Prediction and verification of benignancy and malignancy of pulmonary nodules based on inflammatory related biological markers

Abstract

Objective: Inflammation plays an important role in the transformation of pulmonary nodules (PNs) from benign to malignant. Prediction of benignancy and malignancy of PNs is still lacking efficacy methods. Although Mayo or Brock model have been widely applied in clinical practices, their application conditions are limited. This study aims to construct a diagnostic model of PNs by machine learning using inflammation-related biological markers (IRBMs).

Methods: Inflammatory related genes (IRGs) were first extracted from GSE135304 chip data. Then, differentially expressed genes (DEGs) and infiltrating immune cells were screened between malignant pulmonary nodules (MN) and benign pulmonary nodule (BN). Correlation analysis was performed on DEGs and infiltrating immune cells. Molecular modules of IRGs were identified through Consistency cluster analysis. Subsequently, IRBMs in IRGs modules were filtered through Weighted gene co-expression network analysis (WGCNA). An optimal diagnostic model was established using machine learning methods. Finally, external dataset GSE108375 was used to verify this result.

Results: 4 hub IRGs and 3 immune cells showed significantly difference between MN and BN, C1 and C2 module, namely PRTN3, ELANE, NFKB1 and CTLA4, T cells CD4 naïve, NK cells activated and Monocytes. IRBMs were screened from black module and yellowgreen module through WGCNA analysis. The Support vector machines (SVM) was identified as the optimal model with the Area Under Curve (AUC) was 0.753. A nomogram was established based on 5 hub IRBMs, namely HS.137078, KLC3, C13ORF15, STOM and KCTD13. Finally, external dataset GSE108375 verified this result, with the AUC was 0.718.

Conclusion: SVM model established by 5 hub IRBMs was able to effectively identify MN or BN. Accumulating inflammation and immune dysfunction were important to the transformation from BN to MN.

Keywords: Immune cells infiltration; Inflammation-related biological markers; Nomogram; Pulmonary nodules; Support vector machine.

Fig. 1

Flow of this study.

Fig. 2

Screening of Differentially expressed genes (DEGs) between MN and BN. A-B. Box plot and Heatmap showed that 10 IRGs has significant difference between MN and BN. C-D. Correlation analysis showed that PRTN3 and ELANE was negatively related to CTLA4 and NFKB1.

Fig. 3

Immune cells infiltration between MN and BN. A-B. Monocytes and T cells gamma delta of MN were significantly higher than BN, while B cells memory, T cells CD4 naïve, T cells regulatory (Tregs) and NK cells activated of MN were significantly lower than BN. C. CTLA4 and NFKB1 were positively correlated to T cells CD4 naïve, and negatively correlated to Monocytes. While PRTN3 and ELANE were positively correlated to Monocytes, and ELANE were negatively correlated to T cells CD4 naïve and NK cells activated.

Fig. 4

Identification of IRGs molecular modules by Consistency cluster analysis. A. consensus matrix k = 2. B. consensus index changed with k value. C. Cluster consensus. D. Relative change in area under CDF curve.

Fig. 5

Screening of Differentially expressed genes (DEGs) between C1 and C2 IRGs modules. A-B. Box plot and Heatmap showed that 4 IRGs has significant difference between C1 and C2 IRGs modules. C. PCA analysis showed that patients with PNs were able to distinguish according to the expression of IRGs.

Fig. 6

Immune cells infiltration between C1 and C2 IRGs modules. A-B. B cells naïve, T cells CD8, T cells CD4 naïve, T cells CD4 memory resting, NK cells resting, NK cells activated, Monocytes, Mast cells resting and Eosinophils were significantly different between C1 and C2 modules.

Fig. 7

Weighted gene co-expression network analysis between IRGs molecular modules. A-B. Scale independence and Mean connectivity changed with the soft threshold. C. Gene dendrogram and module colors. D. Module-trait relationships. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)

Fig. 8

Establishment of predictive model of malignancy and benignancy pulmonary nodules by machine learning methods. A. ROC curve showed that the AUC of four models was 0.749, 0.753, 0.694 and 0.500, separately. B. Boxplots of residual. C. Feature importance created for the GLM, RF, SVM and XGB models.

Fig. 9

Construction of nomogram based on 5 IRBMs. A. Prediction of risk of disease according to the scores of 5 IRBMs. B. Net benefit of different decisions. C. Confidence of fit between actual probability and predicted probability.

Fig. 10

Verification of the nomogram by external dataset.