Value of Cytokine Expression in Early Diagnosis and Prognosis of Tumor Metastasis

Abstract

Objective: To investigate the association of the plasma level of cytokines and blood routine indexes with clinical characteristics in patients with cancer.

Methods: We analyzed plasma samples derived from 134 cancer patients. Interleukins (IL) 1β, 2, 4, 5, 6, 8, 10, 12p70, 17, IFN-γ, IFN-α, and TNF-α, and blood routine indexes were measured. The associations of the levels of cytokine and blood routine indexes with demographic and clinical characteristics of cancer patients were analyzed. Partial least-squares discriminant analysis was employed to identify cancer metastasis using these plasma cytokine metrics as input. We compared the predictive effectiveness of numeric machine learning algorithms using these indexes and showed a promising model implemented with random forest.

Results: Plasma levels of IL-6 and IL-8 in cancer patients with metastases were higher than those without metastases (P < 0.05). Cancer patients without metastases had significantly higher levels of plasma IL-12p70 and percentage of lymphocytes as compared with those with metastases (P < 0.05). Our random forest model showed the highest prediction performance (upper quantile AUC, 0.885) among the six machine learning algorithms we evaluated.

Conclusion: Our findings suggest that plasma levels of IL-6, IL-8, and IL-12p70 and the percentage of lymphocytes could predict the recurrence, metastasis, and progression of cancer. Our findings will provide guidance for tumor monitoring and treatment.

Figure 1

Data analysis flowchart.

Figure 2

Association of tumor metastasis and plasma cytokines and blood routine indexes. (a) Cytokines and blood routine indexes related to metastasis and nonmetastasis. Wilcoxon test, ^∗P < 0.05,^∗∗P < 0.01, ^∗∗∗P < 0.001; (b) the volcano plot revealed 5 indicators (red points) were upregulated and 4 (blue points) were downregulated in tumor metastasis; (c) the heatmap showed the relationship between cytokines and metastasis; (d) the heatmap showed the relationship between blood routine indexes and metastasis. Abbreviations. Hb, hemoglobin; LY, lymphocyte; MO, monocyte; NEUT, neutrophil; PLT, platelet; WBC, white blood cell.

Figure 3

Key indicators to distinguish tumor metastases and nonmetastases. (a and b) PLS-DA analysis showed tumor metastasis and nonmetastasis distinguished by certain indicators (PERMANOVA test, P=1e − 04); (c) PLS-DA analysis showed the plasma levels of IL-6, IL-8, lymphocyte percentage, and IL-12p70 were the main influencing factors that distinguished metastases and nonmetastases (VIP scores ≥1). Abbreviation. VIP, variable importance in the projection.

Figure 4

(a) Random forest is the best machine learning algorithm for predicting tumor metastasis. The model with the random forest algorithm showed the highest accuracy (0.84) and its kappa index was 0.38. (b) The relationship between variables and accuracy in the random forest predictor, with the repeated five-fold cross validation. (c) The variable importance rank in the random forest model are lymphocyte percentage, IL-6, IL-12p70, and IL-8. (d) The AUC value of ROC curves is 0.885, which indicated that the random forest model was robust for predicting tumor metastasis. Abbreviation: RF, random forests; LDA, linear discriminant analysis; SVM, support vector machine, KNN, K-nearest neighbours; CART, classification and regression tree; LY, lymphocyte.

Figure 5

The relationship between tumor metastasis with plasma cytokines and blood routine indexes in gender subgroups. Wilcoxon test, ^∗P < 0.05,^∗∗P < 0.01, ^∗∗∗P < 0.001; ns, no significance.