M. pneumoniae isolates

Sixty-seven M. pneumoniae isolates were selected from the historical strain collection stored in the Pneumonia Response and Surveillance Laboratory, Respiratory Diseases Branch, Centers for Disease Control and Prevention (CDC), Atlanta, GA, USA (S1 Table). Isolates were selected to represent various clinical manifestations, broad temporal and geographical distributions, and diverse strain characteristics as determined using existing typing methods, including P1 typing by PCR with HRM, MLVA, and macrolide susceptibility genotyping [20, 25–27]. Reference strains M129 (type 1) and FH (type 2) obtained from American Tissue Culture Collection (ATCC) were also resequenced in this study; two isolates of FH procured in different years were included.

M. pneumoniae isolates were grown in SP4 media (Thermo Fisher Scientific, USA) as previously described [28]. Two mL cultures were used to seed large volume (30 mL) cultures and incubated for 10 days to obtain sufficient material for preparation of genomic gDNA libraries. Genomic DNA was extracted from bacterial cultures using the MasterPure Complete DNA and RNA Purification Kit (Epicentre, Madison, WI) according to manufacturer’s protocol.

Publicly-available M. pneumoniae genomes

Forty M. pneumoniae genomes were available from the National Center for Biotechnology and Information (NCBI) data repository and downloaded on 09/06/2015 (S1 Table). The genomes used as references in the bioinformatics analysis were: M. pneumoniae M129 (Accession: NC_000912.1), M. pneumoniae M129-B7 (Accession: NC_020076.2), M. pneumoniae FH (Accession: NZ_CP010546.1), and M. pneumoniae 309 (Accession: NC_016807.1) (S2 Table).

Illumina whole genome sequencing and assembly

The NEBNext Ultra DNA library prep kit for Illumina (New England Biolabs, Ipswich, MA) was used to prepare gDNA libraries for each isolate according to the manufacturer’s protocol. Whole genome sequencing was performed on the 67 M. pneumoniae isolates using the Illumina MiSeq desktop sequencer (Illumina, San Diego, CA) with Illumina MiSeq version 2.0.

FastQC version 0.10.1 [29] was used to evaluate Illumina sequencing read quality. Sequence read data cleansing was performed with Cutadapt v1.5 [30]. Sequencing reads were removed from the data set if they met one of the following criteria: (i) had low quality (< 25) sequence bases; (ii) trimmed adapter sequences; (iii) had an error rate above 0.03; or (iv) were < 75 base pair in length. De novo assembly was performed with VelvetOptimiser v2.2.5 [31] and Velvet v1.2.10 [32]. The iterative assembly process to identify the optimal kmer value is described in detail in S1 Text.

Pacific Biosciences whole genome sequencing and assembly

A subset (n = 25) of the 67 isolates representing various outbreaks, macrolide resistance genotypes, geographic origins, and patient outcomes/disease characteristics were selected for long-read sequencing using the Pacific Biosciences (PacBio) RS II (Pacific Biosciences, Menlo Park, CA) (S1 Table). One or two SMRT cells were used for each isolate with a movie time ranging from 120–240 minutes. The sequences were assembled with SMRT Analysis v2.2 Hierarchical Genome Assembly Process version 2 (HGAP 2) or 3 (HGAP 3) protocol [33]. The Illumina sequencing data associated with each isolate was aligned for nucleotide accuracy comparison (S1 Text). The final sequences were re-oriented to start with the DNA polymerase III subunit beta gene (dnaN, locus tag–MPN001). The PacBio consensus sequences were used in place of Illumina assemblies for those 25 isolates.

Data availability

All sequencing data and assembled genomic sequences generated in this study were deposited in NCBI Genome and Short Read Archive (SRA) databases and are available under BioProject ID PRJNA328832 (S2 Table).

