A validated pangenome-scale metabolic model for the Klebsiella pneumoniae species complex

Abstract

The Klebsiella pneumoniae species complex (KpSC) is a major source of nosocomial infections globally with high rates of resistance to antimicrobials. Consequently, there is growing interest in understanding virulence factors and their association with cellular metabolic processes for developing novel anti-KpSC therapeutics. Phenotypic assays have revealed metabolic diversity within the KpSC, but metabolism research has been neglected due to experiments being difficult and cost-intensive. Genome-scale metabolic models (GSMMs) represent a rapid and scalable in silico approach for exploring metabolic diversity, which compile genomic and biochemical data to reconstruct the metabolic network of an organism. Here we use a diverse collection of 507 KpSC isolates, including representatives of globally distributed clinically relevant lineages, to construct the most comprehensive KpSC pan-metabolic model to date, KpSC pan v2. Candidate metabolic reactions were identified using gene orthology to known metabolic genes, prior to manual curation via extensive literature and database searches. The final model comprised a total of 3550 reactions, 2403 genes and can simulate growth on 360 unique substrates. We used KpSC pan v2 as a reference to derive strain-specific GSMMs for all 507 KpSC isolates, and compared these to GSMMs generated using a prior KpSC pan-reference (KpSC pan v1) and two single-strain references. We show that KpSC pan v2 includes a greater proportion of accessory reactions (8.8 %) than KpSC pan v1 (2.5 %). GSMMs derived from KpSC pan v2 also generate more accurate growth predictions, with high median accuracies of 95.4 % (aerobic, n=37 isolates) and 78.8 % (anaerobic, n=36 isolates) for 124 matched carbon substrates. KpSC pan v2 is freely available at https://github.com/kelwyres/KpSC-pan-metabolic-model, representing a valuable resource for the scientific community, both as a source of curated metabolic information and as a reference to derive accurate strain-specific GSMMs. The latter can be used to investigate the relationship between KpSC metabolism and traits of interest, such as reservoirs, epidemiology, drug resistance or virulence, and ultimately to inform novel KpSC control strategies.

Keywords: Klebsiella; bacterial metabolism; genome-scale metabolic models; genomics; pangenome.

Fig. 1.

Genome-scale metabolic model (GSMM) overview. (a) Summary of the typical approach for generating a GSMM. (b) Constituent components of the Biomass Objective Function (BOF), which defines the requirements for production of new daughter cells. Optimization of the BOF via constraint-based analysis (c) can be used to predict growth in different conditions with an optimized objective value ≥0.001 indicating growth whereas an objective value <0.001 indicates no growth. (c) Common applications of GSMMs.

Fig. 2.

Approach for generating KpSC pan v2. Blue boxes indicate major analysis steps and data sources, whereas white boxes indicate additional information. The table adjacent to manual curation is a breakdown of the evidence confidence scores (Confidence) and the number of reactions (#Rxns) in each category [13]. The final set of curated gene and reaction information is available in Table S3 and summarized in Table 1.

Fig. 3.

Growth simulation outcomes for strain-specific genome-scale metabolic models derived from four different references. Simulations represent growth in minimal media supplemented with a single source of carbon, nitrogen, sulphur or phosphorus as indicated. The total number and set of substrates simulated is dependent upon those that can be supported by the reference model (see Table S4; note that substrates supported by the iYL1228 reference are a subset of those supported by KpSC pan v1, which are a subset of those supported by KpSC pan v2. Substrates supported by iKp1289 are an independent and overlapping set). Core growth (blue) refers to substrates for which growth was predicted for ≥95 % of the population, whereas accessory growth (red) refers to substrates for which growth was predicted for <95 % of the population. ‘No growth’ (grey) refers to metabolites that are supported by the model, but do not result in positive growth predictions for any strains. These can occur due to incomplete metabolic pathways in the model (e.g. the carbon utilization pathway is complete but the nitrogen pathway is not, resulting in no growth on nitrogen) or where Klebsiella is legitimately unable to utilize the substrate for growth.

Fig. 4.

Pairwise gene and reaction Jaccard distances among strain-specific genome-scale metabolic models (GSMMs) derived from different reference models. Jaccard distances were calculated based on gene (a) and reaction (b) contents of pairs of strain-specific GSMMs derived from each of the iYL1223, iKp1289, KpSC pan v1 and KpSC pan v2 reference models. A Kruskal–Wallis H-test indicated that the distributions were different (Genes: P<2.2×10⁻¹⁶, d.f.=3; Reactions: P<2.2×10⁻¹⁶, d.f.=3). Pairwise comparisons indicated a significant difference between the distributions for KpSC pan v2 and all others (Mann–Whitney test; Genes and Reactions: P<2.2×10⁻¹⁶, Bonferroni correction threshold=0.0167).

Fig. 5.

Comparable growth prediction accuracies for strain-specific genome-scale metabolic models derived from iKp1289, iYL1228, KpSC pan v1 and KpSC pan v2 references. Accuracies, sensitivities and specificities represent combined values for carbon substrate growth predictions compared to true phenotypes generated on the Biolog platform (plates PM1 and PM2) for 37 KpSC isolates. Aerobic conditions are presented in the first row and anaerobic conditions in the second row (raw values for both conditions are available on Figshare). Left panels comprise data for the set of 89 substrates supported by all four models. Middle panels comprise data for the same 89 substrates plus two additional substrates supported by iYL1228, KpSC pan v1 and KpSC pan v2 (91 substrates in total). Right panels comprise data for the same set of 91 substrates plus an additional 14 substrates supported by only KpSC pan v1 and KpSC pan v2. Table 2 shows metrics calculated for the total number of substrates supported by each reference model.