Establishment of Elevated Serum Levels of IL-10, IL-8 and TNF-β as Potential Peripheral Blood Biomarkers in Tubercular Lymphadenitis: A Prospective Observational Cohort Study

Abstract

Background: Tubercular lymphadenitis (TL) is the most common form of extra-pulmonary tuberculosis (TB) consisting about 15-20% of all TB cases. The currently available diagnostic modalities for (TL), are invasive and involve a high index of suspicion, having limited accuracy. We hypothesized that TL would have a distinct cytokine signature that would distinguish it from pulmonary TB (PTB), peripheral tubercular lymphadenopathy (LNTB), healthy controls (HC), other lymphadenopathies (LAP) and cancerous LAP. To assess this twelve cytokines (Tumor Necrosis Factor (TNF)-α, Interferon (IFN) -γ, Interleukin (IL)-2, IL-12, IL-18, IL-1β, IL-10, IL-6, IL-4, IL-1Receptor antagonist (IL-1Ra), IL-8 and TNF-β, which have a role in pathogenesis of tuberculosis, were tested as potential peripheral blood biomarkers to aid the diagnosis of TL when routine investigations prove to be of limited value.

Methods and findings: A prospective observational cohort study carried out during 2010-2013. This was a multi-center study with three participating hospitals in Delhi, India where through random sampling cohorts were established. The subjects were above 15 years of age, HIV-negative with no predisposing ailments to TB (n = 338). The discovery cohort (n = 218) had LNTB (n = 50), PTB (n = 84) and HC (n = 84). The independent validation cohort (n = 120) composed of patients with cancerous LAP (n = 35), other LAP (n = 20) as well as with independent PTB (n = 30), LNTB (n = 15) and HC (n = 20). Eight out of twelve cytokines achieved statistical relevance upon evaluation by pairwise and ROC analysis. Further, variable selection using random forest backward elimination revealed six serum biosignatures including IL-12, IL-4, IL-6, IL-10, IL-8 and TNF-β as optimal for classifying the LNTB status of an individual. For the sake of clinical applicability we further selected a three analyte panel (IL-8, IL-10 and TNF-β) which was subjected to multinomial modeling in the independent validation cohort which was randomised into training and test cohorts, achieving an overwhelming 95.9% overall classifying accuracy for correctly classifying LNTB cases with a minimal (7%) misclassification error rate in the test cohort.

Conclusions: In our study, a three analyte serum biosignatures and probability equations were established which can guide the physician in their clinical decision making and step wise management of LNTB patients. This set of biomarkers has the potential to be a valuable adjunct to the diagnosis of TL in cases where AFB positivity and granulomatous findings elude the clinician.

Fig 1. Overall Study Design including the numbers of subjects included in the Discovery and Validation Cohorts.

The indicated study design also enlists the inclusion criteria for different classes of subjects in the discovery and validation cohorts. The cytokines selected from the discovery cohort were tested in the validation cohort. Among the LAP subjects in the validation cohort, LNTB subjects from discovery cohort were also included for multinomial modelling to check for their accurate classification by the model. Two approaches namely the balanced and proportionate sampling were used. The model was trained using the cytokine levels of the subjects from training set and then applied to a test set to determine their discriminating ability. The LNTB patients were attempted for a follow-up after the completion of their ATT therapy but only eight subjects could be enrolled.

Fig 2. Backward elimination identifies the most important cytokines for the models.

Random Forest backward elimination procedure to further eliminate the cytokines that may be unimportant in the presence of interactions. At each step of the backward elimination procedure, random forest with 10000 trees was run and variable importance and model errors were saved. The variable with the least importance was then purged out, and the procedure repeated until only one variable remained. The cytokines are not individually knocked out, but are sequentially knocked out from right to left as indicated by the arrow. Model errors were then plotted as a function of the order of the purged out variables. The minimum error was obtained with six cytokines (IL-4, IL-10, TNF-β, IL-8, IL-6 and IL-12) as shown in figure. The error rate increases after IL-18 is purged out, therefore implying that all the ones following IL-4 are important. One variable that remains in the end and is not shown in the figure is IL-12.

Fig 3. Heatmap of pairwise discriminatory power of individual cytokine in the validation cohort.

Each panel is representing a cytokine. Phenotypes are on x and y axes. Each combination of phenotypes is filled by -log p value obtained from Tukey’s posthoc testing after ANOVA. If any combination of two phenotypes is filled with red/pink, then the particular cytokine is significantly different between the two groups and thus may help in distinguishing them.

Fig 4. Comparison of mean serum levels of IL-10, IL-8 and TNF-β between the discovery and validation cohorts and the change in levels after treatment with ATT.

“v” added to the group name indicates the value from validation panel. Kruskal-Wallis test followed by Mann-Whitney U test was used to compare the groups as the data was not normally distributed. The results indicated that no significant difference in the mean serum cytokine levels was seen among the derivation and validation panel among the groups (A) for IL-10 [(LNTB vs LNTBv (p = 0.928), PTB vs PTBv (p = 0.265), HC vs HCv (p = 0.215)] and (B) IL-8 [(LNTB vs LNTBv (p = 0.197), HC vs HCv (p = 0.862) except for PTB vs PTBv (p = 0.000),] but for (C) TNF-β LNTB vs LNTBv (p = 0.016), PTB vs PTBv (p = 0.031) except HC vs HCv (p = 0.914)] showed significant difference between the groups, (D-F) Showing the change in levels of the investigated cytokines after completion of ATT. Each dot represent an individual and the change in their corresponding levels (n = 8). A statistically significant reduction in the serum levels was observed. (D) The mean serum levels for the eight followed up individuals went from being 63.06 + 11.6 pg/ml to 20.11 ± 0.08 pg/ml for IL-10 (p = 0.0002), (E) from 776.4 ± 515 pg/ml to 30.17 ± 6.8 pg/ml for IL-8 (p = 0.0002) and (F) from 2251 ± 1721 to 110.5 ± 217. 5 pg/ml for TNF-β p = 0.0009).

Fig 5. Conditional inference tree as applied to cytokine concentrations from all lymphadenopathies.

For each inner node, the Bonferroni adjusted P-values are given. Concentrations are in ng/ml. For example in the terminal node 5, nine samples fulfilled the sequential criteria: (TNF- β < 2739.25) and (IL-8 < 34.659) and (TNF- β < 506) and (TNF β < 219.07). The bars represent probabilities of belonging to respective classes, in this case 60% probability of patient being cancerous, 40% probability of the patient being Other LAP, and 0% chance of being LNTB. Decision tree was made on full data and since the multinomial logistic model was validated on a held-out set.