Introduction

Immune checkpoint inhibitors (ICIs), particularly those targeting programmed cell death 1 (PD-1) and programmed cell death ligand 1 (PD-L1), have significantly improved clinical outcomes for cancer patients1,2. Nonetheless, variability in individual responses and tumor heterogeneity affects the overall effectiveness of these therapies3. In addition to patient-specific physiological characteristics, emerging research suggests that gut microbiota influences the effectiveness of ICIs4. Most studies to date have focused on gut bacteria, with some identified as potential biomarkers for immunotherapy or therapeutic adjuvants5,6,7. However, the gut also hosts fungi, archaea, and viruses, which are increasingly recognized for their roles in disease progression and immune regulation8,9. Consequently, the impact of intestinal fungi on ICI efficacy warrants further investigation. Furthermore, the balance of the intestinal immune system relies on the interactions between gut fungi and bacteria. Yet, the association between these microbial communities and ICI response phenotypes remains unclear. Understanding how gut bacteria and fungi interact may provide valuable insights into enhancing ICI effectiveness and addressing the variability observed in patient responses.

The current landscape of intestinal fungal bioinformatics is limited due to the lack of a comprehensive database for fungal taxonomic sequences10,11. However, researchers have managed to address this gap by integrating multiple fungal reference genome databases, demonstrating a resourceful approach to the problem12,13,14. Moreover, there is considerable variability in the microbial signatures reported in existing studies. This inconsistency arises from a range of factors, including host physiology, dietary habits, geographic location, and environmental conditions, as well as technical aspects such as sample quality, sequencing depth, and the analysis pipeline used for biological data15,16. To increase the consistency and reliability of findings related to the efficacy of ICIs, employing meta-analysis alongside multi-cohort sequencing data represents an effective research strategy. This approach can help identify microbial signatures associated with ICI effectiveness and improve our understanding of their role in treatment outcomes.

Here, we collected and analyzed fecal shotgun metagenomic data from 8 cohorts prior to treatment with ICIs. After performing quality control and filtering, a total of 976 samples were retained in the subsequent analysis. Logistic regression and Cox regression models were used to identify microbiological species associated with ICI efficacy in different cohorts and cancer types, and meta-analysis was then conducted to pinpoint microbial signatures relevant to cross-cohort comparisons. Recognizing that microorganisms may function as communities, we identified three fungal enterotypes, with the Aspergillus-dominated enterotype in melanoma and the Saccharomyces-dominated enterotype in non-small cell lung cancer (NSCLC) associated with improved ICI efficacy. Furthermore, our experiments in mice confirmed that Debaryomyces hansenii can diminish the effectiveness of PD-1 monoclonal antibody (mAb) in melanoma. Overall, our findings highlight and validate the role of specific intestinal fungal species in influencing ICI effectiveness. This underscores the importance of considering intestinal fungi in cancer treatment strategies involving ICIs and provides valuable insights for future research in this area.

Results

Clinical features and microbiological characteristics

The study framework is illustrated in Fig. 1a. After excluding samples that were not treated with ICI and those that were not sequenced at baseline, 1161 samples from eight cohorts were included for further analysis. To minimize technological bias in our bioinformatics analysis, we applied a standard species annotation protocol to all metagenomic sequencing data. The median metagenomic reads count and ratio mapped to fungi were 631 (102.8) and 0.025% (10−3.61, Fig. 1b), respectively, which is lower than the 0.1% reported in previous publications17. We found that there was a positive correlation between the sequencing depth of the samples and the fraction of fungi-read (r = 0.73, Supplementary Fig. 1a) through the Person correlation analysis. Additionally, the sequencing ratio varied across different tumor types, likely due to differences in sequencing techniques and cohort characteristics (Supplementary Fig. 1b, c). A total of 504 fungi and 3216 bacteria were detected across all eight cohorts (Supplementary Fig. 2a, b). The median of species mapped to fungi and bacteria per sample was 124 and 1415, respectively (Supplementary Fig. 2c, d). To ensure the reliability of our edibility follow-up study, we further filtered the samples, with details outlined in Fig. 1c and Supplementary Data 1. This process resulted in a final dataset of 976 samples, including 301 responders (30.8%) and 650 non-responders (66.6%). The results of Chi-square (χ2) test showed that sample filtration did not alter the distribution of ICI response proportions from the original dataset (Fig. 1d). Additional details regarding other clinical indicators are provided in Supplementary Data 2.

Fig. 1: Sample pre-processing and fungal profiles in responders and non-responders.
figure 1

a Flow chart for this study; b Top panel: the count number of bacterial (left) or fungal (right) reads in 8 cohorts (n = 1161). Bottom panel: the ratio of fungal to bacterial read count number in all samples; c Sample quality control procedures and criteria; d The change in the proportion of responder before and after filtration in each tumor; χ2 test was used for evaluation; e Fungal composition at the phylum level for responders and non-responders in each tumor; the relative abundance of each phylum is averaged.

