Background

Bacterial genotyping is a powerful tool to investigate the relationships between individual strains from a single species as well as to study the bacterial population structure and dynamics. Phenotypic and genotypic methods can be applied to distinguish bacteria, with genotyping most often being used these days [1].

For a long time, the method considered a ’gold standard’ for bacterial genotyping in routine practice was pulsed-field gel electrophoresis, where DNA banding patterns are analyzed [2]. The main advantage of this method is its discriminatory power and intra-laboratory reproducibility, but on the other hand, it is time and labor intensive [3].

Another typing method used worldwide is multilocus sequence typing (MLST), a standardized and highly discriminatory technique. Several housekeeping genes’ 450 - 500 bp long fragments are sequenced and analyzed. For each allele of a gene, a unique number is assigned [4]. The combination of numbers for all the genes’ alleles define the allelic profiles represented by sequence types (STs). The MLST schemes are deposited in publicly available databases; thus, the method can be used for comparative epidemiological studies and monitoring the spread of high-risk strains. The MLST is a portable, standardized and reproducible method; however, the sequencing cost is still high. [5].

An alternative to MLST is mini-MLST, where the sequencing is replaced by high resolution melting analysis (HRM) [6]. In the first step, housekeeping gene fragments are amplified by PCR. Next, HRM is performed and as a result, a melting curve is obtained. When 50% of the DNA is denatured, the melting temperature is determined [7]. Each melting curve represents an individual melt allele named according to the GC base content in the amplified region. A combination of melt alleles from each gene defines the so-called melt type. The main advantage is the low cost, which is about 10 – 20% of MLST, low time demands and very high throughput [8]. However, there is no standardized approach to detect genetic markers suitable for mini-MLST, and often the sequences chosen for typing do not have sufficient discriminatory power.

The method with the highest discriminatory power for bacterial strain genotyping is whole genome sequencing [9]. Recently, genome sequencing has become more accessible to routine laboratories; however, the main bottlenecks in post-sequencing data analysis remain. The first bottleneck is the lack of standard protocols for data processing. This is because many tools exist, and each uses different data quality assessments, data processing, and results interpretation. The second bottleneck is in genome assembly, where the outputs are crucial for clinical practice. Two main approaches exist here: de novo assembly and reference-based assembly.

de novo assembly is time-consuming and computationally demanding, high-quality data are also required. High-throughput technologies provide short reads where the assembly is challenging, especially in repetitive regions [10]. As a result, the whole genome is not obtained, but only a large number of short contigs are generated [11]. These drawbacks make it difficult to use de novo assembly in routine clinical practice.

On the other hand, reference-based assembly is faster and less computationally demanding; nevertheless, the choice of an inappropriate reference sequence significantly affects the analyses and final results. In the reference assembly, only the shared genome parts are analyzed; however, the unmapped reads can contain important information, as the bacterial genomes are highly variable [12].

Hybrid assembly can be conducted to overcome de novo and reference-based assembly’s drawbacks. This approach combines two sequencing technologies and can be used to analyze unique genome parts [13]. However, this double-sequencing approach is not used as standard in clinical studies as it requires two sequencing platforms, which means extra cost and time.

Here we present a new approach based on reference-based mapping and de novo assembly, which can be compared to a hybrid assembly approach. The main difference is that data were obtained from only one sequencing platform, specifically Illumina Miseq, one of the most frequently used platforms in clinical practice [14]. Our goal is not to obtain a complete whole genome but only the most variable genomic regions that can be used in mini-MLST. The reference-based assembly is used as a filter to remove the low variable reads, which will map to the reference sequence. From the high variability reads that did not map to the reference, scaffolds are assembled. In the scaffolds, the most variable parts are identified, and these genomic regions can be used to distinguish bacteria in mini-MLST. Thus, more sequencing will not be necessary as further samples of the given bacterium will be classified based on mini-MLST analysis using the identified variable fragments.

In the present study, our goal was to analyze whole genome sequencing data obtained from isolates representing two clinically important bacterial species - Escherichia coli and Klebsiella pneumoniae. E. coli is a Gram-negative bacterium of the Enterobacteriaceae family [15], and includes pathogenic and commensal clones. Commensal clones are natural inhabitants of the human gastrointestinal tract and cause diseases only in immunocompromised patients or those with breached gastrointestinal barriers [16], [17]. Pathogenic strains can cause urinary tract infections, sepsis, or enteric diseases [15]. E. coli genome size varies from 4.2 to 6.0 Mb with an average of about 5 Mb [18]. K. pneumoniae is another bacterium from the Enterobacteriaceae family. It is an opportunistic pathogen that causes serious diseases such as pneumonia, bloodstream infections, urinary tract infections or sepsis, mainly in immunocompromised patients [19, 20]. In recent years, the number of antibiotic-resistant strains has increased [21]; thus, K. pneumoniae has become one of the major threats due to significant morbidity and mortality [8]. The genome has approximately 5.5 Mbp, and encodes about 5500 genes [22].

