Bioinformatics combined with machine learning for potential biomarker screening and immune infiltration analysis in neonatal sepsis.

Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Front Pediatr . 2026 Mar 31;14:1702073. doi: 10.3389/fped.2026.1702073 Search in PMC Search in PubMed View in NLM Catalog Add to search Bioinformatics combined with machine learning for potential biomarker screening and immune infiltration analysis in neonatal sepsis Na Liu Na Liu 1 Department of Pediatrics, Wuhan Fourth Hospital, Wuhan, Hubei, China Data curation, Formal analysis, Investigation, Methodology, Project administration, Software, Writing – original draft, Writing – review & editing Find articles by Na Liu 1 , Zhiyong Tan Zhiyong Tan 2 Informatization Centre, Wuhan Vocational College of Software and Engineering, Wuhan, Hubei, China Conceptualization, Data curation, Resources, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing Find articles by Zhiyong Tan 2, * Author information Article notes Copyright and License information 1 Department of Pediatrics, Wuhan Fourth Hospital, Wuhan, Hubei, China 2 Informatization Centre, Wuhan Vocational College of Software and Engineering, Wuhan, Hubei, China * Correspondence: Zhiyong Tan [email protected] Roles Na Liu : Data curation, Formal analysis, Investigation, Methodology, Project administration, Software, Writing – original draft, Writing – review & editing Zhiyong Tan : Conceptualization, Data curation, Resources, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing Received 2025 Sep 9; Revised 2026 Jan 30; Accepted 2026 Mar 4; Collection date 2026. © 2026 Liu and Tan. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY) . The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms. PMC Copyright notice PMCID: PMC13077849  PMID: 41988156 Abstract Objective Apply bioinformatics combined with machine learning algorithms to screen potential biomarkers of neonatal sepsis (NS) and explore the correlation between biomarkers and immune cells. Methods Expression profiling data containing NS samples and control samples were downloaded from the Gene Expression Omnibus (GEO) database. First, the differentially expressed genes (DEGs) were found. Then gene set enrichment analysis (GSEA) was carried out to reveal the significantly enriched pathways, and the core genes within these pathways were aggregated. Subsequently, weighted gene co-expression network analysis (WGCNA) was used to identify the gene modules significantly correlated with NS, from which the key genes were screened. The feature genes were selected by the use of three machine learning algorithms: least absolute shrinkage and selection operator (LASSO) regression, support vector machine recursive feature elimination (SVM-RFE), and random forest cross-validation (RF-CV), and then the potential biomarkers of NS were identified by logistic regression analyses. The dataset was randomly split into training and test sets. A logistic regression model with biomarkers was constructed using the training set, and its diagnostic value was evaluated separately on both the training and test sets. In addition, immune infiltration analysis was performed using immune cell deconvolution algorithm CIBERSORT to explore the correlation between biomarkers and immune cells. Results Overall, after overlapping the filtered DEGs, the core genes of GSEA, and the key genes of WGCNA, we detected 14 intersecting genes. Three machine learning algorithms further selected 4 feature genes, and logistic regression analyses identified ACSL1 and CD3D as potential biomarkers of NS. The classification accuracies of the logistic regression model were 0.883 and 0.902 on the training set and test set, respectively. The area under the curve (AUC) of the receiver operating characteristic curve (ROC) was 0.961 and 0.966, and the calibration curves were close to the ideal curve. It should be noted that these AUC values were estimated within this discovery GEO cohort and require confirmation in independent external cohorts. Immune infiltration analysis showed significant changes ( P < 0.05) in the infiltration abundance of 13 immune cell types, and a strong correlation (>0.6) existed between ACSL1, CD3D and neutrophils, naive CD4 + T cells, and CD8 + T cells. Conclusion ACSL1, CD3D are potential biomarkers of NS and may play an important role in the pathogenesis of NS by modulating immune cell functions such as neutrophils, naive CD4 + T cells, and CD8 + T cells. These conclusions should be regarded as exploratory biomarker discovery rather than diagnostic readiness. Keywords: neonatal sepsis, bioinformatics, machine learning, biomarkers, immune infiltration 1. Introduction Neonatal sepsis (NS) is a prevalent and severe infectious condition during the neonatal period, representing a major cause of mortality and serious complications in the neonatal intensive care unit. The global incidence of NS is 2,022 (95% CI 1,099–4,360) cases per 100,000 live births, with mortality between 11% and 19% ( 1 ). The incidence of NS is 1.8 times greater in middle-income nations and up to 3.5 times greater in low-income countries compared to high-income ones ( 2 ). In China, despite differing opinions on the evolution of NS incidence over the past thirty years ( 3 , 4 ), even the optimists acknowledge that NS continues to pose a significant public health challenge, necessitating ongoing focus and effective interventions to further mitigate its disease burden ( 4 ). The etiological agents of NS encompass bacteria, viruses, and fungi ( 5 – 7 ). Its clinical manifestation is often atypical, potentially presenting with symptoms such as temperature instability, poor feeding, respiratory distress, oliguria, diarrhea, jaundice, and purpura ( 8 ), thereby complicating early recognition and diagnosis of NS. Blood culture is the gold standard for diagnosing NS; however, its clinical