Analysis of host-pathogen gene association networks reveals patient-specific response to streptococcal and polymicrobial necrotising soft tissue infections

The study is founded in the INFECT study on clinical and pathogenesis in NSTI, where patients were included by prospective enrolment through 4.5 years in five Scandinavian referral hospitals (see Table 1). Study design and presentation of clinical results are detailed elsewhere [6, 19, 20]. The INFECT study is registered at ClinicalTrials.​gov (NCT01790698).

Tissue biopsies and plasma samples on the day of hospital admission (day 0) were obtained from patients diagnosed with NSTI and admitted to Karolinska University Hospital in Stockholm, Copenhagen University hospital Rigshospitalet, Blekinge County Council Hospital in Karlskrona, Sahlgrenska University Hospital in Gothenburg and Haukeland University Hospital in Bergen in the framework of the EU project INFECT (https://​permedinfect.​com/​projects/​peraid/​).

Diagnosis of NSTI was based on the presence of necrotic or deliquescent soft tissue with widespread undermining of the surrounding tissue. Patients were excluded in absence of reports of necrotic or deliquescent tissue. More details on patient characteristics and study design can be found in [19].

The INFECT study was conducted in accordance with the Declaration of Helsinki and was approved by the regional Ethical Review Board at the Karolinska Institutet in Stockholm, Sweden (Ethics Permits: 2012/2110-31/2), the National Committee on Health Research Ethics in Copenhagen, Denmark (Ethics permits: 1151739), the regional Ethical Review Board in Gothenburg, Sweden (Ethics permits: 930-12) and Bergen, Norway (2012/2227/REC West). All experiments were performed in accordance with the approved ethics applications specified above. All patients provided written informed consent. The INFECT study is registered at ClinicalTrials.​gov (NCT01790698).

RNA sequences were mapped against both human and bacterial genomes to obtain gene expression of both the host and resident pathogens. Quality control was performed with the tool FASTQC [21]. The mapping tool Kallisto [22] was used to map the sequences against the human genome (GRCh38 release 91). The same sequences were also mapped against several bacterial genomes using the published pipeline HUMAnN2 [23] from the UniRef database [24]. The final data sets for the Streptococcus monomicrobial classified samples contained 680 human genes and 721 bacterial Uniref90 sequences, and the polymicrobial classified samples contained 680 human genes and 703 bacterial Uniref90 sequences. Details of the methods used can be found in Additional file 1: Section S1.

Gene-Gene Association networks were built with an algorithm developed to be robust against variation in sample size and noise called the Probabilistic Context Likelihood of Relatedness on Correlation (PCLRC) algorithm [25, 26]. We used partial correlations obtained from a Gaussian Graphical Model (GMM, GeneNet R package implementation) as a measure of association\covariation between human and bacterial genes [27, 28]. A significant correlation r_ij between the ith host and jth pathogen gene was established if the corresponding Benjamini-Hochberg adjusted P-value < 0.05 [29].

TopGO R-package v2.42.0 was used for functional category enrichment analysis [30] using the human genes from each bacterial subnetwork as the target sets of interesting genes and the list of all genes from human genome build GRCh38.p12 downloaded from Ensembl Biomart [31] as the background set. The Biological Process ontology from Gene Ontology was used [32], and the Fisher’s exact test was selected to calculate statistical significance of enrichment for the genes of interest.

We used the Linear Interpolation to Obtain Network Estimates for Single Samples (LIONESS) method to infer the single-sample networks [33]. Estimation was done separately for streptococcal and polymicrobial NSTI as described in [34]. For sample q (containing gene expression profiles for host and pathogen from the same biopsy) out of n samples, the corresponding LIONESS single-sample network is obtained as described in Section 1.10 in Additional file 1. The networks have been estimated using the same approach described in Network Inference section using partial correlations.

Each m × m single-sample network can be reduced to a ½m(m-1) × 1 vector containing the perturbations (edges) of the host-pathogen gene correlation. We collapsed these vectors in two matrices of size 2556 × 42 and 2691 × 24. For each matrix, pairwise distances (Euclidean) among samples were calculated, and hierarchical clustering was applied using the Ward linkage method [35].

Single-sample network edges were ranked for each group using the edge relevance as defined in Eq. 4 in Additional file 1: Section 3.12. For each group we retained the 10 most relevant associations.

