Identification and validation of intratumoral microbiome associated with sensitization to immune checkpoint inhibitors

Summary

As a part of the commensal microbiome, the regulatory role of the intratumoral microbiome in tumor immunity is gradually revealed. However, the relationship between the intratumoral microbiome and the efficacy of immune checkpoint inhibitors (ICIs) clinical treatment remains unclear. Here, we collect RNA sequencing (RNA-seq) data and clinical information from publicly available ICIs therapy cohorts. By developing an improved bioinformatics pipeline to identify the intratumoral microbiome and performing a comprehensive association analysis, we find that the intratumoral microbiome is associated with response to ICIs and characteristics of the tumor microenvironment (TME). In vivo experiments demonstrate that intratumoral injection of Burkholderia cepacia, Priestia megaterium, or Corynebacterium kroppenstedtii, which were selected from our analysis results, would synergize with anti-PD-1 therapy to inhibit tumor growth and enhance antitumor immunity. Our findings highlight the essential role of the intratumoral microbiome in the clinical effectiveness differences of ICIs, suggesting its potential in future ICIs combination therapy.

Graphical abstract

Highlights

• •Intratumoral microbiome correlates with immunotherapy response
• •Microbial features linked with efficacy are associated with the characteristics of the TME
• •Microbial species related to improved outcomes enhance antitumor immunity in mice

Chen et al. utilize RNA sequencing data of immune checkpoint inhibitors (ICIs)-treated cohorts and apply weighted gene coexpression network analysis (WGCNA) to construct separate networks for the intratumoral microbiome. Integrative association analysis and validation in mice reveal critical links among the intratumoral microbiome, ICIs efficacy, and tumor immunity.

Introduction

Immune checkpoint inhibitors (ICIs) are now widely used in the treatment of a variety of malignant tumors (e.g., melanoma, non-small cell lung cancer [NSCLC], and renal cell carcinoma [RCC]) and are considered to be a major breakthrough in cancer treatment. However, only a sizable minority of patients benefited from ICIs, and failed treatment is very common due to primary and acquired resistance.¹ There is an urgent need for us to explore the mechanism of resistance and find approaches to improve the clinical benefit of ICIs.

With the development of sequencing technology and bioinformatics, increasing evidence demonstrated the vital role of gut microbiome in ICIs treatment outcomes.^2,3 A few ongoing clinical trials investigating the combination of fecal microbiota transplantation with ICIs have shown exciting results, highlighting the potential application of the targeted commensal microbiome in enhancing immunotherapy efficacy.⁴ Recently, some findings have revealed that the intratumoral microbiome within the tumor microenvironment (TME) is linked with cancer development, prognosis, and treatment response.^5,6,7,8,9 Combined with clinical samples and preclinical models, studies revealed that certain intratumoral microbes could promote cancer progression, metastasis, and chemoradiation resistance.^10,11,12,13 Meanwhile, intratumoral (i.t.) injection of certain probiotics enhanced the efficacy of ICIs in subcutaneous tumor models.^14,15 However, a comprehensive investigation of the associations between the intratumoral microbiome and the clinical benefit of ICIs is still lacking.

In this study, firstly, intratumoral microbial reads from RNA sequencing (RNA-seq) data of 7 ICIs clinical cohorts were extracted with a bioinformatic pipeline. Weighted gene coexpression network analysis (WGCNA) was utilized to construct separate networks for the intratumoral microbial information and conduct association analysis with the efficacy of ICIs. Several microbial modules associated with the response or survival of ICIs-treated patients were identified. Second, correlation analyses between those microbes within associated modules and deconvolved tumor-infiltrating immune cell compositions, as well as immune response signatures, were performed, to determine the associations between the intratumoral microbiome and the TME. Finally, we substantiated the efficacy of Burkholderia cepacia, Priestia megaterium, and Corynebacterium kroppenstedtii from the results of our analysis through validation in mouse models of melanoma and preliminarily explored the potential mechanisms. This study unveils the potential impact of the intratumoral microbiome in clinical ICIs treatment and presents an intratumoral microbial-based adjunct to future immunotherapy.

Results

Extracting intratumoral microbial information from ICIs therapy data

To identify intratumoral microbial features related to the response to ICI treatment, we collected clinical information from 8 ICI therapy cohorts (Table S1).^{16,17,18,19,20,21,22,23} After screening, we chose 7 cohorts and downloaded RNA-seq sequencing data of tumors, including melanoma, metastatic gastric cancer (mGC), NSCLC, resectable esophageal adenocarcinoma (rEAC), and RCC (Figure 1A). The baseline status of the TME determines the efficacy of ICIs. Considering the relationship between the microbiome and immunity, the analysis of pretreatment samples may provide more clues about the intratumoral microbiome in relation to ICI therapy.²⁴ By developing a computational processing flow based on the combination of SAMtools and Kraken2, we extracted "nonhuman" reads and characterized the intratumoral microbiome at the phylum, genus, and species level (Tables S2A–S2U; STAR Methods). We observed that rEAC has the highest fractional microbial reads and RCC has the lowest fractional microbial reads (Figure 1B). Furthermore, after filtering the contaminants and low coverage species (Tables S3A–S3D; STAR Methods), we found that only 74 species have been identified in over 10% of samples of the RCC dataset, probably due to its low sequencing depth. Given the difficulty of acquiring valid results from the sparse microbial information of RCC, we finally selected the remaining 6 cohorts for further analysis (Table S4).

Figure 1.

Extracting intratumoral microbiome from transcriptome of ICI trials

(A) Flowchart of the study that details the samples utilized at each stage of statistical analysis. mGC, metastatic gastric cancer; rEAC, resectable esophageal adenocarcinoma; RCC, renal cell carcinoma; mUC, metastatic urothelial cancer; NSCLC, non-small cell lung cancer; R, responders group; NR, non-responders group; OS, overall survival; PFS, progression-free survival.

(B) Fraction of bacterial-derived reads from the total number of human-mapped reads.

See also Table S1.

The intratumoral microbiome landscape of ICIs treatment cohorts

We first sought to analyze the composition of the intratumoral microbiome. Proteobacteria, Actinobacteria, Firmicutes, and Bacteroidetes occupied a large proportion of the intratumoral microbiome in all 6 cohorts (Figures 2A and 2B). We observed that the profiles of melanoma, NSCLC, and mGC were dominated by Proteobacteria and Actinobacteria, whereas rEAC had a higher level of Bacteroidetes and Firmicutes compared to the other cohorts. It was consistent with the previous reports, demonstrating the reliability of our pipeline for extracting the intratumoral microbiome.^25,26,27 For melanoma, the microbial composition of the melanoma_Gide dataset was similar to the melanoma_Hugo dataset, while the melanoma_Riaz dataset had a higher abundance of Firmicutes and Euryarchaeota, suggesting that the intratumoral microbial composition of the same cancer also varies slightly across different cohorts.

Figure 2.

Profiling the intratumoral microbiome

(A) Profiling the intratumoral microbiome of ICI-pretreatment samples (phylum level).

(B) Average abundance radar plot across datasets and phylum, with Actinobacteria, Bacteroidetes, Firmicutes, Proteobacteria, and other phyla.

(C) Principal coordinate analysis based on the Euclidean distance of the R and NR groups across datasets. p < 0.05 was considered as significant.

See also Figure S1; Table S2. Intratumoral microbial counts from all cohorts (before decontamination), related to STAR Methods, Table S3. Intratumoral microbial counts in the species level (after decontamination), related to STAR Methods, Table S4. Information of pretreatment samples from the cohorts entered the further analysis, related to STAR Methods.

