Unique Microbial Characterisation of Oesophageal Squamous Cell Carcinoma Patients with Different Dietary Habits Based on Light Gradient Boosting Machine Learning Classifier

Abstract

Objectives: The microbiome plays an important role in cancer, but the relationship between dietary habits and the microbiota in oesophageal squamous cell carcinoma (ESCC) is not clear. The aim of this study is to explore the complex relationship between the microbiota in oesophagal tissue and dietary habits in ESCC patients. Methods: 173 ESCC patients were included. The method of 16S rRNA sequencing was used to analyze microbial composition and diversity. The LEfSe and Boruta methods were used to screen important microbes, and the LightGBM algorithm distinguished microbes associated with different dietary habits. PICRUST2 and DESeq2 predicted microbial function and screened differential functions. The Pearson test was used to analyze correlations between microbes and functions, and SPARCC microbial symbiotic networks and Cytoscape were used to determine microbial interactions. Results: Significant differences in microbial composition were observed among ESCC patients with different dietary habits. LEfSe and Boruta identified three, six, and two significantly different bacteria in the FF/FP, FF/PF, and FF/PP groups, respectively, with AUC values of 0.683, 0.830, and 0.715. PICRUST2 and DESeq2 analysis revealed 3, 11, and 5 significantly different metabolic pathways in each group. Eubacterium_B sulci was positively correlated with PWY-6285, PWY-3801, and PWY-5823. PWY-6397 was positively correlated with undefinded (Fusobacterium_C). Microbial network analysis confirmed unique microbial characteristics in different diet groups. Conclusions: Different dietary habits lead to alterations in Eubacterium_B sulci and undefinded (Fusobacterium_C) and related functional pathways.

Keywords: LightGBM; diet; microorganisms; oesophageal squamous cell carcinoma.

Figure 1

Differential analysis of microbiota of ESCC patients in different dietary habit groups. (A) α-diversity violin plots (based on the Shannon diversity index, the Faith phylogenetic diversity index, the observational characterisation index, and the Pielou homogeneity index for the PP, PF, FP, and FF groups; (B) β-diversity PCoA plots (based on unweighted UniFrac distance, Jaccard distance, weighted UniFrac distance, and Bray–Curtis distance) for PP, PF, FP, and FF groups; (C) microbial composition of ESCC patients in PP, PF, FP, and FF groups at the phylum and genus levels; (D) microbial differences in ESCC patients in PP, PF, FP, and FF groups at the genus and species levels; (E) box line plots of important differential bacteria in the FF/FP, FF/PF, and FF/PP groups screened by the Boruta algorithm, red: rejected, green: accepted, blue: shadow, yellow: tentative; (F) ROC curves of Boruta screening results (FF/FP, FF/PF, and FF/PP groups) by LightGBM machine learning algorithm; (G) box line plots of Clr_relative abundance differences among FF, FP, PF, and PP groups for important differential bacteria screened by the Boruta algorithm for FF/FP, FF/PF, and FF/PP groups, red: PP group, blue: PF group, orange: FP group, green: FF group.

Figure 2

Functional and network analysis of microbiota of ESCC patients with different dietary habits. (A) Differential functional volcano plots of FF/FP, FF/PF, and FF/PP groups screened by DESeq2 algorithm; (B) scatter plot of correlation between s__undefined (g__Fusobacterium_C) and PWY-6397 in FF, FP, PF, and PP groups; (C) bubble heat map of correlation analysis of differential function with important bacteria screened by Boruta; (D) bacterial network analysis plots for FF, FP, PF, and PP groups.