Results and discussion

Sequence type determination

The reference-based assembled consensus sequences were used for in silico MLST analysis. For E. coli the Warwick MLST scheme was employed; thus, seven housekeeping genes (adk, fumC, gyrB, icd, mdh, puA, recA) were analyzed. In total 11 sequence types were present in our dataset (1 x ST 69, 4 x ST 131, 1 x ST 95, 2 x ST 404, 2 x ST 38, 2 x ST 1049, 4 x ST 58, 1 x ST 297, 1 x ST 517, 2 x ST 101, 1 x ST UNW). The complete results of MLST typing are published in [23].

The sequence types of K. pneumoniae isolates were determined using seven housekeeping genes (gapA, infB, mdh, pgi, phoE, rpoB, tonB). Overall four STs were identified in the dataset (5 x ST 45, 9 x ST 405, 13 x ST 551, 4 x ST 950). The results of the MLST analysis are attached in Additional file 1: Table S1.

Assembly analysis

The number of obtained reads for E. coli genomes were from 2 514 472 to 5 137 462, and for K. pneumoniae, the number varied from 1 608 984 to 4 001 068. In both datasets, less than 0.14% of reads mapped to the human genome; thus, no serious contamination occurred in the sequencing data. Trimmomatic removed about 5% of E. coli reads and approximately 13% of K. pneumoniae reads, and the remaining 95% and 87% from the total number of reads were used for further assembly. The number of reads mapped to the reference sequence varied from 64% to 81% for E. coli genomes and from 54% to 79% for K. pneumoniae genomes. Thus, around 17% of reads remained unmapped, on average. Complete information on the numbers of total, mapped, and unmapped reads is shown in the Tables 1 and 2.

Table 1 The numbers of all reads obtained from sequencing and number of mapped and unmapped reads to reference sequence for each E. coli genome
Full size table
Table 2 The numbers of all reads obtained from sequencing and number of mapped and unmapped reads to reference sequence for each K. pneumoniae genome
Full size table

The unmapped reads of the 21 E. coli genomes and 31 K. pneumoniae genomes were de novo assembled via SPAdes. The number of scaffolds obtained for E. coli isolates varied from 151 to 296, and after removing the scaffolds shorter than 500 bp, the amount fluctuated from 124 to 230. In K. pneumoniae genomes were assembled from 86 to 323 scaffolds, and after removing short scaffolds, the number varied from 76 to 192. The largest scaffold lengths, N50 and L50, were determined and can be found in Tables 3 and 4 with other statistics. The detailed assembly process is described in the sections Reference-based assembly and De novo assembly of unmapped reads.

Table 3 Results of de novo analysis for E. coli genomes - number of assembled scaffolds, number of assembled scaffolds longer than 500 bp, length of the largest scaffold, N50 and L50 values
Full size table
Table 4 Results of de novo analysis for K. pneumoniae genomes - number of assembled scaffolds, number of assembled scaffolds longer than 500 bp, length of the largest scaffold, N50 and L50 values
Full size table

Variable fragments identification

The scaffolds from genome EC162 and KP1241 were searched for in the remaining E. coli and K. pneumoniae genomes via BLAST+. These genomes were chosen as they contained the largest number of scaffolds. In total, from 230 scaffolds searched, only 25 scaffolds were located in all E. coli isolates and from 192 scaffolds, 78 were identified in all K. pneumoniae genomes.

The isolate sequences from corresponding scaffolds present in all genomes were aligned. Next, the most variable regions were located in the alignments, and in total, 11 variable fragments for E. coli and 244 fragments for K. pneumoniae were identified.

A phylogenetic tree was constructed for each variable fragment, and the number of clusters was determined (see Additional file 1: Table S2 and S3). From the obtained results, only one E. coli variable fragment (labelled as EC_01) was further analyzed as only this one could distinguish the isolates to 11 clusters, according to the MLST results. For K. pneumoniae, five variable fragments (labelled as KP_01, KP_02, KP_03, KP_04 and KP_05) classified the genomes correctly into four clusters according to their STs and were further analyzed. The analyzed variable fragments are shown in Table 5. The variable fragments’ identification is described in the section Detection of variable regions.

Table 5 The identified variable fragments located in assembled E. coli and K. pneumoniae genomes scaffolds with their size and number of clusters obtained by phylogenetic analysis
Full size table

Variable fragment analyses

The variability calculation was used for the preliminary selection of sequences with higher variability. Before further analysis, Web BLAST was used to analyze the fragments and ensure that the sequences were from the bacterial chromosome and not from plasmids. In total, six variable fragments from both datasets were analyzed, and one K. pneumoniae fragment (KP_01) with a length of 1346 bp was removed as it contained only a plasmid sequence.