To improve the computational efficiency of downstream analyses, we removed the batch effect of three melanoma cohorts to increase the sample size (n = 117) and retained their common species (Figure S1B; STAR Methods). To compare the difference in microbial composition between responders (Rs) and non-responders (NRs), we analyzed the beta diversity by calculating the Euclidean distance. The results show that intratumoral microbial profiles significantly differed between Rs and NRs in melanoma (p = 0.025), while there were no significant differences in mGC, rEAC, and NSCLC (p > 0.05), perhaps due to the limitation of sample sizes (Figure 2C). These analyses indicated that our computational flow for extracting intratumoral microbial information from human samples of ICI therapy datasets is reliable, and yet we still need to apply rational methods to mine the key intratumoral microbial features in a limited sample size.

Identifying intratumoral microbial modules associated with response to ICIs

To further identify biological associations between the intratumoral microbiome and clinical outcomes of patients treated with ICIs, we applied WGCNA to construct separate weighted microbiome co-occurrence networks (Figures 3A and S2A–S2D; Table S5; STAR Methods). Using the function modulePreservation in the WGCNA package, we analyzed the stability and preservation of these modules, and significantly conserved modules were selected for further analyses by combining both preservation median rank and Z summary score (Figures S2E–S2H; STAR Methods). First, the associations of the eigenvalues of microbial modules with response to ICIs were evaluated (Figure 3B). For melanoma, brown (odds ratio [OR] 1.42, 95% confidence interval [CI] 1.01–2.02; p = 0.05) and turquoise (OR 1.46, 95% CI 1.04–2.09; p = 0.03) modules were significantly positively associated with response to ICIs. Positive association with response was found for the yellow (OR 2.05, 95% CI 1.07–4.37; p = 0.02) module of mGC. Conversely, the eigenvalues of magenta modules were significantly associated with a lack of response in rEAC (OR 0.45, 95% CI 0.19–0.90; p = 0.04). Based on the median of the eigenvalues of modules related to response, we designated samples into high or low groups. We found that groups of high eigenvalues of brown, turquoise, and yellow modules have more Rs, while the group of high eigenvalues of the magenta module has more NRs (Figure 3C).

Figure 3.

Construction of WGCNA networks for the intratumoral microbiome and analysis of microbial modules with response to ICIs

(A) The workflow of constructing WGCNA networks for the intratumoral microbiome.

(B) Forest plot for logistic regression analysis of microbial module with response to ICIs. ∗ denote ORs with p < 0.05.

(C) Spine plots depict the association between response and response-related microbial modules.

(D) Kaplan-Meier plots show the potential clinical impact of the turquoise and brown modules in OS and PFS. p < 0.05 was considered significant.

(E) Pie charts for the microbial composition of response-related modules.

(F) Co-occurrence subnetworks of the intratumoral microbiome within the turquoise module.

(G) Potential driver species based on the microbial subnetwork of the turquoise module from melanoma. The sizes of each node correspond to their scaled neighbor shift score. Nodes colored red reflect the increase of their betweenness from no response to response, and large red nodes denote essential driver microbes associated with response to ICI therapy within the subnetwork. The colors of the lines indicate the connections present in the response (red edges) and no response groups (purple edges).

See also Figures S2 and S3; Table S5.

Second, we assessed the clinical impact of these modules on survival. In melanoma, 149 patients had overall survival (OS) information, and 73 patients had progression-free survival (PFS) information. 28 patients had PFS information in NSCLC. The hazard ratios and 95% CIs of microbial modules for PFS and OS are illustrated (Figure S3A) with forest plots. Similar to the associations with response, the univariate Cox proportional-hazards models showed that the turquoise module was found to be significantly associated with improved OS, while turquoise and brown modules were significantly associated with improved PFS in melanoma. However, there was no module significantly associated with PFS in NSCLC. Besides, Kaplan-Meier plots showed that the turquoise module had significantly positive correlations with longer OS (p = 0.028) and PFS (p = 0.037) (Figure 3D). These results indicated that the intratumoral microbes present in the turquoise and brown modules in melanoma and the yellow module in mGC were associated with the favorable efficacy of ICIs. Conversely, the microbes within the magenta module in rEAC were associated with an unfavorable efficacy of ICIs.

Next, we tried to identify certain microbes that potentially influence the efficacy of ICIs by analyzing the composition of the response-associated modules at the species level (Figure 3E). For the brown module, Sphingomonas and Methylobacterium had more species. For the turquoise module, Burkholderia had the top number of species. We observed that several species belong to Lactobacillaceae, Lachnospiraceae, and Prevotella, presented in the yellow module. Surprisingly, we found short-chain fatty acids (SCFA)-producing microbes enriched in the magenta module, accounting for 47.2% of the total species. We further constructed a co-occurrence subnetwork for these modules and observed that it was successfully established only in the turquoise and brown modules. As shown in Figure 3F, several species within the Burkholderia genus showed robust associations with other species and occupied a central position in the subnetwork of the turquoise module, suggesting their key roles in this module. Using NetShift analysis, two Burkholderia species and two Ligilactobacillus species were identified as potential driver nodes in the turquoise module contributing to the alteration of networks from the NR to R group (Figure 3G). Similarly, Sinomonas atrocyanea held a pivotal role in the brown module and was also predicted as one of the driver species (Figures S3B and S3C). Taken together, we identified microbial modules associated with the response to ICIs using WGCNA analysis and further recognized several species that played essential roles within these modules.

Correlation analysis of microbes within the associated modules with infiltration immune cells and immune response-related signatures

The efficacy of ICIs depends on the components and features of the TME, and the commensal microbiome has a vital role in influencing the TME.^28,29,30,31 Therefore, we sought to characterize the associations between the TME and the intratumoral microbiome. First, we calculated the proportions of 22 immune cell subsets by CibersortX³² (STAR Methods) and evaluated the relationship between immune cell subsets and response to ICIs. For melanoma, activated memory CD4⁺ T cells (OR 1.41, 95% CI 1.02–1.95; p = 3.76 × 10⁻²), naive B cells (OR 1.51, 95% CI 1.07–2.17; p = 2.19 × 10⁻²), M1 macrophages (OR 1.67, 95% CI 1.17–2.42; p = 5.49 × 10⁻³), Tfh (follicular helper T) cells (OR 1.61, 95% CI 1.14–2.33; p = 8.88 × 10⁻³), CD8⁺ T cells (OR 1.33, 95% CI 0.95–1.89; p = 1.04 × 10⁻¹), and activated natural killer (NK) cells (OR 1.69, 95% CI 1.23–2.37; p = 1.71 × 10⁻³) were positively associated with response, while M2 macrophages (OR 0.52, 95% CI 0.35–0.75; p = 7.02 × 10⁻⁴), resting NK cells (OR 0.64, 95% CI 0.46–0.88; p = 7.84 × 10⁻³), activated dendritic cells (DCs) (OR 0.66, 95% CI 0.46–0.90; p = 1.13 × 10⁻²), and resting memory CD4⁺ T cells (OR 0.69, 95% CI 0.49–0.87; p = 3.45 × 10⁻²) showed a significant association with worse response to ICIs (Figure 4A). In mGC, activated NK cells (OR 1.77, 95% CI 1.05–3.18; p = 4.01 × 10⁻²), M1 macrophages, CD8⁺ T cells, and Tfh cells were significantly associated with a favorable efficacy of ICIs (Figure S4A). However, there were no significant relationships between immune cells and response to ICIs for rEAC, and only activated dendritic cells (OR 2.04, 95% CI 1.03–4.52; p = 5.33 × 10⁻²) had a marginally significant association with better response to ICIs (Figure S4B). Then we performed Spearman correlation analyses of microbial modules with immune cell subsets and focused on efficacy-related immune cells. For melanoma, the turquoise module and the brown module were both significantly associated with naive B cells, activated dendritic cells, M2 macrophages, resting NK cells, activated memory CD4⁺ T cells, and resting memory CD4⁺ T cells (Figure S4C). For mGC and rEAC, we did not observe a significant association between the response-related module and those immune cell subsets associated with response. Considering direct correlation analyses of modules with immune cell subsets may obscure the role of individual microbes in influencing immune cell subsets, we correlated microbial abundance with response-associated immune cells. We found multiple microbes associated with activated memory CD4⁺ T cells, CD8⁺ T cells, resting NK cells, and M2 macrophages for melanoma (Figure 4B; Tables S6A and S6B). In particular, the members of Sphingomonas, Methylobacterium, and Burkholderia, which occupied larger parts of response-related modules, had strong associations with CD8⁺ T cells and M2 macrophages. As for mGC, Lactobacillus delbrueckii was positively correlated with activated NK cells, and Ligilactobacillus salivarius was positively correlated with M1 macrophages, which is consistent with previous reports on the antitumor role of Lactobacillaceae (Figure S4D; Tables S6C and S6D).^33,34 In rEAC, Streptobacillus moniliformis and two species of Microbacterium were found to be negatively associated with activated dendritic cells (Tables S6E and S6F).