Full size image

The relative abundance of the two major fungal phyla, Ascomycota and Basidiomycota, in responders versus non-responders was presented in Fig. 1e. Notably, the fungal Shannon index did not show statistically significant differences between responders and non-responders across various cohorts (Supplementary Fig. 3a). However, across all cohorts, the fungal Shannon index was statistically different among response groups only in melanoma (P = 0.0018) (Supplementary Fig. 3b). The Shannon index of fungi was influenced by covariates such as cohort, age, gender, and body mass index (BMI) (Supplementary Fig. 3c–e). Additionally, we observed differences in β-diversity between responders and non-responders in melanoma (PERMANOVA P = 0.002), NSCLC (PERMANOVA P = 0.44), and renal cell carcinoma (RCC) (PERMANOVA P = 0.63) after adjusting for the influence of covariates (cohort, age, gender, and BMI). Consistent with the α-diversity analysis, the results of the principal coordinate analysis indicated that the heterogeneity of the cohort had a greater impact on fungal diversity than other variables (age, gender, and BMI) (Supplementary Fig. 4).

Microbial signatures associated with ICI efficacy were identified

To identify microbial signatures associated with ICI efficacy, logistic regression was performed. Using the criteria of a meta p-value threshold of less than 0.05 and a heterogeneity p-value greater than 0.05, we identified 27 fungi (Fig. 2a) and 23 bacteria (Supplementary Data 4), of which 13 fungi and 11 bacteria were positively associated with response to ICIs in melanoma. Additionally, there were 34 bacteria (Supplementary Data 5) and 25 fungi (Fig. 2a) that were significantly associated with PFS12 (PFS at 12 months). Notably, high levels of abundance of Rhizophagus irregularis (Fig. 2b) were linked to decreased ICI efficacy (Meta OR = 0.67, Meta P = 0.0036) and PFS12 (Meta OR = 0.71, Meta P = 0.028), while Aspergillus avenaceus (Fig. 2c) was associated with improved ICI efficacy (Meta OR = 1.66, Meta P = 0.002) and PFS12 (Meta OR = 1.66, Meta P = 0.004). In multivariate Cox regression analysis, 15 fungi (Fig. 2d) and 26 bacteria (Supplementary Data 7) were identified as influencing progression-free survival (PFS), organized by calculated risk scores. A high abundance of Hericium alpestre was linked to longer PFS (Meta HR = 0.50, Meta P = 0.003, log-rank P = 0.071 in the merged cohorts, Fig. 2e), while Tilletiopsis washingtonensis showed the opposite effect (Meta HR = 1.71, Meta P = 0.009, log-rank P = 0.0011 in the merged cohorts, Fig. 2f). Due to the limited sample size, results for other indicators, such as PFS6 (PFS at 6 months) and overall survival (OS), are presented in Supplementary Data 6 and 8. When accounting for covariates such as gender, age, and BMI, univariate logistic regression and univariate Cox regression analysis revealed that only BMI significantly affected clinical indicators (Supplementary Data 3). Additionally, confounder analysis indicated that the cohort had a greater influence on fungal abundance (Supplementary Fig. 5a). The bacterial signatures associated with ICI efficacy in melanoma are detailed in Supplementary Data 6–8.

Fig. 2: Fungal signatures associated with response to ICIs in melanoma.
figure 2

a Heatmaps of β value for response and PFS12; *P < 0.05; **P < 0.01; gray: NA; b Details of the association between Rhizophagus irregularis and response (left)/ PFS12 (right) in each cohort; c Details of the association between Aspergillus avenaceus and response (left)/ PFS12 (right) in each cohort; d Risk scores model for PFS; e Survival curve plots for Hericium alpestre versus PFS; P-values are determined by the log-rank tests. f Survival curve plots for Tilletiopsis washingtonensis versus PFS; P-values are determined by the log-rank tests.

Full size image

In NSCLC, 23 fungi and 38 bacteria (Supplementary Data 9) were linked to the effectiveness of ICIs, while 17 fungi and 47 bacteria (Supplementary Data 10) were associated with PFS6 (Fig. 3a). However, detailed information regarding PFS12 (Supplementary Data 11), PFS (Supplementary Data 12), and OS (Supplementary Data 13) was only available from the cohort of DerosaL_2024, preventing a comprehensive meta-analysis, and a p-value of less than 0.05 was considered as an associated species. The fungal markers associated with OS were identified by Cox regression analysis adjusting covariates (age, gender, BMI, and read-depth), and the results of the high and low groups divided according to the median risk score were depicted in Fig. 3b. Notably, high abundance of Aspergillus pseudonomiae was associated with the improved effectiveness of ICIs (Meta OR = 1.28, Meta P = 0.077, Fig. 3c), increased percentage of patients with PFS6 (Meta OR = 1.44, Meta P = 0.004), and longer PFS (HR = 0.75, P = 0.0004, log-rank P = 0.00014) and OS (HR = 0.71, P = 0.0009, log-rank P = 0.00041, Fig. 3d). High abundance of Saccharomyces cerevisiae (Meta OR = 0.74, Meta P = 0.036) was linked to impaired effectiveness of ICIs, while Fusarium flagelliforme (Meta OR = 1.39, Meta P = 0.005) was associated with an increased percentage of individuals with PFS6 (Fig. 3e). Patients with high abundance of Schizosaccharomyces japonicus (HR = 0.85, P = 0.032, log-rank P = 0.029) experienced longer PFS, whereas those with high abundance of Malassezia globosa (HR = 1.21, P = 0.032, log-rank P = 0.0061) had shorter OS (Fig. 3f). The results of confounding variable analysis are presented in Supplementary Fig. 5b, and the bacterial signatures related to ICI treatment in NSCLC were shown in Supplementary Data 9–13.

