F1ALA: ultrafast and memory-efficient ancestral lineage annotation applied to the huge SARS-CoV-2 phylogeny

The unprecedentedly large size of the global SARS-CoV-2 phylogeny makes any computation on the tree difficult. Lineage identification (e.g. the PANGO nomenclature for SARS-CoV-2) and assignment are key to track the virus evolution. It requires annotating clade roots of lineages to unlabeled ancestral nodes in a phylogenetic tree. Then the lineage labels of descendant samples under these clade roots can be inferred to be the corresponding lineages. This is the ancestral lineage annotation problem, and matUtils (a package in pUShER) and PastML are commonly used methods. However, their computational tractability is a challenge and their accuracy needs further exploration in huge SARS-CoV-2 phylogenies. We have developed an efficient and accurate method, called ‘F1ALA’, that utilizes the F1-score to evaluate the confidence with which a specific ancestral node can be annotated as the clade root of a lineage, given the lineage labels of a set of taxa in a rooted tree. Compared to these methods, F1ALA achieved roughly an order of magnitude faster yet with ~12% of their memory usage when annotating 2,277 PANGO lineages in a phylogeny of 5.26 million taxa. F1ALA allows real-time lineage tracking be performed on a laptop computer. F1ALA outperformed matUtils (pUShER) with statistical significance, and had comparable accuracy to PastML in tests on empirical and simulated data. F1ALA enables a tree refinement by pruning taxa with inconsistent labels to their closest annotation nodes and re-inserting them back to the pruned tree to improve a SARS-CoV-2 phylogeny with both higher log-likelihood and lower parsimony score. Given the ultrafast speed and high accuracy, we anticipated that F1ALA will also be useful for large phylogenies of other viruses. Codes and benchmark datasets are publicly available at https://github.com/id-bioinfo/F1ALA.