Chemicals used in this study were acquired from Sigma-Aldrich (St. Louis, MO, USA) in ACS purity unless noted otherwise. Deionized water was prepared using reverse osmosis equipment Aqual 25 (Aqual s.r.o., Brno, Czech Republic) and further purified using Milli-Q Direct QUV equipped with the UV lamp, with 18MΩ resistance. pH was measured using the pH meter WTW inoLab (Weilheim, Germany).

The following extraction protocol was based on MALDI Biotyper™ 3.0 User Manual Revision 2, also used in a previous report.36 500μL of bacterial culture, cultivated overnight, was centrifuged at 14,000×g for 2min. The supernatant was discarded and the pellet was re-suspended in 300μL of deionized water besides adding 900μL of ethanol. After centrifugation at 14,000×g for 2min, the supernatant was discarded and the obtained pellet was air-dried. The pellet was then dissolved in 25μL of 70% formic acid (v/v) and 25μL of acetonitrile and mixed. The samples were centrifuged at 14,000×g for 2min and 1μL of the clear supernatant was spotted in duplicate onto the MALDI target and air-dried at room temperature. Then, each spot was overlaid with 1μL of α-cyano-4-hydroxycinnamic acid (HCCA) matrix solution (20mgmL−1) in organic solvent (50% acetonitrile and 2.5% trifluoroacetic acid, both v/v) and air-dried completely prior to MALDI-TOF MS measurement on UltrafleXtreme MS (Bruker Daltonik GmbH, Bremen, Germany). As matrix solution 2,5-dihydroxybenzoic acid (DHB) was also used in the same concentration and solvent as HCCA. Spectral data were taken in the m/z range of 2000–30,000Da, resulted from the accumulation of 240 laser shots targeted to different regions of the same sample spot. These data were analyzed with the Flex Analysis software (Version 3.4). Final preparation of dendrograms was carried out in the MALDI BioTyper™ 3.1 (Bruker Daltonik GmbH, Bremen, Germany).

Bacterial cells were centrifuged at 4450×g and 20°C for 10min. The pellet was resuspended in 400μL of lysis buffer (6M guanidium hydrochloride, 0.1M sodium acetate) and cells were lysed for 1hour at 20°C and 600rpm on Multi RS-60 (Biosan, Riga, Latvia). Genomic DNA was isolated from lysed bacterial cultures via MagNA Pure Compact (Roche, Mannheim, Germany), using Nucleic Acid Isolation Kit I, protocol DNA_Bacteria.

16S rRNA gene was amplified using Taq PCR Mix (New England Biolabs, Ipswich, MA, USA) and MasterCycler realplex4 epgradient S (Eppendorf, Hamburg, Germany). 100μL of reaction mixture consisted of: 1× Standard Taq Reaction Buffer, 1.6U of Taq DNA polymerase, 200μM Deoxynucleotide Solution Mix, 0.5μM primers and 5μL of isolated genomic DNA. The forward primer E9F 5′-GAGTTTGATCCTGGCTCAG-3′ and reverse primer U1510R 5′-GGTTACCTTGTTACGACTT-3′ were synthesized by Sigma–Aldrich (St. Louis, MO, USA). The reaction profile was as follows: initial denaturation at 94°C for 4min; 30 cycles of denaturation at 94°C for 30s, annealing at 52°C for 30s and elongation at 72°C for 105s; with terminal elongation at 72°C for 10min. Amplified fragments were purified using MinElute PCR Purification Kit (Qiagen, Hilden, Germany).

For sequencing reaction the DTCS Quick Start Kit (Beckman Coulter, Pasadena, CA, USA) was used. To 20μL sequencing reaction mixture, 98ng of purified fragment, 0.75μL of 10μM forward primer, 4μL of DTCS Quick Start Master Mix and 1μL of Sequencing Buffer were added. The conditions of 30 cycle-reactions were as follows: 96°C for 20s; 50°C for 20s and 60°C for 4min. The purification of sequencing product was carried out using CleanSEQ kit (Beckman Coulter). Purified samples in Sample Loading Solution were transferred to the plate and DNA sequencing was performed using Genetic Analysis System CEQ 8000 (Beckman Coulter). After denaturation at 90°C for 2min, a fluorescence-marked DNA fragments were separated in 33cm long capillary with 75μm i.d. (Beckman Coulter) filled with linear polyacrylamide denaturing gel. The separation was run at capillary temperature of 50°C and voltage of 4.0kV for 85min. Sequences were identified by comparison with NCBI database.