Fig. 3: Fungal signatures associated with response to ICIs in NSCLC.
figure 3

a Heatmaps of β value for response and PFS6; *P < 0.05; **P < 0.01; gray: NA; b Risk scores model for OS; c Details of the association between Aspergillus pseudonomiae and response (left)/ PFS6 (right) in each cohort; d Survival curve plots for Aspergillus pseudonomiae versus PFS (left) and OS (right); P-values are determined by the log-rank tests; e Details of the association between Saccharomyces cerevisiae and response (left) in each cohort; Details of the association between Fusarium flagelliforme and PFS6 (right) in each cohort; f Survival curve plots for Schizosaccharomyces japonicus versus PFS (left); Survival curve plots for Malassezia globosa versus OS (right); P-values are determined by the log-rank tests.

Full size image

In RCC, the association analysis results may be biased due to the small sample size and low percentage of responders. The associations between microbial abundance, PFS6, and response are detailed in Supplementary Data 14.

The correlation of microbial pairs varied among groups of response

The presence of microbial interactions is a common phenomenon, so we explore the role of correlations between the bacterial and fungal kingdoms during ICI therapy. Overall, we found that in melanoma, the ecological network was more complex in responders (90 species, 2995 associations) compared to non-responders (90 species, 2149 associations) (Fig. 4a). In contrast, the trend is reversed in NSCLC, where responders exhibited 102 species and 807 associations, while non-responders had 92 species and 1170 associations (Fig. 4b). In addition to the robust connections among species within the kingdom, there are also substantial interactions between fungi and bacteria. Notably, responders demonstrated a higher degree of positive connection, regardless of whether they had NSCLC or melanoma (Fig. 4a, b). Consistently, a similar interaction pattern appears once the weak connection was eliminated in melanoma (|r|<0.55, Fig. 4c). After removing the weakly correlated microbial pairs, a trend was observed in NSCLC where the correlation was slightly stronger in responders than in non-responders (Fig. 4d). Furthermore, in melanoma responders, a higher proportion of fungal interkingdom interactions (Vanrija humicola, Tilletiopsis washingtonensis, Phanerodontia chrysosporium, Pyrenophora seminiperda) could be observed, while in NSCLC, the interactions among the Faecalibacterium (Faecalibacterium prausnitzii, Faecalibacterium duncaniae, Faecalibaculum rodentium) were also more pronounced.

Fig. 4: Coabundance correlations among multi-kingdom species in responders and non-responders.
figure 4

a Summary of type of correlation and microbial pairs involving combined species related to ICI efficacy (Meta P < 0.1) from bacteria and fungi in melanoma; Only significant (FDR < 0.05) absolute correlations above 0.2 are shown; b Summary of type of correlation and microbial pairs involving combined species related to ICI efficacy (Meta P < 0.1) from bacteria and fungi in NSCLC; Only significant (FDR < 0.05) absolute correlations above 0.2 are shown; c Coabundance networks involving combined species related to ICI efficacy (Meta P < 0.1) from bacteria and fungi in melanoma; Only significant (FDR < 0.05) absolute correlations above 0.55 are shown; d Coabundance networks involving combined species related to ICI efficacy (Meta P < 0.1) from bacteria and fungi in NSCLC; Only significant (FDR < 0.05) absolute correlations above 0.55 are shown.

Full size image

We identified significant differences in microbial pairs among responders and non-responders, with 524 pairs in melanoma and 33 pairs in NSCLC (Supplementary Data 15–17), which may contribute to the effectiveness of ICI therapy. In melanoma, this change in correlations between groups was mostly observed between fungi and fungi, as well as between fungi and bacteria (Supplementary Fig. 6a). However, in NSCLC, it was mostly found within the bacterial interkingdom. The cause of this difference cannot be ruled out as being due to the lower proportion of responders or the lower sequencing depth in NSCLC. Among these different microbial pairs, the correlation of responders was stronger than that of non-responders, whether in melanoma or in NSCLC (Supplementary Fig. 6b). Furthermore, we calculated correlations between the abundances of species that met the criteria for ICI prognostic association analysis in the different groups (responders vs. non-responders), then considered the species in differential microbial pairs (empirical P < 10−5) as key nodes. We then overlapped these nodes and microbial signatures associated with ICI response. On the whole, we found the presence of low overlap (Supplementary Fig. 6c, d), which suggests that the abundance changes of the microbiome and the interaction network may be relatively independent. However, a few overlapping species (such as Colletotrichum asianum and Klebsiella oxytoca) may have a dual effect of “abundance and interaction”, which is worthy of verification by subsequent experiments.

