Exploring fecal microbiota signatures associated with immune response and antibiotic impact in NSCLC: insights from metagenomic and machine learning approaches

Abstract

Background: Substantial interstudy heterogeneity in cancer immunotherapy-associated biomarkers has hindered their clinical applicability. To address this challenge, we performed a comprehensive integration of publicly available global metagenomic datasets. By leveraging metagenomic profiling and machine learning approaches, this study aimed to elucidate gut microbial signatures associated with immune response in lung cancer (LC) and to evaluate the modulatory effects of antibiotic exposure.

Methods: A systematic literature search was conducted to identify relevant datasets, resulting in the inclusion of 209 fecal metagenomic samples: 154 baseline samples (45 responders, 37 non-responders, and 72 healthy controls) and 55 longitudinal samples collected during immunotherapy. We performed taxonomic and functional characterization of gut microbiota (GM) differentiating responders from non-responders, delineated microbiome dynamics during treatment, and assessed the impact of antibiotics on key microbial taxa. Among eight machine learning algorithms evaluated, the optimal model was selected to construct a predictive framework for immunotherapy response.

Results: Microbial α-diversity was significantly elevated in responders compared to non-responders, with antibiotic administration further amplifying this difference-most notably at the species level. Integrative multi-omics analysis identified two pivotal microbial biomarkers, s_Bacteroides caccae and s_Prevotella copri, which were strongly associated with immunotherapy efficacy. A random forest-based classifier achieved robust predictive performance, with area under the curve (AUC) values of 0.82 and 0.79 at the species and genus levels, respectively. Notably, P. copri was further enriched in responders with poor progression-free survival (PFS <3 months), indicating a potential deleterious role. Antibiotic exposure significantly influenced the abundance and functional potential of these key taxa. KEGG-based functional analysis revealed the enrichment of amino acid metabolism pathways in responders. Additionally, CARD database annotation demonstrated that the majority of antibiotic resistance genes were associated with Bacteroidetes and Proteobacteria, implicating these taxa in shaping microbial-mediated therapeutic responses.

Conclusions: This study represents the first large-scale, cross-cohort integration of metagenomic data to identify reproducible GM signatures predictive of immune checkpoint inhibitor efficacy in LC. The findings not only underscore the prognostic relevance of specific taxa but also establish a foundation for developing microbiome-informed, personalized immunotherapeutic strategies.

Keywords: antibiotics; gut microbiota; immunotherapy; lung cancer; machine learning; metagenome.

Figure 1

Variance explained by response status (response vs. non-response) is plotted against variance explained by the confounding factor effects for individual microbial species. The significantly differential microbiota are colored in blue, and P-values were from the two-way ANOVA test. The abundance of each taxa is plotted proportionally to the dot size. ANOVA, analysis of variance.

Figure 2

Species composition and difference analysis of response group and non-response group based on total population. (A) α-diversity analysis with richness as the index. (B) Baseline gut microbiota PCOA analysis results of the two groups. (C) Similarity analysis of baseline microbiota between the two groups. (D) Stacking maps of top 10 abundance at the phylum, genus, and species levels between responders and non-responders. (E) Phylogenetic branching map between the two groups based on the results of the difference analysis. (F) Results of LEfSe difference analysis.

Figure 3

Interaction of differential markers between responder and non-responder groups within different populations. (A) Intra-community and inter-community interactions of the responder microbiota in the responder population. (B) Intra-community and inter-community interactions of responders in non-responders. (C) Intra-community and inter-community interactions of non-responsive bacteria in responders. (D) Intra-community and inter-community interactions of non-responsive bacteria in non-responsive population.

Figure 4

Screening and construction of immune efficacy prediction models. (A) Comparison of prediction efficiency and performance of different machine learning methods. (B) The RF model was established and validated based on the genus-level differential markers obtained by LEfSe analysis. (C) RF modeling was performed at the species level based on the differential species obtained by LEfSe analysis.

Figure 5

Key species affecting immunotherapy prognosis and stabilizing biomarkers. (A) To explore the difference of microbiota between the long PFS and short PFS groups based on the total population. (B) In the absence of antibiotic interference, the difference of gut microbiota between the long and short PFS groups. (C) Difference of microbiota between the two groups when they experienced antibiotic intervention. (D) Biomarkers that remained stable in the response group under the results of various bioinformatics analyses. (E) Biomarkers that remained stable in the non-responder group across various bioassays.

Figure 6

Changes in bacterial diversity and marker abundance during immunotherapy. (A) To investigate the changes in alpha diversity of gut microbiota at different time points after immunotherapy and the effect of antibiotics on this trend. (B) Changes of s_Bacteroides caccae in the response group at different time points after immunotherapy. (C) Changes in the marker s_Prevotella copri in the non-response group at different time points after immunotherapy.

Figure 7

Functional analysis of gut microbiota in patients receiving immunotherapy. (A) A level 2 functional pathway statistical map was drawn based on KEGG annotation results. (B) Species attribution analysis of resistance genes between responders and non-responders when the effect of antibiotics was not considered (C) results of differential resistance genes between the two groups of patients.