use is constrained by a low positive detection rate, long turnaround time, and the impact of antibiotic therapy ( 9 ). Other diagnostic tests unrelated to culture, such as complete blood count, C-reactive protein, and procalcitonin, exhibit a lack of specificity and are suboptimal for diagnosing NS ( 10 ). Furthermore, while all newborns with suspected NS are advised to undergo evaluation for cerebrospinal fluid ( 11 ), the collection of cerebrospinal fluid via lumbar puncture is an invasive procedure that is frequently met with apprehension by families. Numerous studies have demonstrated that screening for extremely sensitive and specific biomarkers is crucial for diagnosing NS and may serve as a possible target for its therapy ( 12 , 13 ). High-throughput sequencing facilitates the examination of alterations in disease gene expression and the identification of possibly illness-associated genes, thereby establishing a foundation for the discovery of novel diagnostic and therapeutic strategies ( 14 ). Bioinformatics methodologies, including DEGs analysis, GSEA, and WGCNA, have demonstrated considerable efficacy in identifying biomarkers associated with disease and possible targets for therapy ( 15 – 17 ). Moreover, both supervised and unsupervised learning in machine learning exhibited considerable benefits in elucidating the inherent relationships within high-dimensional data ( 18 , 19 ), and also displayed substantial applicability in analyzing high-dimensional transcriptome data and identifying biologically significant key genes ( 20 , 21 ). This study utilized NS high-throughput sequencing data to analyze and identify potential biomarkers of NS by integrating DEGs, GSEA, WGCNA, and various machine learning algorithms. Additionally, immune infiltration analysis was conducted using the CIBERSORT algorithm to evaluate the correlation between biomarkers and immune cells, aiming to provide a foundational basis for the diagnosis and treatment of NS. 2. Materials and methods 2.1. Materials 2.1.1. Data sources The GEO is an international public database containing high-throughput gene expression and epigenomics datasets obtained by next-generation sequencing and microarray technologies ( 22 ). The GEO database was searched using the formula “NS AND Homo sapiens [Organism] AND Expression profiling by array [Filter],” and the criteria for inclusion were: the dataset contained NS samples and control samples; all samples were free of diseases other than NS; and the number of both NS samples and control samples was greater than 50. The available dataset, GSE25504 , consisted of a subset of data from four different sequencing platforms ( GPL570 , GPL6947 , GPL13667 , and GPL15158 ), with a total of 170 samples. After excluding the samples such as suspected cases, 135 samples remained, comprising 64 NS samples and 71 control neonates. 2.1.2. Data quality All the relevant analyses in this study were done by the R language. Because GSE25504 integrates four expression platforms ( GPL570 , GPL6947 , GPL13667 and GPL15158 ), we mapped probes to official gene symbols within each platform and retained only the genes shared by all platforms, yielding 6,423 common genes. This intersection step avoids platform-specific missingness and ensures that downstream analyses (DEGs, WGCNA, GSEA, and immune deconvolution) are performed on a comparable feature space. Although 6,423 is smaller than the full transcriptome, it still represents a large subset of reliably measured genes and is sufficient for robust biomarker discovery in multi-platform microarray integration. Due to the batch effect between datasets of different platforms, see Figure 1A , only 1/3 of the samples are shown here in steps of 3, using the “removeBatchEffect” function of the “limma” package (v3.58.1) to remove the batch effect ( Figure 1B ). Figure 1. Open in a new tab Batch effect correction normalizes the distribution of gene expression across microarray platforms in the GSE25504 dataset. (A) Distribution of gene expression values across samples prior to batch effect removal. Each box represents one sample, colored by its original microarray platform: GPL570 (red), GPL6947 (orange), GPL13667 (blue), and GPL15158 (green). Substantial differences in median expression and distribution ranges are evident across platforms, indicating a strong batch effect. (B) Distribution of gene expression values after correction for platform-specific batch effects using the “removeBatchEffect” function. The medians and interquartile ranges across all samples and platforms are successfully aligned. For visual clarity, only one-third of the total samples are displayed. 2.2. Methods 2.2.1. Screening of DEGs and calculation of GSEA The gene expression data of the samples were subjected to principal component analysis (PCA) using the “FactoMineR” package (v2.11), revealing the distribution of NS samples and control samples in the principal component space. The DEGs between groups were analyzed using the “limma” package (v3.58.1), and the results of the differential analysis were shown by volcano plot. The DEGs were filtered using an adjusted P < 0.05 and | lo g 2 FC | > 0.5 as the thresholds, in which the DEGs with lo g 2 FC > 0.5 were labeled as up-regulated genes, while those with lo g 2 FC < 0.5 were labeled as down-regulated genes. We used a relatively lenient effect-size cutoff (|log 2 FC| > 0.5) together with multiple-testing control to avoid discarding biologically meaningful, moderately shifted genes in a complex systemic disease such as sepsis. Importantly, this initial DEG list was subsequently subjected to stringent multi-step filtering (WGCNA module relevance, GSEA leading-edge/core genes, and consensus selection across three machine-learning methods), implementing a “broad-in, strict-out” strategy to reduce false positives while retaining potential regulatory signals. GSEA is a computational method used to determine whether a set of a priori genes exhibit statistically significant differences in consistency between two biological states. All DEGs were sorted by lo g 2 FC value from largest to smallest, and “c2.cp.kegg_legacy.v2023.2.Hs.symbols” from the molecular signature database (MSigDB) was used as the set of a priori genes. The enrichment scores for each biological pathway in the a priori gene set were calculated across the sorted DEGs using the “clusterProfiler” package (v4.10.1). The biological pathways were considered statistically significant when adjusted P < 0.05, and the core genes within these significant pathways were aggregated. 