PARNAS: Objectively Selecting the Most Representative Taxa on a Phylogeny

Abstract

The use of next-generation sequencing technology has enabled phylogenetic studies with hun- dreds of thousands of taxa. Such large-scale phylogenies have become a critical component in genomic epidemiology in pathogens such as SARS-CoV-2 and influenza A virus. However, de- tailed phenotypic characterization of pathogens or generating a computationally tractable dataset for detailed phylogenetic analyses requires bias free subsampling of taxa. To address this need, we propose parnas, an objective and flexible algorithm to sample and select taxa that best repre- sent observed diversity by solving a generalized k-medoids problem on a phylogenetic tree. parnas solves this problem efficiently and exactly by novel optimizations and adapting algorithms from operations research. For more nuanced selections, taxa can be weighted with metadata or genetic sequence parameters, and the pool of potential representatives can be user-constrained. Motivated by influenza A virus genomic surveillance and vaccine design, parnas can be applied to identify representative taxa that optimally cover the diversity in a phylogeny within a specified distance radius. We demonstrated that parnas is more efficient and flexible than current approaches, and applied it to select representative influenza A virus in swine genes derived from over 5 years of genomic surveillance data. Our objective selection of 4 to 6 strains selected every two years from the 16 distinct genetic clades were sufficient to cover 80% of diversity circulating in US swine. We suggest that this method, through the objective selection of representatives in a phylogeny, provides criteria for rational multivalent vaccine design and for quantifying diversity. PARNAS is available at https://github.com/flu-crew/parnas.

Detection and spread of high pathogenicity avian influenza virus H5N1 in the Antarctic Region

“…All H5N1 HPAIV clade 2.3.4.4b sequences available in the EpiFlu database between 1 st September 2020 and 22 nd January 2024 were downloaded to create a sequence dataset. As North America and Europe were over-represented in this dataset, these were sub-sampled to maintain representative sequences using PARNAS 78 . The remaining dataset was separated by segment and aligned using Mafft v7.520 79 , and manually trimmed to the open-reading frame using Aliview version 1.26 80 The trimmed alignments were then used to a infer maximum-likelihood phylogenetic tree using IQ-Tree version 2.2.3 81 along with ModelFinder were downloaded to create a sequence dataset.…”

“…The remaining dataset was separated by segment and aligned using Mafft v7.520 79 , and manually trimmed to the open-reading frame using Aliview version 1.26 80 The trimmed alignments were then used to a infer maximum-likelihood phylogenetic tree using IQ-Tree version 2.2.3 81 along with ModelFinder were downloaded to create a sequence dataset. As North America and Europe were over-represented in this dataset, these were sub-sampled to maintain representative sequences using PARNAS 78 . The remaining dataset was separated by segment and aligned using Mafft v7.520 79 , and manually trimmed to the open-reading frame using Aliview version 1.26 80 The trimmed alignments were then used to a infer maximum-likelihood phylogenetic tree using IQ-Tree version 2.2.3 81 along with ModelFinder 8 and 1,000 ultrafast bootstraps 82 .…”

“…All H5N1 HPAIV clade 2.3.4.4b sequences available in the EpiFlu database between 1 st September 2020 and 27 th October 2-23 were downloaded to create a sequence dataset. As North America and Europe were over-represented in this dataset, these were sub-sampled to maintain representative sequences using PARNAS 68 . The remaining dataset was separated by segment and aligned using Mafft v7.520 69 , and manual trimmed to the open-reading frame using Aliviewversion 1.26 70 The trimmed alignments were then used to a infer maximum-likelihood phylogenetic tree using IQ-Tree version 2.2.3 71 along with ModelFinder 7 and 1,000 ultrafast bootstraps 72 For the time-resolved and mugration analysis, all HA sequences available from South America, and representative from North America were combined with the sequences from Bird Island and the Falkland Islands and used to infer a maximum-likelihood phylogenetic tree as described above.…”

DNA damage drives antigen diversification through mosaic VSG formation inTrypanosoma brucei

“…Threshold distances of 0.24659 and 0.21677 were used for EATRO1125 and Lister427, respectively. Down sampling and subtree generation of representative sequences was performed using PARNAS (0.1.4) 59 . The command used for PARNAS included flag: --cover –radius <THRESHOLD distance>.…”

Rapid mortality in captive bush dogs (Speothos venaticus) caused by influenza A of avian origin (H5N1) at a wildlife collection in the United Kingdom

“…The remaining sequences were aligned and trimmed to the ORF for each segment before generating a concatenated alignment using SeqKit (44) and then used to a infer maximum-likelihood phylogenetic tree using IQ-Tree version 2.2.3 (45). This resultant phylogenetic tree contained over 2,000 sequences and was therefore sub-sampled to cover 98% of the diversity within using PARNAS (46), which reduce the dataset down approximately 300 sequences whilst still containing representatives of the predominant UK genotypes. The sub-sampled dataset was then used to infer maximum-likelihood phylogenies for each gene segment using IQ-Tree along with ModelFinder (47) and 1,000 ultrafast bootstraps (48).…”