What we talk about when we talk about species

Apurva Narechania, Shyam Gopalakrishnan, Tom Gilbert

Genome annotation, alignment, and phylogenetics are at the center of most work in evolutionary genomics. These techniques function best when rooted in prior work. Genes are mined from new genomes using evidence from old gene models. These genomes are aligned to well-worn references to create matrices for tree reconstruction. And trees are often populated with well characterized genomes to add context to the newly sequenced. Genome inference traces a line back to model organisms, yoking the analysis of new genomes to layers of previous knowledge. We instead highlight methods that use unannotated and unaligned sequence to understand the information diversity of sequence ensembles. Any set of genomes can comprise our sequence ensemble. In a pandemic context, a sequence ensemble might be clinically isolated strains from one day. In a systematic context, a sequence ensemble could be the pangenome available for a clade. The normal bioinformatics playbook would have us align. But we instead compress. A sequence ensemble that compresses easily contains lower information diversity. For pandemics, we can use curves of information diversity to trace genomic novelty and monitor selective sweeps in new strains. For systematics, we can calculate compressibility quickly across all known bacterial taxa, leveling the criteria for species across clades. If we tolerate data loss, we can go one step further and capture structural evolution as we compress. Our approach sacrifices a lot. We skip many of the products of modern bioinformatics like variation anchored to known genes or genome alignment to prescribed references or pangenome graphs. But we gain speed, breadth, and the ability to respond to novelty.