Fungal enterotypes are associated with the response to ICI therapy

Microbes may function as classes or groups, whereas the enterotype is a stable microbial community that remains unaffected by covariates18. In this study, we identified three distinct fungal enterotypes using the Bray–Curtis distance by PAM (PERMANOVA P = 0.001, Fig. 5a). Based on the genera with the highest average relative abundance as the characteristics of enterotype, they were divided into Asper_high, Sacc_type, and Asper_low (Fig. 5b). Aspergillus and Saccharomyces, as the dominant genera in all enterotypes, exhibited significant differences in relative abundance across the different enterotypes (Fig. 5c). Meanwhile, there were significant differences in species diversity among the three fungal enterotypes, with the Asper_low having a higher Shannon index (Fig. 5d).

Fig. 5: Association between fungal enterotypes and ICI efficacy.
figure 5

a PCoA diagram of fungal enterotypes based on Bray-Curtis distance by PAM; b Genera in the top 10 in mean relative abundance among the three fungal enterotypes; c Comparison of abundance difference of dominant genera in the three fungal enterotypes by Wilcoxon test; left: Aspergillus; right: Saccharomyces; d Comparison of Shannon index difference in the three fungal enterotypes by Wilcoxon test; e Association between fungal enterotypes and ICI efficacy in melanoma (left) and NSCLC (right); χ2 test was used for evaluation; f, g Survival curve plots for fungal enterotypes versus PFS (left) and OS (right) in melanoma; P-values are determined by the log-rank tests; *P < 0.05; **P < 0.01; ***P < 0.001.

Full size image

There were differences in the proportions of responders among the three fungal enterotypes (Fig. 5e). Among them, the Asper_high had a higher proportion of responders in melanoma (P = 0.0027), while in NSCLC, the proportion of the Sacc_type was higher (P = 0.043). The fungal enterotypes had no significant association with PFS6 (Supplementary Fig. 7a), but the Sacc_type showed a higher proportion of PFS12 individuals (P = 0.0062, Supplementary Fig. 7b) in melanoma. Due to the limitation of sample size, no similar trend was observed in RCC (Supplementary Fig. 7c). Meanwhile, patients with the the Asper_low have shorter PFS (log-rank P = 0.019, Fig. 6f) and OS (log-rank P = 0.016, Fig. 6g) in melanoma, while there was no significant difference in PFS and OS among these three fungal enterotypes in NSCLC (Supplementary Fig. 7d, e). Additionally, we analyzed bacterial enterotypes, classifying them into two groups based on the abundance of Bacteroides: Bac_high and Bac_low (Supplementary Fig. 8a–c). However, these bacterial enterotypes were not statistically associated with the efficacy of ICIs (Supplementary Fig. 8d). We also did not find an association between bacterial enterotypes and fungal enterotypes (Supplementary Fig. 8e). Interestingly, we noted varying distributions of fungal enterotypes across different cancers and cohorts, with the Sacc_type predominant in NSCLC and the Asper_low more common in melanoma (Supplementary Fig. 8f).

Fig. 6: Functional validation of Debaryomyces hansenii in mice.
figure 6

a Details of the association between Debaryomyces hansenii and response in each cohort of melanoma; b The experimental design of mice included the dose of Debaryomyces hansenii and the treatment plan of PD-1 mAb; c The relative abundance of Debaryomyces hansenii in the stool of the gavage and the non-gavage at the end of the experiment by ITS sequencing; P-values are determined by Wilcoxon test; d Tumor growth curves measured every two days; Tumor volume (mm3) was calculated as: (shortest diameter)2 × (longest diameter)/2; e The picture of Tumor entity; f Representative immunohistochemistry images of CD8-positive cells and quantitative data analysis; scale bar: 100 μm; g The relative expression of immune factors (CCL5, CXCL10) and immune checkpoint (PD-L1) by qPCR; Unless otherwise specified above, one-way ANOVA was used for statistical analysis, while a two-tailed unpaired Student’s t-test was used to compare the two groups; Data are the mean ± s.e.m; n = 6 per group; *P < 0.05; **P < 0.01; ***P < 0.001.

Full size image

Debaryomyces hansenii influenced the efficacy of ICI in the mouse model

Debaryomyces hansenii was found to be associated with the decreased effectiveness of ICIs in melanoma, as demonstrated by logistic regression analysis (Meta OR = 0.69, Meta P = 0.02) (Fig. 6a). To further investigate the impact of Debaryomyces hansenii on ICI efficacy, we utilized a subcutaneous xenograft mouse model implanted with the melanoma cell line B16F10 (Fig. 6b). In addition, we performed Internal Transcribed Spacer (ITS) sequencing of microbial DNA in mouse feces and discovered that the abundance of Debaryomyces hansenii was significantly higher in the daily gavage group compared to the control group (P = 0.008, Fig. 6c). Following four intraperitoneal injections of either IgG2a or PD-1 mAb, we observed that Debaryomyces hansenii not only promoted tumor growth (P = 0.03, Fig. 6d), but also reduced the effectiveness of PD-1 mAb treatment (Fig. 6e).

