Classification of invasive bloodstream infections and Plasmodium falciparum malaria using autoantibodies as biomarkers

Abstract

A better understanding of disease-specific biomarker profiles during acute infections could guide the development of innovative diagnostic methods to differentiate between malaria and alternative causes of fever. We investigated autoantibody (AAb) profiles in febrile children (≤ 5 years) admitted to a hospital in rural Ghana. Serum samples from 30 children with a bacterial bloodstream infection and 35 children with Plasmodium falciparum malaria were analyzed using protein microarrays (Protoplex Immune Response Assay, ThermoFisher). A variable selection algorithm was applied to identify the smallest set of AAbs showing the best performance to classify malaria and bacteremia patients. The selection procedure identified 8 AAbs of which IFNGR2 and FBXW5 were selected in repeated model run. The classification error was 22%, which was mainly due to non-Typhi Salmonella (NTS) diagnoses being misclassified as malaria. Likewise, a cluster analysis grouped patients with NTS and malaria together, but separated malaria from non-NTS infections. Both current and recent malaria are a risk factor for NTS, therefore, a better understanding about the function of AAb in disease-specific immune responses is required in order to support their application for diagnostic purposes.

Figure 1

Variable importance of the eight autoantibodies based on the selected random forest model. Identifiers: FBXW5 F-box-containing protein 5, IFNGR2 Interferongamma receptor 2, HAUS8 8-Subunit Human Augmin Complex, ODAM Odontogenic, ameloblast associated, TAF6 Transcription initiation factor TFIID subunit 6, BTN2A2 Butyrophilin subfamily 2 member A2, CCDC134 Coiled-coil domain containing 134, GP9 Glycoprotein IX (platelet).

Figure 2

Cluster analysis. (A) Multidimensional scaling (MDS) map summarizing patient’s proximity in the final random forest model. Clusters are numbered and indicated by dashed lines. (B) The bars show the proportion of diagnoses allocated to the three clusters.

Figure 3

AAb induction levels. The induction levels of the eight selected autoantibodies are shown for controls, non-NTS, NTS and malaria patients. Identifiers are explained in Table 1. RFU relative fluorescence unit, NTS non-Typhi Salmonella, non-NTS bacterial species other than NTS.

Figure 4

Classification error. The plot displays the robustness of the variable selection models by summarizing the distribution of classification errors over 100 marker selection algorithms. The x-axis shows the number of markers in a model and the y-axis the classification error of the respective random forest models. The change in classification error within the repeated marker selection algorithms are shown by the gray lines. The applied model is displayed by the red line. The boxplots show summary statistics about the number of AAb in finally selected models (y-axis) and the lowest classification errors (y-axis) in the repeated models.

Figure 5

AAb selection model. Number of times where AAbs were selected in a repeated model. Markers selected by the applied model are colored dark gray and markers only selected in the repeated models are colored in white. Identifiers of AAbs not listed in Table 2: TEP4 (Transducin-like enhancer protein 4), KBTBD7 (kelch repeat and BTB (POZ) domain containing 7).