CO-phylum: An Assembly-Free Phylogenomic Approach for Close Related Organisms

1. Introduction to algorithm

• The method indexes each nucleotide as an Object paired with its surrounding short flank sequence, the Context, and treats matching Contexts as evidence of homologous Objects.For raw sequencing data, Context–Object structures are derived from reads and filtered for low quality.Source: §1 · pp. 5–6
• The probability of a Context appearing randomly is estimated under a random-genome assumption, with K ≥ 8 given p < 0.01 for a 5M-length prokaryotic genome.The estimate is used to avoid contexts that appear randomly.Source: §1 · p. 5
• For each genome pair, CO-phylum counts shared Contexts and shared Contexts with non-identical Objects, then uses their ratio as the distance.Repeated Contexts are excluded unless all associated Objects are identical.Source: §1 · pp. 5–6
• The resulting pairwise distances form a matrix that the neighbor-joining method uses to construct a phylogenetic tree.Indexing uses a unified, fixed-length Context–Object structure for computational efficiency.Source: §1 · p. 5

2. Low quality value tuple filtering policies

• Tuples with a minimum quality value below 10 are filtered, and tuples appearing only once in an organism’s dataset are also removed.Tuple quality is represented by the minimum quality value among its constituent bases.Source: §2 · p. 6

3. NGS data Simulation

• Perfect simulated reads are 75-bp genome substrings sampled randomly from either strand.The reads were generated with an in-house Perl script.Source: §3 · p. 6
• A separate FastQ simulation used Maq with a haploid model and mutation and 1-bp indel rates set to zero.The study’s simulation parameters are listed in Table S1.Source: §3 · pp. 6–7

1. Test CO-phylum by ‘CK,KO1’ of varying K

• The test compared trees for 13 Brucella organisms against an alignment-based reference tree.The reference was based on previously studied phylogenetic relationships.Source: §1 · p. 7
• Across tested K values, CO-phylum trees closely matched the Brucella reference in topology and branch length, showing robustness to K choice.The reference relationships had been explored using genome-wide orthologous SNP alignments and multiple tree-building methods.Source: §1 · p. 7

2. Compare the accuracy of CO-phylum tree with that of other alignment-free methods

• CO-phylum produced topology and branch lengths highly similar to the Escherichia/Shigella reference tree.The reference was inferred from concatenated alignments of 2034 core genes using maximum likelihood.Source: §2 · p. 7
• CVtree placed Shigella in a way that violated the monophyleticity of Escherichia, unlike the other compared methods.The comparison concerns the 26 Escherichia/Shigella phylogeny.Source: §2 · pp. 7–8
• Kr trees were less accurate than CO-phylum in both topology and branch length, while CVtree branch lengths were not proportional to the reference.The comparison also notes that FFPs achieved a similar result to CVtree.Source: §2 · p. 8
• The CO-phylum and reference trees differed at the E. coli CFT073/536/ED1a branch, a discordance also seen with two other alignment-free methods.The authors report that deleting CFT073 yielded a tree without this inconsistency; those data were not shown.Source: §2 · p. 8
• Extensive genome reorganization, including duplication or recombination, might slightly affect the accuracy of CO-phylum, CVtree, and Kr distance-based trees.The authors note that distance-based methods cannot identify which genome segments underwent reorganization.Source: §2 · p. 8

Simulated NGS data

• At 2X coverage, datasets contained only 73.67% as many contexts as complete genomes, and inadequate context counts could bias distances and trees.The 2X-coverage tree was the exception to the otherwise identical simulated-tree topologies.Source: §3 · p. 9
• The filtering policy captured most genomic tuples when sequencing depth exceeded 6X.The authors describe this depth as exceeded by most prokaryotic NGS projects.Source: §3 · pp. 9–10
• Simulated datasets above 16X coverage produced trees with identical topologies and close branch lengths to the complete-genome reference, including mixed-coverage data.The mixed-coverage tree also matched the reference in topology and branch lengths.Source: §3 · p. 10