Subsequently, we performed CD8 immunohistochemistry on mouse tumor tissues, revealing that a high abundance of Debaryomyces hansenii was associated with a reduced positive proportion of CD8+ T cells (P = 0.012, Fig. 6f). Additionally, we observed decreased expression levels of CCL5 (P = 0.04), CXCL10 (P = 0.04), and PD-L1 (P = 0.007) when comparing treatment with PD-1 mAb alone to treatment with PD-1 mAb combined with Debaryomyces hansenii (Fig. 6g).

Discussion

A large number of studies have recently highlighted the significant role of gut bacteria and other host factors in the effectiveness of ICIs19. However, the specific gut bacterial profiles associated with ICI efficacy differ across studies, influenced by geographic location, dietary practices, and other factors16. Furthermore, the predictive ability of gut bacteria for the effectiveness of ICIs was inadequate, although multi-cohort analyses were integrated to detect bacterial alterations20. Recent studies have shown that combining intestinal bacteria and fungi can enhance predictions regarding tumor response to ICIs. However, the majority of these investigations have concentrated on a specific cancer type or a broad array of cancers12,13,14. Herein, we focused on the role of intestinal fungi in the context of ICI treatment across different tumors. We analyzed metagenomic sequencing data of 976 samples from eight cohorts, covering three cancer types (melanoma, NSCLC, and RCC). This included 301 responders and 650 non-responders. Through a meta-analysis of these diverse cohorts, we present a comprehensive profile of gut microbiota’s response across various tumor types and identify species-level markers from different kingdoms that are associated with ICI efficacy.

Herein, we identified 27 fungi in melanoma and 23 fungi in NSCLC associated with immunotherapy across a multi-cohort. Notably, most of them are plant pathogenic or involved in food fermentation fungi, and their role in mice or humans has been infrequently documented. Numerous investigations have demonstrated that intestinal fungi can significantly influence the host’s physiological and pathological conditions8,21, and our findings may guide future investigation in this field. One notable fungus identified in our analysis is Aspergillus avenaceus, which has the potential to enhance the effectiveness of ICIs in melanoma treatment. This fungus produces avenaciolide, a compound known for its antifungal and antibacterial effect, which also inhibits peptidoglycan biosynthesis in methicillin-resistant Staphylococcus aureus22,23. This suggests that Aspergillus avenaceus may play a role by influencing the growth of other microbes. The main components of the fungal cell wall include mannan or β-glucan, which can be recognized by the pattern recognition receptors of immune cells, like natural killer T cells, dendritic cells, and macrophages, to induce an immune response against fungal infection24. Aspergillus sydowii induces the expression of the CARD9 signaling pathway in host cells through β-glucan binding to Dectin-1 in the cell wall, thereby inhibiting the killing ability of CD8+ T cells and ultimately promoting the development of lung cancer25. Malassezia globosa is strongly associated with the high expression of PD-L1 in patients with gastric cancer26 and can accelerate the development of pancreatic cancer by stimulating the mannose-binding lectin-complement C3 pathway27. It’s also worthwhile to look into whether Malassezia globosa and other fungi have comparable effects in the treatment with ICIs. Furthermore, related research has shown that certain fungus species alter immune checkpoint expressions during infection. For example, Candida albicans escapes the killing of natural killer T cells by binding to the checkpoint receptor TIGIT with lectin- like sequence protein28. PD-L1, a common immune checkpoint, is highly abundant in phagosomes that contain yeast, such as Saccharomyces cerevisiae, and can directly bind to the yeast surface ribosomal protein Rpl20b, controlling the inflammatory response of phagocytes and the release of cytokines like IL-1029. Candida auris infection causes upregulation of the PD-1/PD-L1 pathway in mice, suggesting the potential to improve the efficacy of immune checkpoint blocking therapy30. In short, we cannot ignore the important role of intestinal fungi and their metabolites in intestinal immunity, nor can we ignore that they may change the expression of immune checkpoints and then affect the efficacy of ICIs in the treatment of tumors.

The dynamics and resilience of the entire community are regulated by the gut microbiome, an ecosystem made up of intricate microbial interactions10. The development and course of disease are influenced by interactions between fungi and bacteria31. Current studies have confirmed that intestinal bacteria and fungi can interact through the secretion of metabolites to affect colonization and growth32,33, biofilm formation34 and other processes. While antibiotic-mediated bacterial depletion lowers reactivity and is linked to the expansion of symbiotic fungi, antifungal-mediated fungal depletion increases the effectiveness of radiation therapy, indicating that gut bacteria and fungi collaborate to maintain antitumor immunity35. Here, we constructed intestinal bacteria-fungus interaction networks in melanoma and NSCLC. In melanoma patients, we observed that both intra-kingdom (bacteria-bacteria, fungi-fungi) and inter-kingdom (bacteria-fungi) correlations were stronger among responders compared to non-responders, a trend consistent with earlier reports by Huang et al.14. Conversely, in NSCLC, we found a reversal of this trend, likely due to the low microbial abundance influenced by sequencing depth and the limited percentage of responders (20% of the total samples) on the other. Interestingly, most associations in both melanoma and NSCLC were positive, with a slightly higher proportion observed among responders. Overall, this also implies that we should modulate the interaction between bacteria and fungi through the use of antibiotics, probiotics, antifungal medications, or dietary strategies, which will enable us to successfully increase the efficacy of ICIs. Just as probiotic bacteria like Bifidobacterium bifidum (favorable for response to ICI in melanoma in our study) can prevent intestinal fungi from colonizing by generating lipopeptides10. The probiotic fungi, such as Saccharomyces boulardii, can enhance humoral immunity and innate immunity, improve the integrity of intestinal epithelium, and restructure the composition of the gut microbiota36.

