What are the considerations for studying population genetics of bacterial pan-genomes?

Question

I am beginning to work in the field of human gut microbiome, and wondering how (and if) the concepts of population genetics could be applied there.

• Considering the competition between the species seems straightforward, although ecological models are probably more appropriate in this case.
• Considering the pan-genome of single species however seems next to impossible, as it cannot be divided into independent alleles. One could probably define a number of alleles as the number of copies of a particular gene in the community. However, as it confers fitness on the whole community, it is not clear how to define selection. The notion of genetic drift also seems vague.

I will appreciate clarifications and references to relevant literature.

Update
To clarify the question:

• The question is rather theoretical, motivated more by my previous work than by what I am about to do in the field of metagenomics. The latter will be analysis of shotgun sequencing data, where the focus is on identifying new species and comparing the species composition in many patients. I doubt that we will have any data permitting to study the intra-species evolution within a single patient. However, understanding how bacteria evolve to be a particular strain in a particular patient could be a plus: if members of a bacterial community with different genomes live as a group, we can't meaningfully say that some of them are more fit or less fit.
• My past experience with population genetics is in the field of HIV, which for most purposes can be treated as a diploid hermaphrodite (each virion carries two copies of the genome, and the genomes are shuffled when a cell is co-infected with several virions). Thus, many basic population genetics concepts are easily applied, although the speed of evolution and the need for constant adaptation add other difficulties (motivating such approaches as fitness edge).
• Finally, as was pointed in the comments, the problem is not specific to bacteria - altruism does take place in many communities, and should raise the same kinds of issues as I mentioned previously. I admit that I am not familiar with the necessary theoretical frameworks, and this can be considered a part of the question. I would guess however that most such frameworks do not account for direct gene exchange.

Update 2
A relatively recent article by Eduardo Rocha (2018) discusses the shortcomings of the neutral model (as the principal null model for modern genomics) when applied to procaryotes. This has clarified a lot for me, and I give below a short list of their points:

1. Mutations in bacteria cannot be treated as neutral.
2. Polymorphism concept is poorly defined for genomes of varying length, which requires new mathematical framework.
3. Promiscuity of bacteria, freely recombing and exchanging genes via the horisontal gene transfer makes it difficult to define species and apply many concepts of the traditional population genetics.
4. Difficulties in separating demographics and genetics (essentially, one needs a combined genetico-ecological approach).

Update 3
Inclusive fitness theory / kin theory (also here) probably fits into this context.

Answer 1

It is possible to perform gene-level population genetics analyses on microbiome-derived bacterial species.^{1,2 †} However, the biology of bacteria and the nature of metagenomic sequencing data make such analyses difficult and intractable to many of the well-worked tools of population genetics. Here are some of the things you should keep in mind as you build a conceptual framework for understanding the field:

Bacterial genomes are flexible

The term "allele" connotes gene variants, and the underlying assumption of analyses involving alleles is that each (diploid) individual in a population has one or two alleles of a given gene. When considering a bacterial population, specifically all the cells that constitute a species, some genes will not be present in all cells. For this reason, "allele" is not frequently used in discussions of complex bacterial communities. For well-studied species like E. coli and K. pneumoniae with many complete strain genomes, plotting the number of genes found across genomes gives a characteristic U-shaped distribution.³ For a given species, this suggests that many genes are present in most strains, and many genes are present in only one or two strains. The simple thing to do is to constrain your analysis to the gene set that you observe to appear in all ^‡ individuals in a population (the core genome, to use a pangenomics term), but this likely excludes genes that are important for niche-specific fitness like virulence factors, antibiotic resistance genes, carbohydrate degradation operons, etc. Recently, this problem has been addressed computationally by representing pangenomes as weighted graphs with genomic features at nodes ⁴ and by building composite reference genomes from strain genomes data encoded as reference graphs.^{3 #}

Metagenomes are messy (but getting cleaner!)

Horizontal gene transfer is likely a frequent occurrence in microbiomes,⁵ confounding analyses that rely on strict inheritance. Since microbiome sequencing is often short-read, the problem of horizontal gene transfer is exacerbated by the difficulty in reliably assembling transferred genes with specific metagenome-assembled genomes. As noted by Maximillian Press, there are several ways to address this issue. At the bench, proximity ligation sequencing can supplement traditional metagenomic assembly-by-alignment using the ground truth of physical closeness.^5,6 Likewise, single-cell isolation combined with multiplexed whole-genome amplification greatly reduces complexity of assembly.⁷ At the computer, sequence composition (i.e. kmer content) and coverage profiles can be leveraged to associate metagenomic contigs in bins,^8,9 which are like pseudo-genomes.^§ If you're lucky enough to have long-read sequencing data, many of the problems associated with assembly of short reads are no longer an issue,¹⁰ with the added bonus that SMRT and nanopore sequencing produce methylation data -- unique methylation profiles can be used to infer which contigs likely belong to the same genome or to associate extragenomic contigs (e.g. plasmids) with their hosts.^11,12 All of this to say that getting complete genomes from microbiome sequencing is non-trivial, but there are many techniques and tools at your disposal.