Genomic structure variation

BLAST Ring Image Generator (BRIG) was used to visualize comparisons of reference genomes and complete closed genomes [34]. Mauve v2.4.0 [35] was used to identify differences between the genomic structures of type 1 and type 2 isolates. Seven permutations of alignments were performed to observe structural arrangements to ensure that type-specific genome structures were captured with draft genome content as described in S3 Table. Alignment using the first dataset was performed to observe the genomic structure between closed genomes only, and alignments using second and third datasets were used to confirm consistency of regions across isolates. The remaining four datasets were constructed with draft genomes. Closed genomes from Xiao et.al [11], assembled from PacBio in this study, and NCBI references M. pneumoniae M129-B7 (Accession: NC_020076.2) and M. pneumoniae FH (Accession: NZ_CP010546.1) were aligned with Mauve v2.4.0 [36]. Backbone files derived from progressiveMauve [35] were analyzed to identify regions that were present in all type 1 genomes and absent in all type 2 genomes and the reciprocal. Each of these regions identified with backbone file were compared using BLAST against a sequence database constructed with whole genome sequences in order to verify that these sequences were type-specific. The criterion to identify these regions were as follows: >75% identity and >75% query coverage in any inter-type genome (compared to the reference type). We further used BLAST results to identify any hits that were identified to be only partial query hit (less than 40% coverage) compared to the region of interest. Any genes present in type-specific genomic regions that encoded hypothetical proteins were analyzed using Interproscan for prediction of motifs and protein function [37].

Gene prediction and phylogenetic analysis

Prodigal v2.60 [38], an ab initio gene finder, was used to predict protein coding sequences. Since M. pneumoniae has a non-traditional translation code [39], parameter “–g” was changed to 4 to allow for translation suited for Mycoplasma species. Also, protein predictions were restricted to closed ends by using parameter “-c”, which does not allow for partial gene predictions on the ends of contig sequences. To identify recombinant sites in shared genome, an alignment (n = 107) was used with Gubbins v.1.4.1 with default settings [40]. The whole genome alignment was not used because the dataset included both broken contigs and closed genomes and could result in over-representation of missing intergenic regions or create large misaligned regions.

PanOCT v3.23 [41], a sequence clustering tool designed for closely related species or strains, was used to determine the orthologous relationships between the isolates. A distinctive feature of PanOCT is the ability to differentiate paralog sequences with gene synteny [41]. Orthologous clusters were stringently defined as all sequences in a cluster having shared sequence identity and coverage ≥ 75% (S1 Text). The core genome was extracted from the pan-genome dataset to construct a phylogenetic tree. ClustalOmega v1.2 [42] was used to perform multiple sequence alignments on each set of genes in a cluster defined by PanOCT in order to avoid potential gene rearrangements in the sequence concatenation step. All individual aligned protein sequences were concatenated together to construct the core sequences for phylogenetic analysis using the method described by Hasan et al. [43]. A ML phylogenetic tree was constructed using RAxML v7.3.0 [44] with the following parameters: JTT matrix-model with the г model for rate heterogeneity, GAMMA model for four discrete rate categories, and 1000 bootstrapping replicates. Figtree [45] and the MEGA6 application [46] were used to visualize the tree. The core genome determination and subsequent phylogenetic analysis was performed on two datasets: (i) all available genomes regardless of origin and method of sequencing (n = 107) and (ii) closed genomes only (n = 34) (S1 Table).

single nucleotide polymorphism (SNP) analysis

SNPs were identified using Bowtie2 and Freebayes for the isolates that had Illumina data only. First, Illumina sequence reads from each isolate were aligned to the appropriate P1 type reference sequence using Bowtie2 with default parameters [47]. Secondly, Freebayes was used to identify SNPs based on allele coverage (> 95%) and nucleotide quality on the Phred score (>25%). Positive variant calls with likelihood score of zero were classified as false positive variants and were removed from the remainder of the analysis. Finally, all corresponding SNPs were merged based on the type reference genome position. For the 25 isolates sequenced using PacBio RSII and the isolates included in Xiao et al. [11], NUCmer and the show-snps utility from the MUMmer version 3.0 application were used to identify SNPs. For show-snps a–C parameter was chosen to remove the variants with ambiguous mapping as well as–I to exclude indel calls.

Synonymous and non-synonymous mutation rate analysis

Gene clusters were codon aligned using Multiple Alignment of Coding SEquences (MACSE) version 1.01b [48]. After alignment, a custom Perl script was used to strip in-frame stop codons and the alignments were analyzed with Phylogenetic Analysis by Maximum Likelihood package version 4.5 [49]. The codeml program was run on each single gene cluster as well as the concatenated 464 gene core genome set using the tree topology generated above and a free ratio model (settings model = 1, NSsites = 0, icode = 3, fix_omega = 0, omega = 0.4, fix_alpha = 1, alpha = 0, ncatG = 3, cleandata = 1, fix_blength = 1, method = 1). The result files were analyzed with a custom Perl script to locate branches with sufficient information to generate a meaningful estimate of ω (defined here as S * d_S ≥ 2).

