Microarray dataset

We utilized the publicly available GEO database (www.ncbi.nlm.nih.gov/geo/), which contains high-throughput gene expression data, to identify relevant studies related to sepsis in Homo sapiens. Our search in GEO was conducted using specific keywords such as "sepsis", "S. aureus", "E. coli", and "array". Studies that have used the microarray technique, whole blood samples; involved adult sepsis patients with positive culture results for both E. coli and S. aureus as well as normal controls were included in this study. One dataset, with an accession number of GSE33341 [14], met the selection criteria (15 April 2023) and was consequently retained for further analysis. The dataset includes patients with confirmed bloodstream infections (BSI), comprising 19 cases of E. coli-induced, 32 cases of S. aureus-induced sepsis, and 43 healthy, uninfected control samples (Table 1). Individual participant’s identification was not feasible during or after data collection.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Cohort characteristics by age and gender.

https://doi.org/10.1371/journal.pone.0305920.t001

Data preparation

The primary data for the study with accession number GSE33341 [14] was obtained from the GEO database and analyzed using R software (v4.4.0) [33] and Bioconductor packages [34]. Normalization of Affymetrix microarrays was conducted using Robust Multichip Average (RMA) [34]. For input, all transcripts that were detected in at least one sample were used, without screening for differential expression. Following the conversion of the probe ID to unique official gene symbols, the dataset was subjected to a transformation process resulting in the arrangement of the 94 samples in rows and 22,277 genes in columns.

Model selection and imbalance mitigation strategies

To ascertain the appropriate model for identifying gene markers which classify E. coli- and S. aureus-induced sepsis and healthy controls, we employed three well-established machine learning algorithms: Ridge [35], Elastic Net [36] and Lasso regression [31]. Briefly, utilizing normalized expression data, we conducted model construction by partitioning the dataset into training (80%) and testing (20%) sets. Through tenfold cross-validation with three repeats, we fine-tuned the lambda hyperparameters across the models. Subsequently, Lasso regression emerged as the most suitable model for the transcriptome data and was utilized for further testing, as it maintained comparable performance while selecting fewer genes compared to other methods (S1 File). Detailed steps for the application of Lasso regression are delineated in the following section, with identical configurations applied for Ridge and Elastic Net regression. Furthermore, in order to address the effects of imbalanced data during model training, we explored different upsampling techniques, including the Synthetic Minority Oversampling Technique (SMOTE) [37] and random upsampling specifically for minority groups, implemented on the normalized expression data before parameter tuning. We then reassessed the Ridge, Elastic Net and Lasso models with the same configuration as previously on the upsampled data. Following evaluation of mean square error (MSE) metric, random upsampling of minority groups was identified as the most suitable method for our dataset and was subsequently utilized throughout our study (S1 File). Further elaboration on the methodology employed for random upsampling will be presented in subsequent sections of this paper.

Predictive genes discovery

In our study, we utilized two approaches of unsupervised learning using Principal Component Analysis (PCA); to reduce the dimensionality of our dataset and visualize the relationship between samples in a reduced dimensional space. Moreover, we used a supervised learning method; Lasso regression [31], to select genes that could differentiate between E. coli- and S. aureus-induced sepsis. To ensure the robustness of the analysis, 10-fold cross-validation with three repeats was performed while maintaining the same proportion of each group at each fold. A penalty was applied, and the penalty constant was determined by tuning the lambda hyperparameter within the range of 0 to 1, with a step size of 0.01. The Lasso model was employed with a fine-tuned lambda parameter for the selection of genes. To predict the performance of the Lasso model, the dataset was split into two sets of training and test sets, with each set consisting of 80% and 20% of the samples, respectively. A Lasso regression model was generated using the training dataset with a tuned lambda value. Subsequently, the model’s predictive capability was assessed using the MSE metric. Furthermore, we investigated the variations in MSE induced by Lasso in both smaller and larger sample groups. To study the smaller sample size, we utilized the training dataset, applied robustness through cross-validation with the same configurations, and determined the lambda parameter with tuning as described. Subsequently, the dataset was partitioned into training and test sets with 80% and 20% of samples respectively, and the Lasso model was built using the training dataset, followed by an examination of the MSE. For the larger sample data, we employed an upsampling technique, as described in the following section. We maintained the same configuration for cross-validation, penalty constant tuning, Lasso model building, and prediction as detailed previously.

