FastMLST: A Multi-core Tool for Multilocus Sequence Typing of Draft Genome Assemblies

Abstract

Multilocus Sequence Typing (MLST) is a precise microbial typing approach at the intra-species level for epidemiologic and evolutionary purposes. It operates by assigning a sequence type (ST) identifier to each specimen, based on a combination of alleles of multiple housekeeping genes included in a defined scheme. The use of MLST has multiplied due to the availability of large numbers of genomic sequences and epidemiologic data in public repositories. However, data processing speed has become problematic due to the massive size of modern datasets. Here, we present FastMLST, a tool that is designed to perform PubMLST searches using BLASTn and a divide-and-conquer approach that processes each genome assembly in parallel. The output offered by FastMLST includes a table with the ST, allelic profile, and clonal complex or clade (when available), detected for a query, as well as a multi-FASTA file or a series of FASTA files with the concatenated or single allele sequences detected, respectively. FastMLST was validated with 91 different species, with a wide range of guanine-cytosine content (%GC), genome sizes, and fragmentation levels, and a speed test was performed on 3 datasets with varying genome sizes. Compared with other tools such as mlst, CGE/MLST, MLSTar, and PubMLST, FastMLST takes advantage of multiple processors to simultaneously type up to 28 000 genomes in less than 10 minutes, reducing processing times by at least 3-fold with 100% concordance to PubMLST, if contaminated genomes are excluded from the analysis. The source code, installation instructions, and documentation of FastMLST are available at https://github.com/EnzoAndree/FastMLST

Introduction

Multilocus Sequence Typing (MLST) was a revolutionary attempt initially proposed for molecular typing of Neisseria meningitidis ¹ and subsequently replicated for multiple pathogens with global relevance.^2-5 This technique assigns a sequence type (ST) to bacterial isolates based on a combination of alleles from an optimal set of housekeeping genes defined for each species. Even though there are methods with higher discrimination power (cgMLST/wgMLST), MLST offers several advantages, such as (1) a common language for typing, (2) simple implementation, and (3) a vast amount of information collected in dedicated databases throughout history. ⁶

Currently available programs for ST determination from assembled genomes, that is, mlst, ⁷ CGE/MLST, ⁸ MLSTar, ⁹ and the online tool PubMLST, ¹⁰ deliver results in 1 to 9 minutes when dealing with hundreds of genomes. However, as most of them do not (or inefficiently) implement parallel computing, processing of tens of thousands of genomes can be extremely time-consuming. In addition, most current programs do not generate a file with concatenated allele sequences, despite its common requirement in downstream processes, such as evaluation of genetic diversity and phylogenetic analyses. ¹¹

In response to these limitations, we developed FastMLST, a tool focused on parallel computing via a divide-and-conquer approach ¹² that performs BLASTn searches ¹³ against the sequences of MLST schemes deposited in the PubMLST database ¹⁰ using genome assemblies of isolates in (multi-) FASTA format as input. FastMLST delivers 2 main outputs: (1) a table with the allelic profile detected for each genome query and (2) a (multi)-FASTA file with the (concatenated) sequences of the housekeeping genes included in each MLST scheme, suitable for downstream multilocus sequence analyses (MLSAs).

Methods

Implementation

FastMLST is an open-source, stand-alone software implemented in Python 3, that only requires BLASTn, Biopython, ¹⁴ tqdm, ¹⁵ and pandas ¹⁶ as an external program. It is easy to install via the Anaconda Package Manager (https://anaconda.org/bioconda/fastmlst) and is in constant development. A divide-and-conquer app tasks but in High-Performance Computing (HPC) clusters. ¹⁷ Instead, our specific implementation for MLST typing makes it available in HPC clusters, workstations, or personal computers.

Key analysis steps

PubMLST database setup

FastMLST can retrieve and update the PubMLST database ¹⁰ with its “–update-mlst” option. It automatically processes the 153 currently available schemes (September 2021) and prepares the system for subsequent deployment.

Allele searches

Allele searches are locally performed with BLASTn using the public, regularly updated, and centralized database maintained at PubMLST. ¹⁰ FastMLST is based on a divide-and-conquer approach, whereby each genome is processed in parallel and BLASTn runs in only one central processing unit (CPU). This is a major difference from some of the other programs that use BLASTn for MLST typing. In other programs, the genome processing is serial and uses multiple processes to do the BLASTn search in multiple CPUs.