Population structure analysis

The hierarchical Bayesian Analysis of Population Structure (hierBAPS) utility [50] was used to analyze the population structure of all genomes (n = 107), closed genomes (n = 34), type 1 genomes (n = 56), and type 2 closed genomes (n = 20). Independent core nucleotide genome alignments were generated for type 1 and type 2 isolates based on type determined by established laboratory methods as previously described for the core genome. The maximum cluster number was set to 15, and maximum hierarchical runs were set to 10–20.

Genome classification with Random Forest

The whole genome SNP matrix and the classifications identified in the population structure analysis were used as input for the Random Forest (RF) analysis to construct a composite matrix (S4 and S5 Tables) in which population subgroups identified in the hierBAPS analysis were in columns and allele variants for each site were in rows. The input matrix was reduced to 659 lists each containing 1–520 variant positions within the genome that shared the same SNP profile across 106 genomes (S5 Table). Variant calls were converted to a binary representation for presence (1)–absence (0) for each location (S4 Table).

Four-fold cross-validation method was used to build a classification model; the input matrix was randomly divided into three-fourths for training and one-fourth for testing the resulting model. The dataset included >3000 variant features and approximately 70 genomes depending on randomly divided three fourths dataset in each fold. The RF training model consisted of randomized feature selection with 1000 trees using RandomForest v4.6–12 [51] with R version 3.2 [52]. Ten iterations of four-fold cross validation were performed with randomly assigned testing and training datasets at each fold. At each iteration, decrease in Gini index was used to identify features of highest importance.

Genome assembly and characteristics

Assembly characteristics and genome characterization features for each genome included in this study are summarized in S1 Table. As expected, Pacific Biosciences (PacBio) Single-Molecule Real-Time (SMRT) sequencing resulted in longer read lengths and fewer contigs but lower average genome depth coverage compared to Illumina sequence data. Complete closed genomic sequences were generated for 17 of 25 (68%) isolates sequenced with the PacBio platform; the remaining eight consisted of ≤ 6 contigs. Nucleotide accuracy of PacBio sequences was > 99.85% as determined by alignment with Illumina reads. On average, 759 genes were predicted in the resulting genome sequences, which was comparable to the results of available M. pneumoniae reference genomes [11, 53]. Using Gubbins, 19 regions were identified as potential recombination sites (all <300 bp), yet the phylogenetic tree remained unchanged after masking these regions (data not shown) and thus was not considered supportive evidence of recombination driving the evolution of M. pneumoniae.

We initially compared available genome sequences of the prototypical reference strains of M. pneumoniae, M129 (type 1) and FH (type 2), to those re-sequenced in the current study in terms of both genomic content and SNPs (S1 Fig). The original FH reference genome (NC_017504.1) lacked a 6 kb region shown to be present among newer constructs of the FH genome and all type 2 isolates [11], including those in the current study (n = 51). Thus, we used the sequence reported by Xiao et al. [11] (NZ_CP010546.1) as the reference genome for type 2 isolates for the remaining analysis. The type 1 strain M129 was >99.99% identical in nucleotide sequence to recent reference sequence (NC_020076.2) that was used as type 1 reference genome in the current study (S6 Table).

Type-specific genomic content

Initial genome-wide alignments of type 1 and 2 reference genomes revealed six segments ranging in size from 13 bp to over 5400 bp in length that were unique to one type. The largest segment was previously identified by Xiao et al. [11] as a type 1-specific insertion sequence consisting of genes encoding hypothetical proteins and a single tRNA gene. Three small indels (13–15 bp) previously identified as unique to either type 1 or type 2 isolates [54] were confirmed in all newly sequenced genomes (n = 67). To confirm the apparent absence of these segments was not due to lack of Illumina sequencing reads, we performed multiple sequence alignments of only closed genomes (n = 34) (Table 1). We identified three larger regions present in all type 1 and absent in all type 2 genomes and one region present in all type 2 and absent from all type 1 genomes. However, BLAST analysis showed that these regions contained partial repetitive genomic content that were not unique to that type; hence, we examined these regions more closely to identify the sections of unique genomic content of at least 100 bp (Table 1). Portions of the single large type 2-specific region were also present in type 1 genomes, thus this segment was divided into three smaller regions that were truly unique to type 2 isolates. These regions were examined further to identify gene content and putative or known function of affected genes, which included lipoproteins, hypothetical proteins, and pseudogenes (Table 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Type-specific genomic regions and gene content.

