2.1 Study design and sample collection

Between January 2018 and December 2019 at the Department of Lung Cancer Surgery, Tianjin Medical University General Hospital, we collected cancerous and adjacent normal tissues from 30 consenting patients with primary NSCLC who underwent surgical resection for primary lung cancer.

Inclusion criteria were as follows: (1) a definitive, primary pathological NSCLC diagnosis; (2) patients in the 30–80 year old age range; (3) patients not receiving any anti-tumour therapy, such as chemotherapy, targeted therapy, immunotherapy or radiotherapy; (4) patients showing no traces of acute, exacerbated chronic obstructive pulmonary disease, complex pneumonia, infectious bronchiectasis, acute asthma or bronchitis, no purulent or fever or grey phlegm and no other malignant diseases; (5) patients with no history of hormone use and antibiotic-related treatments at least 1 month before surgery.

Tissue samples were aseptically collected at surgery, transferred to cryovials, promptly snap-frozen in liquid nitrogen and stored in the Tianjin Lung Cancer Institute. Matched adjacent normal tissue samples were situated at >5 cm from the tumour tissue edge. Overall survival (OS) was defined as the period extending from surgery to death or the last follow-up time for patients who remained alive. Progression-free survival (PFS) was the time from surgery to disease progression or death resulting from any cause.

2.2 16S rRNA sequencing

Microbiota gDNA was extracted from tissue samples. Polymerase chain reaction (PCR) was performed using primers targeting the V3–V4 region of 16S rRNA (338F: 5′-ACTCCTACGGGAGGCAGCAG-3′ and 806R: 5′-GGACTACHVGGGTWTCTAAT-3′). After purification and quality assessment, final products were sequenced on an Ion S5™ XL platform (Thermo Fisher Scientific, USA). Clean data were generated from the preliminary quality control of raw data (removing barcodes, primers and chimeras, and splicing and filtering low-quality sequences). Uparse software was used to cluster clean reads into operational taxonomic units (OTUs) with 97% identity. Mothur software and the SILVA132 SSUrRNA database were used for OTU annotation analysis.^{14, 15}

2.3 Transcriptome sequencing

All samples underwent RNA extraction in accordance with manufacturer's protocols of Trizol (Invitrogen, Carlsbad, California, USA). RNA quantity and quality were measured by gel electrophoresis and a NanoPhotometer spectrophotometer (Implen GmbH, Munich, Germany). Total RNA quantity for library construction was not <1 µg. Next, using the NEBNext UltraTM RNA Library Prep Kit (New England Biolabs, New Jersey, USA), and after mRNA enrichment, fragmentation processing, reverse transcription to cDNA, adapter addition, PCR amplification and product purification, libraries were generated. After insert size quantification and examination, libraries were sequenced on the Illumina sequencing platform (Illumina, California, USA). After filtering, verifying sequencing error rates and inspecting GC content, clean reads were acquired for analyses. Subread (version 1.5.0, https://subread.sourceforge.net/) was used for quantitative analysis. Reads with mapping quality values lower than 10, reads from non-paired alignments and reads mapped to multiple regions of the genome were filtered out respectively. Finally, an original read count expression matrix was generated for bioinformatics analysis.

2.4 Comparison of microbial diversity

The Chao1 and Shannon indices are used to evaluate alpha diversity. Permutational multivariate analysis of variance tests were used to compared beta diversity (Bray–Curtis distances) across tissue samples. Principal co-ordinates analysis was used for the visualisation of beta diversity.

2.5 Constructing a diagnosis model

DEseq2 was used to perform differential genus analysis on NSCLC and its subtypes (LUAD and LUSC) and matched normal tissues. A generalised linear model was used to calculate if a single or microbial model could discriminate between tumour and normal tissues. The pROC R package was used to demonstrate diagnostic model accuracy.

2.6 Screening prognosis-related microbiota

Univariate Cox regression analysis was used to identify microbiota that were associated with OS and PFS (threshold; p < .05). Forest plots were also generated to show hazard ratios (HR) for prognosis-related microbiota. Venn diagrams were also used to analyse any compositional interrelationships between microbiota related to OS and PFS.