Through Lasso regression, a set of predictive genes were selected. These genes were further evaluated for predictive performance using logistic regression [32] and the area under the curve (AUC) and precision, recall, F1-score, and accuracy metrics. Moreover, we applied hierarchal clustering to discover distinct gene clusters and their underlying pattern within our predictive genes.

Upsampling for group balance

Age over 65 and gender have been reported to be a risk factor for developing sepsis [25, 26, 38, 39]. To further explore the stability of the predictive genes, we stratified the samples by age and gender, creating four subgroups: young male (< 65), young female (< 65), old male (> 65), and old female (> 65). An imbalance in sample sizes across these subgroups was observed, with the old male subgroup having the fewest samples (Table 1). To mitigate the effect of this imbalance, we used an upsampling technique.

In the first step, we addressed age imbalance by randomly duplicating existing samples in the minority age subgroups using ‘resample’ method until the number of samples in each age subgroup was equal. Secondly, the gender imbalance was addressed by randomly upsampling existing samples in the minority gender subgroup. As a result, we obtained an equal number of female and male samples within each bacterial group with a total of 151 samples and 22,277 genes (Table 1), which allowed us to train and test our predictive genes using the same number of age-gender balanced samples for each group. We did not perform age-gender-based upsampling for the healthy control group as the subgroups were already balanced.

In the third step, the imbalance between the bacterial and the healthy control groups was addressed by randomly upsampling the minority groups of E. coli and healthy control to match the number of samples in the S. aureus-induced sepsis group. This resulted in a total of 228 samples with 22,277 genes (Table 1), allowing us to train and test the predictive genes using an equal number of samples for each group.

External validation in independent datasets

To ensure the generalizability of our results, we validated our predictive gene set using two additional datasets, GSE13015 [40] and GSE65088 [17]. In GSE13015, we examined data from 4 E. coli (female, ages 63–72), and 2 S. aureus (female, ages 52–74) samples of sepsis patients, and 5 healthy control samples (2 females and 3 males, ages 35–67), all obtained from whole blood and RNA expression levels were assessed using the Illumina HumanHT-12 v3 Expression BeadChip array (Illumina, San Diego, CA, USA). Similarly, GSE65088; a BSI resembling dataset, consisted of 10 E. coli and 10 S. aureus samples cultured in whole blood samples from healthy donors (male, age ≥ 40 years) and analyzed using the Illumina HumanHT-12 v4 Expression BeadChip array (Illumina, San Diego, CA, USA). For consistency, we normalized both datasets using the quantile method of normaliseIllumina within the beadarray package (v2.52.0) [41] in R (v4.4.0) [33]. In GSE13015, pre-normalization involved background correction as described in Schmid et al. [42] using the bgAdjust.affy method of lumiB() function, Lumi package (v2.54.0) [43], enabling subsequent log2-transformation of expression values. Following the conversion of probe IDs to unique official gene symbols, the datasets underwent a transformation process. The intersection of each dataset with the set of predictive genes was then determined. We averaged the expressions of genes with multiple annotations in the validation datasets.

For validation purposes, we utilized GSE13015 to assess the predictive performance of the selected genes in distinguishing between the three groups: E. coli-, S. aureus-induced sepsis, and healthy control. In parallel, GSE65088 was leveraged to validate the predictive performance of the selected gene model in discriminating between E. coli- and S. aureus-induced samples. The efficacy of our gene-based diagnostic approach was evaluated using the AUROC curve as the primary metric.