2.2.2. Construction of co-expression gene modules The gene co-expression network was constructed using the “WGCNA” package (v1.73) to identify the gene modules that are significantly related to NS, and then the key genes were selected from them. Firstly, hierarchical clustering was used to cluster all samples and eliminate outliers. Then a suitable soft threshold was selected to make the connections between genes conform to the scale-free network characteristics. After constructing the topological overlap matrix with the soft threshold, gene modules were generated by hierarchical clustering and dynamic shear tree methods. The first principal component of each module was extracted as the module eigengene (ME), and the eigengene significance (ES) of each module was calculated, then the module was considered as a key module when | ES | > 0.6 and P < 0.05. The module membership (MM) and gene significance (GS) of genes within the key modules were calculated to screen key genes, which were defined as those with | MM | > 0.8 and | GS | > 0.2 ( 23 ). 2.2.3. Machine learning algorithms for selecting feature genes Take the intersection of the filtered DEGs, the core genes of GSEA, and the key genes of WGCNA. Then we selected three complementary and widely used feature-selection approaches—LASSO (embedded, sparse linear model) ( 24 ), SVM-RFE (wrapper method optimizing classification margin) ( 25 ), and Random Forest (non-linear ensemble with variable importance) ( 26 )—to select feature genes and reduce method-specific bias. We retained only the intersection of genes selected by all three methods to obtain a conservative, high-confidence set of candidate biomarkers. The corresponding R packages for these three algorithms are, in order, “glmnet” package (v4.1.8), “e1071” package (v1.7.14), and “randomForest” package (v4.7.1.2). λ , which minimizes the cross-validation error, was taken as optimal in LASSO regression; a radial basis function kernel was used in SVM-RFE; for Random Forest, the final number of trees was chosen at the point where the out-of-bag (OOB) error stabilized at its minimum, and the importance of feature gene was calculated by linear scaling in RF-CV. To reduce the risk of overfitting, both SVM-RFE and RF-CV use 5-repeated 3-fold cross-validation. These three algorithms select their corresponding feature genes and then take the intersection to get the feature genes related to NS. 2.2.4. Logistic regression for biomarkers identification and diagnostic value evaluation Univariate logistic regression analysis was conducted with the following independent variables: the feature genes selected by the machine learning algorithms, and the potential influencing factors for NS, including birth weight, birth gestational age, gender, and whether antibiotics were used. Birth weight and gestational age were transformed into dichotomous variables: according to the World Health Organization (WHO) criteria ( 27 ) to determine whether the birth weight was low birth weight (birth weight < 2,500 g was low birth weight) and whether the birth gestational age was preterm (birth gestational age < 37 weeks was preterm). Variables with statistical significance ( P < 0.05) in the univariate analysis were tested for multicollinearity before being included in multivariate logistic regression analysis, and genes characterized by statistical significance ( P < 0.05) in the multivariate analysis were considered as potential biomarkers of NS. The dataset was randomly split into a training set (70%) and a test set (30%). A logistic regression model was constructed on the training set, with NS potential biomarkers as independent variables. The diagnostic value of the model was evaluated on the training and test sets, respectively: decision boundaries were plotted to demonstrate the model's classification ability; using the “pROC” package (v1.18.5), the predictive power of the model was evaluated by plotting receiver operating characteristic (ROC) curves and calculating the area under the curve (AUC); using the “rms” package (v6.8.1), the reliability of the model predictions was assessed by performing an internal calibration with 1,000 resamples and plotting calibration curves. 2.2.5. Immune infiltration analysis CIBERSORT immune infiltration analysis was performed based on the gene expression data of all samples. Since CIBERSORT used a linear support vector regression model ( 28 ), and the downloaded gene expression data had been log-transformed, we applied exponential transformation to restore the original expression values. The infiltration matrix of 22 immune cell types was calculated using the “CIBERSORT” package (v0.1.0), and samples with P < 0.05 were screened. Differences ( P < 0.05) in infiltration abundance between groups of immune cells were analyzed, and the correlation coefficient R between potential biomarkers and infiltration abundance of immune cells was calculated. 