2.7 Microbial ecological network analyses

An integrated network analysis pipeline was used to analyse microbial ecological networks.¹⁶ The sparse correlations for compositional data (SparCC) method was adopted for microbial ecological network analyses.^{17, 18} The parameters of SparCC are as follows: (1) The number of inference iterations to average over was 20; (2) the number of exclusion iterations to remove strongly correlated pairs was 10; (3) the number of shuffled times was 100, and the two-sided p value was calculated. Cytoscape (v.3.9.1) was used to generate a final network diagram.¹⁹

2.8 Constructing a prognosis model

OTU counts were normalised using log₂ (counts + 1), and a LASSO-Cox regression approach to construct a microbiota-related prognosis model based on OS using the ‘glmnet’ R package and to also predict PFS. A penalty parameter (λ) value was decided in accordance with the minimum partial likelihood deviation. The lambda.1se was 0.2586844 and the lambda.min was 0.1624613 in LASSO regression analyses. Risk scores were calculated as follows: risk score = sum (each genus count × corresponding coefficient). Kaplan–Meier survival curves and receiver operating characteristic (ROC) analyses were conducted to test prognosis model accuracy. In Peptococcus analyses for prognosis outcomes, GSE31210 and GSE50081 datasets were used to verify relationships between Peptococcus-related genes and prognosis.^{20, 21}

2.9 Mendelian randomisation

Summary-level data from genome-wide association studies (GWAS) associated with the gut microbiota were retrieved from the MiBioGen website (https://mibiogen.gcc.rug.nl/).^{22, 23} The studies encompassed 24 cohorts consisting of 18 340 participants and 211 taxa (131 genera, 35 families, 20 orders, 16 classes and nine phyla). Summary-level GWAS data associated with lung cancer (ieu-a-a966), LUAD (ieu-a-a965) and LUSC (ieu-a-a967) were sourced from the International Lung Cancer Consortium, which is an international organisation for lung cancer researchers.²⁴ We used all summary data from published studies and publicly available GWAS abstracts; therefore, no additional ethical approval or consent was required. To satisfy the three core Mendelian randomisation (MR) assumptions: relevance assumption (the genetic variant is associated with the exposure), independence assumption (the genetic variant is independent of confounding factors that affect the relationship between the exposure and outcome) and exclusion restriction assumption (the genetic variant affects the outcome only through the exposure), we conducted the following single nucleotide polymorphism (SNP) screen associated with exposure: (1) p value < 1e−5; (2) SNP removal from the major histocompatibility region; (3) linkage disequilibrium analysis (r² < 0.001, kb = 10 000); (4) SNP removal with outcome p values < .05; (5) SNP removal associated with confounders (PhenoScanner); (6) the F statistic was calculated using: F = $$\frac{{{{\beta }^2}}}{{s{{e}^2}}},$$²⁵ and SNPs were removed at F < 10; and (7) SNPs with palindromic structures were removed. Finally, we obtained 12 SNPs for MR analysis. In MR exposure (Peptococcus) and outcome analysis (lung cancer), ‘TwoSampleMR’ and ‘MRPRESSO’ R packages were used.²⁶ Inverse variance weighted analysis was the primary analysis method. Weighted median and PRESSO were validation methods. Result sensitivity and heterogeneity were also analysed. Reverse analysis conditions were the same as for forward MR analysis.

2.10 Statistical analysis

T-tests were used to analyse statistical variance between groups for which specific methods were not delineated. The Pearson method was used for correlation analysis between the microbiota and genes. The edgeR R package was used to identify differentially expressed genes (DEGs). The algorithm genewise negative binomial generalised linear model with quasi-likelihood tests was also used. All statistical tests were two-tailed, and statistical significance was determined at p < .05. The false discovery rate (FDR), used as a sole method, was used to adjust p values. R (v.4.3.2) was used to conduct analyses. Non-significant (NS) values were indicated by p > .05, * indicated .01 < p < .05, ** indicated .001 < p < .01 and *** indicated p < .001.

3.1 Population demographics and clinical characteristics

We gathered 30 cancerous and adjacent normal tissue pairs from primary NSCLC patients, encompassing 17 LUAD and 13 LUSC patients. Of these, 23 were male and seven were female. The median age was 64 years and the age range was 46–83 years. Participants predominantly had stages II–IV NSCLC, with 19 lymph node metastasis and 19 smoker cases. At final follow-up, 16 patients experienced relapse (53.3%) and 14 died (33.3%). The median OS was 952 days, while the median PFS was 526.5 days. Patient clinical characteristics are shown (Tables S1 and S2).

3.2 16S rRNA and RNA sequencing

From 16S rRNA sequencing, the average number of reads from tumour and adjacent normal tissue samples was 77 192 and 76 981, respectively. All effective sample tags were clustered into OTUs with 97% identity, after which 10 665 OTUs were acquired. After annotation and screening, 1022 genera underwent bioinformatics analysis. From transcriptome sequencing, average raw counts numbered 91 193 198, and average clean counts after filtering numbered 88 007 549. Finally, 57 299 RNAs were annotated and used in analyses.