The interaction networks between human and bacterial genes are shown in Fig. 2. The network specific to monomicrobial S. pyogenes NSTI comprises the interaction of 20 human and 24 Streptococcus pyogenes genes, while the network for polymicrobial NSTI consists of 69 human and 79 bacterial genes.

We observed NSTI type-specific responses in the host, with different sets of human genes highly correlated with bacterial gene expression depending on whether the infection is caused by S. pyogenes or by multiple bacteria. We found the polymicrobial correlation network to be divided into subnetworks with genes from three bacterial species (S. pyogenes, E.coli and P. asaccharolytica) that have a high relative abundance over the samples. While the input gene expression matrix also contained several gene sequences for Parvimonas micra and Prevotella oris, only a single gene interaction with a host gene for each of these species was observed in the resulting networks.

The genes from these association networks were isolated based on species for enrichment analysis. The set of human genes in the S. pyogenes monomicrobial network (consisting of S. pyogenes genes and associated human genes) were significantly enriched (adjusted P-value <0.05) in GO terms for cytokine production (GO:0080134) and regulation of response to stress (GO:0001816), which include genes coding for the interleukin receptors (IL1R2, IL18R1), CD55 and the heat shock proteins, HSPA5 and HSP90B1 (Fig. 3). The set of genes in the S. pyogenes subnetwork from the polymicrobial samples were significantly enriched in the GO term, cellular response to cytokine (GO:0034097), involving the genes for STAT1, IL18R1, POSTN, CXCL9, CXCL5, demonstrating different responses to mono- versus polymicrobial S. pyogenes NSTI.

We observed a strong negative association between the human gene Zinc-Finger Protein 354B (ZFN354B) and three streptococcal genes, sic1 (Q1J9L2), SpyM3_0968 (Q8K763) and MGAS9429_Spy1542 (Q1JK93) in monomicrobial S. pyogenes. We also observed strong associations of S. pyogenes gene SpyM3_0408 (Q7CFC6) with TATA Box-binding protein-associated factor1D (TAF1D) in polymicrobial infections and with mitochondrially encoded NADH dehydrogenase 4L (MT-ND4L) in monomicrobial infections. In polymicrobial infections, we found correlations between the human pseudogene Ferritin (FTH1P2) and two S. pyogenes genes sagF (Q1JHQ0) and sagG (A2RFD6). S. pyogenes genes of known significance are given in Table 2 and the human genes in Table 3. For all human and bacterial gene interactions and their known functions from Uniprot, we refer you to the Additional file 1: Tables S1, S2, S3 and S4.

We used the patient-specific gene-gene correlations to characterise host-pathogen response at a patient level and to stratify patients based on such responses. Patient clusters based on single-sample network edges are shown in Fig. 4 for streptococcal (monomicrobial) (A) and polymicrobial (B) NSTI.

In the case of S. pyogenes NSTI, we found 6 distinct groups with 4 to 10 patients each, while for the polymicrobial NSTI, we individuated 4 distinct groups with 5 to 7 patients. For each one of these groups, we retrieved the top 10 most relevant host-pathogen gene associations, characterising the particular response to the infection of each patient group. These are shown in Table 4 and Table 5 for S. pyogenes (monomicrobial) and polymicrobial NSTI, respectively. For the polymicrobial case, the top relevant associations involve genes that were mapped to five bacteria species namely S. pyogenes, E. coli, P. asaccharolytica, P. micra and P. oris. We were unable to ascertain associations between the groups and clinical outcomes of patients with significant statistical power due to the lack of sufficient samples per group.

In the S. pyogenes monomicrobial interaction network, the most enriched GO categories for the human genes in the network were cytokine production (GO:0080134) and regulation of response to stress (GO:0001816) suggesting immune system defensive mechanisms in the host response to changes in expression of specific streptococcal genes. The host heat shock protein (HSPA5) is associated with streptococcal acyl carrier protein (P63443, acpP) which is involved in fatty acid biosynthesis in lipid metabolism. Eraso et al. have demonstrated the selection of mutations in the fabT gene, another gene involved in fatty acid biosynthesis during necrotising myositis infections in a non-human primate model [56]

Q1J9L2 is 90% identical to Sic1 [36]. Sic has several different mechanisms of actions, interference with complement and other host defences, and has been proposed to play a significant role in streptococcal infections [57]. It was recently shown that the Sic protein from M1 S. pyogenes, a type over-represented among severe invasive cases of NSTI [20], interacts with TLR2 resulting in release of pro-inflammatory cytokines [58]. A study by Kachroo et al. [59] revealed a positive correlation between sic and genes involved in the host immune response and inflammation when they examined the dual RNA-seq transcriptomes of S. pyogenes and host skeletal muscle from infected non-human primates. In addition, vaccination-induced anti-sic antibodies were effective in bacterial clearance in rabbit, mice, and in an ex vivo whole body assay [60].