2.2.6. Statistical analysis Statistical analysis of the data was done in RStudio (v2023.12.1.402) using R language version 4.3.3. Differences between two groups were compared using the Mann–Whitney U -test, and correlations were analysed using the Spearman correlation coefficient, with a correlation coefficient >0.6 deeming the presence of a strong correlation. In the multicollinearity test, a tolerance >0.1 and variance inflation factor (VIF) < 10 were considered as no multicollinearity. P < 0.05 was considered as statistically significant difference. To ensure reproducibility, a fixed random seed (seed = 1,234) was used for all procedures involving randomness (including the 70/30 train–test split, resampling-based calibration, and model fitting steps where applicable). The overall analytical workflow is summarized in Supplementary Figure S1 . 3. Results 3.1. Concurrent innate immune activation and adaptive immune suppression PCA result showed that NS samples were separated from control samples in the three-dimensional principal components of gene expression ( Figure 2A ). There were 992 DEGs after screening by threshold, of which 475 were up-regulated and 517 were down-regulated in expression ( Figure 2B ). The 10 biological pathways significantly enriched for a total of 129 core genes were obtained by GSEA ( Figure 2C ). The enriched pathways showed a clear bidirectional immune pattern in neonatal sepsis. Pathways related to pathogen sensing and innate inflammation were enriched in the sepsis group, whereas T-cell–associated pathways (e.g., T-cell receptor signaling) tended to be suppressed. This pattern is consistent with the concept that sepsis involves not only hyper-inflammation but also early and/or progressive adaptive immune dysfunction (immunosuppression/immunoparalysis). Figure 2. Open in a new tab DEGs screening and GSEA significantly enriched pathways. (A) Three-dimensional principal component analysis (3D PCA) of global gene expression across all samples. Each point represents one sample, colored by its group. The first three principal components (PC1, PC2, PC3) explain 24.39%, 10.88%, and 6.83% of the total variance, respectively, indicating clear separation between groups. (B) Volcano plot of DEGs comparing NS vs. Control. Genes with a |log 2 FC| > 0.5 and an adjusted P < 0.05 are considered significant. Significantly down-regulated and up-regulated genes are highlighted in red and green, respectively, while non-significant genes are shown in black. (C) Plot of the significantly enriched pathways from GSEA, ranked by normalized enrichment score (NES). Pathways with an adjusted P < 0.05 were considered significant. 3.2. Sepsis-associated WGCNA modules The result of hierarchical clustering showed that there was one outlier sample among 135 samples ( Figure 3A ), then the outlier sample was excluded. A soft threshold β = 9 was chosen, at which point the fitting index was >0.85 for first time and the change of the average connectivity was smooth ( Figure 3B ). The number of genes in the module was defined to be at least 30, and the threshold of modules merging was set to 0.25, resulting in the generation of 11 gene modules ( Figure 3C ). The results of the ES calculation for the modules ( Figure 3D ) showed that the key modules were green ( | ES | = 0.798 , P < 0.001), black ( | ES | = 0.794 , P < 0.001), and blue ( | ES | = 0.621 , P < 0.001). The key genes in these three modules were filtered according to the criteria of | MM | > 0.8 and | GS | > 0.2 ( Figures 3E–G ), and 256 key genes were obtained in total. Figure 3. Open in a new tab WGCNA identifies key gene modules associated with NS. (A) Sample clustering dendrogram based on global gene expression, used to detect outliers. One sample (above the red line) was identified as an outlier among the initial 135 samples and was excluded from subsequent network construction. (B) Analysis of network topology for various soft-thresholding powers ( β ). The left panel shows the scale-free fit index, and the right panel shows the mean connectivity. β = 9 was selected as the optimal value, where fit index first exceeded 0.85. (C) Hierarchical clustering dendrogram of genes and the resulting co-expression modules identified by WGCNA. Genes are clustered based on topological overlap, with each module assigned a unique color. (D) Module-trait associations. Each cell contains the eigengene significance (ES) value and its corresponding P (in parentheses) for the correlation between a module's eigengene and NS. The key modules—green, black, and blue—show the strong significant associations with NS. (E–G) Scatter plots of gene significance (GS) for NS vs. module membership (MM) in the (E) green, (F) black, and (G) blue modules, respectively. Genes with high intramodular connectivity (|MM| > 0.8) and strong association with NS (|GS| > 0.2) are considered key genes. 3.3. Consensus ML feature selection The intersection of the filtered DEGs, the core genes of GSEA, and the key genes of WGCNA yielded 14 overlapping genes ( Figure 4A ). Machine learning algorithms were applied to these 14 genes for feature selection, in which 9 feature genes were selected by LASSO regression ( Figure 4B ), 13 feature genes were selected by SVM-RFE ( Figure 4C ). In the Random Forest step, the final number of trees was 287 (selected based on the OOB error curve, Figure 4D ), and 4 feature genes were selected through cross-validation ( Figure 4E ). Finally, 4 feature genes were obtained by taking the intersection of the three ( Figure 4F ): ACSL1, CD8B, CD3D, and JAK2. Figure 4. Open in a new tab Integrative machine learning-based feature selection identifies a robust 4-gene signature for NS. (A) Venn diagram illustrating the overlap of candidate genes derived from three approaches: DEGs, GSEA, and WGCNA. The intersection yielded 14 consensus candidate genes for subsequent analysis. (B) Feature selection using LASSO. The left vertical dashed line indicates the optimal λ ( λ = 0.008) at which the model achieves minimum cross-validation error, selecting 9 feature genes (non-zero coefficients). (C) Feature selection using SVM-RFE. A total of 13 feature genes were selected (indicated by a vertical dashed line) at the point where both accuracy and kappa were maximized. (D) Determination of the optimal number of trees (ntree) for the RF model. The OOB error stabilizes at a minimum when ntree = 287 (indicated by a vertical dashed line), which was chosen as the optimal parameter. (E) Feature selection using RF-CV. The model achieved the lowest mean cross-validated error with 4 genes (indicated by a vertical dashed line). (F) Final consensus feature selection. The Venn diagram shows the overlap of the feature genes independently selected by LASSO (9 genes), SVM-RFE (13 genes), and RF-CV (4 genes). The intersection of all three methods yielded a highly robust set of 4 core feature genes. 