Generation and Validation of the iKp1289 Metabolic Model for Klebsiella pneumoniae KPPR1

Abstract

Klebsiella pneumoniae has a reputation for causing a wide range of infectious conditions, with numerous highly virulent and antibiotic-resistant strains. Metabolic models have the potential to provide insights into the growth behavior, nutrient requirements, essential genes, and candidate drug targets in these strains. Here we develop a metabolic model for KPPR1, a highly virulent strain of K. pneumoniae. We apply a combination of Biolog phenotype data and fitness data to validate and refine our KPPR1 model. The final model displays a predictive accuracy of 75% in identifying potential carbon and nitrogen sources for K. pneumoniae and of 99% in predicting nonessential genes in rich media. We demonstrate how this model is useful in studying the differences in the metabolic capabilities of the low-virulence MGH 78578 strain and the highly virulent KPPR1 strain. For example, we demonstrate that these strains differ in carbohydrate metabolism, including the ability to metabolize dulcitol as a primary carbon source. Our model makes numerous other predictions for follow-up verification and analysis.

Klebsiella pneumoniae has a well-deserved reputation for causing pneumonia, bloodstream infections, surgical-site infections, and urinary tract infections in debilitated individuals in hospitals and long-term care facilities [1–5]. Pneumonia caused by K. pneumoniae may be particularly severe and is associated with mortality rates of up to 23% [6]. In a prevalence study of intensive care units in 75 countries on 5 continents, K. pneumoniae was found to be the fourth most common pathogen [7]. Management of K. pneumoniae infections presents a serious therapeutic challenge owing to the increasing incidence of antibiotic resistance. Starting in the 1980s, K. pneumoniae isolates harboring extended-spectrum β-lactamases (ESBLs) active against nearly all β-lactams except carbapenems were increasingly reported [8–10]. ESBLs are now produced by 8%–44% of all K. pneumoniae clinical isolates [11]. There was therefore much concern when carbapenem-resistant Enterobacteriaceae (CRE) strains of K. pneumoniae were identified [12]. CRE strains have now spread across the globe [13]. The incidence of carbapenem resistance in K. pneumoniae has risen in the United States, from 1.6% in 2001 to 10.4% in 2011 [14], and is predicted to continue to rise [15]. Because patients infected with these strains frequently receive inadequate empirical and definitive antibiotic therapy, they experience mortality rates of 23%–75% [16]. For these reasons, K. pneumoniae has been identified by the Infectious Disease Society of America as an ESKAPE organism [17], one of 6 bacteria pathogens most in need of novel therapies, and CRE strains have received the Centers for Disease Control and Prevention's highest threat rating [18].

One approach to developing novel therapies for K. pneumoniae is to target metabolic pathways required for growth. Advances in systems biology approaches that use genetic, genomic, and metabolomics data now allow for the detailed modeling of bacterial metabolism [19]. Metabolic modeling has the potential to unmask the complex enzymatic reaction networks used by bacteria under different growth conditions, including those encountered in the mammalian hosts. By defining key metabolic pathways essential for growth under specific conditions, modeling can uncover the strategies used by bacteria to survive and grow under these conditions. Data generated through global approaches such as transposon-insertion sequencing screens, transcriptomics, proteomics, and phenotypic screening can be used to test, inform, and refine these models. In turn, accurate models allow identification of bacterial vulnerabilities in the form of nonobvious metabolic genes and combinations thereof that are critical for growth under the specific conditions of infection. These types of genes are proven candidates for drug targets, as evidenced by the success of trimethoprim-sulfamethoxazole, a combination of antibiotics that inhibit synthesis of tetrahydrofolic acid [20, 21]. Additional targets are currently being sought, and preclinical studies are ongoing for several inhibitors of metabolic pathways [22, 23]. For example, inhibitors of enzymes involved in biosynthesis of branched-chain amino acids have been shown to slow the growth of Salmonella enterica serovar Typhimurium [24]. Likewise, inhibitors of the glyoxylate shunt and tryptophan biosynthesis have activity against Pseudomonas aeruginosa [23] and Mycobacterium tuberculosis, respectively [25]. Targeting key metabolic enzymes has the potential to lead to sorely needed new antibiotics active against K. pneumoniae, but a more thorough understanding of this bacterium's metabolism is a prerequisite.