If the lengths of the identified fragments were long (more than 150 bp [7]), the sequences’ analysis via commonly used laboratory methods such as mini-MLST would be complicated. For this reason, the fragments can be shortened to identify the region with the highest variability rate. The process of shortening is described in the Fragment shortening section.

The original length of the variable fragment (EC_01) with sufficient discriminatory power to distinguish E. coli genomes was 4,762 bp. After fragment erosion and decomposition, three fragments with a length of 120 bp were obtained (EC_01_1, EC_01_2, EC_01_3). The phylogenetic tree, constructed based on the three fragments with the highest variability, is depicted in Fig. 1.

Fig. 1
figure 1

Cladogram of 21 E. coli isolates based on three variable fragments (EC_01_1, EC_01_2, EC_01_3) with highlight clusters obtained from MLST analysis created by Evolview [38]

Full size image

For K. pneumoniae genomes, the identified fragments lengths were 12 bp (KP_02), 13 bp (KP_03), 21 bp (KP_04) and 26 bp (KP_05); thus, no erosion was needed. Again, the phylogenetic trees were constructed, and one of them created based on variable fragment KP_02 is depicted in Fig. 2. The other trees are shown in Additional file 1: Fig. S1 A-C.

Fig. 2
figure 2

Cladogram of 31 K. pneumoniae isolates based on the variable fragment (KP_02) with highlight clusters obtained from MLST analysis created by Evolview [38]

Full size image

All trees were constructed according to the process described in the section Phylogenetic analysis.

Melting temperatures analysis

For E. coli, the melting temperatures for three variable fragments (EC_01_1, EC_01_2, EC_01_3) for each genome were calculated based on the nearest neighbor method [24] and are shown in the Table 6. A melting temperature cluster analysis was conducted, and the obtained dendrogram is depicted in Fig. 3.

Fig. 3
figure 3

Dendrogram of 21 E. coli isolates obtained from cluster analysis of variable fragments (EC_01_1, EC_01_2, EC_01_3) melting temperatures with highlight clusters obtained from MLST analysis created

Full size image
Table 6 Calculated melting temperatures for each E. coli genome for three variable fragments
Full size table

The melting temperatures were also calculated for variable K. pneumoniae isolate fragments, and cluster analysis was conducted. It was found that three (KP_02, KP_03, KP_05) out of four identified fragments could distinguish the genomes based on the calculated melting temperatures. The calculated values are shown in Table 7, and one obtained dendrogram based on fragment KP_02 is shown in Fig. 4. See other dendrograms in Additional file 1: Fig. S2 A-B.

Fig. 4
figure 4

Dendrogram of 31 K. pneumoniae isolates obtained from cluster analysis of variable fragment (KP_02) melting temperatures with highlight clusters obtained from MLST analysis created

Full size image
Table 7 Calculated melting temperatures for each K. pneumoniae genome for three variable fragments
Full size table

The way the melting temperatures were determined is described in the Melting temperature calculation chapter.

The melting temperature calculation can be used to distinguish E. coli and K. pneumoniae strains. Thus, the fragments can be analyzed in mini-MLST. However, the difference in calculated temperature from different sequence types fluctuated only in the range of one degree is some cases. Therefore, the mini-MLST parameters, such as salt concentration, must be carefully determined.

Conclusions

Bacterial genotyping is an essential process in epidemiology as it helps to find an infection’s source and monitor outbreaks. Results delivery should be done in the shortest possible time. However, typing methods are often laborious and computationally demanding, and financial costs are also high, especially for local clinical laboratories. The solution to the mentioned problems can be using mini-MLST, a cost-effective and efficient laboratory method.

The pipeline to identify variable sequences in the next-generation sequencing data that can be used in mini-MLST for bacterial typing was proposed and tested on 21 E. coli and 31 K. pneumoniae genomes. The hybrid assembly approach consists of reference-based mapping, and de novo assembly of unmapped reads was used. The preliminary location for variable fragments in the assembled scaffolds was carried out using variability rate calculation. In the selected fragments, the most variable parts were identified.

The melting temperatures were calculated to verify that the variable segments can be used in mini-MLST. The calculated melting temperature cluster analysis showed that distinguishing of individual strains is possible. Also, in contrast with MLST, mini-MLST does not use sequencing; thus, the bacterial genotyping cost will be significantly lower.

The proposed approach can be used to identify genomic regions that are not presented in the chosen reference sequence and can be specific to analyzed bacterial strains. Nevertheless, analyzing unmapped reads must be done carefully as the assembled sequences can contain parts of plasmids extracted with the genomic DNA.