Figure 4.

Association of intratumoral microbiota with characteristics of the TME

(A) Forest plot for logistic regression analysis of immune cell subsets with response to ICIs in melanoma. ∗ denote ORs with p < 0.05.

(B) Heatmap of significant correlations between microbes within the turquoise and brown modules and response-related immune cells in melanoma. This plot only shows the microbes that were significantly associated with immune cell subsets related to the response. p < 0.05 and a correlation coefficient > 0.2 or < −0.2 were considered significant.

(C) Violin plots for immune response-related signatures of the R and NR groups. p < 0.05 was considered significant.

(D) Circos correlation plots for intratumoral microbes within the response-related modules, and immune response-related signatures in melanoma (left), and mGC (right). p < 0.05 and a correlation coefficient > 0.2 or < −0.2 were considered significant.

See also Figure S4; Tables S6 and S7.

Several genes involved in certain signaling pathways, such as antigen presentation and cytotoxicity, and the expression of immune checkpoint have been suggested to determine the patient’s sensitivity to ICI therapy.²⁹ Except for immune checkpoints (PD-L1, PD-1, and CTLA-4), we assessed interferon (IFN)-γ pathway activity, major histocompatibility complex (MHC)-I score, and cytotoxicity score (CYT score) by calculating the mean expression levels of related genes for patients with melanoma, mGC, and rEAC (STAR Methods). Then we compared these signatures between the R and NR groups across these datasets (Figure 4C). All signatures were elevated in Rs for melanoma, while in mGC, the Rs group exhibited higher levels of these signatures except for CTLA-4. In contrast, only the MHC-I score was significantly increased, and the IFN-γ score showed a non-significant trend toward increase in Rs with rEAC. There were also several associations between microbes within response-related modules and the signatures (Figure 4D; Tables S7A and S7B) in melanoma and mGC. We observed that more of the associations focused on immune checkpoints, especially PD-L1 (Figures 4D and S4E). However, only Gemmatirosa kalamazoonesis and Streptobacillus moniliformis were found to be significantly negatively correlated with the IFN-γ score for rEAC (Table S7C). We speculated that this may be due to the limited sample size or that the intratumoral microbiome did not act by affecting immune infiltration and immune-related signal pathways in rEAC. Interestingly, we noted that P. megaterium was the sole species significantly associated with the IFN-γ score, MHC-I score, and CYT score in melanoma, which is consistent with the correlation between P. megaterium and CD8⁺ T cells, indicating that it may affect tumor immunity through promoting CD8⁺ T cells. Moreover, both Fusobacterium nucleatum and L. salivarius were significantly associated with the IFN-γ score and MHC-I score in mGC, with F. nucleatum previously shown to enhance CD8⁺ T cell function.^35,36 Accordingly, our data revealed that the intratumoral microbiome is closely related to the TME and could affect immunotherapy efficacy by reshaping the TME.

Validation of the response-associated microbes using a mouse melanoma model

To validate the effectiveness of our bioinformatics analysis, we performed logistic regression analysis between the individual microbes within response-related modules and ICIs response in melanoma and integrated the aforementioned analyses to identify potential species for experimental validation (Figure 5A). Microbial reads extracted from RNA-seq had limited accuracy in resolving species within the same genus. Hence, we retained the genus where its species had OR > 1.2 in the merged dataset and OR > 1.2 in at least two individual datasets. As a result, 8 genera were included in the final selection. Notably, P. megaterium and C. kroppenstedtii were the only two species with OR > 1.2 across all three individual datasets and the merged dataset, with C. kroppenstedtii also being reported to be enriched in the gut of Rs in patients with melanoma treated with anti-PD-1 therapy.³⁷ We used TCMbio (https://microbiomex.sdu.edu.cn/), an interactive platform that provides the curated intratumoral microbiome data, which is extracted from RNA-seq of The Cancer Genome Atlas, to conduct survival analysis with the above microbes in melanoma cohorts.³⁸ Among them, patients with a high abundance of P. megaterium had significantly longer OS, and a high level of Burkholderia and Sphingomonas was associated with a better progression-free interval (PFI) (Figure 5B). Considering the results of the previous analysis, the accessibility of experimental resources, and operational feasibility, we finally selected B. cepacia (a type species of Burkholderia), P. megaterium, and C. kroppenstedtii for further verification in C57BL/6 mice bearing subcutaneous B16-F10 melanoma tumors. B16-F10 tumor-bearing mice were i.t. injected with B. cepacia, P. megaterium, or C. kroppenstedtii every 3 days for three times in the presence or absence of intraperitoneal (i.p.) injection of an anti-PD-1 antibody every 3 days for three times (Figure 6A). B16-F10 melanoma cells form aggressive and poorly immunogenic tumors; as expected, anti-PD-1 alone did not inhibit tumor growth. Treatment with B. cepacia, P. megaterium, or C. kroppenstedtii combined with anti-PD-1 all significantly reduced tumor growth, especially the combination of B. cepacia and anti-PD-1, which exhibited the strongest antitumor activity (Figures 6B, 6C, S5A, and S5B). Treatment with P. megaterium or C. kroppenstedtii alone did not inhibit tumor growth, while the B. cepacia group showed an apparent tumor suppressor effect although it did not reach significance (p = 0.12). Besides, no difference in mice body weight was observed between different groups, indicating that i.t. selected bacteria had no distinct toxicity to mice (Figure S5C). These results suggested that i.t. B. cepacia, P. megaterium, or C. kroppenstedtii enhance the antitumor efficacy of anti-PD-1 therapy, thereby validating the effectiveness of our bioinformatics analysis.

Figure 5.

Search for potential microbes that affect the efficacy of ICIs

(A) Phylogenetic tree of the species within the turquoise and brown modules and the association with response to ICIs. The color of the tree was based on the phylum classification of each species node. The heatmap illustrates the ORs of individual species for the association with response in each dataset, as well as the batch-adjusted dataset. A colored square represents the OR value of the indicated species greater than 1.2, while a gray square reflects the value less than 1.2.

(B) Kaplan-Meier curve of OS or PFI in melanoma in TCMbio using the potential species.

Figure 6.

Validation of selected microbes with a mouse melanoma model

(A) The workflow of B16-F10 tumor-bearing mice experiment.

(B) Tumor volume of mice measured every other day during the indicated treatment regimen; PBS + IgG (n = 8), PBS + anti-PD-1 (n = 8), P. megaterium + IgG (n = 9), B. cepacia + IgG (n = 9), C. kroppenstedtii + IgG (n = 9), P. megaterium + anti-PD-1 (n = 8), B. cepacia + anti-PD-1 (n = 9), and C. kroppenstedtii + anti-PD-1 (n = 6).

(C) Tumor weight of mice measured at the end of the experiment; PBS + IgG (n = 8), PBS + anti-PD-1 (n = 7), P. megaterium + IgG (n = 7), B. cepacia + IgG (n = 7), C. kroppenstedtii + IgG (n = 9), P. megaterium + anti-PD-1 (n = 7), B. cepacia + anti-PD-1 (n = 9), and C. kroppenstedtii + anti-PD-1 (n = 6).

(D) Flow cytometry analyses on the proportions of tumor-infiltrating CD8⁺ T cells, IFN-γ⁺ CD8⁺ T cells, and NK1.1⁺ NK cells and the expression of MHC-II and PD-L1 by CD11b⁺F4/80⁺ macrophages and CD11c⁺ DCs shown as mean fluorescent intensity. n = 6–7 per group.

(E) Representative images showing immunofluorescence staining of CD8 (green), IFN-γ (red), and DAPI (blue) in tumor tissues. Scale bars, 50 μm.

(F) Representative images showing immunofluorescence staining of CD8 (green) and DAPI (blue) and FISH staining of B. cepacia or P. megaterium (red) in tumor tissues. Scale bars, 100 μm.

(G) Quantification of CD8 in tumor nests (n = 16 for bacteria-enriched nests and n = 16 for bacteria-poor nests pooled from four B16F10 tumor tissues, in the P. megaterium + anti-PD-1 or B. cepacia + anti-PD-1 group).

(H) Relative mRNA expression of tumor CXCL9, CXCL10, PD-L1, CCL5, TNF-α, and PRF1 from B16-F10 bearing mice. n = 4–5 per group.

(I) Immune cell subsets of CibersortX from RNA-seq analysis of B16-F10 tumors. n = 4–5 per group.

(J) Immune response-related signatures score from RNA-seq analysis of B16-F10 tumors. n = 4–5 per group.

(K) The workflow of A375 tumor-bearing huPBMC-C-NKG mice experiment.

(L) Tumor volume of mice measured every other day during the indicated treatment regimen; PBS + IgG (n = 4), PBS + anti-PD-1 (n = 4), and B. cepacia + anti-PD-1 (n = 4).

(M) Tumor weight of mice measured at the end of experiment; PBS + IgG (n = 4), PBS + anti-PD-1 (n = 4), and B. cepacia + anti-PD-1 (n = 4).

(N) Flow cytometry analyses on the proportions of tumor-infiltrating IFN-γ⁺ CD8⁺ T cells. n = 4 per group.

Data are shown as mean ± SEM (B–D, G–J, and L–N). Statistics are analyzed by two-way ANOVA with Tukey’s multiple comparison test (B and L), one-way ANOVA with Tukey’s multiple comparison test (C, D, H–J, M, and N), and unpaired parametric t test (G); ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001.

See also Figures S5 and S6.

B. cepacia, P. megaterium, and C. kroppenstedtii augment anti-PD-1 efficacy by activating the antitumor immunity

To assess the impact of the combination treatment on the TME, we examined different immune cell subsets of tumors in B16-F10 tumor-bearing mice (Figure 6D). Flow cytometry analysis revealed that all three combination groups increased the CD8⁺ T cells in tumors, and the combination of P. megaterium with anti-PD-1 had the most significant effect, which is consistent with the results of the correlation analysis. Moreover, only the B. cepacia + anti-PD-1 group showed a significant increase in the proportions of IFN-γ⁺CD8⁺ T cells compared to the control group. There were no significant differences between different groups in the infiltration of CD4⁺ T cells, TNF-α⁺CD8⁺ T cells, TNF-α⁺CD4⁺ T cells, and regulatory T cells (Tregs) (Figures S6A and S6C–S6E). To our surprise, although the anti-PD-1 treatment alone group elevated the percentages of IFN-γ⁺CD4⁺ T cells compared to the control group, all the combination groups reversed it. Interestingly, the trend in the proportion of CD19⁺ B cells was similar to that of IFN-γ⁺CD4⁺ T cells (Figures S6B and S6F). In addition, the infiltration of NK cells significantly increased in the B. cepacia + anti-PD-1 group, demonstrating a superior role of B. cepacia in enhancing cytotoxicity to tumors. Immunofluorescence staining verified that both B. cepacia and P. megaterium combined with anti-PD-1 expanded the infiltration of CD8⁺ T cells, and the B. cepacia + anti-PD-1 group further promoted IFN-γ production in the TME (Figure 6E). In myeloid cell subsets, the expression of MHC-II and PD-L1 on F4/80⁺CD11b⁺ macrophages of the C. kroppenstedtii + anti-PD-1 group was higher than that of the anti-PD-1 alone treatment group. Moreover, the combination of B. cepacia or C. kroppenstedtii with anti-PD-1 significantly downregulated the expression of PD-L1 on CD11c⁺ DCs compared to the anti-PD-1 alone treatment group, while a lower level of MHC-II expression on CD11c⁺ DCs was observed in the C. kroppenstedtii + anti-PD-1 group (Figures 6D and S6H). These data implied that C. kroppenstedtii may enhance the antitumor effects after anti-PD-1 treatment by affecting the function of myeloid cell subsets. To test for a potential interaction between CD8⁺ T cells and both B. cepacia and P. megaterium, we employed fluorescence in situ hybridization (FISH) staining against the two species in the tumor tissues. We observed that CD8⁺ T cell infiltration was associated with P. megaterium distribution but not with B. cepacia (Figures 6F and 6G). It indicated that P. megaterium may directly promote the infiltration or function of CD8⁺ T cells, whereas B. cepacia could potentially boost CD8⁺ T cells through affecting other immune cells or reprogramming the metabolism of the TME.

Next, we further explored the influence of selected bacteria on overall antitumor immunity. Real-time PCR revealed that in the presence of anti-PD-1, B. cepacia, P. megaterium, and C. kroppenstedtii all increased the levels of Cxcl9, Pdl1, Ccl5, Tnf, and Prf1 (Figure 6H). Consistent with the strongest antitumor activity, Cxcl9 was most significantly upregulated in the B. cepacia + anti-PD-1 group, and Cxcl10 was dramatically upregulated exclusively in this group, suggesting that B. cepacia might promote the production of chemokines to increase the infiltration of activated CD8⁺ T cells and NK cells. In addition, the B. cepacia + anti-PD-1 group had the most obvious ability to upregulate the expression of Pdl1. Treatment of C. kroppenstedtii with anti-PD-1 had the highest expression level of Tnf, further suggesting that C. kroppenstedtii enhanced anti-PD-1 therapy by promoting macrophages toward an immunostimulatory phenotype. RNA-seq analysis was performed on the tumor tissue of these mice. After calculating the proportions of immune cell subsets by CibersortX and the three signature scores, the results showed that all three combination treatment groups had elevated infiltration of CD8⁺ T cells and increased CYT score (Figures 6I and 6J). Furthermore, only the B. cepacia + anti-PD-1 group increased the proportion of activated NK cells, and the MHC-I score was significantly elevated in the combination of P. megaterium or C. kroppenstedtii with anti-PD-1. However, no differences were observed in the IFN-γ score, and the proportion of M1 macrophage decreased in the C. kroppenstedtii + anti-PD-1 group. Overall, most of the aforementioned results lined up with our previous analyses of humans. Collectively, these findings suggested that B. cepacia, P. megaterium, and C. kroppenstedtii improve the antitumor effects of anti-PD-1 therapy by activating the antitumor immunity.

To investigate the translational relevance, we constructed humanized immunocompetent mice (huPBMC-C-NKG) reconstituted with human PBMCs to mimic human immune responses and implanted them with human A375 melanoma cells. A375 tumor-bearing mice were i.t. injected with B. cepacia or PBS every 3 days for three times in the presence or absence of i.p. injection of an anti-PD-1 antibody every 3 days for three times (Figure 6K). Consistent with findings from the B16-F10 bearing model, the combination therapy of B. cepacia with anti-PD-1 antibody markedly improved the antitumor role compared with anti-PD-1 monotherapy (Figures 6L, 6M, and S5D). Similarly, B. cepacia + anti-PD-1 significantly increased the percentage of IFN-γ⁺CD8⁺ T cells, while had no effect on the total CD8⁺ T cells (Figures 6N and S6I). These data further corroborate the translational relevance of specific intratumoral microbes in modulating immunotherapy efficacy.

Discussion

In our study, we profiled the intratumoral microbiome from RNA-seq data of ICI therapy cohorts and performed a comprehensive analysis of the microbiome with clinical benefit and TME characteristics. Subsequently, we have validated three response-related species: B. cepacia, P. megaterium, and C. kroppenstedtii in mouse tumor-bearing models, further demonstrating the effectiveness of our bioinformatic pipeline and the essential role of the intratumoral microbiome in the efficacy of ICI treatments.

Recent studies suggest that the intratumoral microbiome can reshape the landscape of the TME, including immune infiltrations and function, the expression level of immune checkpoints, the activity of immune-related signaling pathways, and DNA damage repair.^{10,11,39,40,41,42,43,44} Following these suggestions, the intratumoral microbiome may influence the clinical response to ICIs in the same ways. However, due to the low biomass of intratumoral microbiome and the difficulty in obtaining pure tumor samples from patients treated with ICIs, a thorough analysis of the associations between intratumoral microbiome, treatment response, and TME features in ICI therapy is still lacking.⁴⁵

To assess the relationships between the intratumoral microbiome and ICIs response from a global view, we employed WGCNA to generate microbial modules. Our analysis revealed several modules associated with improved response in melanoma and mGC, while a module was associated with worse response in rEAC. Within these modules, some microbes have been reported to be related with survival or treatment in cancer. Sphingomonas was found to be enriched in the tumor tissue of patients with long-term survival pancreatic cancer.⁴⁶ Several species of the Lactobacillaceae and Lachnospiraceae family could boost antitumor immune to improve the efficacy of anti-PD-1 in different cancer types of mice.^{33,34,47,48,49,50,51} Interestingly, our analyses revealed that L. salivarius was associated with enhanced ICI efficacy and stronger antitumor immune responses in both melanoma and mGC. Despite a recent study reporting that L. salivarius did not interfere with the suppression of ipilimumab (anti-CTLA-4) on B16 tumor when it colonized in the gut, its potential to modulate antitumor immunity as intratumoral microbe warrants further investigation.⁵² Notably, we observed SCFA-producing microbes, especially the Lachnospiraceae family, enriched in the module with worse response in rEAC. Some research has revealed the role of SCFAs in enhancing immune cell infiltration and function to improve the anticancer effect of anti-PD-1 treatment.^36,50,53,54 Nevertheless, recent studies have proposed the opposite role of SCFAs. Excessive butyrate could directly promote the proliferation of colorectal cancer (CRC) in the gut or induce immune suppression in the TME.^10,55,56 Given the butyrate paradox, the specific role of these SCFA-producing microbes in cancer therapy needs to be further investigated. Moreover, we took a deeper exploration into these modules by constructing subnetworks. Several species of Burkholderia were identified as diver nodes in the subnetwork of the turquoise module, which had the best performance in the association with long-term survival. A previous study has shown the effect of Burkholderia in recovering germ free (GF) mice’s response to anti-CTLA-4 therapy.³ However, it is noteworthy that not all microbes colonizing the gut that improved the efficacy of ICIs exhibited the same effect when they colonized within tumors.³³ Hence, the microbes present in these modules may be the underlying cause of the difference in efficacy.

Our study further reveals the complex associations between intratumoral microbiome and TME characteristics. For instance, P. megaterium had significant positive correlations with CD8⁺ T cells, IFN-γ score, CYT score, and MHC-I score in melanoma. This highlights the potential of P. megaterium to enhance the infiltration or function of CD8⁺ T cells. In rEAC, only activated DCs were associated with response, and only the MHC-I score had a significant difference between Rs and NRs in rEAC. Indeed, few connections between microbes and TME characteristics were observed in rEAC. Considering that a combined radiotherapy and immunotherapy regimen was adopted in rEAC, the SCFA-producing microbes may facilitate the resistance to radiotherapy by promoting DNA damage repair or tumor metabolic reprogramming.¹⁷ Our comprehensive association analysis may provide more information to further explore the interaction between the intratumoral microbiome and TME.

Using a mouse tumor model, we successfully validated the synergistic effect of selected bacteria and anti-PD-1 therapy. The purified peptidoglycan from P. megaterium was reported to inhibit the growth of fibrosarcoma when injected intravenously to mice, while only i.t. injection of B. cepacia apparently suppresses B16F10 tumor, indicating that P. megaterium enhanced the effect of anti-PD-1 through another way.^57,58 Aligning with our bioinformatics analysis, as the driver taxa in the turquoise module, B. cepacia combined with anti-PD-1 had the strongest antitumor effect. Besides, the P. megaterium + anti-PD-1 group had the highest CD8⁺ T frequencies in flow cytometry analysis and CibersortX analysis and exhibited the highest MHC-I score. These results confirmed that our bioinformatics methods of extracting microbial reads and association analysis are reliable. By integrating the results of bioinformatical analysis, flow cytometry, and immunofluorescence image analysis, we assumed that P. megaterium may induce tumor cells to enhance chemokine production or reduce immunosuppressive cytokine secretion. Concurrently, it upregulated MHC-I expression, thereby promoting CD8⁺ T cell infiltration and enhancing tumor cell recognition by these lymphocytes. According to the results of flow cytometry and quantitative reverse-transcription PCR, we speculated that the superior effect of B. cepacia was caused by the synergy of multiple pathways or that there are still unknown immune cell subsets that have not been identified; the specific mechanism needs to be further investigated. Moreover, we noticed that the combination of C. kroppenstedtii and anti-PD-1 significantly upregulated PD-L1 and MHC-II on F4/80⁺CD11b⁺ macrophages, while downregulated PD-L1 on CD11c⁺ DCs. Recent studies reveal that the high expression of PD-L1 on macrophages predicted the better efficacy of immunotherapy and survival.^59,60,61,62 The interaction between PD-L1 and PD-1 mainly occurs between PD-1⁺CD8⁺ T cells and PD-L1⁺ DCs, while PD-L1⁺ tumor-associated macrophages (TAMs) rarely interact with T cells.⁶³ Hence, C. kroppenstedtii may regulate the functions of macrophages and DCs simultaneously to promote antitumor immunity. Additionally, we employed humanized immunocompetent mice to verify that B. cepacia could induce human immune response, confirming its translational potential.

In conclusion, our study provides evidence for the impact of intratumoral microbiome on patients’ response to ICIs, offers another insights from the tumor microbial perspective to improve the anticancer efficacy of ICIs, and suggests the potential prognostic and therapeutic value of the intratumoral microbiome.

Limitations of the study

Several shortcomings of this study require consideration. First, despite consideration of decontamination, potential contaminants might persist throughout the analysis process. Second, the sample size of ICI therapy cohorts is small, which results in a lack of statistical power. Third, there is a lack of validation cohorts and clinical samples that could further confirm our findings. More publicly available sequencing data about intratumoral microbiome are still needed, and the development of new sequencing techniques and coordinated research approaches for uncovering host-intratumoral microbiome interactions is imperative.

Resource availability

Lead contact

Further information and requests for resources should be directed to and will be fulfilled by the lead contact, Wei Zhang (csuzhangwei@csu.edu.cn).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •Code has been uploaded to GitHub at https://github.com/chenjhybyb/Intratumoral-microbiome. DOIs are listed in the key resources table.
• •Raw reads of RNA-seq were uploaded to the National Center for Biotechnology Information Sequence Read Archive database (accession number: PRJNA1288353).
• •Any additional information required to reanalyze the data reported in this work paper is available from the lead contact upon request.

Acknowledgments

This study was partly supported by the National Key Research and Development Program of China (2021YFA1301200), the National Natural Science Foundation of China (82073945, 82373961, 82474022, and 82373960), and the Hunan Provincial Science and Technology Project (2023SK2083 and 2022RC1022). The authors acknowledge the authors from published studies for sharing their metagenomics sequencing data of immunotherapy trials and are grateful for resources from the High Performance Computing Center of Central South University and the Bioinformatics Center, Xiangya Hospital, Central South University.

Author contributions

W.Z. and Y. Zou conceived and supervised the study. Y. Zou contributed to the study design, the collection of RNA-seq data, and analysis in the high-performance computing center. J.C. performed statistical analysis, carried out the experiment, and drafted the original manuscript. Q.W., Y. Zhang, Y.H., X.X., and D.Z. discussed data. Y.G., Y.C., R.L., and H.Z. revised the manuscript. All authors read and approved the final version of the manuscript.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Anti-mouse in vivo mAb PD-1(CD279)	Bio X Cell	Cat#BP0146; RRID: AB_10949053
Anti-mouse in vivo IgG2a isotype	Bio X Cell	Cat#BP0089; RRID: AB_1107769
Anti-human in vivo mAb PD-1(CD279)	Bio X Cell	Cat#BE0188; RRID: AB_10950318
Anti-human in vivo IgG2a isotype	Bio X Cell	Cat#BE0083; RRID: AB_1107784
FITC Hamster anti-mouse CD3e PE-Cy7 Rat anti-mouse CD25 BV421 Rat anti-mouse CD4 Alexa Fluor 647 Rat anti-mouse Foxp3 PE-Cy7 Rat anti-mouse TNF APC Rat anti-mouse IFN-γ Leukocyte Activation cocktail Alexa Fluor™ 700 anti-mouse MHC Class II PE Rat anti-mouse MHC Class I (H-2Kb) PE-Cy5 anti-mouse F4/80 PerCP-Cy5.5 Rat anti-mouse CD8α BV605 anti-mouse CD11c BV650 anti-mouse CD19 BV785 anti-mouse PD-L1	BD Biosciences BD Biosciences BD Biosciences BD Biosciences BD Biosciences BD Biosciences BD Biosciences Biolegend BD Biosciences Biolegend BD Biosciences Biolegend Biolegend Biolegend	Cat#561827; RRID: AB_394595 Cat#561780; RRID: AB_394509 Cat#743088; RRID:AB_2741280 Cat#560402; RRID:AB_1645201 Cat#561041; RRID:AB_396761 Cat#562018; RRID:AB_398551 Cat#550583; RRID:AB_2868893 Cat#107621; RRID:AB_493726 Cat#562010; RRID:AB_396546 Cat#123111; RRID:AB_893494 Cat#561109; RRID:AB_394081 Cat#117333; RRID:AB_11204262 Cat#115541; RRID:AB_11204087 Cat#124331; RRID:AB_2629659
Bacterial and virus strains
Burkholderia cepacia	DSMZ	DSM: 7288
Priestia megaterium	DSMZ	DSM: 32
Corynebacterium kroppenstedtii	DSMZ	DSM: 44385
Chemicals, peptides, and recombinant proteins
RPMI 1640 Medium	Gibco	Cat#11875093
Fetal bovine serum	Gibco	Cat#A5256701
LB Broth	OXOID	Cat#12780052
TSB Broth	Millipore	Cat#22092
Dulbecco’s Modified Eagle Medium (DMEM)	Gibco	Cat#10569010
Phosphate-buffered saline (PBS) tablets PrimeScript™ RT Reagent Kit Foxp3/Transcription Factor Staining Buffer Set Matrigel Matrix	Thermo Fisher Takara eBioscience Corning	Cat#18912014 Cat#RR037B Cat#00-5523-00 Cat#356234
Cell Stimulation Cocktail	eBioscience	Cat#00-4970-03
TRIzol Reagent	Invitrogen	Cat#15596026CN
Deposited data
RNA sequencing of human tumor tissue sample	Gide et al.20	EBI: PRJEB23709
RNA sequencing of human tumor tissue sample	Hugo et al.18	EBI: PRJNA312948
RNA sequencing of human tumor tissue sample	Riaz et al.21	EBI: PRJNA356761
RNA sequencing of human tumor tissue sample	Kim et al.19	EBI: PRJEB25780
RNA sequencing of human tumor tissue sample	Tom et al.17	NCBI: PRJNA694637
RNA sequencing of human tumor tissue sample	Kim et al.19	NCBI: PRJNA557841
RNA sequencing of human tumor tissue sample	David et al.23	NCBI: EGAD00001004183
RNA sequencing of human tumor tissue sample	Mariathasan et al.22	NCBI: EGAD00001003977
Raw reads of RNA-sequencing	This paper	SRA: PRJNA1288353
Experimental models: Cell lines
Mouse: B16-F10 cells	ATCC	Cat#CRL-6475; RRID: CVCL_0159
Human: A375 cells	ATCC	Cat#CRL-1619; RRID: CVCL_A375
Experimental models: Organisms/strains
Mouse: C57BL/6J	The Jackson Laboratory	N/A
Mouse: huPBMC-C-NKG	Cyagen	Cat#C001329
Software and algorithms
FlowJo, version 10.4	Tree Star	www.flowjo.com
GraphPad Prism, version 8	GraphPad	https://www.graphpad.com/features
FastQC	Babraham Institute	https://www.bioinformatics.babraham.ac.uk/projects/fastqc/
R-4.2.3	R-project	https://www.r-project.org/
Kraken2	Wood et al.64	https://github.com/DerrickWood/kraken2
Bracken	Wood et al.64	https://github.com/DerrickWood/kraken2
SAMtools (v.1.13)	Li et al.65	https://github.com/samtools/samtools
Biorender	Biorender	https://www.biorender.com
Analysis code	This paper	https://doi.org/10.5281/zenodo.15829835

Experimental model and subject details

Cell lines

Murine-derived melanoma cell line B16-F10 and human melanoma cell line A375 were obtained from the American Type Culture Collection; B16-F10 cells were cultured in RPMI 1640 (Gibco, USA) medium containing 10% fetal bovine serum (Gibco, USA); A375 cells were cultured in DMEM (Gibco, USA) medium containing 10% fetal bovine serum (Gibco, USA).

Animals

Six-week-old female C57BL/6J mice were purchased from SLAC Laboratory Animal Co. Ltd. HuPBMC-C-NKG (Catalog C001329) mice were purchased from Cyagen. The animals were housed in the specific pathogen-free environment of the Center of Laboratory Animals, Xiangya hospital Central South University. All animal experiments were approved by the Animal Ethics Committee of Xiangya hospital of Central South University (No. 202410178).

Bacterial strains

Corynebacterium kroppenstedtii DSM 44385, Priestia megaterium DSM 32 and Burkholderia cepacia DSM 7288 bacterial strains were obtained from the Deutsche Sammlung von Mikroorganismen und Zellkulturen (DSMZ); LB medium (Luria Broth Medium) for B. cepacia and P. megaterium, and TSB medium (Tryptic Soy Broth Medium) for C. kroppenstedtii. These strains were cultured in an aerobic Incubator.

Method details

Data accession

To identify the intratumoral microbial signature associated with response to ICIs, we collected 8 cohort datasets comprising pre-treatment samples treated with ICIs (Figures 1A; Table S1). For the Melanoma_Hugo,¹⁸ Melanoma_Riaz,²¹ Kim_2018,¹⁹ and Melanoma_Gide²⁰ datasets, the processed mRNA datasets were downloaded from the EBI. The Mariathasan_2018²² dataset was obtained from the R package“IMvigor210CoreBiologies” (http://researchpub.gene.com/IMvigor210CoreBiologies). For the Kim JY_2020¹⁶ and Tom_2021¹⁷ were downloaded from Gene Expression Omnibus under accession number GSE135222 and GSE165252. The mRNA and clinical information of McDermott_2020²³ were available at the clinical study data request platform (https://clinicalstudydatarequest.com/). As the tumor tissues in Mariathasan_2018 originated from different tissues of the urinary system, given the tissue-specificity of the microbiome, the dataset was first excluded. To be consistent with the above datasets, log-CPM transformation was applied to the counts data and the processed expression matrix was utilized for further analyses.

Intratumoral microbiome profiling

Raw RNA-sequencing data underwent quality control with Fastp.⁶⁶ Subsequently, we used Minimap2 (v.2.22)⁶⁷ to compare the high quality reads with the GRCh38.p7 human genome, and then we applied SAMtools (v.1.13)⁶⁵ to extra reads mapped to the human genome. After the host depletion of sequencing data, the non-human reads were mapped to microbes by Kraken2⁶⁴ based on the NCBI Reference genome. Then the successfully aligned sequences were reestimated by Bracken⁶⁴ to obtain a highly accurate counts matrix. To obtain more purified matrix of the intratumoral microbiome, we established a rigorous decontamination process: 1. Potential contaminants derived from a related study were filtered out from the counts matrix at the genus level⁶⁸ (Tables S8); 2. We retained only genera shared between the TCMbio³⁸ and our counts matrix; 3. To minimize the impact of zero-inflated and over-dispersed abundance data on downstream analyses, we further excluded low coverage species with the prevalence of <10% in samples from each dataset.

Removal of batch effects in melanoma datasets

To increase the sample size to improve the computational efficiency and test efficacy of the downstream analyses, we combined the transcriptome data and the relative abundance of microorganisms from these cohorts separately. Transcriptome data from different research sources may have errors due to different experimental conditions, which are reflected as batch effects at the gene expression level. Sva, an R package, provides two functions, Combat() and Combat_seq(), which can be used to correct for batch effects in high-dimensional data.^69,70 Among them, Combat() is based on Bayesian principle for data with decimals (e.g., microarray data), while Combat_seq() is based on negative binomial distribution regression for RNA-seq counts data. In this study, we used the Combat() function to correct for microbial relative abundance, and Combat-seq() to correct for counts of transcriptional data before combining them separately for subsequent analyses. During the de-batch effect correction process, if the sample size of one of the datasets is much larger than the other datasets, it may bias the correction effect in favor of that dataset, resulting in a bias in the merge. In addition, different treatment protocols may also have some impact on the correction results. To minimise confounding factors in the de-batch effect process, we retained only the samples receiving anti-PD-1 monotherapy in the Gide dataset for the merging of the melanoma dataset.

Beta diversity estimations

Beta diversity was analyzed using principal coordinate analysis (PCoA) based on Euclidean distance. The Euclidean distance between two samples was first calculated by the statistical algorithm of the vegan R package to obtain the distance matrix, and then the pcoa function in the ade4 R package was used to perform the PCoA.

Generating a coexpression network by WGCNA

Weighted gene co-occurrence network analysis (WGCNA) is an analytical method for analyzing gene expression patterns across multiple samples, which can cluster genes with similar expression patterns and identify the associations between modules and specific traits or phenotypes.⁷¹ In this study, we replaced the expression of genes with the abundance of intratumoral microbes to generate the coexpression network of intratumoral microbiome. First, we transformed the counts matrix of species into log-counts per million (log-cpm) maxtrix. Second, we applied WGCNA analysis with appropriate parameters (Table S5) to cluster the species into microbial modules. Finally, using the WGCNAmodulePreservation() function, we assessed the preservation of these modules. The modules with a higher preservation median rank than the gold module (a reference module) and Z summary score greater than 2 were selected for further analysis.

Logistic regression analysis

To determine the association between the intratumoral microbiome and response to ICIs treatment, logistic regression analysis was performed. To derive more easily interpretable Odd ratios (ORs), quartiles (25%, 50%,and 75%) of the abundance of individual microbes or the eigenvalues of microbial modules were modeled as continuous variables in the logistic regression models. The resulting ORs, along with their associated p values and 95% confidence intervals (CIs), were subsequently visualized using forest plots in R. P < 0.05 was considered significant.

Survival analysis

We obtained curated survival phenotypic data including overall survival (OS) and progression-free survival (PFS) from the melanoma cohorts (Figure 1A). Then, patients from each cohort were stratified into high and low groups based on the median of the eigenvalues of microbial modules. We evaluated the clinical impact of the microbial modules on OS and PFS by Kaplan-Meier analysis and univariate Cox Analysis in R. P < 0.05 was considered significant.

Taxonomic co-occurrence analysis within modules and identification of driver microbes

We constructed sparse taxonomic co-occurrence networks with SPIEC-EASI using the Meinshausen-Buhlmann neighborhood selection method.⁷² The nodes and edges produced by SPIEC-EASI were visualized using Gephi 0.10.1 (https://gephi.org/). NetShift,⁷³ which can quantify the directional changes among intermicrobial associations of the individual node between Rs and NRs, was used for identifying potential driver microbes according to differences in subnetwork interactions between microbiome of different groups.