A previous study developed the iYL1228 metabolic model for the K. pneumoniae strain MGH 78578 [26], but many molecular genetic studies use KPPR1 [29–31]. KPPR1 (a rifampin-resistant derivative of ATCC 43816) is a human K. pneumoniae isolate with high levels of virulence in several models of infection [11, 12]. In this study, we developed a metabolic model for KPPR1 by translating and expanding the iYL1228 model. We converted the gene associations in the iYL1228 model to KPPR1 on the basis of sequence homology and added reactions for metabolic genes in KPPR1 that lack orthologs in MGH 78578. We validated our new KPPR1 model by using a combination of high-throughput growth data and published transposon insertion sequencing (INSeq) results [31]. Finally, we demonstrated the value of our model in predicting variations between KPPR1 and MGH 78578 in the use of a subset of nutrient sources. Specific predictions of the model were tested by growing KPPR1 in defined media.

METHODS

Model Generation and Gap Filling

The KPPR1 model was generated using the Propagate Model to New Genome application [32] in the Narrative interface of the Department of Energy Systems Biology Knowledgebase (KBase). In short, the model propagation process involves translating the genes in the source model into genes in the target genome on the basis of the closest predicted ortholog. Further details are in the Supplementary Materials.

The model was gap filled in glucose minimal medium, as K. pneumoniae strain KPPR1 is known to grow in this condition. The gap-filling process involves the identification of a minimal set of reactions that must be added to the model to enable it to produce biomass in a specific growth condition. The challenge in gap filling is that we lack evidence for the addition of these reactions, and often multiple alternative sets of reactions can be added to permit growth. We rely on the assumption that the simplest solution involving the fewest gap-filled reactions is most likely to be the best. Further details are in the Supplementary Materials.

Testing Growth Conditions in KPPR1 and MGH 78578

K. pneumoniae strain MGH 78578 (ATCC 700721) was obtained from the American Type Culture Collection; K. pneumoniae strain KPPR1 [29] and Escherichia coli strain MG1655 are from laboratory stocks. Biolog and M9 medium experiments were performed as described in the Supplementary Materials.

RESULTS

Comparison of KPPR1 and MGH 78578 Genomes

To construct a genome-scale metabolic model of K. pneumoniae strain KPPR1, we started with a model for the similar MGH 78578 strain, called iYL1228 [26]. To facilitate the translation of the iYL1228 model to KPPR1, we conducted a detailed comparison of the gene content in KPPR1 [30] and MGH 78578 (GenBank accession number CP000647). Of the 5184 genes composing the MGH 78578 genome, 4540 (88%) have close homologs (e value, <1 × 10⁻¹⁰) in the KPPR1 strain. Almost all of these were orthologs and will therefore be called “orthologs” for the remainder of this report. Similarly, 4528 of the 5081 genes (89%) in KPPR1 have orthologs in MGH 78578. There is also significant conservation of chromosomal synteny between the 2 strains [33]. Diving more deeply into the variation in these genomes (Table 1), we find that many of the genes that lack orthologs are redundant copies of other genes that have orthologs, meaning that the absence of these genes appears to have no impact on functional capacity at this level of resolution. Additionally, many of the distinctive genes that are present in MGH 78578 are located on 5 plasmids, which together encode 409 genes. The KPPR1 genome, in contrast, has no plasmids [30]. KPPR1 and MGH 78578 each have >500 unique genes, annotated with 332 and 360 distinct functions, respectively (Table 1). However, many of these distinct functions appear to be poorly characterized, exemplified by the fact that only approximately 21% of these functions appear in curated subsystems in the SEED (a database of genomic-based metabolic models [34]), as opposed to an average of >50% of the functions shared by both genomes occurring in subsystems. The unique functions in KPPR1 and MGH 78578 that do occur in subsystems fall into a broad set of functional categories (Figure 1), although most of these functions relate to either virulence or carbohydrate use. Only 18 and 16 of the unique functions in KPPR1 and MGH 78578, respectively, are metabolic (Table 1). These results collectively indicate that these 2 strains are similar metabolically, meaning that a direct translation of the iYL1228 model of MGH 78578 to KPPR1 should yield an accurate model.

Table 1.

Functional Roles and Families in KPPR1 and MGH 78578

Variable	Distinct Coding Sequences	Distinct Functions/Families	In SEED Subsystems	Metabolic Functions
Shared	4528	3710	1924 (52)	922 (25)
KPPR1 only	553 (10.9)	332 (8.2)	71 (21)	18 (5)
MGH 78578 only	656 (12.7)	360 (8.8)	83 (23)	16 (4)

Data are no. or no. (%) of genes or families.