3.3 Distinct microbial composition in cancerous tissues when compared with normal tissues

Alpha diversity was higher in cancerous tissues (NSCLC, LUSC and LUAD) when compared with matched normal tissues, irrespective of Chao1 or Shannon indices. This difference was not statistically significant and was probably due to low sample size. Intriguingly, the Chao1 index of LUSC samples tends to be higher than that of LUAD samples, while the trend of the Shannon index is the opposite (Figure 1A).

We next analysed beta diversity. Statistically significant differences were observed between NSCLC and matched normal tissues (p = .017; Figure 1B). After differentiating subtypes, we observed beta diversity differences between LUSC and matched normal tissues (LUSC vs. normal_SC: p = .036; Figure 1C). However, no diversity disparities were identified for LUAD (LUAD vs. normal_AD: p = .266; Figure 1D). Therefore, in contrast to normal tissues, LUSC tissues had a more distinctive microbial composition.

3.4 Microbiota comparisons across tissue types

To further determine microbial composition, we listed high-abundance microbiota (HAM) (top 10 in average rankings) at phylum, class, order and family levels, and also low-abundance microbiota (LAM) (others) across NSCLC histology types. HAM composition at the four levels was completely different across tissue types. At the phylum level, Proteobacteria, Cyanobacteria, Acidobacteria and Chloroflexi in LUAD were all higher than in other types but total LAM levels in LUAD were the lowest. Conversely, Acidobacteria and Verrucomicrobia in LUSC were substantially lower than in other types, but total LAM levels were the highest (Figure 2A). Unidentified microbiota were regarded as non-important research subjects.

At the class level, eight HAM classes mainly originated from four phyla: Proteobacteria (including Gammaproteobacteria, Alphaproteobacteria and DeltaproteobacteriaI classes), Firmicutes (including Clostridia, Bacilli and Erysipelotrichia classes), Bacteroidetes (including the Bacteroidia class) and Verrucomicrobia (including the Verrucomicrobiae class). Moreover, when compared with other types, Gammaproteobacteria and Alphaproteobacteria classes and the Proteobacteria phylum to which the two classes belonged were both the most abundant in LUAD (Figure 2B).

By comprehensively analysing microbial composition at these four levels, ‘Other’ levels in LUAD were consistently lower across tissue types (Figure S1). To further examine microbial composition differences between LUAD and LUSC samples, we identified 19 differential microbial genera (p < .01). The top 16 genera are shown, of which 14 were highly abundant in LUAD and two in LUSC (Figure 2C). Our microbial ecological network also showed that correlations for these differential genera in LUAD and LUSC were significantly different (Figure 2D). This indicates that the interconnections of differential microbial genera in LUAD and LUSC are different. Therefore, we believe that LUAD microbial composition is dissimilar to LUSC.

3.5 Differential microbial genera analyses between cancerous and normal tissues

We also conducted microbial genus analyses between NSCLC, LUAD, LUSC and normal tissues. Ultimately, 56 genera showed abundance disparities, of which 12 displayed significant differences (differences were present across all types) between cancer and normal tissues. These consisted of 11 HAM genera in tumour (Anaerovorax, Marivivens, Donghicola, Lachnospira, Dubosiella, Lactobacillus, Methylobacterium, Akkermansia, Paenibacillus, Aerococcus and Cloacibacterium) and 1 HAM genera in normal tissue (Campylobacter, p < .05; Figure 3A). We next assessed genera accuracy in differentiating NSCLC and normal tissues. Critically, area under the curve (AUC) values for five genera exceeded 0.6, including Aerococcus (AUC = 0.629), Paenibacillus (AUC = 0.643), Lachnospira (AUC = 0.626), Cloacibacterium (AUC = 0.622) and Dubosiella (AUC = 0.633; Figure 3B–F). Any individual genus is unable to accurately predict whether the sample is NSCLC tumour tissue or normal benign tissue. From these observations, we have attempted to provide a microbial NSCLC diagnostic model (Table S3), which had a diagnostic accuracy = 0.862 (Figure 3G). This model still requires validation by more cohorts in the future.

To understand the roles of 11 HAM genera in NSCLC, we performed correlation and functional analyses. Our ecological network showed complex correlations among the 11 genera, but no closely related clusters were observed (Figure 3H). We then integrated differential expression and correlation analyses (Figure 3I) and identified 580 genera-related genes. From Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis, these aforementioned microbiota were potentially implicated in epithelial cell differentiation and metabolism-related functions (Figure 3J).

