Current statistical methods have reached breaking point and are unable to scale adequately to the enormous quantities of data which can now be generated by high-throughput genome sequencing. This has led to either an overly simple analysis of all the data or a more biologically realistic model being used on a much smaller dataset.

Inferring trees from genome data, phylogenomics is a critical first step in many modern analyses of biological data such as trees being used to track the spread of anti-bacterial resistance in hospitals. In the medium term, this new approach should make the rapid analysis of very large datasets computationally tractable.

A machine learning pipeline for phylogenomics

We use several quick-to-compute statistics to compare the input gene sequences and summarise the distances between them. These are then fed into a neural network that has been trained to relate these statistics to an evolutionary tree relating the set of species (or genomes) being studied which is demonstrated in the diagram below: