Background <p>Sepsis is mostly caused by bacterial infections and requires a prompt diagnosis. There is a need for improved diagnostics by differentiating between gram-negative and gram-positive bacterial infections.</p> Methods <p>The plasma levels of 285 unique proteins in patients with gram-negative infection (<i>n</i> = 154), gram-positive infection (<i>n</i> = 92), and in healthy controls (<i>n</i> = 35) were quantified using proximity extension assay. Three machine learning algorithms; random forest, recursive feature elimination, and adaptive least absolute shrinkage and selection operator (Lasso) were employed to identify discriminative proteins, with their effectiveness assessed using accuracy metrics. The selected proteins were further evaluated for their ability to differentiate between gram-negative and gram-positive infections through logistic regression and area under the receiver operating characteristic curve.</p> Results <p>We identified 55 discriminative proteins differentiating between gram-negative and gram-positive infections using the Lasso, the best performing algorithm. The discriminative proteins achieved AUROC values of 0.69 for gram-negative infections and 0.66 for gram-positive infections, both compared to the remaining groups, and 0.58 for differentiating between the two infection groups. Comparative statistical analysis revealed no significant differences in protein expression between gram-negative and gram-positive patients.</p> Conclusions <p>We identified 55 proteins with some discriminative potential between gram-negative and gram-positive infections. However, the overall predictive performance was low and did not exceed that of established single biomarkers. These findings highlight the challenges of applying a multimarker approach in infection classification and emphasize the need for further studies using larger and more diverse cohorts, as well as broader analytical methods, to explore their potential clinical utility.</p> Clinical trial <p>Not applicable.</p>