3.4. Performance of the two-gene model The results of the univariate logistic regression analysis showed that the coefficients of ACSL1, CD8B, CD3D, JAK2, birth weight, and birth gestational age were statistically significant, as shown in Table 1 ; the multicollinearity test showed that there was a small risk of high correlation between these variables, and so they were included in the multivariate logistic regression analysis. The final results showed that the coefficients of ACSL1 and CD3D were statistically significant. The results of the multicollinearity test and multivariate logistic regression analysis are shown in Table 2 . We observed an extremely large and unstable odds ratio for antibiotic use in the univariate model, consistent with quasi-separation (antibiotics being administered predominantly after clinical suspicion/diagnosis). Therefore, we excluded this post-diagnosis intervention variable from the multivariate diagnostic model and explicitly acknowledge the instability of such clinical covariates in retrospective, small-sample settings. Table 1. Results of univariate logistic regression analysis. Variable * Coefficient SE * Statistic P OR * 95% CI * ACSL1 2.548 0.434 5.875 <0.001 12.782 [6.058, 33.852] CD8B −3.483 0.709 −4.913 <0.001 0.031 [0.006, 0.104] CD3D −2.800 0.487 −5.748 <0.001 0.061 [0.020, 0.140] JAK2 3.996 0.709 5.640 <0.001 54.380 [15.516, 253.8] Birth weight 2.817 0.495 5.696 <0.001 16.727 [6.761, 48.239] Birth gestational age 2.554 0.431 5.920 <0.001 12.858 [5.716, 31.335] Gender −0.484 0.355 −1.364 0.173 0.616 [0.305, 1.232] Antibiotics 19.428 1,118.623 0.017 0.986 2.738 × 10 8 [0, 3.928 × 10¹⁵⁶] Open in a new tab SE, standard error; OR, odds ratio; CI, confidence interval. Variables with P < 0.05 are highlighted in bold. Table 2. Results of multivariate logistic regression analysis. Variable * Multicollinearity Multivariate logistic regression Tolerance VIF Coefficient SE * Statistic P OR * 95% CI * ACSL1 0.566 1.766 1.529 0.683 2.239 0.025 4.614 [1.432, 21.568] CD8B 0.481 2.080 −0.369 0.992 −0.372 0.710 0.691 [0.085, 4.558] CD3D 0.395 2.531 −1.711 0.831 −2.058 0.040 0.181 [0.027, 0.791] JAK2 0.700 1.429 1.528 0.921 1.658 0.097 4.609 [0.830, 34.378] Birth weight 0.189 5.294 1.391 2.229 0.624 0.533 4.019 [0.067, 311.046] Birth gestational age 0.189 5.290 2.202 2.099 1.049 0.294 9.043 [0.264, 577.669] Open in a new tab SE, standard error; OR, odds ratio; CI, confidence interval. Variables with P < 0.05 are highlighted in bold. Using the potential NS biomarkers ACSL1 and CD3D as predictors, a logistic regression model was constructed on the training set. The model's output probabilities p were subsequently calculated for both the training and test sets. Taking a classification threshold of 0.5 as an example (i.e., diagnosing NS when p > 0.5), the decision boundaries on both the training and test sets showed that the model had good classification ability, with a classification accuracy of 0.883 on the training set and 0.902 on the test set ( Figure 5A ). ROC curves visualize the trade-off between sensitivity and specificity across all possible classification thresholds. The model demonstrated good predictive performance, with AUC values of 0.961 on the training set and 0.966 on the test set ( Figure 5B ). The calibration curves on both the training and test sets were essentially close to the ideal curve ( Figure 5C ), indicating that the model's prediction was reliable. Figure 5. Open in a new tab Evaluation of the model based on ACSL1 and CD3D expression. (A) Decision boundary analysis of the logistic regression model on the training set (top) and the test set (bottom). The dashed black line represents the decision boundary at a classification threshold of 0.5. The model achieved a classification accuracy of 88.3% on the training set and 90.2% on the test set. (B) ROC curves of the model on the training set (top) and the test set (bottom). The diagonal dashed line represents the performance of a random classifier (AUC = 0 .5). The AUC was 0.961 for the training set and 0.966 for the test set. (C) Calibration curves for the training set (top) and the test set (bottom). The diagonal dashed line (ideal line) represents the ideal scenario where predicted probabilities perfectly match observed frequencies. The apparent calibration curve (red) was derived from the original dataset, while the bias-corrected curve (green) was obtained after bootstrap internal validation (1,000 repetitions). The close alignment of the model's bias-corrected curves to the ideal line across both datasets. 