media.

Approach

The team, led by first author Chrats Melkonian (Chr. Hansen and Wageningen University), made Cheddar with the full commercial starter culture and with single strains left out, and followed it through a year of ripening; shorter milk fermentations left out single Lactococcus strains. They combined genome sequencing, genome-scale metabolic models, metatranscriptomics and metabolite measurements. S. thermophilus breaks down milk protein, which relieves nitrogen limitation for Lactococcus, and a Lactococcus cremoris strain keeps the off-flavour compounds diacetyl and acetoin in check.

My contribution

Second author. I did most of the published metabolic modelling, together with first author Chrats Melkonian: genome-scale metabolic models of the starter strains built with metaGEM's modelling module (CarveMe, SMETANA), and the community simulations. This built on earlier modelling by Daniel Machado. I also helped write the paper. The cheese-making experiments were run at Chr. Hansen.

What I used

metaGEM, CarveMe, SMETANA, ReFramed, COBRApy, Python

Press coverage

• Chemistry World
• Phys.org
• ZME Science
• MicroBites
• C&EN (ACS)

Links

• Paper (Nat. Commun.)
• Chr. Hansen, now part of Novonesis (industry partner)
• Code (GitHub)

Problem

Recurrent C. difficile infection responds to faecal transplant, but the active ingredients of "colonisation resistance" are unclear. Which metabolic interactions actually keep the pathogen out?

Approach

Joy Scaria's lab (Oklahoma State and South Dakota State universities) challenged a 14-strain gut community, and communities grown from stool samples, with antibiotics and then C. difficile in a continuous-flow bioreactor; the Patil lab in Cambridge did the modelling. Population models and genome-scale metabolic models pointed to interactions between E. coli and Bacteroides/Phocaeicola, and metabolomics pointed to fructooligosaccharide use, vitamin B3 synthesis and competition for the amino acids C. difficile ferments. Patient metagenomes supported these signals.

My contribution

Third author. With Arianna Basile I annotated the genomes and built and simulated genome-scale metabolic models of the community members and C. difficile (CarveMe, SMETANA). I analysed the metabolomics data comparing communities that suppressed C. difficile with those that did not, checked the genomes for the genes behind those differences, and helped analyse the patient metagenomes.

What I used

CarveMe, KEMET, SMETANA, ReFramed, R

Links

• bioRxiv preprint

Problem

Building and curating a genome-scale metabolic model is slow and tedious, and Hansenula polymorpha (Ogataea polymorpha), an industrially relevant methylotrophic yeast, had no published high-quality model.

Approach

A step-by-step protocol chapter that builds the model from the models of related yeasts with the RAVEN toolbox (MATLAB), and releases the result as hanpo-GEM.

My contribution

First author, with Eduard Kerkhoven (Chalmers). We built the draft model together as a side project during my MSc, and I wrote the protocol chapter; I published the model's v1.0 release on GitHub in 2024.

Tools

RAVEN, MATLAB, SBML, Homology-based reconstruction

Links

• Chapter (Methods Mol. Biol.)
• hanpo-GEM repo
• bioRxiv preprint

Problem

Plastic pollution is growing faster than the list of enzymes known to break plastic down, and finding new ones in the lab is slow. The question was whether global metagenomic data could point to candidates.

Approach

The team (first author Jan Zrimec, Chalmers) built profile hidden Markov models from 95 experimentally verified plastic-degrading enzymes and searched ocean and soil metagenomes from 236 sites, finding over 30,000 candidate enzymes for 10 plastic types. Enzyme counts correlated with measured ocean plastic pollution and, for soils, with country-level data on poorly managed plastic waste.

My contribution

Fourth author. I generated all the metagenome assemblies and genome bins (MAGs) for the study with metaGEM; Jan Zrimec then ran the enzyme searches and all the statistics.

What I used

metaGEM, MEGAHIT, CONCOCT, MetaBAT2, MaxBin2, metaWRAP

Press coverage

• The Guardian (also in Yale E360 and National Observer)
• Smithsonian Magazine
• Chalmers press release (also on EurekAlert and Phys.org)

Links

• Paper (mBio)