Sic has been shown to bind to extracellular histones, a group of danger signals released during necrotising tissue damage. The aggregates formed from this interaction have been shown both in vitro and in co-localised biopsies from NSTIs resulting in the neutralisation of host antimicrobial activity [57]. The study by Frick et al. showed that Sic enhances bacterial survival in an animal model of subcutaneous infection [61]. .The increase in the expression of Sic (Q1J9L2) and its role in the inhibition of complement, accompanied by a downregulation of ZFN354B may be a Streptococcal strategy to evade the host innate immune response.

Different bacteria are known to target steps of host gene expression during pathogenic invasion, potentially as a mechanism to modify the expression of inflammatory genes [62]. The ZFN354B gene is also negatively correlated with the gene MGAS9429_Spy1542 (Q1JK93). Although, the sequence of this gene is unannotated, it is located immediately upstream to the gene encoding EndoS. The gene immediately downstream from the gene encoding EndoS, MGAS10270_Spy1608 (H8HD54), is also found in our analysis correlated with the human gene Neurensin (NRSN1). Even though the exact function of these two gene sequences is unclear, it should be noted that the protein EndoS displays endoglycosidase activity on immunoglobulin G (IgG) by hydrolysing the chitobiose core of the asparagine-linked glycan. EndoS modification of IgG antibodies results in impaired Fc-dependent effector function involved in phagocytic killing and elimination of antibody-antigen complexes from circulation [37, 63]. The strong connectivity of Sic, the EndoS-region, and others in the monomicrobial NSTI networks underlines the importance of the immune evasive strategies in S. pyogenes infections. Interestingly, we find that the transcription of both these virulence factors is abated by clindamycin, lending support to contemporary guidelines advocating adjunctive clindamycin treatment in streptococcal NSTIs [64].

LRRFIP1 gene was found associated with the S. pyogenes gene M1GAS476_1767 (J7MBD1) encoding Fba. Fba is a cell-wall-anchoring, surface-located protein that is found in M-type 1, 2, 4, 22, 28 and 49. These M-types constitute 55 out of the 95 sequenced isolates in the INFECT study [18]. On studying the effects of Fba in relation to the bacterial invasion of and adhesion to HEp-2 cells, Terao et al. inferred that the presence of both Fba and M-protein are required for the most efficient bacterial adhesion and invasion. In addition, the report showed that a Fba mutant displayed lower mortality in a murine skin infection model [38].

In the S. pyogenes subnetwork from the polymicrobial interaction network, the most enriched GO category for the human genes was GO:0034097 (Response to Cytokine) with 6 genes (STAT1, IL18R1, KYNU, POSTN, CXCL9, CXCL5) out of 20 annotated with this GO term. IL18R1, the IL18 receptor complex and the chemokine CXCL5 are associated with bacterial Spx which has an important role in growth, general stress protection and biofilm formation in S. aureus [65]. Therefore, the transcriptional response to cytokines in the host is associated with a change in the regulation of transcription in S. pyogenes that may impact its ability to form biofilm and modulate its response to stress.

The human gene Ferritin heavy chain 1 Pseudogene 2 (FTH1P2) was found associated with three S. pyogenes genes lbp (Q9ZHG8), sagF (Q1JHQ0) and sagG (A2RFD6). Although fth1p2 is a pseudogene, studies have recognised it to compete with fth1, the main intracellular iron-storage protein in the cytoplasm [52]. The study by Terao et al. on adhesion of S. pyogenes to the HEp-2 cells showed that the absence of Lbp significantly lowered the efficiency of adhesion to epithelial cells [45]. More studies have indicated that the primary function of Lbp is as a zinc-scavenger and referred to the gene as adcA [66, 67]. The genes sagF and sagG are two out of the nine genes that form the Streptolysin S-operon. The toxin Streptolysin S has been shown to be responsible for the Beta-haemolysis observed by S. pyogenes and has also been implicated in NSTI pathogenesis [46, 68‐70]. The association of these S. pyogenes genes to fth1p2 is unclear.