3.5. Immune-deconvolution trends The results of CIBERSORT infiltration analysis were significant for all 135 samples. The infiltration abundance of 13 out of 22 immune cell types, including naive B cells, CD8 + T cells, etc., differed significantly between NS and control samples ( Figure 6A ). Because CIBERSORT relies on the LM22 reference signature matrix derived primarily from adult immune transcriptomes, we interpret the immune-deconvolution results as relative shifts/trends rather than exact quantification of neonatal immune-cell fractions. Figure 6. Open in a new tab Differential immune infiltration landscape and its correlation with ACSL1/CD3D. (A) Violin plots comparing the infiltration abundance of 22 immune cell types between the NS and Control groups, as estimated by the CIBERSORT algorithm. Each plot shows the distribution (median, density) of immune cell fractions. Significance levels are indicated as: ns ( P ≥ 0.05), * ( P < 0.05), ** ( P < 0.01), *** ( P < 0.001), **** ( P < 0.0001). (B,C) Scatter plots showing the correlation between the expression levels of biomarkers ACSL1/CD3D and the infiltration abundance of significantly altered immune cell subsets. The results of the correlation calculation ( Figures 6B,C ) showed that ACSL1 was positively correlated with neutrophils ( R = 0.770, P < 0.001), and negatively correlated with naive CD4 + T cells ( R = −0.754, P < 0.001), CD8 + T cells ( R = −0.649, P < 0.001), and that CD3D was positively correlated with naive CD4 + T cells ( R = 0.861, P < 0.001), CD8 + T cells ( R = 0.603, P < 0.001), and negatively correlated with neutrophils ( R = −0.670, P < 0.001). 4. Discussion Neonates, being a vulnerable population with underdeveloped immune, skin, and mucosal barriers, are prone to sepsis and may swiftly advance to infectious toxic shock and multi-organ failure upon infection. The symptoms and signs of NS resemble those of premature newborns, complicating clinical diagnosis ( 7 ). To enable prompt and efficient clinical identification of NS cases, it is essential to analyze and identify potential biomarkers of NS. This study screened 992 DEGs, determined 10 significantly enriched pathways comprising 129 core genes through GSEA, recognized 3 key modules totaling 256 key genes via WGCNA. Utilizing these three bioinformatics methods, 14 genes potentially associated with NS were isolated from a pool of 6,423 genes. The three machine learning algorithms were employed to select 4 feature genes from a set of 14 genes, then a “univariate followed by multivariate” logistic regression analysis was conducted. Ultimately, ACSL1 and CD3D were identified as potential biomarkers of NS. The CIBERSORT immune infiltration study indicated substantial variations in the abundance of 13 immune cell types, implying that various immunological mechanisms contribute to the pathogenesis of NS. ACSL1 showed strong correlations with neutrophils, naive CD4 + T cells, and CD8 + T cells, while CD3D exhibited similar associations but in the opposite directions. ACSL1 and immunometabolic activation. ACSL1 encodes a long-chain acyl-CoA synthetase involved in fatty-acid activation and immunometabolic reprogramming ( 29 ). In bacterial sepsis, innate immune cells—particularly neutrophils/monocytes—undergo metabolic rewiring that supports rapid effector functions. The upregulation of ACSL1 in our analysis is consistent with heightened innate immune activation and may reflect a metabolically primed inflammatory state in neonatal sepsis. Numerous transcriptome datasets have shown that ACSL1 expression is markedly elevated in patient blood, neutrophils, and mice macrophages following bacterial infection and inflammatory stimuli ( 30 , 31 ), corroborating the findings of the current work. CD3D, adaptive immune suppression, and “immunoparalysis”. CD3D is a structural component of the T-cell receptor (TCR)/CD3 complex and is critical for T-cell activation signaling ( 32 ). Its downregulation, together with the observed suppression of T-cell–associated pathways (e.g., TCR signaling) and reduced T-cell–related signals in immune deconvolution, supports the concept that neonatal sepsis involves adaptive immune dysfunction (T-cell exhaustion/immune paralysis) in addition to hyper-inflammation. A study has demonstrated the evidence of T cell exhaustion in spleens from sepsis patients ( 33 ). Thus, CD3D is not merely a generic T-cell marker; it provides a mechanistic window into sepsis-related immune suppression that may have value for immune-status monitoring. Comparison with existing biomarkers. Traditional biomarkers used in neonatal sepsis, such as CRP and procalcitonin, mainly capture non-specific acute inflammation and can be confounded by non-infectious conditions in neonates. In comparison with recently reported sepsis biomarker candidates, including targeting matrix metalloproteinase-9 (MMP9) to alleviate T cell exhaustion ( 34 ), a single-cell sequencing and machine learning-derived marker LILRA5 ( 35 ), B cell-derived ELL2 identified through multi-omics integration ( 36 ), and a monocyte differentiation-related signature developed via machine learning-based integration ( 37 ), our ACSL1 + CD3D pair possesses its own distinct advantages. It integrates two complementary dimensions—innate immunometabolic activation and adaptive immune suppression—potentially offering improved biological specificity. Limitations and future validation. Several limitations warrant emphasis. First, the model was trained and tested within a single GEO dataset ( GSE25504 ); even with a train/test split, this is not true external validation, and the reported AUC may represent an optimistic upper bound. Second, the sequential filtering strategy (DEGs → GSEA/WGCNA → ML → logistic regression) can introduce information leakage/optimism bias when candidate selection uses the full dataset; we therefore toned down claims and explicitly state this risk. Third, detailed clinical metadata (e.g., early- vs. late-onset sepsis and culture status) are not available in the public dataset, limiting phenotypic stratification. Fourth, antibiotic use is a post-diagnosis intervention and shows quasi-separation; such covariates are unstable in retrospective small cohorts. Finally, immune deconvolution based on adult-derived reference signatures (LM22) may be less accurate in neonates. Future work should include independent multi-center cohorts, platform-level replication, and wet-lab validation (e.g., qPCR) to