Isolation of genomic DNA was performed using the same method as described in section DNA sequencing. The mecA and fnbA genes were amplified using polymerase chain reaction (PCR) as previously reported.37 The primers were synthesized by Sigma-Aldrich and the sequences of forward and reverse primers for mecA gene were 5′-CCCAATTTGTCTGCCAGTTT-3′, and 5′-TGGCAATATTAACGCACCTC-3′ and for fnbA gene were 5′-GATACAAACCCAGGTGGTGG-3′, and 5′-TGTGCTTGACCATGCTCTTC-3′. The volume of PCR reaction mixture was 100μL containing 1× Taq reaction buffer, 0.2mM dNTP, 1.6U of Taq DNA polymerase (New England Biolabs) and 0.5mM of each primer. The reaction profile was as follows: initial denaturation at 94°C for 4min, 30 cycles of denaturation at 94°C for 30s, annealing at 53°C for 30s and extension at 72°C for 1min with a final extension of 7min. The amplification generated a 223bp for mecA and 191bp for fnbA.

Agarose gel (2% (v/v), high melt, Mercury, San Diego, CA, USA) was prepared with 1× TAE buffer (40mM Tris, 20mM acetic acid and 1mM ethylenediaminetetraacetic acid) and ethidium bromide (5μL per 100mL of the gel) as described elsewhere.38 100bp DNA ladder (New England Biolabs) within the size range from 100 to 1517bp was used to monitor the size of the analyzed fragment. The electrophoresis (Bio-Rad, Hercules, CA, USA) was run at 60V and 6°C for 160min. The bands were visualized by UV transilluminator at 312nm (Vilber-Lourmant, Marne-la-Vallée, France).

Automated neuronal network was used as a predictive model. Classification analysis automated neuronal network was used for the estimation of categorical data. The dataset was randomly divided as follows: 80% for learning, 10% for testing, and 10% for validation. Following network types were tested using automated network search: multilayer perceptron network (MLP), and radial basis function (RBF). Number of hidden units to search was determined as follows: 8–24 and 8–11 for MLP and RBF, respectively. Total 1000 networks were trained, and activation functions were searched for identity, logistic, tanh, exponential. Weight decay of 0.0001–0.001 was used for hidden layer and output layer. Weight of input variables for learning was used based on MALDI-TOF classification score. Unless noted otherwise, p-value less than 0.05 was considered significant. Software Statistica 12 (StatSoft, CA, USA) was used for analysis.

For the identification of bacterial entities we employed complementary methods for independent evaluation of different biomolecules – proteins and DNA.36,41,42 Sanger sequencing was utilized as a confirmation method, based on sequencing of 16S rRNA gene. This gene contains hypervariable regions, providing species-specific sequences, hence it can provide enough information for a confident discrimination, and thus became popular in medical microbiology to classify bacteria.43,44

When compared to sequencing, MALDI-TOF MS offers much shorter analysis time. By using this technique, wound microbiome could be discriminated within one hour of incubation, and thus this will likely become the method of choice for future microbiome identification. Nevertheless, the classification is based on a still developing database34; hence MALDI-TOF MS identification of non-databased bacteria has still to be connected with other confirmation methods. From this reason we firstly employed MALDI-TOF MS with a condition: If score <2.00=16S rRNA sequencing.

As shown in S2, 108 bacterial strains were identified37 of them had to be confirmed by sequencing and confirmed strains were immediately databased to increase future classification success. Strains of S. aureus were the most often identified (n=35). Thus, methicillin-resistant S. aureus (MRSA) is highly associated with severe infections in post-surgical wounds45; we further analyzed the mecA gene, encoding a modified penicillin binding protein (PBP) known as PBP2a, with decreased affinity toward β-lactams.46 The mecA positivity was determined in four isolates. Since 67% of patients had deep wound infections and were treated for more than 8 weeks after admission to infectious Department of Trauma Hospital of Brno, presence of mecA was shown to be a crucial microbiological factor, affecting patients prognosis. Further, we determined the presence of fnbA gene, responsible for adhesins production. Adhesion to human extracellular matrix components and serum proteins is an important facet in the interaction between bacteria and its host cells.47 Lim and coworkers identified the presence of fnbA in 96% of all isolated MRSA strains.35 In our case, fnbA presence was confirmed in all MRSA isolates and in 89% of methicillin-sensitive S. aureus isolates. Similarly to mecA, fnbA was found to be associated with infection severity. In patients with negative fnbA and mecA the treatment duration was less than four weeks in 75% of cases, despite the fact that patients had deep wound infections. This finding suggests that the severity of staphylococcal infections does not depend solely on antimicrobial resistance, but also on adhesins expression, which enhance the interaction with the target host cells.

According to duration of treatment, the patients were divided into specific subgroups, where each sector represents one bacterial strain.

The subdivision of patients was based on surgical wounds classification SSI. As shown, patients were divided into two groups - deep and superficial wounds and the associated bacterial strains are depicted in Fig. 1A and B.

Fig. 1.

Representation of microorganism species present in patients’ wounds. Patients were grouped based on infection severity. The graphs show bacterial cultures grown on different selective nutrient media. (A) Infection severity – deep wounds and (B) infection severity – superficial wounds.

(0.56MB).