Scheme identification

The MLST scheme to use for each query genome is inferred in FastMLST by an automatic scoring system, similar to previous software. ⁷ This system assigns known, full-length alleles a maximum score of 100 points, 70 points to alleles of the expected length but with Single Nucleotide Polymorphisms (SNPs) (95% identity by default), or 20 points to alleles with unexpected lengths (less than 99% of coverage by default). At the end, the MLST scheme with the highest overall score is selected for further analyses. Alternatively, it is possible to manually choose the scheme to use with the argument “–scheme.” When multiple perfect hits to a single scheme are found, FastMLST reports all hits to alert the user of potential contaminations.

Outputs

FastMLST delivers 2 output types. The first one is a table with the ST, allele classification, and—when available—the clonal complex or clade to which the ST belongs. The second type of output is a FASTA file of the MLST alleles concatenated in alphabetical order, or a series of independent files containing one gene of the scheme each.

Concordance and speed analysis

Concordance analysis

We assessed the capability of FastMLST to correctly assign STs for 33 464 isolates of 91 bacterial species that were selected and downloaded using ncbi-genome-download v0.3.0. ¹⁸ Typing by PubMLST was also done and considered a gold standard. Cohen’s kappa was calculated with the cohen_kappa_score method implemented in scikit-learn V0.24.2. ¹⁹ The code to reproduce this analysis is available in a Jupyter Notebook in the FastMLST-Concordance repository (https://github.com/EnzoAndree/FastMLST-Concordance), along with the intermediate results of FastMLST and PubMLST typing and the genome assembly stats. The quality of the assemblies used was assessed using assembly-stats v1.0.1. ²⁰

Speed analysis

The Unix time program was used to compare the processing speed of FastMLST with that of the other 4 widely used software: mlst v2.11, CGE/MLST v2.0.4, MLSTar v1.0, and the online PubMLST tool. Only the MLSTar run time was measured with the R function proc.time(). The species on which the speed test was done were carefully chosen to evaluate the impact of genome size on speed performance: Burkholderia cepacia, median genome length (MGL) = 8 544 148 bp; n = 186; scheme = bcepacia_complex; Clostridium botulinum (MGL = 3 943 908 bp; n = 250; scheme = cbotulinum); and Mycoplasma pneumoniae (MGL = 816 465 bp; n = 175; scheme = mpneumoniae). The code to reproduce the speed test is available in a Jupyter Notebook in the FastMLST-Concordance repository (https://github.com/EnzoAndree/FastMLST-Concordance).

Results

Compared with PubMLST, FastMLST reached an overall concordance of 99.06% (n = 27 101 of 27 359) (Table 1). The 258 inconsistencies detected were due to incorrect ST assignment by PubMLST. These 258 genomes were flagged as suspected contaminated genomes and were not typed by FastMLST because they had more than one perfect hit for an allele (Supplementary Tables S1-S6). This situation remained unnoticed by PubMLST and could be misleading for users. After removal of these possibly contaminated genomes, the overall concordance of FastMLST increased to 100% (Cohen’s kappa = 1.000).

Table 1.

Concordance analysis between FastMLST and PubMLST using 91 different species.

Species	Scheme	%GC	Median genome length (bp)	Total number of isolates	ST assigned by PubMLST a	ST assigned by FastMLST b	New ST detected by FastMLST	Genomes with new alleles detected by FastMLST	Concordance (%) c	Cohen’s kappa	Possibly contaminated genomes d	Concordance when possibly contaminated genomes are removed (%)	Cohen’s kappa when possibly contaminated genomes are removed
Clostridioides difficile	cdifficile	29	4 147 183	2044	1991	1991	–	1	100.00	1.000	–	100.00	1.000
Enterococcus faecium	efaecium	38	2 923 827	2339	1953	1953	120	113	100.00	1.000	–	100.00	1.000
Enterococcus faecalis	efaecalis	37	2 964 918	1828	1645	1645	62	106	100.00	1.000	–	100.00	1.000
Burkholderia pseudomallei	bpseudomallei	68	7 124 009	1697	1567	1567	96	17	100.00	1.000	–	100.00	1.000
Streptococcus suis	ssuis	41	2 106 098	1539	1359	1359	68	84	100.00	1.000	–	100.00	1.000
Vibrio parahaemolyticus	vparahaemolyticus	45	5 112 626	1574	1254	1254	103	141	100.00	1.000	–	100.00	1.000
Vibrio cholerae	vcholerae	47	4 025 109	1476	1239	1239	65	142	100.00	1.000	–	100.00	1.000
Other 78 species	–	25-67	761 051-8 544 149	13 548	9416	9416	610	2765	100.00	1.000	–	100.00	1.000
Listeria monocytogenes	lmonocytogenes	38	2 969 897	3849	3813	3812	6	8	99.97	1.000	1	100.00	1.000
Campylobacter jejuni	cjejuni	31	1 676 498	1763	1704	1703	6	15	99.94	0.999	1	100.00	1.000
Enterobacter cloacae	ecloacae	55	4 991 140	341	275	272	6	25	98.91	0.989	3	100.00	1.000
Helicobacter cinaedi	hcinaedi	39	2 131 745	41	34	33	–	2	97.06	0.961	1	100.00	1.000
Porphyromonas gingivalis	pgingivalis	48	2 334 097	67	14	13	19	24	92.86	0.915	1	100.00	1.000
Streptococcus agalactiae	sagalactiae	36	2 093 532	1358	1 095	844	2	10	77.08	0.758	251	100.00	1.000
Total				33 464	27 359	27 101	1163	3453	99.06	0.990	258	100.00	1.000