https://doi.org/10.1371/journal.pone.0174701.t001

Phylogenetic analysis of core genomic content

The M. pneumoniae core genome dropped from 595 proteins to 464 upon inclusion of assembled contigs from Lluch-Senar et. al [8] (S2 Fig). The core used for the primary analysis of all 107 isolates included 464 proteins compared to 642 proteins for analysis of closed genomes only (S2 Fig). The lower number for the non-closed genomes was most likely due to mis-assembly, particularly at repetitive regions. The 464 core protein sequences were used to construct a maximum likelihood (ML) phylogenetic tree to represent the genomic relatedness of isolates included in this study (Fig 1). The phylogenetic analysis revealed a strong separation of isolates into clades corresponding to the known P1 type classifications based on previous laboratory testing methods. Two branches were evident within the type 1 clade; one included the majority of type 1 isolates (n = 49, 87.5%), which we designated as type 1U (Ubiquitous), and the other consisted of three genome representations of type 1 reference strain M129, along with four isolates, 303 (AL, USA), 549 (WA, USA), 4802 and 4807 (Tunisia), which we designated type 1Ref (Fig 1). The two isolates from Tunisia were previously reported as a distinct subtype designated 1d [8]. Within the larger type 1 branch, four isolates from the current study (EPC83, EPC104, EPC122, and NM1) formed a distinct subgroup designated 1N (New) (Fig 1). These four isolates originated from clinical cases occurring in the United States between 2010 and 2012. Although initially identified as MLVA type 4570 or 5570, WGS analysis revealed the presence of three tandem repeats at mpn16, the final VNTR included in the MLVA scheme, instead of zero (S3 Fig). Review of the original MLVA electropherograms for these isolates revealed the appropriate size peak corresponding to three repeats at this position, which was masked in the initial analysis by a peak of nearly identical size corresponding to repeats at a separate locus. The three repeats at this locus are a distinguishing feature from all other isolates in this study.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Phylogenetic tree of all M. pneumoniae isolates in the current study (n = 107).

Maximum likelihood phylogenetic tree of 107 M. pneumoniae isolates based on core protein sequences (n = 464) identified through orthologous clustering. Bootstrapping values over 50 are represented on the tree.

https://doi.org/10.1371/journal.pone.0174701.g001

Further phylogenetic separation of type 2 genomes within the large clade was also observed. Two main clusters were identified with one branch including all versions of the type 2 reference strain FH along with 27 isolates, all previously identified as type 2 through laboratory testing methods (designated here as type 2Ref), and the other containing the type 2 reference strain 309, which previously has been proposed as a new variant type, 2a [7] and 13 other isolates (Fig 1). Three isolates (3896, 5393, and 6009) from Germany and France previously designated as type 2 variants designated 2d by Lluch-Senar et al. [8] were also present in the clade with 309 and were identified as type 2a for the purpose of this study. A number of isolates reported as a variant of type 2 by high-resolution melt analysis of a 1900 bp amplicon within P1 were also included in this analysis [17]. Interestingly, these did not form a distinct separate cluster in the tree based on 464 core proteins; however, clear separation of these isolates was observed when the phylogenetic tree was generated using a kmer based SNP method (S4 Fig). This group consisted of isolates from the United States (n = 5), South Africa (n = 1), and France (n = 2), all collected since 2000. The two isolates from France (4911 and 6282) were previously reported as a separate subgroup designated 2c as determined by SNP/indel analysis [8]. Consistent with previous nomenclature used in our laboratory, this branch was designated as type 2 variant (2v).

The tree based on 464 core proteins (using closed and non-closed genomes) was found not to separate one of the subgroups (2v) from other type 2 genomes. Hence, phylogenetic analysis was performed using only the 34 closed genomes (S1 Table) based on the core proteins determined from this subset of genomes (n = 642, 80% of the total genes). This analysis also revealed the two expected clades consisting of P1 types 1 and 2 (Fig 2). The clade consisting of type 2v isolates from our laboratory was evident in the tree generated from closed genomes, similar to the kmer based SNP analysis (S4 Fig). The distinction of this type may be attributed to the approximately 178 additional core proteins included in the closed genome analysis (S7 Table), which represents approximately 30% more sequence data compared to the core protein set used for the primary phylogenetic analysis.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Phylogenetic tree of closed M. pneumoniae genomes in the current study (n = 34).