Infiltrating immune cell subsets

To estimate the proportions of immune cell subsets in a mixed cell population from RNA-seq data, the CIBERSORTx³² (https://cibersortx.stanford.edu/) deconvolution algorithm was used (Submitted job type: ‘Impute cell fractions’). The LM22 signature was set as a reference, and the relative immune cell proportions from each processed dataset were estimated in 1000 permutations with B mode batch correction, with quantile normalization disabled for RNA-seq data. The B mode batch correction mode was utilized to correct for batch effects between the LM22 signature generated from the microarray-based dataset and the dataset utilized here, which consisted of RNA-seq data. A p value that measures the reliability of the deconvolution results was computed, and samples with a p > 0.05 were discarded from downstream association analyses.

Immune response–related signatures

Herein, four kinds of immune response related gene signatures were assessed. The MHCI score was calculated as the average expression level of the “core” MHCI gene set, which contains NLRC5, TAP1, TAP2, HLAA, HLA-B, HLA-C, PSMB9, PSMB8, and B2M.⁷⁴ Gene lists of the cytolytic activity (CYT score) were accessed from the Immune Cell Abundance Identifier (ImmuCellAI) tool.⁷⁵ Gene lists of the IFN-γ pathway were accessed from the Molecular Signatures Database.⁷⁶ The pathway score was calculated as the mean expression levels of the genes included in this pathway. For the assessment of immune checkpoints including PD-L1, PD-1 and CTLA-4, their expression levels were used directly.

Correlation analysis

To explore the correlation between intratumoral microbiome and characteristics, we performed Spearman’s correlation analyses in R. To maintain strong associations, p < 0.05 and correlation coefficient >0.2 or < −0.2 were considered significant.

Tumor models and treatments

100 μL of B16-F10 cells containing 10⁵ cells were implanted under the skin of the right dorsal side of C57BL/6J mice. When the tumor size reached about 80–100 mm³, 10⁷ colony-forming units (CFU) of selected bacteria or an equal volume of PBS was administered by multipoint intratumoral injection every 3 days until the end of the experiments. The day after intratumoral injection, 200 μg of PD1 mAb (BioXCell, BE0146, USA) or isotype control IgG (BioXCell, BE0089, USA) was given in the anti-PD-1 treatment group and selected bacteria + anti-PD-1 groups every 3 days. Tumor volume was measured by vernier calipers every other day, and body weight was measured every other day. When the tumor grew to nearly 1,000 mm³, the tumor and serum were taken from mice for follow-up experiments.

100 μL of A375 cells containing 10⁶ cells were implanted under the skin of the right dorsal side of huPBMC-C-NKG mice. When the tumor size reached about 80–100 mm³, 10⁷ colony-forming units (CFU) of B. cepacia or an equal volume of PBS was administered by multipoint intratumoral injection every 3 days until the end of the experiments. The day after intratumoral injection, 200 μg of PD1 mAb (BioXCell, BE0188, USA) or isotype control IgG (BioXCell, BE0083, USA) was given in the anti-PD-1 treatment group and selected bacteria + anti-PD-1 groups every 3 days. Tumor volume was measured by vernier calipers every other day, and body weight was measured every other day. When the tumor grew to nearly 700 mm³, the tumor and serum were taken from mice for follow-up experiments.

Multicolor flow cytometry analysis

The spleen or tumor tissues were prepared as single-cell suspensions; 100 μL of prepared Zombie Aqua Fixable Viability Kit (anti-BV510, BioLegend, USA) was added and incubated for 10 to 15 min at room temperature; configured CD16/32 was added, incubated for 15 min at 4°C, and protected from light; and 100 μL of prepared surface antibody (CD45, CD3, CD4, CD8, NK1.1, CD25, F4/80, CD11b, CD11c, CD19, PD-L1, MHC-I, and MHC-II) was added and incubated for 30 min at 4°C. To label intracellular molecules, cells were incubated for 30 min from light after fixation and permeabilization (eBioscience, USA); 100 μL of the prepared intracellular antibody (FOXP3, IFN-γ, and TNF-α) was added and incubated for 30 min at 4°C. Antibodies used in flow cytometry were listed in key resources table.

Immunofluorescence staining and 16S rRNA FISH assay

For immunofluorescence staining, slides were incubated with anti-mouse CD8 antibody and anti-mouse IFN-γ antibody and counterstained with DAPI. The EUB338 16S rRNA gene probe labeled with the fluorophore Cy5 were used to detect the bacterial colonization within mouse tumor specimens (n = 4 per group) by FISH. Nonspecific complement probe was used as a control for hybridization protocol. The images were performed using a fluorescence microscope (EVOS M7000, Invitrogen). To explore the association between intratumoral bacteria distribution and CD8⁺ T cells, eight 30 × fields were captured from each tumor specimen. Based on the percent area of FISH, these fields were defined as ‘enriched’ and 'poor’ groups. Subsequent quantification was performed to assess the intergroup differences in percent area of CD8.

RNA-sequencing

Total RNA from tumors was isolated using TRIZOL reagent. High-throughput sequencing was performed by PANOMIX Biomedical Tech Co., LTD (Suzhou)., utilizing 1 μg of total RNA for library preparation. Low-quality reads were removed, and the clean reads were subsequently aligned to the mouse reference genome (GRCm39) and normalized to FPKM.

Quantitative reverse transcription PCR

Total RNA from tumors was isolated using TRIZOL reagent and reverse transcribed to cDNA using the Two-Step gDNA Removal and cDNA Synthesis Kit (Takara). Relative mRNA expression of CXCL9, CXCL10, PD-L1, CCL5, TNF-α, and PRF1 were examined by quantitative real-time PCR. Real-time PCR was performed using SYBR Purple Master Mix and quantified by the DDCT method. The expressions of genes were normalized to Actb. Primers used in this assay were listed in key resources table.

Quantification and statistical analysis

Downstream bioinformatics analysis and plots were performed with R version v4.2.3. For animal experiments, statistical significance was calculated with the two-way ANOVA or one-way ANOVA test in GraphPad Prism 8. All data are expressed as mean ± standard error of mean (SEM). Unless otherwise noted, a p < 0.05 was regarded as statistically significant. The statistical details of experiments can be found in the figure legends.

Published: August 26, 2025

Supplemental information

Table S2. Intratumoral microbial counts from all cohorts (before decontamination), related to STAR Methods

(A-G) Phylum level, (H-N) Genus level, (O-U) Species level

Table S3. Intratumoral microbial counts in the species level (after decontamination), related to STAR Methods

(A) Melanoma-Merge species level (Before batch effect adjusting), (B) Melanoma-Merge species level (After batch effect adjusting), (C) PRJEB25780 species level, (D) PRJNA694637 species level

Table S6. Results of Spearman correlation analysis between microbes and immune cell subsets, related to Figures 4 and S4

(A) Spearman correlation analysis between microbes and immune cell subsets in melanoma (P value), (B) Spearman correlation analysis between microbes and immune cell subsets in melanoma (cor), (C) Spearman correlation analysis between microbes and immune cell subsets in mGC (P value), (D) Spearman correlation analysis between microbes and immune cell subsets in mGC (cor), (E) Spearman correlation analysis between microbes and immune cell subsets in rEAC (P value), (F) Spearman correlation analysis between microbes and immune cell subsets in rEAC (cor)

Table S7. Results of Spearman correlation analysis between microbes and immune response-related signatures, related to Figures 4 and S4

(A) Spearman correlation analysis between microbes and immune response related signatures in melanoma, (B) Spearman correlation analysis between microbes and immune response related signatures in mGC, (C) Spearman correlation analysis between microbes and immune response related signatures in rEAC