Abbreviations: MLST, Multilocus Sequence Typing; ST, sequence type.

^aA genome was included in this category if an ST was assigned using PubMLST.

^bA genome was included in this category if an ST was assigned using FastMLST and PubMLST.

^cAn ST assignment was considered concordant if the STs assigned by PubMLST and FastMLST were identical.

^dFastMLST found evidence of contamination due to multiple alleles for a single gene.

FastMLST reported 1163 novel STs and 3453 isolates with novel alleles that PubMLST overlooked. The number of downloaded and typed genomes are different and is explained by the fact that some genomes are not typable because of missing alleles due to genome fragmentation. Therefore, ST assignment in these isolates was not possible by either PubMLST or FastMLST. An extended version of Table 1 is shown in Supplementary Table S7.

Our results with real-world data regarding isolates of different %GC, genome lengths, fragmentation levels (Supplementary Table S8), and possibly contamination levels confirm that MLST typing by FastMLST and PubMLST is of comparable quality. Nonetheless, FastMLST provides additional information about possible contaminations and identifies new alleles and STs.

The speed of FastMLST, MLSTar, PubMLST, and mlst (expressed in Genomes/minute) was determined with a workstation equipped with 2 Intel^(R) Xeon^(R) CPU E5-2683 v4 processors @ 2.10 GHz using 1, 2, 4, 8, 16, 32, or 64 cores. Each program was run in triplicate for each test datasets. The PubMLST’s results were obtained remotely because it does not offer a stand-alone version. In all evaluated programs, processing speed and genome size were inversely related (Figure 1A). When a single core was used, mlst v2.11 achieved the best performance for all 3 datasets, with a max speed of approximately 145 genomes/min (a total run time of approximately 1.2 minutes) (Figure 1A). However, mlst v2.11 did not speed up when its multiple processing option was enabled (Figure 1A), as it inefficiently implements parallel computing (Figure 1C). MLSTar can also use multiple processors, but it was at least 3-fold slower than FastMLST when 64 CPUs was used (Figure 1A and B). FastMLST’s processing speed grew exponentially as more cores were deployed (Figure 1A and B), reaching a max speed of 2855 genomes/min using 64 cores (a total run time of approximately 3.7 seconds). This allows the processing of 28 000 genomes in less than 10 minutes.

Figure 1.

Processing speed and simplified comparative-workflow. (A) Processing speed (genomes/minute) of FastMLST and standard software for MLST determination (mlst, PubMLST, CGE/MLST, and MLSTar) was compared by triplicate using 3 Burkholderia cepacia (n = 186), Clostridium botulinum (n = 250), and Mycoplasma pneumoniae (n = 175) datasets and 1, 2, 4, 8, 16, 32, and 64 CPUs. Error bars represent the standard error of 3 replicates. Only mlst, MLSTar, and FastMLST natively support parallel processing. (B) FastMLST and MLSTar workflow. (C) mlst, PubMLST, and CGE/MLST workflow. CPU indicates central processing unit; MLST, Multilocus Sequence Typing.

Although speed will depend on the length of the genome queries, the speed of FastMLST in the worst-case scenario will be at least 14 times higher than the speed offered by mlst (the fastest single-core program). Raw data of run time for each speed test are available in Supplementary Table S8.

Along with the improvement in processing speed, FastMLST allows the user to generate a file with concatenated MLST gene fragments that can be used without further modification in subsequent phylogenetic analyses.

Conclusions

Compared with PubMLST, FastMLST assigns STs to thousands of genomes in minutes with 100% concordance in genomes without suspected contamination in a wide variety of species with different genome lengths, %GC, and assembly fragmentation levels. In addition, FastMLST can generate a multi-FASTA file with concatenated allele sequences suitable for downstream phylogenetic analysis. This addition allows the user to focus on downstream analyses and the biological aspect of the data much sooner. These unique features have the potential to boost future research.