Maximum likelihood phylogenetic tree of 34 M. pneumoniae isolates based on core protein sequences (n = 642) identified through orthologous clustering. Bootstrapping values over 50 are represented on the tree.

https://doi.org/10.1371/journal.pone.0174701.g002

Further distinction of epidemiologically-related isolates was also observed. For example, the dominant type 1 subgroup (1U) contained the vast majority of isolates obtained from surveillance and outbreak investigations between 2010 and 2013, a period during which increased M. pneumoniae activity was observed worldwide [55–57]. The four isolates noted as type 1N (Fig 1) were an exception to this finding, having all been collected during this same period. Notably, isolates related by close association of patients, such as household contacts with M. pneumoniae infection (CO36/37 with CO58 and CO59; CO3 with CO13) [58] or cases from an outbreak in a long-term care facility (NE4 and NE26) [59] were found to be closely related.

Inherent limitations of short-read sequencing may impact resolution of repetitive sequences, several of which are known to exist in M. pneumoniae, including repetitive elements within P1. To explore the added value of whole genome sequencing data, we performed a phylogenetic analysis on a portion of the P1-encoding sequence extracted from closed genomes (S5 Fig). We observed a clear separation of the same three variant subgroups within the type 2 clade identified in the primary phylogenetic analysis using WGS data; however, no clear separation of subgroups was observed for type 1 isolates (S5 Fig).

SNP variant analysis

We identified a total of 3,206 SNPs present in at least one isolate. Intra-type examination of SNPs revealed 889 SNPs present in at least one type 1 isolate relative to the type 1 reference and 942 SNPs in one or more type 2 isolates relative to the type 2 reference genome. However these SNPs were not consistent amongst all isolates within the type designations. Comparing all 107 isolates, 520 SNPs were identified as consensus alleles in all isolates within one type group as compared to all isolates of the other type (Fig 3, S8 Table). Of these, 470 (90.4%) were located in coding regions. These 520 SNPs led to clear separation of isolates corresponding to known P1 type based on laboratory methods. Interestingly, the primary gene encoding P1 (mpn141) was not included in the core genome identified using all 107 isolates, although it was present in the core based on closed genomes only. Thus, the separation observed in our analysis must result from genomic variation outside of this gene. Other subgroups identified in the phylogenetic analysis varied from the FH reference genome by 70 (type 2a), 59 (type 1Ref), 56 (type 1N), or 7 (type 2v) SNPs that are unique to that subtype (Fig 3B). When comparing closed genomes only, the number of SNPs between the large type 1 and 2 groups was 744, presumably due to the inclusion of a larger number of core proteins; subtype-specific SNPs were similar using either core dataset (Fig 3B).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Clusters of M. pneumoniae isolates sharing unique SNPs.

(A) Number of shared unique SNPs in isolate clusters ranging from 1–106 isolates relative to reference genome FH. Only the group of isolates sharing the largest number of SNPs is shown. (B) Number of shared SNPs in each subtype relative to type 2 reference FH identified among all genomes (black bars) or closed genomes only (grey bars). *No closed genomes were available for Type 1N.

https://doi.org/10.1371/journal.pone.0174701.g003

SNPs associated with macrolide resistance were identified only in strains previously reported to harbor these mutations or have a laboratory-determined resistant phenotype. Verification of these known SNPs, including A2063G, A2064G, and C2611G mutations in the 23S rRNA gene, increased our overall confidence in variant calls.

We confirmed the presence of a type-specific non-synonymous SNP in the CARDS toxin gene (T1112G) resulting in the I371S substitution that was previously reported [8, 11] (S6 Fig); this SNP was consistently found in all 67 newly sequenced isolates. In addition, one other non-synonymous SNP (C217T) was identified in the CARDS toxin gene of four isolates from the current study (CO36, 37, 58, and 59) (S6 Fig). This SNP results in a P73S substitution that is located at the beginning of the S4 beta-sheet region in the D1 mART domain. These four isolates include two recovered from upper respiratory tract specimens collected sequentially from a patient with M. pneumoniae infection during an epidemic in Colorado in 2013 [60] (CO36 and 37) as well as two isolates recovered from household contacts of this patient (CO58 and 59) [58]. Interestingly, several other isolates obtained from patients during the same outbreak period did not harbor this same mutation. In addition, ten other SNPs outside of the CARDS toxin gene were shared by these four isolates relative to all other isolates in the study (data not shown).