3.6 The identification of prognosis-related microbiota

We performed a prognosis-related analysis of multi-level microbiota using univariate Cox regression; one phylum, one class, five orders, eight families and 18 genera were significantly associated with OS, while four phyla, five classes, nine orders, 16 families and 28 genera were significantly related to PFS (Figure 4A,B). Therein, one phylum (Rokubacteria), one class (Thermoplasmata), five orders (Sphingomonadales, Methanomassiliicoccales, Frankiales, Micropepsales, Glycomycetales), seven families (Sphingomonadaceae, Methanomassiliicoccaceae, Micropepsaceae, Barnesiellaceae, Heliobacteriaceae, Kiloniellaceae and Glycomycetaceae) and 10 genera (Sphingobacterium, Xanthobacter, Pantoea, Methylophilus, Oscillospira, Hydrogenispora, Woeseia, Pusillimonas, Peptococcus and Glycomyces) were significantly related to both OS and PFS (Figure 4C). These microbiota also exhibited consistent beneficial or detrimental relationships with OS and PFS, further signifying significant role in NSCLC prognosis outcomes.

3.7 Functional analysis of prognosis-related microbial clusters

We initially conducted a correlation analysis on 36 prognosis-related microbial genera. After eliminating weak correlations, 23 genera were presented in the ecological network (Figure 5A). Of these, a positive correlation was identified among nine genera, all of which were associated with a longer PFS (HR < 1). Therefore, these nine genera were called the protective microbial cluster (PMC). We also discovered three positively correlated genera, all of which were associated with shorter OS and PFS times (HR > 1); therefore, these were known as the harmful microbial cluster (HMC).

To further explore potential cluster functions, we used univariate Cox regression and correlation analysis. The PMC was related to 214 PFS genes (p < .01) and the HMC was related to 186 OS genes (FDR < 0.05) and 153 PFS genes (p < .01, Tables S4–S6). After KEGG enrichment analyses, the PMC was mainly related to metabolic-related pathways, including glucagon signalling, arginine and proline metabolism, lipoic acid metabolism, biosynthesis of nucleotide sugars and pyruvate metabolism (Figure 5B). The HMC was mainly related to immunity and inflammation (Th17 cell differentiation, T cell receptor signalling, cytokine–cytokine receptor interactions, nuclear factor-kappa B (NF-κB) signalling, natural killer cell-mediated cytotoxicity, neutrophil extracellular trap formation), microbial infections (Yersinia infection and coronavirus disease—COVID-19) and tumour-related signalling (phosphatidylinositol 3-kinase-protein kinase B (Akt) signalling; Figure 5C,D). Putatively, the PMC mainly exerted protective effects via metabolic-related functions, while the HMC mainly affected prognosis via infection invasion, thereby inducing changes in inflammation-, immunity- and tumour-related pathways.

3.8 Constructing a microbial prognostic model and analysing prognostic difference mechanisms

Ten genera showed correlations with both OS and PFS outcomes. From screening (LASSO and multivariate Cox regression analyses), six genera (Xanthobacter, Pantoea, Oscillospira, Hydrogenispora, Peptococcus and Glycomyces) were used to construct a prognostic model in a cohort of 28 NSCLC patients (Figure 6A,B and Table S7). The model accurately distinguished OS in patients (p < .0001), with AUC values for 1, 3 and 5 years = 0.9916, 1.0000 and 0.9649, respectively (Figure 6C,D). The model also forecasted PFS in patients (p < .0001), with AUC values for 1, 3 and 5 years = 0.9487, 0.9066 and 0.8623, respectively (Figure 6E,F). The model also performed well in predicting OS and PFS outcomes in patients with LUAD and LUSC (LUAD-OS: p < .0001, LUSC-OS: p = .0006, LUAD-PFS: p = .0001, LUSC-PFS: p = .0020; Figure S2). Therefore, our model, composed of six genera, showed good predictive accuracy for NSCLC prognoses in patients.

In comparing high and low-risk groups, no significant disparity was identified in alpha and beta diversity indices. Alpha diversity in the low-risk group was marginally higher than in the high-risk group (Figure S3A,B), suggesting that patients with a poorer prognosis may have had a more detrimental TME. This unfavourable TME may not be suitable for the survival of certain microbiota. And we hypothesise that the TME may play an active role in screening intratumoural microbiota. Hence, the microbiota may potentially function as a valuable and novel biomarker. Additionally, correlation analyses showed that among the six genera (all detrimental factors), Hydrogenispora and Pantoea, Peptococcus and Xanthobacter, and Oscillospira and Glycomyces exhibited tighter correlations (Figure S3C). These genera were considerably prevalent in the high-risk group, in accordance with prognosis results (Figure S3D).