Microbiome data is relative

Much of population genetics deals in counted data -- $n$ individuals with allele $a$, $m$ individuals with allele $b$, and so on. Concerning microbiomes, the absolute abundances of individual microbes are not recoverable from sequencing alone. Microbiome data is therefore said to be compositional. Like RNA-seq, gene or organism abundances derived from metagenomic sequencing are relative, and, importantly, many of the assumptions underlying the statistical analyses applied to absolute counts do not hold for compositional data.^13,14 Thankfully, statisical tools for population-level analysis of microbiomes have been developed,^15,16 though their widespread adoption has been slow.

Microbiome members are often interdependent

As hinted at in the question update, bacteria display a type of community altruism, where an individual cell with a specific gene can influence the fitness of neighboring cells that lack the gene. For an example, see my answer to Will all bacteria become resistant against all antibiotics in the long term? concerning secreted β-lactamases. Therefore, spatial association is an added factor when considering population dynamics of a genetically heterogeneous bacterial species. Some methods that address cell-cell spatial proximity in microbiomes include sequencing of cryofractured fragments ¹⁷ and probe-based spectral imaging.¹⁸ Even if microbes are not spatially associated, different microbes may play complementary roles in the iterative metabolism of large carbohydrates into small metabolites.^19,20

Surely, this discussion is incomplete, though I hope my answer has given you the footing you need to continue your own exploration to find the appropriate resources for your research. For a more in-depth discussion of the points I've addressed here, see What Is Metagenomics Teaching Us, and What Is Missed?,²¹ particularly the sections titled Strain-Level Analyses and Ecoevolutionary Modeling.

---

^{† I recommend keeping up with the publications of Ran Blekhman, Peer Bork, and Katie Pollard.}

^{‡ Core genome membership seldom requires that 100% of individuals represented by that genome contain the gene in question, and setting such a strict cutoff would likely exclude truly essential genes that were not found associated with a given strain due to gaps. In practice, core genome cutoff thresholds are on the order of 90%.}

^{# The senior author for reference 3 made a great explanatory Twitter thread.}

^{§ Contigs and Bins and MAGs, oh my! As with any field, metagenomics has field-specific jargon. Coursera seems to have a good lecture that includes these concepts and how they relate to eachother.}

---

References

1. Garud NR, Pollard KS. Population Genetics in the Human Microbiome. Trends Genet. 2020 Jan;36(1):53-67. doi: 10.1016/j.tig.2019.10.010.
2. Priya S, Blekhman R. Population dynamics of the human gut microbiome: change is the only constant. Genome Biol. 2019 Jul 31;20(1):150.
3. Colquhoun RM et al. Nucleotide-resolution bacterial pan-genomics with reference graphs. bioRxiv 2020.11.12.380378.
4. Gautreau G et al. PPanGGOLiN: Depicting microbial diversity via a partitioned pangenome graph. PLoS Comput Biol. 2020;16(3):e1007732.
5. Kent AG et al. Widespread transfer of mobile antibiotic resistance genes within individual gut microbiomes revealed through bacterial Hi-C. Nat Commun. 2020 Sep 1;11(1):4379.
6. Stalder T et al. Linking the resistome and plasmidome to the microbiome. ISME J. 2019 Oct;13(10):2437-2446.
7. Chijiiwa R et al. Single-cell genomics of uncultured bacteria reveals dietary fiber responders in the mouse gut microbiota. Microbiome 8, 5 (2020).
8. Alneberg J et al. Binning metagenomic contigs by coverage and composition. Nat Methods. 2014 Nov;11(11):1144-6.
9. Kang DD et al. MetaBAT, an efficient tool for accurately reconstructing single genomes from complex microbial communities. PeerJ. 2015 Aug 27;3:e1165.
10. Kolmogorov M et al. metaFlye: scalable long-read metagenome assembly using repeat graphs. Nat Methods 17, 1103–1110 (2020).
11. Beaulaurier J et al. Metagenomic binning and association of plasmids with bacterial host genomes using DNA methylation. Nat Biotechnol. 2018;36(1):61-69.
12. Tourancheau A et al. Discovering and exploiting multiple types of DNA methylation from individual bacteria and microbiome using nanopore sequencing. bioRxiv. 2020.02.18.954636.
13. Tsilimigras MC and Fodor AA. Compositional data analysis of the microbiome: fundamentals, tools, and challenges. Ann Epidemiol. 2016 May;26(5):330-5.
14. Gloor GB et al. Microbiome Datasets Are Compositional: And This Is Not Optional. Front Microbiol. 2017 Nov 15;8:2224.
15. Shi P et al. Regression analysis for microbiome compositional data. Ann. Appl. Stat. 10 (2016), no. 2, 1019--1040.
16. Kurtz ZD et al. Sparse and compositionally robust inference of microbial ecological networks. PLoS Comput Biol. 2015 May 7;11(5):e1004226.
17. Sheth RU et al. Spatial metagenomic characterization of microbial biogeography in the gut. Nat Biotechnol. 2019 Aug;37(8):877-883.
18. Shi H et al. Highly multiplexed spatial mapping of microbial communities. Nature. 2020 Dec.
19. Kundu P et al. Species-wide Metabolic Interaction Network for Understanding Natural Lignocellulose Digestion in Termite Gut Microbiota. Sci Rep 9, 16329 (2019).
20. Selber-Hnatiw S et al. Metabolic networks of the human gut microbiota. Microbiology. 2020 Feb;166(2):96-119.
21. New F, Brito IL. What Is Metagenomics Teaching Us, and What Is Missed? Annual Review of Microbiology. 2020 74:1, 117-135.