Figure 1.

Functional categories of genes unique to MGH 78578 or KPPR1. The unique functions assigned by RAST to the genes that are unique to either MGH 78578 or KPPR1 are distributed among 21 distinct SEED subsystem classes. Most distinct genes fall in virulence and carbohydrate metabolism.

Metabolic Model Reconstruction for Strain KPPR1

With our genome comparison complete, we translated the previously published iYL1228 model to the KPPR1 strain (see Methods). The resulting model, which we call iKp1289, included 2145 reactions and 1289 genes (Supplementary Tables 1 and 2). Five reactions in the original iYL1228 model were associated with 14 genes in the MGH 78578 genome that had no orthologs in the KPPR1 genome. Thus, these reactions were not included in our KPPR1 model. These reactions are primarily related to carbohydrate metabolism, including 1 key step in the use of dulcitol (also called galactitol). In contrast, 178 reactions in our KPPR1 model were associated with 75 genes in KPPR1 that lacked orthologs in the MGH 78578 genome. These genes and reactions primarily relate to carbohydrate metabolism, cell wall biosynthesis, and transmembrane transport.

Our entire genome comparison and model translation analysis was conducted using the DOE Knowledgebase (KBase) Narrative interface. This analysis can be viewed in the K. pneumoniae KPPR1 model reconstruction narrative [33].

Variation in Growth Phenotypes of KPPR1 and MGH 78578

To validate our new iKp1289 model (and specifically its differences from the original iYL1228 model), we ran a Biolog Phenotypic Microarray assay on KPPR1, which we compared to a previously published assay for MGH 78578 [26]. For our analysis, we focused on 163 carbon sources and 85 nitrogen sources. Both MGH 78578 and KPPR1 were able to use 60 of the carbon sources and 42 of the nitrogen sources, while neither MGH 78578 nor KPPR1 could use 66 of the carbon sources and 28 of the nitrogen sources. There were 7 carbon and 6 nitrogen sources that only MGH 78578 could use, and there were 30 carbon and 9 nitrogen sources that only KPPR1 could use (Table 2 and Supplementary Tables 3 and 4). We used the iKp1289 and iYL1228 models to simulate growth on all of these conditions. In these simulations, the iYL1228 model had an overall accuracy of 71%, while our iKp1289 model had an accuracy of 75%. We then applied a gap-filling algorithm (see Methods) to add reactions to the models as needed to correct false-negative growth predictions for carbon/nitrogen use. This gap filling successfully corrected 10 and 15 false-negative predictions in MGH 78578 and KPPR1, respectively, while simultaneously introducing 3 and 4 new false-positive predictions. This gap filling added 29 and 18 reactions to the iYL1228 and iKp1289 models, respectively, filling gaps in transporters, sugar derivative use, dipeptide use, and fatty acid use pathways (Supplementary Table 5). After this gap-filling process, the overall accuracy for each model improved to 74% and 79% for iYL1228 and iKp1289, respectively. Overall, both models were effective at predicting the use of carbon and nitrogen sources from the Biolog assay. Despite >52 conditions in which our Biolog analysis revealed differences in growth between MGH 78578 and KPPR1, our initial models correctly predicted only 1 condition as different (dulcitol use), and our gap-filled models correctly predicted only 4 conditions as different.

Table 2.

Accuracy in Predicting Carbon and Nitrogen Use With Models

Variable	Model, Proportion (%) of Sources
	iYL1228	iKp1289	Gap-Filled iYL1228	Gap-Filled iKp1289
Carbon sources both MGH 78578 and KPPR1 can use	49/60 (82)	49/60 (82)	49/60 (82)	50/60 (83)
Carbon sources neither MGH 78578 nor KPPR1 can use	58/66 (88)	57/66 (86)	56/66 (85)	53/66 (80)
Nitrogen sources both MGH 78578 and KPPR1 can use	24/42 (57)	24/42 (57)	34/42 (81)	34/42 (81)
Nitrogen sources neither MGH 78578 nor KPPR1 can use	24/28 (86)	24/28 (86)	24/28 (86)	24/28 (86)
Carbon sources only MGH 78578 can use	1/7 (14)	7/7 (100)	1/7 (14)	7/7 (100)
Nitrogen sources only MGH 78578 can use	3/6 (50)	3/6 (50)	3/6 (50)	3/6 (50)
Carbon sources only KPPR1 can usea	13/30 (43)	17/30 (57)	12/30 (40)	20/30 (67)
Nitrogen sources only KPPR1 can use	5/9 (56)	4/9 (44)	5/9 (56)	5/9 (56)
Overall accuracy	177/248 (71)	185/248 (75)	184/248 (74)	196/248 (79)

^a Subsequent growth assays on select carbon sources are listed in Supplementary Table 6.