Protein-Protein Interaction (PPI) network creation

After identifying the predictive genes using the Lasso regression algorithm, a protein-protein interaction (PPI) network was constructed by integrating the data with the STRING database [44]. The Cytoscape software (v3.9.1) [45] was used to visualize the interactive relationship between the predictive genes. To identify densely connected regions in the network, the Molecular Complex Detection (MCODE) plug-in of Cytoscape was utilized to cluster the proteins [46].

Functional annotation analysis of candidate genes

To investigate the biological significance of the predictive genes, functional and pathway enrichment analyses were conducted using the PANTHER (Protein ANalysis THrough Evolutionary Relationships) [47], and Reactome [48] resources within the Gene Ontology (GO) interface [49–51]. An FDR p-value <0.05 was considered statistically significant for enrichment.

Statistical analysis

For all statistical tests, we utilized Jupyter Notebook v6.0.3 [52], the Scikit-Learn [53], and the Seaborn library [54]. We employed the Mann-Whitney test (two-sample Wilcoxon test) to assess the differences in gene expression levels between the E. coli- and S. aureus-induced sepsis groups as needed. The Pearson correlation coefficient was utilized to measure the linear association between predictive genes.

Identification of predictive genes

To identify the predictive genes within the pool of 22,277 genes associated with the diagnosis of E. coli- and S. aureus-induced sepsis in 94 samples, we employed Lasso regression with a lambda value of 0.07. The Lasso regression model yielded a nonzero coefficient for 25 genes (S1 Table) which exhibited an MSE of 0.20, implying an 80% predictive accuracy.

To investigate the variation of MSE, we employed Lasso regression to a more restricted dataset, specifically the training dataset with a sample size of n = 76, using the same cross-validation parameters, a lambda value of 0.07 was achieved, resulting in an MSE of 0.15. The Lasso model identified 20 genes, with 14 genes overlapping the 25 selected genes (S1 Table).

Expanding the investigation to include a larger sample size through upsampling to 151-sample and 228-sample datasets, Lasso selected all available genes with a lambda parameter set to 0 in both conditions. The MSE for the 151-sample dataset reduced to 0.07, while for the 228-sample dataset, it further decreased to 0.04. This underscores the heightened predictive performance of the gene selection model when provoked with a more extensive dataset obtained through upsampling.

Classification analysis of three groups of participants

To assess the predictive performance of the selected genes in classifying the three groups of E. coli-, S. aureus-induced sepsis, and healthy controls, we established a logistic regression model utilizing the 25 selected genes. Subsequently, we generated receiver operating characteristics (ROC) curves to measure the effectiveness of these predictive genes, as illustrated in Fig 1A. The 25-gene model exhibited an AUC of 0.96 when distinguishing E. coli-induced sepsis from other conditions, and an AUC of 0.98 for S. aureus-induced sepsis from other circumstances. The model achieved a perfect AUC of 1 for distinguishing healthy controls from other groups, indicating perfect separation of the healthy group from sepsis patients (metrics available in S2 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Assessment of the performance of the selected genes.

(A) Receiver operating characteristic (ROC) curves (E. coli; green dot line, S. aureus; orange dot line, healthy control; blue dot line) (B) Principal Component Analysis (PCA) plot for the 25-gene model predicting E. coli- and S. aureus-induced sepsis in the patient (E. coli; purple, S. aureus; light blue, healthy control; green). The selected genes showed a high predictive performance for classifying each group from other conditions.

https://doi.org/10.1371/journal.pone.0305920.g001

The 25 predictive genes were analyzed by group (E. coli vs S. aureus vs healthy control) using a PCA plot (Fig 1B). The PCA plot indicated the clustering of samples based upon groups, with the separation of healthy controls from E. coli- and S. aureus-induced sepsis cases. Partial overlap between E. coli- and S. aureus-induced sepsis, particularly along PC2 and PC3 indicates some similarities in gene expression between these two patient groups or that there is some variability within each group. Furthermore, PC3 could indicate a subpopulation within the three groups of E. coli-, S. aureus-induced sepsis, and healthy controls.