Answer 2

I will answer this as best as I am able, the updates are helpful for providing context. I'll also note that there are several questions packed in here, which makes it a bit more laborious to answer. I am going to try to answer the following questions:

1. How can I infer selection on microbiota within single hosts (patients)?
2. How can I resolve genomes of related organisms from a complex mixture?
3. How can I do population genetics in the presence of horizontal gene transfer (HGT)?

Question 1

I think that there are some assumptions here that rest on methodological tractability that are, as far as I can tell, not accurate. For example: "...if members of a bacterial community with different genomes live as a group, we can't meaningfully say that some of them are more fit or less fit."

This seems to be the crux of the issue. I would argue that if you are willing to take time-courses from the same patient, it is not at all difficult to measure changes in abundance. Causal inference of e.g. selection is a somewhat harder problem, but one that has been treated exhaustively elsewhere.

If you are willing to postulate that you can deconvolute lineages from a single sample, and measure multiple samples from the same patient, then you can try to measure selection over time. Some steps of this are hard (e.g. if lineages are closely related), but those are merely technical problems that can be solved.

There are also methods that try to infer bacterial growth rates from a single sample, but they're a bit more complicated.

Question 2

Resolution of complex mixed populations into genomes is a well-studied problem. THere are a number of approaches to do this (all of which require relatively high investment and sequence assembly probably):

• Long-read single-molecule sequencing: simply use Nanopore or high-coverage PacBio reads to directly ascertain the genome. Expensive (especially for complex samples) but perfectly possible. If you use Nanopore you may not even need to assemble!
• Assembly binning: Statistical approaches such as MetaBAT2 attempt to deconvolute metagenomic assemblies based on sequence composition and contig coverage. Limited by the resolution of your assembly- if a contig is collapsed in the assembly, you can't put it in two places (as in the case of very closely related genomes).
• Proximity-ligation binning: similar to assembly binning, but using complementary single-cell information to place contigs into the same bin. Some additional refs. More accurate and higher-resolution, but involves an additional data type/expense. Full disclosure: I work for a company that provides this as a service and sells kits to do it.

All of these methods, in principle, could be applied to a subset of samples (e.g. beginning and end) and reference genomes could be imputed to the other samples, where you can just measure them by low-coverage shotgun.

Question 3

This is a big one, I would suggest simply googling it to learn more than I could hope to tell you about population genetic theory in the presence of HGT.

The shorter version is that detectable HGT is not super common at the time scales that you are likely to be dealing with. Most HGT ends quickly with the transferred DNA getting digested for energy or selected out as genomic junk.

Nonetheless there is a clear case that selection of lineages can be driven by HGT events, especially under very strong challenges like antibiotics.

I would argue that if you can measure abundance of a lineage and assign HGTs to genomes in specific samples (readily possible with proximity ligation at least), there is no barrier to doing population genetics and measuring selection. It's just another kind of mutation that happens to have a high rate and to occur in parallel. People measure the selective advantage of e.g. antibiotic resistance all the time.

You may be using heuristics and ad hoc methods for some steps, but genomics is just heuristics and ad hoc methods anyways.