For several growth conditions (ie, carbon sources), the Biolog results showed differences in growth between KPPR1 and MGH 78578 even though the models predicted that metabolic pathways for using these compounds were present in both strains. To further evaluate these findings, we conducted several growth assays in supplemented M9 medium and confirmed that both strains exhibit growth on M9-proline, M9-aspartate, M9-arginine, M9-asparagine, and M9-succinate, despite Biolog data for MGH 78578 that suggested an absence of growth on these carbon sources (Supplementary Table 6). The model did, however, predict a bona fide difference in growth on dulcitol as a carbon source, and our targeted assay confirmed that only MGH 78578 grows on M9-dulcitol. These results reveal the insights that can be gained from model-based analyses, which can identify inconsistent experimental data that requires further confirmation. This is particularly important when dealing with high-throughput experimental assays, which can be prone to error [35]. We note that Biolog phenotype arrays in particular can produce a small fraction of false-negative results for slow-growing strains, as was the case in this study. This analysis was conducted using the KBase Narrative interface and can be viewed in the Klebsiella KPPR1 model validation narrative [36].

Validating Essential Gene Predictions for KPPR1

We also wanted to test the capacity of our iKp1289 model for predicting essential metabolic genes. Here, we define “essential” as genes required for growth in rich (LB) medium. To do this, we used our model to simulate KPPR1 growth on LB medium with each individual gene knocked out, resulting in 1289 distinct simulations. In 57 simulations, the model failed to produce biomass, resulting in the prediction of 57 essential metabolic genes (Supplementary Table 7) and 1232 nonessential metabolic genes. To validate these predictions, we compared our results to those from a recent INSeq study of KPPR1 during growth in LB medium (Supplementary Table 8) [31]. In the INSeq analysis, 3880 genes in KPPR1 were observed to have at least 1 transposon insertion, meaning that these genes are likely nonessential for growth in LB medium. Of these 3880 likely nonessential genes, 961 were included in our iKp1289 model, and 953 (99%) were predicted by our model to be nonessential. This confirms the capacity of our model to identify nonessential genes. Since INSeq does not distinguish between essential genes and unsampled genes, we were not able to do a similar comparison of genes found to be essential by both the model and the INSeq study.

The 9 genes that were incorrectly predicted to be essential were associated with reactions in cell wall biosynthesis, amino acid biosynthesis, and transmembrane transport pathways. Our model will incorrectly predict the essentiality of biosynthesis pathways when transporters are missing from the model, when some components of the cell wall are nonessential, or when the representation of LB medium used by our model lacks some essential biomass precursors that may be used by KPPR1. Essentiality prediction errors can also occur when the model fails to account for the capacity of an alternative enzyme to take over the metabolic activity of a knocked out gene.

The 49 remaining unconfirmed essential gene predictions are associated with reactions involved in cell wall biosynthesis, fatty acid and lipid biosynthesis, and amino acid biosynthesis pathways. All of these pathways lead to essential cofactors and structural components in the cell, making them logical predictions for essential metabolic functions.

Our model essentiality predictions were generated using the KBase Narrative interface. The analysis can be viewed in the Klebsiella KPPR1 model validation narrative [36]. It is important to note that our model can be used to predict essential genes in media conditions in addition to LB, and the KBase Narrative shared with this article offers this capability. This can be used in the future to identify genes that are essential for growth under other conditions, including conditions that mimic the host environment.