Enrichment of immune-related host genes was also found in 6 out of the 12 host genes in the P. asaccharolytica subnetwork which are annotated with the GO category - Immune system process (GO:0002376). E. coli gene interactions with host genes did not reveal any enrichment of genes with immune-related functions. In this subnetwork, enrichment of host genes in cellular localisation and intracellular protein transport was observed and several of the host genes have roles in transcriptional regulation. The majority of the co-expressed bacterial genes in the E. coli subnetwork are uncharacterised proteins.

In the P.asaccharolytica subnetwork from the polymicrobial interaction network, the most highly connected human gene is KYNU which is correlated with eight P. asaccharolytica genes, five of which code for ribosomal proteins (rpmF, rplU, rpsR, rplF, rpsG), one is the small heat shock protein Hsp20 (encoded by Poras_0808) and two are proteins of unknown function [71]. Heat shock proteins are chaperones that can interfere with the uncontrolled protein unfolding that occurs under stress, such as the immune response of the host [72]. An increased kynurenine pathway activity has been linked to inflammation and immune activation and has been implicated in diverse diseases such as depression and cancer [73‐75].

Expression of the host gene, TGFBI, an important cytokine with broad regulatory role in the immune system was also positively correlated with five ribosomal genes (rplU, rpsR, rplF, rpsG, rpsO) in P. asaccharolytica. Two other host genes, SIAE and HINT3, are associated with the ribosomal genes (rplU, rplF) and Hsp20 in this species. SIAE has been functionally associated with autoimmune diseases and preeclampsia [76], and although it has a role in the immune system, it has not previously been studied in the context of an infection. The host gene, SLC11A1, is also associated with the ribosomal gene rplF. SLC11A1 controls natural resistance to infection with intracellular parasites. Pathogen resistance involves sequestration of Fe(2+) and Mn(2+), cofactors of both prokaryotic and eukaryotic catalases and superoxide dismutase, not only to protect the macrophage against its own generation of reactive oxygen species, but also to deny the cations to the pathogen for the synthesis of its protective enzymes.

The expression of S. pyogenes gene SpyM3_0408 (Q7CFC6) was found to be correlated with the expression of human genes in both monomicrobial and polymicrobial infections. In monomicrobial infections, it was positively correlated with the gene mitochondrially encoded NADH dehydrogenase 4L (MT-ND4L). In polymicrobial infections, it was associated with the gene TATA Box-binding protein-associated factor 1D (TAF1D). The gene SpyM3_0408 (Q7CFC6) itself is transcribed as part of the three-gene scfAB-operon. It was hypothesised by Breton et al. that the scfAB-operon plays an integral role in enhancing adaptation and fitness of S. pyogenes during localised skin infection and potentially in the propagation to other deeper tissue in a genome-wide Tn-seq analysis to identify S. pyogenes genetic determinants necessary for in vivo fitness using a murine model of skin and soft tissue infection. The scfAB-operon is part of S. pyogenes core genome, and the gene encodes for a putative transmembrane protein and was found important at the subepithelial site of infection and for the dissemination of into the bloodstream. Homologues of the scfAB-operon are also found in other pathogenic streptococci and closely related gram-positive pathogens [44].

Although cytokine gene expression was associated with S. pyogenes gene expression in both polymicrobial and monomicrobial infections, the only common host gene involved was IL18R1. This occurrence of IL18R1 in both types of infections stands to reason as the Random Forest (RF) models built from cytokine concentrations in the study by Palma Medina et al. found IL18 to be the least important cytokine to differentiate between monomicrobial and polymicrobial NSTI [77].

The genes representing chemokines CXCL5 and CXCL9 were only found in the S. pyogenes subnetwork in polymicrobial infections and not in monomicrobial infections. Thänert et al. found CXCL9, CXCL10 and CXCL11 to be overexpressed in streptococcal NSTI [18] and a recent study analysing cytokines and chemokines in plasma samples from NSTI patients by Palma Medina et al. found CXCL10 to be a robust biomarker for differentiating between monomicrobial and polymicrobial NSTI infections [77]. In accordance with these studies, our analysis shows a difference in the involvement of these chemokines between monomicrobial and polymicrobial S. pyogenes response. The nature of this network analysis renders any information regarding over-expression or under-expression unascertainable.