Synonymous and non-synonymous mutation rate analysis

Our preliminary analysis of 1408 SNPs located in the 464 gene core genome showed a bias towards non-synonymous substitutions (61.93%), but this was expected based on the high degree of relatedness of M. pneumoniae genomes [61]. To identify whether these non-synonymous substitutions were accumulating in a manner suggestive of adaptive evolution on some branches, we analyzed gene clusters for evidence of selection (S1 Text). We identified two genes, DNA Ligase (ALA35866.1) and DNA Polymerase III subunit alpha (ALA35550.1), with a higher number of non-synonymous substitutions, potentially suggesting that purifying selection may be relaxed on these genes or that positive selection may be acting upon a subset of residues in these genes.

Population structure analysis and predictive model

Based on the high-degree of shared genomic content between the various types of M. pneumoniae genomes, we evaluated the genetic structure to elucidate novel features that could be used to refine the M. pneumoniae classification schema. Consistent with divergent lineages of type 1 and 2 isolates documented previously [8, 11], the phylogenetic analysis revealed two population substructures that followed the primary P1 type classifications (Fig 1, Fig 2, S4 Fig). Initial hierarchical Bayesian Analysis of Population Structure (hierBAPS) using core sequences from all 107 isolates revealed 3 subgroups corresponding to type 1, type 2, and type 2a (S7 Fig). Considering the larger core consisting of 642 genes, two subtypes corresponding to the larger type 1 and type 2 groups were identified (S7 Fig).

In order to reveal hidden substructures masked by the dominant population structure, we performed hierBAPS on type 1 and 2 datasets separately (Fig 4). For type 2, core nucleotide sequences (n = 642) obtained from closed genomes (n = 34) were used whereas, for type 1, the core genes (n = 464) identified in all isolates (n = 107) were used as none of the type 1N subgroup genomes were completely closed by WGS and thus would not have been represented in the dataset. Using hierBAPS we identified a total of six subpopulation structures, including three groups within type 1 (Fig 4A) and three within type 2 (Fig 4B) corresponding to the groups identified in the phylogenetic analysis. The groups within type 1 each contained the same isolates as the subtypes described in the phylogenetic analysis: type 1U (containing the majority of type 1 isolates), type 1Ref (containing genome representations of M129), and type 1N (containing newly emerging isolates with 3 repeats in the MLVA locus mpn16) (Fig 4A). Similarly, the three groups within type 2 corresponded to the previously identified subtypes, and thus were designated as type 2Ref (containing genome representations of FH), type 2a (containing isolate 309 originally designated as variant type 2a), and type 2v (containing isolates designated by our laboratory as variants of type 2) (Fig 4B).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Hierarchical Bayesian Analysis of Population Structure (hierBAPS) of type 1 and type 2 groups separately.

Three subpopulations were identified in (A) Type 1 and (B) Type 2 genome groups. Green, T; blue, A; red, G; yellow, C.

https://doi.org/10.1371/journal.pone.0174701.g004

In order to identify the SNPs that were used to classify isolates in the hierBAPS utility, we constructed a predictive classification model for whole genome SNP output using the Random Forest supervised machine learning algorithm with the subtypes as an isolate classifier and lists containing at ≥1 SNP as features (S4 and S5 Tables). We performed ten iterations of four-fold cross validation to train and test the model using a different seed for each iteration. The accuracy of the models ranged from 92.5% to 100%. In the represented model, only three misclassifications were observed (training, n = 2; testing, n = 1); each involved the subgroups containing the smallest number of representative isolates, including type 1N and type 2v. Of the 659 lists that contributed to the model (S5 Table), the relative importance of each feature list for classifying the data was determined as indicated by the mean decrease in Gini index value (Fig 5A, S9 Table). The presence/absence of features with the greatest contribution to differentiating the subtype classifications in all iterations of the model were identified (Fig 5B).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. High importance features for classification of M. pneumoniae subtypes based on Random Forest model.

(A) Representative plot of mean decrease in Gini value for top 50 feature lists. Each list consisted of ≥ 1 variant position. (B) Heatmap of presence/absence of the 25 features of highest relative importance for separation of six subtypes resulting from all iterations of the model (n = 40). Other variant sites also contributed to the final model.

https://doi.org/10.1371/journal.pone.0174701.g005