Internal validation of the predictive genes

To validate the outcomes of Lasso feature selection, we assessed the distribution and performance of the selected 25 genes, aiming to determine the distributional properties of genes between samples. Skewness and kurtosis metrics were employed as indicators of symmetry and peakedness, respectively, with the normal range defined as -2 to +2 for skewness and -7 to +7 for kurtosis [55, 56]. The predictive genes showed positive and negative skewness within the range of +/-2 for E. coli-induced sepsis patients, except for the genes EIF1AY, APOBEC3B, with positive and GUSBP3 (also known as SMA3) with negative skewness, and positive or limited negative kurtosis within the range of +/-7, except for GUSBP3 with considerable positive kurtosis. The predictive genes for S. aureus-induced sepsis patients demonstrated positive and negative skewness and kurtosis within the range of +/-2 and +/-7, respectively (Fig 2A and 2B).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Internal validation of selected predictive genes.

(A) Skewness and (B) Kurtosis of 25 genes selected by Lasso regression (E. coli; blue line, S. aureus; red line). (C) Violin plot of three skewed predictive genes in patients with E. coli- and S. aureus-induced sepsis, as assessed by Mann-Whitney U test. (D) PCA plot of predictive genes after removing the EIF1AY gene (n = 24). EIF1AY: Eukaryotic Translation Initiation Factor 1A Y-Linked, APOBEC3B: Apolipoprotein B MRNA Editing Enzyme Catalytic Subunit 3B, GUSBP3: Glucuronidase, Beta Pseudogene 3. The predictive genes were robust within the range of +/-2 and +/-7 for skewness and kurtosis respectively.

https://doi.org/10.1371/journal.pone.0305920.g002

We performed a Mann-Whitney U test to evaluate the differential expression of the three genes (EIF1AY, APOBEC3B, and GUSBP3) in E. coli- and S. aureus-induced sepsis groups. The corresponding p-values for EIF1AY, APOBEC3B, and GUSBP3 were 0.0002, 0.038, and 0.027, respectively, with average log2-fold changes of 1.57, 1.16, and 0.95. These results suggest that the observed differences in gene expression between the patients with E. coli- and S. aureus-induced sepsis are statistically significant (Fig 2C).

Following filtration of the three genes EIF1AY, APOBEC3B, and GUSBP3 from the list of 25 predictive genes and reassessment of the model in classifying two patient groups and healthy control, the AUC increased to 1 for all groups, with the same accuracy, precision, recall, and f1-score as before the filtration of these genes from the dataset (S2 Table). This indicates that the three genes in E. coli-induced sepsis may have had a negative impact on the performance of the model. This finding was also supported by the PCA analysis which demonstrated an improvement in PC1 (38.40%), and a decrease in PC3 (6.69%) after filtration of these genes (S2 Table). As a result of the filtration of the three genes, the subpopulation within each group was less pronounced, with the PCA plot demonstrating that EIF1AY played a key role in this modification (Fig 2D). As it can be inferred from Fig 2C and Table 1, the gene was expressed in the female group with E. coli-induced sepsis.

Classification analysis of two groups of patients

To evaluate the performance of the 25 and 22 predictive genes (unfiltered vs filtered genes of EIF1AY, APOBEC3B, and GUSBP3) in classifying samples from the two patient groups, E. coli-induced sepsis vs. S. aureus-induced sepsis, we conducted a two-group classification analysis using logistic regression algorithm. The AUC values for the 25-gene set and the 22-gene set were 0.75 and 0.89, respectively. The higher AUC value obtained with the 22-gene set suggested improved performance in discriminating between the two patient groups (Fig 3A). Furthermore, we performed a hierarchical heatmap clustering using Ward’s method on 25 predictive genes to discover gene expression patterns. The analysis led to the formation of two distinct gene clusters (Fig 3B) suggesting unsynchronized behavior in response to the E. coli and S. aureus pathogens between clusters. This was further supported by the lack of significant correlations between genes, as evidenced by the prevalence of values in the low positive range (0 to 0.25) and low negative range (-0.25 to 0) in Pearson correlation analysis (S1 Fig). This suggests genes do not exhibit a strong association with each other, aligning with their high predictive performance.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Classification performance of E. coli- and S. aureus-induced sepsis.

