Metabolic biomarkers of neonatal sepsis: identification using metabolomics combined with machine learning

Background: Sepsis is a common disease associated with neonatal and infant mortality, and for diagnosis, blood culture is currently the gold standard method, but it has a low positivity rate and requires more than 2 days to develop. Meanwhile, unfortunately, the specific biomarkers for the early and timely diagnosis of sepsis in infants and for the determination of the severity of this disease are lacking in clinical practice.

Methods: Samples from 18 sepsis infants with comorbidities, 25 sepsis infants without comorbidities, and 25 infants with noninfectious diseases were evaluated using a serum metabolomics approach based on liquid chromatography‒mass spectrometry (LC‒MS) technology. Differentially abundant metabolites were screened via multivariate statistical analysis. In addition, least absolute shrinkage and selection operator (LASSO) and support vector machine recursive feature elimination (SVM-RFE) analyses were conducted to identify the key metabolites in infants with sepsis and without infections. The random forest algorithm was applied to determine key differentially abundant metabolites between sepsis infants with and without comorbidities. Receiver operating characteristic (ROC) curves were generated for biomarker value testing. Finally, a metabolic pathway analysis was conducted to explore the metabolic and signaling pathways associated with the identified differentially abundant metabolites.

Results: A total of 189 metabolites exhibited significant differences between infectious infants and noninfectious infants, while 137 distinct metabolites exhibited differences between septic infants with and without comorbidities. After screening for the key differentially abundant metabolites using LASSO and SVM-RFE analyses, hexylamine, psychosine sulfate, LysoPC (18:1 (9Z)/0:0), 2,4,6-tribromophenol, and 25-cinnamoyl-vulgaroside were retained for the diagnosis of infant sepsis. ROC curve analysis revealed that the area under the curve (AUC) was 0.9200 for hexylamine, 0.9749 for psychosine sulfate, 0.9684 for LysoPC (18:1 (9Z)/0:0), 0.7405 for 2,4,6-tribromophenol, 0.8893 for 25-cinnamoyl-vulgaroside, and 1.000 for the combination of all metabolites. When the septic infants with comorbidities were compared to those without comorbidities, four endogenous metabolites with the greatest importance were identified using the random forest algorithm, namely, 12-oxo-20-trihydroxy-leukotriene B4, dihydrovaltrate, PA (8:0/12:0), and 2-heptanethiol. The ROC curve analysis of these four key differentially abundant metabolites revealed that the AUC was 1 for all four metabolites. Pathway analysis indicated that phenylalanine, tyrosine, and tryptophan biosynthesis, phenylalanine metabolism, and porphyrin metabolism play important roles in infant sepsis.

Conclusion: Serum metabolite profiles were identified, and machine learning was applied to identify the key differentially abundant metabolites in septic infants with comorbidities, septic infants without comorbidities, and infants without infectious diseases. The findings obtained are expected to facilitate the early diagnosis of sepsis in infants and determine the severity of the disease.