It has been suggested that enterotypes can efficiently stratify populations, summarize the traits of the human gut microbiome, and give a general picture of the variations in gut microbiome composition among individuals18,37. In this study, we identified three fungal enterotypes: Asper_high, Sacc_type, and Asper_low. Compared with the classification by Lai et al.38, we lacked enterotypes characterized by Candida, which may be due to geographical location. Most of our samples were from Europe, where the flora was characterized by the expansion of Saccharomyces and Penicillium but the consumption of Candida38. We found that patients with the Asper_high had a higher proportion of response to ICI treatment, followed by those with the Sacc_type, which was a rare perspective on the association between fungal enterotypes and ICI efficacy. The majority of current research centers around the possible connection between bacterial enterotypes and illness or drug response39,40, which also serves as a theoretical guide for investigations involving fecal microbiota transplantation (FMT)41. An aging-enriched enterotype with abundant Bacteroides has been linked to a positive response to ICI therapy42, which is consistent with the characteristics of bacterial enterotypes in this study. Thus, it also offers a study approach for boosting the effectiveness of ICIs through FMT and improving prognosis dependent on the fungal enterotypes. Additionally, we found that fungal enterotypes differ across cohorts and cancer types, which may be influenced by sequencing platforms and geographic locations.

Considering the culturability and safety of fungi in the laboratory, we selected the top 10 associated Debaryomyces hansenii with response to ICIs in melanoma for in vivo validation. Debaryomyces hansenii is an uncommon human fungal pathogen that is frequently present in dairy products, particularly processed meats and cheeses43, but it is frequently detected in stool metagenomic sequencing12,14,44. At present, the relevant literature shows that Debaryomyces hansenii not only can enrich the damaged part of the intestine and hinder intestinal repair45, but also is associated with autoimmune diseases46. However, its impact on tumor immunotherapy remains unclear. It appears to decrease the effectiveness of ICIs in treating melanoma by lowering the percentage of CD8+ T cells, according to both our association findings and in vivo confirmation in mice. Nevertheless, further investigation is needed to elucidate the underlying mechanisms of this interaction.

The study had some limitations. First, the small number of cohorts for NSCLC and RCC and the fact that most of the samples were sourced from France or Canada limited the assessment of the impact of geographic differences on gut microbiota and affected the accuracy of the results, with some heterogeneity present. The microbial signatures in various tumors fluctuate greatly, and there are comparatively few common microbial signatures because of the imbalance in cohort numbers, as well as variations in sequencing platforms and depths. Additionally, this causes some discrepancies in the findings of enterotypes and microbial interaction analyses between melanoma and NSCLC. It is impossible to rule out that the fraction of responders, sequencing depth, and cohort size had an impact on this. Second, because fungal DNA is difficult to extract and its abundance in the gut is low, more precise sequencing methods and sequencing depth are required. The sequencing depth of some cohorts in this study was low, resulting in sparse fungal abundance, and the obtained association results were biased to a certain extent. Third, some of the fungi associated with the efficacy of ICIs are difficult to culture or have a certain risk, and it cannot even be determined whether they are contaminated strains. Therefore, its role cannot be verified. Next, it is impossible to tell whether the low-abundance fungi are actual biological signals in the intestinal fungus or just technical noise because this study does not have blank control or negative control samples. Finally, it is currently unknown how exactly intestinal fungi, such as Debaryomyces hansenii, impact the efficacy of ICIs; further research is required to elucidate this mechanism.

Overall, we conducted a meta-analysis to identify microbial signatures that influence the effectiveness of ICIs across various cancers using fecal metagenomic sequencing data. We also validated the role of Debaryomyces hansenii in melanoma immunotherapy through functional validation in vivo experiments. Furthermore, we examined the connection between fungal enterotypes and ICI treatment, as well as the interaction between microbial kingdoms. Altogether, our findings indicate that we should not overlook the role of intestinal fungi in tumor immunotherapy but rather highlight the crucial role of multi-kingdom microbes in ICI therapy, thereby providing insights for improving tumor immune efficacy and providing a reference for further studies on intestinal fungi.

Methods

Collection of metagenomic files and metadata from patients treated with ICIs

The National Center for Biotechnology Information (NCBI) was searched for studies on human fecal shotgun metagenomic sequencing related to ICIs. Our analysis included data from 1161 patients across eight cohorts, comprising 489 patients with melanoma, 571 patients with NSCLC, and 101 patients with RCC. Metagenomic sequencing data were downloaded from the European Nucleotide Archive (ENA) for the following projects: PRJEB22863 for RoutyB_201847, PRJEB43119 for LeeKA_20225, PRJNA397906 for FrankelAE_201748, PRJNA399742 for MatsonV_201849, PRJNA541981 for PetersBA_202050, PRJNA1023797 for DerosaL_20246, PRJNA762360 for McCullochJA_202251, and PRJNA770295 for SpencerCN_202152. Additionally, metadata were manually curated from the respective published studies.