(A) AUROC curves generated using the 25-gene model (left) and the 22-gene model (right) (B) Hierarchical clustering of E. coli- and S. aureus-induced sepsis with the 25-gene model. The selected genes showed high predictive performance for classifying two groups of patients.

https://doi.org/10.1371/journal.pone.0305920.g003

Stability analysis of predictive genes

To address dataset imbalance, we employed upsampling techniques, ensuring 16 samples per gender subgroup for E. coli-induced sepsis and 38 for S. aureus-induced sepsis, minimizing age-gender discrepancies (Table 1, S2A Fig). For broader predictive gene consistency, we evenly distributed participants among E. coli, S. aureus, and healthy control groups, resulting in 76 samples per group (Table 1). Stability assessment of predictive genes using PCA, AUC, and classification metrics demonstrated improved separation of E. coli- and S. aureus-induced cases (S2A and S2C Fig, S2 Table) with an AUC of 1 for all groups (S2B and S2D Fig).

Diagnostics power assessment of 25-gene classifier

To further assess the diagnostic efficacy of the model in distinguishing between E. coli- and S. aureus-induced sepsis, we applied the AUROC curve analysis to both validation datasets. In the first validation dataset (GSE13015), 23 genes overlapped with the 25-gene model (except for IGHG1 and YLPM1, 92% identity). As shown in Fig 4A, using the 23 shared genes, the model demonstrated strong discriminative performance, with an AUC of 0.79 for E. coli-induced sepsis cases, 0.72 for S. aureus-induced sepsis cases, and a notable AUC of 0.87 for healthy control cases. In the second validation dataset (GSE65088), we found 24 overlapped genes (except for IGHGI, 96% identity) which distinguished E. coli and S. aureus infection cases with an AUC of 0.62 (Fig 4B).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Model performance in two distinct scenarios using AUROC.

(A) Distinguishing between E. coli-, S. aureus-induced sepsis, and healthy control using the GSE13015 dataset, and (B) Separating E. coli and S. aureus infections using the GSE65088 dataset. The selected genes showed high diagnostic performance in distinct external datasets for classifying the three groups of participants.

https://doi.org/10.1371/journal.pone.0305920.g004

Integrative analysis of gene associations

Analyzing associations among these genes could reveal mechanistic differences between E. coli- and S. aureus-induced sepsis. Using STRING [44], the PPI network analysis on proteins corresponding to the 25 genes revealed a sparse network (23 nodes, 6 edges) with a protein-protein interaction enrichment p-value of < 0.0187 for 8 proteins (IFIT1, IFI27, GBP1, FCGR1A, EIF1AY, DDX3Y, HIST1H1T (H1-6), and HIST1H1AD (H2AC7)), indicating partial functional or physical connections (Fig 5A). In Cytoscape with the MCODE plug-in, we identified a densely connected protein cluster (IFI27, IFIT1, GBP1, and FCGR1A) (Fig 5B) and noted functional associations among EIF1AY, and DDX3Y, as well as HIST1H1T and HIST1H1AD. However, other proteins showed no associations, indicating that genes associated with each bacterium may activate distinct pathways. Concurrently, we conducted a functional enrichment analysis on the 25 predictive genes using the GO interface and PANTHER resource. In the GO ontology analysis, the categories within the biological process were notably enriched in immune system processes, defense responses, and immune responses (S3A Table). Moreover, Reactome pathway analysis revealed immune system-related pathways including cytokine signaling, innate immune system, and interferon signaling as the most significant pathways associated with the predictive genes (S3B Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Networking and clustering of predictive genes.

(A) Protein-Protein interaction network of 25 predictive genes using STRING (B) One cluster detected by Cytoscape consisting of 4 nodes and 4 edges. The lack of significant association between the selected genes supports the high predictive performance in the classification of the groups.

https://doi.org/10.1371/journal.pone.0305920.g005