confirm diagnostic and prognostic utility. Accordingly, our results should be interpreted as exploratory, hypothesis-generating biomarker discovery in a single public cohort, and diagnostic readiness will require validation across independent cohorts, platforms and institutions. Funding Statement The author(s) declared that financial support was not received for this work and/or its publication. Footnotes Edited by: Abdoul Karim Ouattara , University Norbert Zongo, Burkina Faso Reviewed by: Jingyuan Ning , Chinese Academy of Medical Sciences and Peking Union Medical College, China Deepika Kainth , All India Institute of Medical Sciences, India Data availability statement The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found in the article/ Supplementary Material . Ethics statement Ethical approval was not required for the study involving humans in accordance with the local legislation and institutional requirements. Written informed consent to participate in this study was not required from the participants or the participants’ legal guardians/next of kin in accordance with the national legislation and the institutional requirements. Author contributions NL: Data curation, Formal analysis, Investigation, Methodology, Project administration, Software, Writing – original draft, Writing – review & editing. ZT: Conceptualization, Data curation, Resources, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing. Conflict of interest The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. Generative AI statement The author(s) declared that generative AI was not used in the creation of this manuscript. Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us. Publisher's note All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher. Supplementary material The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fped.2026.1702073/full#supplementary-material Image1.pdf (151.3KB, pdf) References 1. Fleischmann-Struzek C, Goldfarb DM, Schlattmann P, Schlapbach LJ, Reinhart K, Kissoon N. The global burden of paediatric and neonatal sepsis: a systematic review. Lancet Respir Med. (2018) 6(3):223–30. 10.1016/S2213-2600(18)30063-8 [ DOI ] [ PubMed ] [ Google Scholar ] 2. Fleischmann C, Reichert F, Cassini A, Horner R, Harder T, Markwart R, et al. Global incidence and mortality of neonatal sepsis: a systematic review and meta-analysis. Arch Dis Child. (2021) 106(8):745–52. 10.1136/archdischild-2020-320217 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 3. Wu Y, Xia F, Chen M, Zhang S, Yang Z, Gong Z, et al. Disease burden and attributable risk factors of neonatal disorders and their specific causes in China from 1990 to 2019 and its prediction to 2024. BMC Public Health. (2023) 23(1):122. 10.1186/s12889-023-15050-x [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 4. Yin ML, Yin YL, Sang YF, Fu TL, Tian XL, Yang HQ, et al. Trend and prediction analysis of the changing disease burden of neonatalsepsis in China and globally from 1990 to 2021. Med J Wuhan Univ. (2025) 46(10):1307-14+1320. 10.14188/j.1671-8852.2024.0628 [ DOI ] [ Google Scholar ] 5. Singer M, Deutschman CS, Seymour CW, Shankar-Hari M, Annane D, Bauer M, et al. The third international consensus definitions for sepsis and septic shock (Sepsis-3). JAMA. (2016) 315(8):801–10. 10.1001/jama.2016.0287 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Shane AL, Sánchez PJ, Stoll BJ. Neonatal sepsis. Lancet. (2017) 390(10104):1770–80. 10.1016/S0140-6736(17)31002-4 [ DOI ] [ PubMed ] [ Google Scholar ] 7. Odabasi IO, Bulbul A. Neonatal sepsis. Sisli Etfal Hastan Tip Bul. (2020) 54(2):142–58. 10.14744/SEMB.2020.00236 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. Dai W, Zhou W. A narrative review of precision medicine in neonatal sepsis: genetic and epigenetic factors associated with disease susceptibility. Transl Pediatr. (2023) 12(4):749–67. 10.21037/tp-22-369 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Kim F, Polin RA, Hooven TA. Neonatal sepsis. Br Med J. (2020) 371:m3672. 10.1136/bmj.m3672 [ DOI ] [ PubMed ] [ Google Scholar ] 10. Cantey JB, Lee JH. Biomarkers for the diagnosis of neonatal sepsis. Clin Perinatol. (2021) 48(2):215–27. 10.1016/j.clp.2021.03.012 [ DOI ] [ PubMed ] [ Google Scholar ] 11. Subspecialty Group of Neonatology, the Society of Pediatrics, Chinese Medical Association; Editorial Board, Chinese Journal of Pediatrics. Expert consensus on diagnosis and management of neonatal bacteria sepsis (2024). Chin J Pediatr. 2024;62(10):931–40. 10.3760/cma.j.cn112140-20240505-00307 [ DOI ] [ PubMed ] [ Google Scholar ] 12. Kosyakovsky LB, Somerset E, Rogers AJ, Sklar M, Mayers JR, Toma A, et al. Machine learning approaches to the human metabolome in sepsis identify metabolic links with survival. Intensive Care Med Exp. (2022) 10(1):24. 10.1186/s40635-022-00445-8 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Zhang WY, Chen ZH, An XX, Li H, Zhang HL, Wu SJ, et al. Analysis and validation of diagnostic biomarkers and immune cell infiltration characteristics in pediatric sepsis by integrating bioinformatics and machine learning. World J Pediatr. (2023) 19(11):1094–103. 10.1007/s12519-023-00717-7 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Bacchelli C, Williams HJ. Opportunities and technical challenges in next-generation sequencing for diagnosis of rare pediatric diseases. Expert Rev Mol Diagn. (2016) 16(10):1073–82. 10.1080/14737159.2016.1222906 [ DOI ] [ PubMed ] [ Google Scholar ] 15. Zeng X, Qing J, Li CM, Lu J, Yamawaki T, Hsu YH, et al. Blood transcriptomic signature in type-2 biomarker-low severe asthma and asthma control. J Allergy Clin Immunol. (2023) 152(4):876–86. 10.1016/j.jaci.2023.05.023 [ DOI ] [ PubMed ] [ Google Scholar ] 16. Powers RK, Goodspeed A, Pielke-Lombardo H, Tan AC, Costello JC. GSEA-InContext: identifying novel and common patterns in expression experiments. Bioinformatics. (2018) 34(13):i555–64. 