Given the clinical tumor radiological data and the Response Evaluation Criteria in Solid Tumors (RECIST) v1.1, the following classifications of response to ICIs are made in clinical practice.: complete response (CR), partial response (PR), stable disease (SD), and progressing disease (PD). In our study, responders (R) were classified as either PR or CR, while non-responders (NR) included all other categories. We also defined additional metrics, including PFS, OS, PFS6, and PFS12, which align with the original published studies. These metrics have been thoroughly verified for consistency.

Sequence preprocessing and taxonomic profiling

The KneadData v.0.6.1 tool (http://huttenhower.sph.harvard.edu/kneaddata) was used to process our data, ensuring the microbial reads were of high quality and free from contaminants. Low-quality reads were filtered out using Trimmomatic. The remaining reads were mapped to mammalian genomes, bacterial plasmids, complete plastomes, and UniVec sequences with Bowtie2 v.2.3.5.153. This mapping process allowed us to isolate microbial sequences by removing those associated with host or laboratory origins.

For taxonomic classification of bacteria, archaea, fungi, and viruses, we employed Kraken2 v.2.1.254, an advanced metagenomic taxonomy classifier. A custom database consisting of 38,278 bacterial reference genomes, 580 archaeal reference genomes, and 14,972 viral reference genomes from the NCBI RefSeq database (downloaded on September 15, 2023), along with 1919 fungal reference genomes from the NCBI RefSeq database, FungiDB (http://fungidb.org, v65), and Ensemble (http://fungi.ensembl.org, v57), was built. To estimate taxonomic abundance, especially at the species and genus level, Bracken (v2.8) was used based on the results from Kraken255. The read counts for each species were then converted into relative abundance. Pearson correlation analysis was used to analyze the correlation between sequencing depth and fungal reads.

Sample filtration

To ensure data consistency and quality, we conduct a sample filter process based on the data features of this study and the guidelines established by Lin et al.12. Firstly, we removed the samples with low alignment read counts (<1 million), which were attributed to low sequencing depth and host contamination. Additionally, given that the genetic material of fungi is less than 1% of the total microbial material in the intestine44,56, and the extremely low fungal reads make the abundance overly sparse, we retained samples where the total microbial read counts showed a ratio of fungal reads between 0.01% and 1%. To assess potential bias in the sample filtering process, we employed the χ2 test, confirming that samples were evenly distributed among different groups.

Microbial diversity analysis

The metrics of α-diversity of each sample, specifically the Shannon index, were calculated using the R package “otusummary”. The Wilcoxon test was utilized to assess differences in alpha diversity between responders and non-responders, taking into account stratification variables such as cohort, gender, age, and BMI. Age and BMI were initially continuous variables, but for the convenience of analysis and the completeness of clinical information collection, we divided age into the old (≥60 years old) and the young (<60 years old). BMI was classified as lean (<25), overweight (25–30), and obese (>30) according to conventional cut-offs for obesity. Likewise, β-diversity was evaluated using the Bray–Curtis distance. A permutational multivariate analysis of variance (PERMANOVA) was conducted with 999 permutations to examine the differences in the microbial community between responders and non-responders, adjusted by cohort, age, and gender.

Identification of microbial features associated with efficacy of ICIs

To standardize the study, batch effect correction of microbial relative abundance was carried out using the MMUPHin approach57, given the sparse attributes of microbial data and heterogeneity among cohorts. Fungal species with a relative abundance of >0.1% and mapped in at least 10% of samples were included in subsequent association analyses. The criterion for the relative abundance of bacterial species was set at 0.001% to accommodate the significant variability in the number and abundance of gut fungal and bacterial species observed. The relative abundance data of microbial species were ranked from large to small, with the first 33% of samples classified as high group, the last 33% as low group, and the remaining samples as middle group, and multivariate logistic regression or Cox regression was employed to investigate their associations with the response to ICIs, PFS6, PFS12, OS, and PFS. Meanwhile, in the cohorts of FrankelAE_2017 and RoutyB_2018, we adjusted for the effects of covariates such as gender, age, and read-depth. In the cohort of MatsonV_2018, gender and read-depth corrections were employed. For the remaining cohorts, gender, age, BMI, and read-depth were used for correction. A meta-analysis was carried out to identify the associated microbial signatures, applying a p-value threshold of greater than 0.05 for heterogeneity (het P) and less than 0.05 for significance in the meta-analysis (Meta P).