DISCUSSION

We have constructed and validated a metabolic model for K. pneumoniae strain KPPR1, called iKp1289, that has the precision necessary to be of practical use in identifying novel antibiotic strategies. Our new model relied on (1) the KPPR1 genome sequence, (2) a model of the closely related K. pneumoniae strain MGH 78578, (3) biochemical growth experiments, and (4) INSeq essentiality experiments. Our validation with growth experiments revealed a predictive accuracy of 75%, while the model predicted nonessential genes identified from INSeq experiments with 99% accuracy. The relatively high accuracy was achieved through a combination of starting with a previously curated model (rather than building a draft from the sequence only) and the optimization of the model to fit growth data from diverse conditions. We performed all analysis in the KBase platform, sharing all of our data and work flows in an integrated manner through the narrative interface (available at: http://narrative.kbase.us). Any researcher may view, copy, and edit our narratives, facilitating the review, reuse, analysis, and extension of our data and model.

Many high-throughput studies are often accomplished at the expense of the precision of traditional assays [35]. This is true for genomic annotation of genes, Biolog-type growth assays, and INSeq functional genomics approaches. Using the metabolic modeling framework, we were able to identify inconsistencies and conflicts that existed between these data sets (eg, Biolog data indicating no growth when all necessary pathways were fully annotated). Once these inconsistencies were identified, targeted follow-up experiments could be performed to resolve each conflict. Our study underscores the need for modeling in the era of high throughput biology to provide for quality control and reconciliation.

This approach does have limitations, which should be considered. The predictive power of metabolic models is confined to metabolic capabilities. They cannot capture or forward-predict bacterial growth inhibition involving factors such as ribosomes, DNA polymerases, and transcriptional regulation. For example, the 57 essential genes predicted by our model only include metabolic genes and will not include genes that one could predict from orthogonal approaches (eg, projection of homologs of well-curated essential genes from related organisms). Additionally, the use of gap filling to correct conditions that were incorrectly predicted as no growth resulted in new (although fewer) incorrect growth predictions; this demonstrates the balancing act associated with model refinement and the risk for overfitting.

K. pneumoniae is a remarkably fit organism, in that it has an extremely rapid rate of growth both in vitro and during respiratory infection. Indeed, the ability of K. pneumoniae to use multiple carbon sources and other host-derived nutrients suggest that this pathogen has become highly adapted to a variety of environments. In terms of infection, this is reflected by the number of diseases caused by K. pneumoniae and the wide variety of host tissues in which K. pneumoniae can infect and proliferate. While many studies have used KPPR1 in infection models, our data suggest that there may be differences between isolates in their fitness within the host, owing to their different functional metabolic pathways. For instance, genomic comparisons suggested several interesting metabolic differences between the highly virulent KPPR1 strain and the MGH 78578 strain. KPPR1 encodes gene products that may (1) enable catabolism of a larger range of carbohydrates (and in a more flexible manner), (2) uptake nutrients more robustly, and (3) invoke more-extensive stress responses (Figure 1). These abilities may endow KPPR1 with greater fitness during respiratory infection as compared to MGH 78578, and it is interesting to speculate that these differences may be defining features of hypervirulent K. pneumoniae.

Our model additionally identified relevant metabolic differences between MGH 78578 and KPPR1. MGH 78578 was capable of using dulcitol as a carbon source, whereas KPPR1 was not. This finding confirms previous reports indicating that the genes responsible for dulcitol metabolism are found in some K. pneumoniae strains but not others [37].

In other cases, our Biolog data revealed interesting differences between KPPR1 and MGH 78578. For example, KPPR1 catabolizes acetate and butyrate, 2 short chain fatty acids (SCFAs), while MGH 78578 does not. Previous studies show that these SCFAs are important chemoattractants for neutrophils and also suppress inflammation in the lung [38]. It is possible that KPPR1 can catabolize lung acetate/butyrate, therefore reducing the SCFAs in the environment. This would result in reducing the attraction by neutrophils and increasing toxicity associated with inflammation. Surfactant proteins (particularly SP-A) are highly abundant and important to both the surface-tension reduction and innate immune properties of the lung. Previous studies have shown mutations in SP-A that eliminate N-glycosylation and result in nonfunctional proteins that aggregate. If KPPR1 has the ability to cleave these glycans and consume them, then KPPR1 could inflict unique damage on the lungs, compared with MGH 78578. Indeed, KPPR1 is able to induce significant tissue damage during respiratory infection, and the lungs become highly inflamed in a short period. This is reflected by the rise in the number of polymorphonuclear leukocytes in the airspace, the increased abundance of cytokines and chemokines, the activation of the acute phase response, and the leakage of plasma into the respiratory compartment. As this environment evolves, K. pneumoniae must adapt not only to combat innate immune responses, but also to respond to the changing composition of nutrients from which K. pneumoniae must derive energy. The fact that, according to our Biolog data, KPPR1 shows an enhanced substrate range for glycan molecules (fucose, galactosides, mannose, and the reduced/modified sugars glucuronamide, D-galactosamine, D-mannosamine, N-acetyl-D-mannosamine) suggests this isolate may be more adapted to these changes than other K. pneumoniae isolates. We stress that the data presented here do not conclusively support these conjectures, but they do generate interesting hypotheses that can be further studied.