10.1093/bioinformatics/bty271 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Song C, Wang L, Zhang F, Lv C, Meng M, Wang W, et al. DUSP6 protein action and related hub genes prevention of sepsis-induced lung injury were screened by WGCNA and Venn. Int J Biol Macromol. (2024) 279(Pt 1):135117. 10.1016/j.ijbiomac.2024.135117 [ DOI ] [ PubMed ] [ Google Scholar ] 18. Uddin S, Khan A, Hossain ME, Moni MA. Comparing different supervised machine learning algorithms for disease prediction. BMC Med Inform Decis Mak. (2019) 19(1):281. 10.1186/s12911-019-1004-8 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 19. Choi RY, Coyner AS, Kalpathy-Cramer J, Chiang MF, Campbell JP. Introduction to machine learning, neural networks, and deep learning. Transl Vis Sci Technol. (2020) 9(2):14. 10.1167/tvst.9.2.14 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 20. Handelman GS, Kok HK, Chandra RV, Razavi AH, Lee MJ, Asadi H. Edoctor: machine learning and the future of medicine. J Intern Med. (2018) 284(6):603–19. 10.1111/joim.12822 [ DOI ] [ PubMed ] [ Google Scholar ] 21. Greener JG, Kandathil SM, Moffat L, Jones DT. A guide to machine learning for biologists. Nat Rev Mol Cell Biol. (2022) 23(1):40–55. 10.1038/s41580-021-00407-0 [ DOI ] [ PubMed ] [ Google Scholar ] 22. Clough E, Barrett T, Wilhite SE, Ledoux P, Evangelista C, Kim IF, et al. NCBI GEO: archive for gene expression and epigenomics data sets: 23-year update. Nucleic Acids Res. (2024) 52(D1):D138–44. 10.1093/nar/gkad965 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 23. Tang J, Kong D, Cui Q, Wang K, Zhang D, Gong Y, et al. Prognostic genes of breast cancer identified by gene co-expression network analysis. Front Oncol. (2018) 8:374. 10.3389/fonc.2018.00374 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 24. Kang J, Choi YJ, Kim IK, Lee HS, Kim H, Baik SH, et al. LASSO-based machine learning algorithm for prediction of lymph node metastasis in T1 colorectal cancer. Cancer Res Treat. (2021) 53(3):773–83. 10.4143/crt.2020.974 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 25. Nakao H, Imaoka M, Hida M, Imai R, Nakamura M, Matsumoto K, et al. Determination of individual factors associated with hallux valgus using SVM-RFE. BMC Musculoskelet Disord. (2023) 24(1):534. 10.1186/s12891-023-06303-2 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 26. Hu J, Szymczak S. A review on longitudinal data analysis with random forest. Brief Bioinform. (2023) 24(2):bbad002. 10.1093/bib/bbad002 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 27. Ledinger D, Nußbaumer-Streit B, Gartlehner G. WHO-Leitlinie: versorgung von frühgeborenen und neugeborenen mit niedrigem geburtsgewicht [WHO recommendations for care of the preterm or low-birth-weight infant]. Gesundheitswesen. (2024) 86(4):289–93. 10.1055/a-2251-5686 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 28. Newman AM, Liu CL, Green MR, Gentles AJ, Feng W, Xu Y, et al. Robust enumeration of cell subsets from tissue expression profiles. Nat Methods. (2015) 12(5):453–7. 10.1038/nmeth.3337 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 29. Qu G, Liu H, Li J, Huang S, Zhao N, Zeng L, et al. GPX4 Is a key ferroptosis biomarker and correlated with immune cell populations and immune checkpoints in childhood sepsis. Sci Rep. (2023) 13(1):11358. 10.1038/s41598-023-32992-9 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 30. Misheva M, Kotzamanis K, Davies LC, Tyrrell VJ, Rodrigues PRS, Benavides GA, et al. Oxylipin metabolism is controlled by mitochondrial β-oxidation during bacterial inflammation. Nat Commun. (2022) 13(1):139. 10.1038/s41467-021-27766-8 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 31. Roelands J, Garand M, Hinchcliff E, Ma Y, Shah P, Toufiq M, et al. Long-chain acyl-CoA synthetase 1 role in sepsis and immunity: perspectives from a parallel review of public transcriptome datasets and of the literature. Front Immunol. (2019) 10:2410. 10.3389/fimmu.2019.02410 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 32. Yang Q, Feng Z, Ding D, Kang C. CD3D And CD247 are the molecular targets of septic shock. Medicine (Baltimore). (2023) 102(29):e34295. 10.1097/MD.0000000000034295 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 33. Hotchkiss RS, Monneret G, Payen D. Sepsis-induced immunosuppression: from cellular dysfunctions to immunotherapy. Nat Rev Immunol. (2013) 13(12):862–74. 10.1038/nri3552 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 34. Wang X, Ning J, Zhou L, Li H, Cui J, Wang J, et al. Targeting matrix metalloproteinase-9 to alleviate T cell exhaustion and improve sepsis prognosis. Research. (2025) 8:0996. 10.34133/research.0996 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 35. Ning J, Fan X, Sun K, Wang X, Li H, Jia K, et al. Single-cell sequence analysis combined with multiple machine learning to identify markers in sepsis patients: lILRA5. Inflammation. (2023) 46(4):1236–54. 10.1007/s10753-023-01803-8 [ DOI ] [ PubMed ] [ Google Scholar ] 36. Wang X, Li J, Ning J. Comprehensive multi-omics integration reveals B cell-derived ELL2 as a novel diagnostic and prognostic biomarker in sepsis. Med Res. (2025) 1(2):170–4. 10.1002/mdr2.70010 [ DOI ] [ Google Scholar ] 37. Ning J, Sun K, Wang X, Fan X, Jia K, Cui J, et al. Use of machine learning-based integration to develop a monocyte differentiation-related signature for improving prognosis in patients with sepsis. Mol Med. (2023) 29(1):37. 10.1186/s10020-023-00634-5 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Image1.pdf (151.3KB, pdf) Data Availability Statement The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found in the article/ Supplementary Material . Articles from Frontiers in Pediatrics are provided here courtesy of Frontiers Media SA ACTIONS View on publisher site PDF (2.8 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top