As it is obvious from Fig. 1A, in the more serious infections (deep) S. aureus was the main bacterium of microbiome composition (28% of identified strains), followed by Enterococcus faecalis (15%), and Escherichia coli (11%). On the other hand, E. coli was not so often identified in surficial wounds (5% – Fig. 1B). Taken together, the microbiome composition in both groups exhibits substantial differences, and thus it can be hypothesized that presence of minority representatives as Hafnia alvei, Proteus vulgaris, Staphylococcus lugdunensis, or Enterobacter cloacae in the wound can significantly influence the infection severity. It can be also stated that increasing duration of treatment leads to increased number of identified Enterobacteriaceae and opportunistic pathogens (Pseudomonas, Enterococcus).

As was shown by Rettinger and colleagues,48 MALDI-TOF mass spectra-based phylogenetic analysis is considered equivalent to 16S rRNA gene sequencing. Therefore, we employed MALDI Biotyper™ for preparation of dendrograms for our groups, divided by treatment duration (Fig. 2). Dendrograms showed similarity of same bacterial strains (low distance level), but in some cases larger differences were found – usually among bacterial strains from different patients. These differences were caused probably by modifications of bacterial proteins. Karger et al. found methylation as a cause of higher distance level in dendrogram between Burkholderia pseudomallei and other types of B. pseudomallei.49 Thus it can be concluded that not only changes in microbiome representatives affect treatment duration and success, but also small changes in protein post-translation modifications can be highly important for patients’ recovery.

Fig. 2.

Dendrograms from protein mass profiles of microorganisms in different groups based on treatment duration. Created in MALDI Biotyper™. (A) Treatment duration less than four weeks. (B) Treatment duration 4–7 weeks. (C) Treatment duration eight and more weeks. (D) Exitus.

(0.81MB).

Two neuronal networks were created: (1) for the prediction of time-to-heal, and (2) for the prediction of infection severity. The following input parameters were used for the construction of networks: from 2000 networks five were retained and one was used for further final custom neuronal network. The settings of the network created using automated algorithm and used for the custom final learning were Multilayer perceptron 89-13-3 (input-hidden-output neurons), Broyden–Fletcher–Goldfarb–Shanno (BFGS) training algorithm, sum of squares error function, identity function for hidden layer, and then for output layer. The design of the network is displayed in Fig. 3A. With stopping conditions enabled (Fig. 3B), a final network was created in the 17th training cycle with performances of 91.4%, 85.7%, and 71.4% for training, testing, and validation (accuracy in prediction up to 85% – Fig. 3C), respectively.

Fig. 3.

Design and performance of the neuronal networks. (A) Design of classification network for the prediction of time-to-heal. The number of neurons/inputs is indicated by n. *Note the number of input and hidden neurons is not displayed exactly. (B) Training process of the classification network with stopping conditions activated. (C) Accuracy of the final network for classification of time-to-heal. (D) Design of classification network for the prediction of infection severity. (E) Training process for creation of this network with stopping criteria activated. (F) Accuracy of the network for the prediction of infection severity.

(0.6MB).

Consequently, a second neuronal network for the prediction of infection severity was created using an automated algorithm. The best-performing network was trained under following settings: multilayer perceptron 89-19-2 (input-hidden-output neurons) (Fig. 3D) BFGS training algorithm, cross entropy error function, and exponential and softmax activation function for hidden and output layer. The training process is depicted in Fig. 3E (accuracy in prediction up to 85% – Fig. 3F).

The performances of the network were 100.0%, 85.7%, and 85.7% for training, testing, and validation (Table 1), respectively. The accuracy for individual cases is displayed in Table 2.

Sensitivity analysis of input variables for both networks was carried out. For the prediction of infection severity, the mean sensitivity level was 4.66, ranging from 0.63 to 20.1, and a total sensitivity=179.9 (Fig. 4) The highest level of sensitivity (thus highest impact on prediction of a network) was observed for hypertension (20.15), ulcerative colitis (17.06), obesity (10.13), E. coli (8.87), E. faecalis (8.50), and other factors. The factors with sensitivity <1 were P. vulgaris, Neisseria mucosa, E. cloacae, S. aureus 2 mecA genes, Staphylococcus haemolyticus, paresis/plegia, Alzheimer's disease, age.

Fig. 4.

Sensitivity analysis of all factors for prediction of time-to-heal and infection severity. Sensitivity of individual factors depicted as a percentage of total sensitivity. (A) Sensitivity of individual factors for the prediction of time-to-heal. (B) Sensitivity of individual factors for the prediction of infection severity.

(0.48MB).

For the prediction of time-to-heal, the sensitivity was distinctly more homogeneous for the input factors with mean sensitivity=1.35, (0.99–2.35), total sensitivity=52.49. The factors with higher sensitivity included ulcerative colitis (2.35), Parkinson's disease (2.27), osteomyelitis (2.15), Pseudomonas mendocina (2.11), S. haemolyticus (1.84) and others (Fig. 4). The factors with sensitivity <1 include Streptococcus agalactiae, Corynebacterium striatum, and Stenotrophomonas maltopilia.