One new variable fragment was located in E. coli isolates, and three variable fragments were identified in K. pneumoniae genomes. The identified fragments’ discriminatory power was the same as the seven housekeeping genes used in the MLST analysis. The most variable fragment regions were identified to ensure that it will be possible to perform mini-MLST. As only three regions located in the variable fragment in E. coli and one region in K. pneumoniae can be analyzed in mini-MLST instead of sequencing seven housekeeping genes, the analyzing time and cost will be significantly lower.

Material and methods

Dataset

The 21 E. coli and 31 K. pneumoniae isolates were collected in the Internal Hematology and Oncology Clinic at the University Hospital of Brno between 5/2019 and 7/2019. KAPA HyperPlus Kits (Roche, Switzerland) were used for sequencing library preparation, and a 2100 Bioanalyzer (Agilent Technologies, USA) was employed as a quality check. The prepared sequencing libraries were quantified with a KAPA Library Quantification kit (Roche, Switzerland) and the sequencing process using a MiSeq Reagent Kit v2 (500-cycles) was performed on an Illumina MiSeq platform. As a result, paired-end reads about 250 bp long were acquired.

Reference-based assembly

Before genome assembly, the sequenced data quality was checked by FastQC (v0.11.5, [25]) combined with MultiQC (v1.7, [26]). BBMap (v38.71, [27]) software was used to map reads to the human genome (GRCh38.p13) to remove possible contamination. Then the Trimmomatic (v0.36, [28]) was employed for adapters and low quality read trimming. Reference-based mapping was applied, and NC_002695.2 [29] and NC_012731.1 [30] obtained from the RefSeq database [31] were chosen as the reference genomes for E. coli and K. pneumoniae assembly. The assembly was performed via Burrows-Wheeler Aligner MEM (v0.7.17-r1188, [32]).

De novo assembly of unmapped reads

The reads that did not map to the reference genome were extracted, and the PCR and optical duplicated reads and low-quality reads were removed using Samtools (v1.9, [33]). In the next step, the St. Petersburg genome assembler (SPAdes) (v3.11.1, [34]) was employed for the unmapped reads’ de novo assembly. The assembly was run together with MismatchCorrector, and the Phred quality offset for the input reads was set to 33.

From genome assembly, the scaffolds were further analyzed. Due to a large number of short scaffolds, only those with a length greater than 500 bp were examined.

Detection of variable regions

The variable fragments should be present in all isolates of the analyzed bacterium. For that reason, the scaffolds from one genome of E. coli (EC162) and one genome of K. pneumoniae (KP1241) were searched for in remaining genomes via BLAST+ (v2.6.0+, [35]).

The scaffolds that corresponded to the same region and were found in all genomes were aligned. From the alignment, the parts present in all genomes were analyzed. The variability rate was determined for each fragment of alignments. The variability V was calculated as

$$ V = \frac{n_{vp}}{n_{p}}\cdot 100, $$
(1)

where nvp is the number of variable positions in the alignment, and np is the number of all positions in the alignment. If the variability of an examined fragment is more than 10%, the fragment is further analyzed. The proposed process is shown in Fig. 5.

Fig. 5
figure 5

Scheme of variability regions’ calculation for a set of aligned sequences

Full size image

Phylogenetic analysis

The evolutionary distances based on the Kimura [36] model were calculated for the segments of alignment with high variability. Using the distances, phylogenetic trees were constructed via UPGMA. Then a cluster analysis was conducted, and as a result, the number of clusters was calculated and compared with the number of sequence types obtained from the MLST analysis.

Fragment shortening

If the variable fragment was too long to use in standard laboratory methods, it was shortened by two methods: erosion and dyadic decomposition. Firstly, erosion was performed. The nucleotides from the beginning of the sequence alignment were removed one by one. In each iteration, the phylogenetic tree was constructed, and controlling the number of clusters was performed to ensure that the discriminatory power was still preserved. As soon as the number of clusters decreased, the erosion was stopped. The same process was done from the end of the alignment. Secondly, fragment decomposition was carried out. As it was not possible to shorten the alignment using erosion, the fragment was split into two separate, equal halves. The first half of the alignment was shortened from the end as it was not possible to shorten it further from the beginning. For the second half of the alignment, the nucleotides were removed from the beginning. After each time it was shortened, the number of clusters was determined. If it was impossible to shorten the fragment further, each part was again split in the middle, and the process was repeated. Shortening was stopped when the length of the fragment was 120 bp.

Melting temperature calculation

The melting temperatures were calculated for the variable fragments using the Oligo Calc [37]. The calculation parameters were left by default, and the melting temperatures computed using the nearest neighbor method were used for further analysis.