Identification of Novel Serodiagnostic Signatures of Typhoid Fever Using a Salmonella Proteome Array

Abstract

Current diagnostic tests for typhoid fever, the disease caused by Salmonella Typhi, are poor. We aimed to identify serodiagnostic signatures of typhoid fever by assessing microarray signals to 4,445 S. Typhi antigens in sera from 41 participants challenged with oral S. Typhi. We found broad, heterogeneous antibody responses with increasing IgM/IgA signals at diagnosis. In down-selected 250-antigen arrays we validated responses in a second challenge cohort (n = 30), and selected diagnostic signatures using machine learning and multivariable modeling. In four models containing responses to antigens including flagellin, OmpA, HlyE, sipC, and LPS, multi-antigen signatures discriminated typhoid (n = 100) from other febrile bacteremia (n = 52) in Nepal. These models contained combinatorial IgM, IgA, and IgG responses to 5 antigens (ROC AUC, 0.67 and 0.71) or 3 antigens (0.87), although IgA responses to LPS also performed well (0.88). Using a novel systematic approach we have identified and validated optimal serological diagnostic signatures of typhoid fever.

Keywords: Salmonella Typhi; antibody response; controlled human infection model; enteric fever; fever diagnostics; machine learning; rapid diagnostic tests; serodiagnostics.

Figures

FIGURE 1

Structure of the controlled human infection models of typhoid fever and endemic cohort. In both (A) the discovery set and (B) the validation set, study participants ingested 103–104 CFU Salmonella Typhi Quailes strain suspended in oral sodium bicarbonate solution on day 0 (D0). Sera samples were collected and probed at the time points indicated. Participants developing an oral temperature ≥38°C sustained for ≥12 h or evidence of bacteremia after challenge were diagnosed with typhoid (TD) and commenced on antimicrobial treatment. All remaining participants not diagnosed during the 14-day period (nTD) were commenced on the same treatment on day 14. (C) Samples (serum and blood culture) in the endemic setting cohorts were collected on one occasion at point of hospital presentation. Pathogens isolated from blood cultures collected from other, non-S. Typhi bacteraemia cases are listed in the box.

FIGURE 2

Reactivity of 4445 antigens in samples from a human challenge study performed in Oxford (Discovery set). (A) Mean number of creactive antigen per participant at time points D0 – D28 following challenge. Blue: nTD group. Red: TD group. Gray shaded area: acute disease (TD48 and TD96 h). (B) Number of reactive antigens in number of samples at TD+96 h, and D14 for both groups.

FIGURE 3

Time course of responses to four selected antigen/antibody isotype combinations by participants challenged and subsequently diagnosed with typhoid fever. (A) IgA responses. (B) IgG responses. (C) IgM responses. (D) Responses to purified S. Typhi flagellin (0.1 μg) and lipopolysaccharide (LPS, 0.1 μg) as additional antigens included on the array. Vertical black dashed line, TD time point; vertical green dashed line, TD+48hr time point; vertical blue dashed line, TD+96hr time point.

FIGURE 4

Reactivity and diagnostic performance of antigen/antibody isotype combinations selected from the discovery set and applied to the validation set. Antigen names are given in Supplementary Table S1. Boxplots of fold-change in reactivity between TD+96hr and day 14 time points in TD or day 14 in nTD participants and individual baseline FI values, to antigen/antibody isotype combinations selected from the discovery set. Paired t-tests were performed between the time point featured and baseline values. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001.

FIGURE 5

Selection of an antigen/antibody isotype signature using machine learning algorithms applied to the controlled human infection typhoid datasets. (A) Test set prediction performance measures AUC receiver operator characteristic (ROC) and balanced accuracies (BalAcc) for four different machine learning models using 500 bootstrap samples of the data. (B) Frequency of features selected in each of 500 iterations by the partial least squares (PLS) algorithm. (C) Proportions of features selected across all 500 bootstrap samples using the PLS algorithm. Features had to be selected in at least 10% of the bootstrap samples (column ‘overall’). Proportions are split by classifier size. The last column represents the overall proportion across all 500 bootstrap samples.

FIGURE 6

Multivariable analysis to find optimal antigen/antibody isotype signature. (A) Fold-change values of 12 target antigens plus flagellin and LPS in the combined Oxford data. Antigens included in one of the final models are indicated by the colored squares below the antigens. (B) Risk scores for the Oxford samples based on the antigens and coefficients in the four final models. (C) ROC curves for each of the final model based on the Oxford data comparing participants diagnosed with typhoid fever (day 14 or TD96 h; brown) with those who stayed well (day 14; blue). (D) Fold-change values of 12 target antigens plus flagellin and LPS in the Nepal data. Fold-changes were generated against the median of the healthy population (n = 50). Antigens included in one of the final models are indicated by the colored squares below the antigens. (E) Risk scores for the Nepalese febrile control (orange) and typhoid (brown) samples based on the antigens and coefficients in the final model. (F) External validation ROC curves for each of the final risk equation fitted to the Nepali data comparing febrile controls against typhoid cases.