Construction of bacteria-fungus interaction network

Given that some marginal significantly correlated signatures may be involved in the interaction between bacteria and fungi, we chose to the species with Meta P less than 0.1 to constructed the interaction network by calculating the Spearman correlation coefficient in the different groups (responders vs. non-responders). The p-value was corrected for multiple comparisons using the “BH” method to obtain the FDR. When FDR is less than 0.05 and the absolute value of the correlation coefficient is greater than 0.2, it is considered that there is a correlation between the microbial abundances. Particularly, the empirical p-value with less than 0.05 was utilized to assess whether the correlation of microbial pairs was different between responders and non-responders using permutation testing in the R package “DCGA”58. The resulting network was visualized with Gephi v0.10.1.

Identification of microbial enterotype

The samples were organized into enterotypes with partitioning around medoid (PAM) based on the Bray–Curtis distance estimated at the genus level. This analysis focused on genera with an abundance greater than 0.1% in at least 40% of the samples. The optimal number of enterotypes was determined using Silhouette scores, with the maximum average relative abundance representing the defining characteristic of each enterotype. To test the association between enterotypes and response to ICIs, we employed the χ2 test or two-sided test for Fisher’s Exact test.

Validation of characteristic fungi in mice

Debaryomyces hansenii NRRL Y-833 was obtained from the China General Microbiological Culture Collection Center (depository number: CGMCC 2.493) and cultured on Sabouraud Dextrose Broth (SDB) at 28 °C for 36 h by oscillating. The count was determined by the dilution-coated plate method.

Six-week-old female conventional C57BL/6 mice were housed in an air-filtered and pathogen-free condition. After one week of acclimation, mice were randomly divided into four groups based on body weight with 6 mice in each group. Murine melanoma B16F10 cells were resuspended in phosphate-buffered saline (PBS; 5 ×105 cells) and injected into the right flank of each mouse. Intragastric administration of 200 μL of Debaryomyces hansenii (5 ×108 colony-forming units per mL) or PBS one week in advance until the end of the experiment. One week after tumor inoculation, each mouse was intraperitoneally injected with 200 μg of anti-mouse PD-1 mAb (BioXCell, BE0146, USA) or IgG2a (BioXCell, BE0089, USA) every 3 days for a total of 4 times. Tumor volume was measured every two days using vernier calipers. At the end of the experiment, fungi in the feces of mice were detected by ITS sequencing to determine the abundance of Debaryomyces hansenii. All animal experiments were approved by the Experimental Animal Ethics Review Committee of Xiangya Hospital, Central South University (No. 202407132). All mice were anesthetized by inhalation of isoflurane (RWD, R5102210, China) and euthanized by cervical dislocation.

Immunohistochemistry

For IHC, tissue sections were incubated overnight at 4 °C with primary antibodies overnight (CD8, Abcam, code #ab209775). Following this, they were treated with a secondary antibody (Abcam, code #ab205718) conjugated to horseradish peroxidase (HRP) at room temperature for 45 min. The sections were then counterstained with hematoxylin. Statistical analysis of the results was performed using ImageJ.

RNA isolation and real-time quantitative PCR

Total RNA was extracted from tumor tissues using Trizol (Taraka, code #9109), and the quality and concentration of RNA were measured using the Nanodrop 2000 (ThermoFisher Scientific, Massachusetts, USA). 1 μg of total RNA was reverse transcribed to determine relative mRNA expression with PrimeScriptTM RT reagent kit (Takara). Quantitative PCR was performed using TB qPCR Premix (Takara #639676) on an Eppendorf Mastercycler (Eppendorf), utilizing the following primers:

β-actin: Forward: (5’-ATATCGCTGCGCTGGTCGT-3’, Reverse: (5’-CCTTCTGACCCATTCCCACC-3’);

CCL5: Forward: (5’-GCTGCTTTGCCTACCTCTCC-3’), Reverse: (5’-TCGAGTGACAAACACGACTGC-3’);

CXCL10: Forward: (5’-TCCCTATGGCCCTCATTCTCA-3’), Reverse: (5’-CTTACTGACTGGCATGAGGATCA-3’);

PDL1: Forward: (5’-GCTCCAAAGGACTTGTACGTG-3’), Reverse: (5’-TGATCTGAAGGGCAGCATTTC-3’).

For quantification of immune-related genes in tumor tissue, relative abundance was calculated by the ΔΔCt method and normalized to the amount of β-actin.

Statistical analysis

Multivariate logistic regression was used to examine the association between binary categorical variables (response to ICIs, PFS6, and PFS12) and microbial abundance. Multivariate Cox regression models were used to determine the association between outcomes (OS, PFS) and microbial abundance. Survival curves were estimated using the Kaplan-Meier method, and the log-rank test was applied for comparisons. The relationship between enterotypes and response to ICIs was determined by the χ2 test or two-sided test for Fisher’s Exact test. One-way analysis of variance (ANOVA) for the entire analysis with Tukey’s post hoc test for multiple comparisons was used to analyze the results of multiple groups, while a two-tailed unpaired Student’s t test was used to compare the two groups in animal experiments. A p-value of less than 0.05 was considered statistically significant, and two-sided tests were used unless otherwise indicated. Furthermore, we also employed the Benjamini–Hochberg (BH) method for multiple hypothesis correction to obtain the corrected p-values. All analyses were performed on R 4.3.3 and GraphPad Prism 10.1.2.