Overall, we observed different associations of genes coding for collagen proteins, consistent with the study by Singh et al. which showed that various human pathogens utilise the proteins found in the extracellular matrix (ECM) such as collagen proteins for the invasion of the host. Invasive pathogens infract the basal lamina and degrade the ECM proteins employing various proteases drafted from the host. Pathogens use these abilities to adhere to and invade the host tissue [47].

Group 2 differs from other groups only for perturbation of the association between COL3A1 and A2RGM6. In group 3, we find the association between collagen type V and the gene (F5U6Q2) encoding for a hypothetical protein with high similarity (91.4%) to an immunogenic secreted protein (Isp). This isp gene is located immediately downstream of the ihk-irr TCS and the gene is highly conserved among S. pyogenes strains [41]. The function of Isp remains elusive but Kachroo et al. showed that the isp gene contributed to virulence in a necrotising myositis model in non-human primates. Kachroo et al. also postulated the potential role of isp in cell-wall metabolism based on the CHAP domain located at the carboxy terminus [59].

The P0C0H1 (HasA) is required for hyaluronic acid capsule. The capsule represents an important virulence factor of S. pyogenes [43]. We find this gene associated with collagen VI in groups 2, 3, 5 and 6. Despite Collagen VI having antimicrobial properties [78], it has been shown to be a target of adherence by S. pyogenes for persistent infections and inducing invasions [79]. Immunodetection, in vitro binding assays and electron microscopy have shown S. pyogenes to have a strong affinity to Collagen VI and evolved adhesins with the ability to mediate interactions between S. pyogenes and the host [79]. Studies have reported the possibility of hyaluronic capsule to be a major virulence determinant and is also listed in the PATRIC database as a known virulence factor [80]. Dinkla et al. [81] demonstrated that in rheumatic fever, S. pyogenes had the unique capability to bind and aggregate membrane collagen type IV via the M3 protein or hyaluronic acid capsule in M18 serotype. It has also been shown that the upregulation of hyaluronic acid by S. pyogenes in blood is used as a mechanism to mask surface immunogenic determinants and evade antigen-specific antibodies to avoid bacterial death in blood. Dinkla et al. [82] managed to demonstrate that only the S. pyogenes abundant in hyaluronic acid capsule were capable of surviving in human blood containing high levels of antibodies directed against highly conserved bacterial surface proteins. This association may point to a mode of entry and immune evasion by S. pyogenes.

Group 4 differs from other groups only for perturbation of the association between the host gene TMSB4X and the P. micra gene A8SND5. The TMSB4X gene plays an important role in the organisation of the cytoskeleton and binds to and sequesters actin monomers (G actin) inhibiting actin polymerisation. A common target of bacterial pathogens is the host cell actin cytoskeleton, a dynamic system of filaments that is central to shape determination, movement, phagocytosis and intracellular trafficking [83]. After invasion, some pathogens remain within a membrane-bound compartment and target actin to subvert membrane trafficking [83] by polymerising actin on their surface and use filament assembly to power intracellular actin-based motility, generating actin comet tails that trail the moving bacteria [83‐85]. The associated P. micra gene A8SND5 (rplC) highlights the importance of the expression of ribosomal proteins to increase bacterial protein synthesis indicating a possible remodulation of this interplay between host and pathogen.

The association of the E. coli gene encoding for protein E1IVY0 and several human genes are present in all 4 groups. The E1IVY0 is an uncharacterised protein with high similarity (86%) to a serine acetyl transferase from Pantoea ananatis, a plant pathogenic gram-negative, facultatively anaerobic gamma proteobacterium which has been shown to help the bacterium survive oxidative stress conditions [86, 87].

Limitations of this study include the (relatively) small number of samples, heterogeneity among patients with respect to their comorbidities, the time of infection and treatments, along with the differences in biopsy sampling, type of tissue, depth of infection and lack of longitudinal samples. Many of these limitations are unavoidable as the biopsies are obtained at a timing when surgery is clinically indicated and in areas where the tissue pathology warrants surgical removal. However, using important quality aspects such as the collection of biopsies by dedicated teams of clinicians using standardised SOPs, a prospective observational study design similar in all participating clinical centres and the employment of highly stringent statistical approaches strengthen this study.