To further explore if the microbiota affected tumour gene expression, we performed DEG analyses and observed distinct gene expression profiles between high and low-risk groups. Of these, 82 genes were up-regulated in the low-risk group and 63 up-regulated in the high-risk group (FDR < 0.05 and |LogFC| > 1; Figure 6G). To further determine potential mechanisms, a correlation analysis was performed between the six genera and DEGs, with ultimately 37 related genes identified. KEGG enrichment results demonstrated that infections (malaria, Staphylococcus aureus and Hepatitis B), inflammation (interleukin (IL)-17 and NF-κB signalling) and immune regulation (B cell receptor signalling) were the principal pathways whereby the microbiota affected prognosis outcomes (Figure 6H). This was analogous to HMC prognosis-related pathways. Thus, the microbiota may impact prognosis outcomes by causing infection and inducing inflammation and immune responses.

3.9 The Peptococcus genus may affect prognosis via tumour necrosis factor signalling

Given that LUAD and LUSC microbial composition varied, we separately conducted a prognostic screen of microbial genera in LUAD and LUSC (Figure 7A–D), which showed compositional disparities in prognosis-related microbial genera across tissue types. In particular, Peptococcus shows the most positive results in terms of prognosis for NSCLC, LUAD and LUSC (p < .05; Figure 7E). Also, Peptococcus levels in the high-risk group were significantly higher than those in the low-risk group (p < .01; Figure S3D). Hence, Peptococcus showed a significant association with a poorer prognosis.

Using MR, we identified a significant causal relationship between Peptococcus and lung cancer (p = .017, odds ratio (OR) = 1.149) and LUSC (p = .014, OR = 1.246), but no relationship with LUAD (p = .602, OR = 1.047; Figure 8A). Similarly, in a reverse MR analysis, no causal relationship was identified between lung cancer and Peptococcus. The results of the MR analysis indicate that Peptococcus may have a promoting effect on NSCLC, but NSCLC has no influence on Peptococcus. Therefore, Peptococcus is very likely a new genus that affects NSCLC prognosis outcomes.

To explore potential pathways whereby Peptococcus influenced prognosis, we divided 28 patients into high-abundance (abundance > the median value) and low-abundance groups (abundance ≤ the median value) based on Peptococcus levels. In a prognostic multivariate Cox regression analysis of these levels together with age, smoking, stage, subtype and lymph node status, Peptococcus was identified as an independent prognostic factor (Figure 8B). We then used DESeq2 and edgeR methods to identify 108 and 107 DEGs between groups, respectively, of which 49 were common DEGs (Figure 8C). After KEGG enrichment analysis, Tumour necrosis factor (TNF) signalling was the main enriched pathway (Figure 8D). We then conducted correlative analyses between enriched genes in the TNF pathway and Peptococcus abundance and found that Peptococcus exhibited a marked positive correlation with C-X-C motif chemokine ligand 1 (CXCL1, R = 0.46, p = .013), Interferon regulatory factor 1 (IRF1, R = 0.51, p = .0056) and Prostaglandin-endoperoxide synthase 2 (PTGS2, R = 0.50, p = .0068), and a substantially negative correlation with C-X3-C motif chemokine ligand 1 (CX3CL1, R = −0.49, p = .0086; Figure S4). These observations suggested that TNF signalling may be a pathway whereby Peptococcus impacts on prognosis outcomes. We also identified 13 core gene products using the protein–protein interaction (PPI) network, with aforementioned gene products showing core positions in this network (Figure 8E). Next, in univariate Cox regression analysis, these genes were associated with prognosis outcomes (CXCL1: HR = 1.63, 95% confidence interval (CI): 1.08–2.45, p < .05; IRF1: HR = 2.63, 95% CI: 1.00–6.90, p < .05; PTGS2: HR = 1.65, 95% CI: 1.04–2.64, p < .05 and CX3CL1: HR = 0.56, 95% CI: 0.33–0.93, p < .05; Figure 8F).

Considering that most patients were in stages II–IV, we also conducted the same analysis in two stages I–II NSCLC datasets and found that CXCL1 was significantly associated with prognosis in the GSE31210 dataset (HR = 1.42, 95% CI: 1.12–1.81, p < .01) and PTGS2 in the GSE50081 dataset (HR = 1.18, 95% CI: 1.05–1.32, p < .01; Figure 8G,H). Thus, Peptococcus may lead to an unfavourable prognosis in NSCLC patients by influencing TNF signalling.