Microbial Phylogenomics

Amplifying one gene tells you who is present. It says nothing about what they do. The next step was to sequence everything in a sample at once and reassemble genomes out of the mixture. In 2004 Jillian Banfield's group did this for an acid mine drainage biofilm, simple enough — five dominant organisms — that near-complete genomes came out, and Craig Venter's group did it for Sargasso seawater, returning more than a billion base pairs and over a million genes, most without any known relative.

Metagenome-assembled genomes are a strange kind of object. They are consensus sequences over a population rather than a clone; they can be contaminated by binning errors; they describe organisms that have never been in a flask. They are also now the majority of known microbial genomes, and they revealed a large part of the tree that culture had missed entirely. Filtering groundwater to remove anything cell-sized left a residue of bacteria with genomes under a megabase, lacking the pathways to make their own amino acids and nucleotides, belonging to dozens of lineages as distinct from each other as the familiar phyla are. Laura Hug's 2016 tree of life, built from over 3,000 genomes, has more than half of its bacterial branches populated only by sequences.

Whole genomes brought a harder problem than diversity. Genes move sideways in microbes. A gene for antibiotic resistance, for a metabolic step, for a toxin, can pass between unrelated organisms by conjugation, by virus, or by uptake of DNA from the water, and once arrived it is inherited normally. W. Ford Doolittle argued in 1999 that if this happens often enough, the universal tree is not an approximation to microbial history but the wrong shape for it. A genome is a mosaic, and each tile may have its own ancestry.

The partial reply is that not all genes travel equally. Those of the ribosome and the transcription machinery are embedded in so many interactions that a transferred copy rarely works, and trees built from them agree with one another far more than chance allows. That core is what deep phylogenies use, and it is why a universal tree is still drawn — with the understanding that it is the history of a conserved core, not of every gene in any genome.

That core tree then produced the field's largest surprise. In 2015, metagenomes from marine sediment near Loki's Castle, a vent field on the Mid-Atlantic Ridge, yielded archaea carrying genes that were supposed to be eukaryotic: relatives of actin, components of the membrane-remodelling ESCRT machinery, small GTPases. More Asgard lineages followed, and in 2020 one was cultured, after twelve years of coaxing, as a small cell with branching protrusions. When these organisms are included, phylogenies place eukaryotes not as a sister group to the archaea but inside them. If that is right, there are two primary domains of life, not three, and our own lineage is a branch of the archaea that acquired a bacterium — the event discussed under endosymbiosis.

Take a millilitre of coastal seawater. Direct counts give roughly $5 \times 10^{5}$ cells; plating gives a few hundred colonies, so culture reaches of order

$$\frac{5 \times 10^{2}}{5 \times 10^{5}} = 0.1\%.$$

Now sequence instead. A Sanger-era survey cloned perhaps $n = 100$ ribosomal genes. If a taxon makes up a fraction $f$ of the community, the chance that at least one of $n$ independent reads comes from it is

$$P(\text{detected}) = 1 - (1-f)^{n}.$$

For a taxon at 10% abundance and 100 clones, detection is a certainty: $1 - 0.9^{100} \approx 1 - 3 \times 10^{-5}$. At 1%, $1 - 0.99^{100} = 0.63$ — better than a coin flip. At 0.1%,

$$1 - 0.999^{100} = 0.095,$$

so nine times out of ten the organism is simply absent from the result. A survey of a hundred clones is blind below about one part in a thousand, and it is blind in a way that leaves no trace: the output looks like a complete community of a dozen abundant types.

Modern amplicon sequencing returns $n = 10^{5}$ reads from one sample. For a taxon at $f = 10^{-4}$ the expected number of reads is $nf = 10$, and

$$P(\text{detected}) = 1 - (1 - 10^{-4})^{10^{5}} \approx 1 - e^{-10} = 0.99995.$$

Three orders of magnitude deeper, and the long tail becomes visible: thousands of low-abundance types in a single sample, the "rare biosphere". Nothing about the ocean changed between 1990 and 2006. The detection limit moved, and with it the apparent shape of microbial diversity — which is a warning worth carrying. Every claim about how many kinds of organism are present is a claim about sampling depth, and the curve of new types against reads has, in most environments, not yet flattened.

The scale underneath these numbers is worth stating plainly. Estimates put the number of bacterial and archaeal cells on Earth at $4$ to $6 \times 10^{30}$, with a large fraction in sediments and the deep subsurface, and bacteria alone hold some 70 gigatonnes of carbon, about fifteen times the mass of all animals. Most of that is in lineages described in the last twenty years, by people who never saw a cell of it.

The practical response to all this has been to stop asking what a microbial species is and to define an operational unit instead. Two genomes sharing more than about 95% average nucleotide identity are treated as one species; ribosomal sequences above 98.7% identity likewise. The Genome Taxonomy Database rebuilt bacterial and archaeal classification on these rules in 2018, renaming a great many organisms and producing a taxonomy in which ranks are defined by evolutionary distance rather than by judgement.

It works, and it is an admission. The units are stipulated, not discovered, which is the species problem of systematics in its sharpest form: for most of life on Earth, the kinds are drawn by threshold. What the resulting lineages have done over geological time — whether they diversify, saturate, or turn over — is a question the fossil record can barely address for microbes, and that macroevolution answers mainly for animals and plants.