Well-validated metabolic models also have clinical utility in that they can help in the design of novel antibiotic strategies and identify which taxa will be vulnerable to a given antibiotic. Two examples of genes predicted to be essential by the model support that this approach may be successful. Thymidylate kinase (TMK; VK055_1369), essential for DNA nucleotide synthesis, was predicted to be essential, and TMK inhibitors have been developed in some bacteria [39]. In a similar manner, the lipid A disaccharide synthase LpxB (VK055_2370) was predicted to be essential. LpxB catalyzes a step in lipid A biosynthesis downstream from LpxC [40] and is broadly required for lipopolysaccharide biogenesis. Inhibitors of LpxC are now under development as antimicrobial agents [41]. These are just 2 examples to illustrate how model predictions can support the identification of future targets for antibacterial therapeutics. These targets are also highly conserved among Enterobacteriaceae, demonstrating that model predictions can identify targets with broad applications to many pathogens.

Using metabolic models, it is also possible to identify nonintuitive synthetic lethal pairs of genes (ie, lethality in a double mutant for a pair of genes that individually are nonessential). Likewise, metabolic models should allow for prediction of metabolic genes essential for survival in the environment of an infected lung. More and perhaps different genes would become essential in the lung environment, compared with rich growth medium. Refinement and validation of the model under easily controlled laboratory conditions enhance its predictive abilities. As more is learned about metabolic conditions in the host and gene requirements in host tissue, the model had be refined to understand signaling and metabolism under those constraints. The conditions under which essential genes are predicted by the model can also be adapted to account for variations in the host environment that occur due to host genetics and diet. The iKp1289 model presented here is a framework for KPPR1 metabolism that can be further adapted once additional information about the lung environment is obtained.

It is anticipated that metabolic modeling will become increasingly valuable in dissecting pathogenic bacteria. The tools available for carrying out this type of analysis are becoming more accessible and user friendly, which should lower the barrier to entry. The availability of multi-omics data will allow these models to become better refined and validated. These models will be useful in the identification of metabolic genes essential for growth in host tissues and amenable to targeting by small-molecule inhibitors. For these reasons, metabolic modeling is becoming an important addition to the strategies being developed to combat multidrug-resistant bacteria for which few active antibiotics are currently available.

Supplementary Data

Supplementary materials are available at http://jid.oxfordjournals.org. Consisting of data provided by the author to benefit the reader, the posted materials are not copyedited and are the sole responsibility of the author, so questions or comments should be addressed to the author.

Notes

Acknowledgments. We acknowledge the SEED and PATRIC research teams at Argonne National Laboratory, University of Chicago, and the Virginia Bioinformatics Institute for advice, feedback, and suggestions relating to this work.

Financial support. This work was supported by the National Institute of Allergy and Infectious Diseases (to C. S. H. PATRIC Bioinformatics Resource Center to C. S. H.; grants AI053674, AI04831, and AI118257 to A. R. H.; and grant R21AI117262 to M. J. M.); the Office of Biological and Environmental Research, Department of Energy (contract DE-AC02-06CH11357 to C. S. H. via the DOE Knowledgebase project); the National Science Foundation (grant 1452549 to K. E. J. T. and grant IOS-1456963 to M. J. M.); the Chicago Biomedical Consortium, with support from the Searle Funds at The Chicago Community Trust (funding to M. J. M.); the National Institute of General Medical Sciences (R35GM119627 to M. J. M.); and the Northwestern University Feinberg School of Medicine (seed grant to W. W. L., A. R. H., and M. J. M.).

Potential conflicts of interest. All authors: No reported conflicts. All authors have submitted the ICMJE Form for Disclosure of Potential Conflicts of Interest. Conflicts that the editors consider relevant to the content of the manuscript have been disclosed.