Real NGS data

• Using the assembled-genome tree as a reference, comparisons indicated the NGS raw-data tree was almost identical to the assembled-genome tree across data from all three platforms: 454, Illumina, and SOLiD.Source: §3 · p. 10

Distance comparison

• 0.9919 and 0.9907 were the correlation coefficients for CO-distance versus p-distance with ‘C9,9O1’ and ‘C12,12O1’, respectively.The fitted models were CO-distance = 0.70p-distance - 0.00038 and CO-distance = 0.66p-distance - 0.00045, respectively.Source: §4 · p. 11
• 0.3464 and 0.7796 were the correlation coefficients for CVtree-distance and Kr-distance versus p-distance, respectively.The fitted models were CVtree-distance = 4.27p-distance + 0.081 and Kr-distance = 0.85p-distance - 0.003, respectively.Source: §4 · p. 11
• CO-distance significantly outperformed CVtree-distance and Kr-distance in fitting p-distance for the tested Escherichia/Shigella genome pairs.The analysis compared all 325 pairwise distances among 26 genomes.Source: §4 · p. 11
• The linear relationship between CO-distance and p-distance was also reported for other close-related organisms, including primate mitochondrial DNA.The primate test data were not shown.Source: §4 · p. 11

Family-level comparison

• Enterobacter, Citrobacter, and Yersinia formed single clades in the CO-phylum tree, unlike their polyphyletic or paraphyletic groupings in the 16S rRNA tree.The trees agreed well in general, while CO-phylum had higher resolution at family level.Source: §5 · p. 12

Higher taxonomy levels

• In tests of higher taxonomy levels, CO-phylum accuracy was reported not to diminish significantly until some higher level was reached; the passage does not name that level, so no class-level endpoint is asserted.The higher-level test included 63 organisms from 18 genera.Source: §5 · p. 12
• At class level, the tree generally retained clades for Enterobacteriales, Xanthomonadales, Pasteurellales, and Thiotrichales, while other orders were paraphyletic or polyphyletic.The tree used 70 genomes from 14 orders within Gammaproteobacteria.Source: §5 · p. 12
• Performance at taxonomy levels above class was worse than at class level.The results were not shown.Source: §5 · p. 12

Discussion

• CO-phylum was developed for unassembled NGS data, addressing the stated lack of an assembly-free phylogenomic method.The discussion describes it as applicable to unassembled sequencing data.Source: Discussion · p. 12
• CO-phylum is described as having ultrahigh resolution through use of whole-genome information.The discussion also describes the method as producing unique trees.Source: Discussion · p. 13
• For close-related organisms, CO-phylum produces highly accurate relationships and distance almost linear with p-distance, which the authors link to agreement with alignment-based trees.The discussion gives organisms within the same family as an example of close-related scope.Source: Discussion · p. 13
• CO-phylum indexes ‘Object’ by ‘Context’, unlike CVtree’s normalized K-mer frequencies, while both methods have equal algorithm efficiency.Both take O(L) steps to compute a distance between two genomes and O(N^2) steps for all pairwise distances among N organisms.Source: Discussion · p. 13
• A CO-phylum ‘Object’ could represent a SNP, whereas CVtree’s normalized K-mer frequencies have no explicit biological meaning.The discussion presents this distinction as explaining why CO-phylum is more accurate than CVtree and similar methods.Source: Discussion · pp. 13–14

• Figure 1 is labeled as an overview of the CO-phylum algorithm.Source: Figure 1
• Figure 2 is labeled as a Brucella phylogeny.Source: Figure 2 · p. 16
• The supplied Figure 3 entry contains only its figure number.Source: Figure 3 · p. 17

• Figure 4 is labeled as a test of CO-phylum using real NGS data.Source: Figure 4 · p. 18
• Figure 5 is labeled as linear regression of p-distance using distances generated by different methods.Source: Figure 5 · p. 19
• Figure 6 is labeled as a test of CO-phylum performance at family level.Source: Figure 6 · p. 20