Graphical abstract graphic file with name fx1.jpg [65]Open in a new tab Highlights * • Microbial signatures are associated with immunotherapy resistance in MSI-H GI cancers * • Machine learning model based on microbiome data can predict primary resistance * • During the process of acquired resistance, several microbes are changed gradually __________________________________________________________________ Cheng et al. report the microbiota signature associated with primary/acquired resistance of ICI in MSI-H GI cancers and validate these findings in an independent validation cohort and mouse model. Based on clinical microbiota data, a machine learning model is built to predict drug resistance. Introduction The deficient mismatch repair (dMMR) subtype accounts for 15%–22% patients with gastrointestinal (GI) cancers[66]^1^,[67]^2^,[68]^3 and ∼5% metastatic/recurrent GI cancers.[69]^4^,[70]^5^,[71]^6 Patients with the dMMR subtype are unable to recognize and repair certain spontaneous mutations, resulting in a quite high tumor mutation burden and microsatellite-instability-high (MSI-H) status, and therefore are more likely to benefit from anti-PD-1/PD-L1 (antibodies targeting programmed cell death protein-1 [PD-1] or its ligand PD-L1) immunotherapy. According to existing evidences, MSI status is one of the most effective biomarkers of cancer immunotherapy. However, patients with MSI-H/dMMR GI cancer present highly heterogeneous responses to immunotherapy, and roughly 30% MSI-H patients exhibit primary resistance to immunotherapy.[72]^7 Moreover, ∼17% MSI-H patients present acquired resistance following a 2-year treatment, and the proportion usually increases with the prolonged treatment.[73]^7 Considering the rather consistent molecular patterns within the MSI-H/dMMR subtype, we believe understanding other factors that impact efficacy heterogeneity is of great importance to enhance immunotherapy response in dMMR/MSI-H patients. In our previous work, we demonstrated that gut microbiome composition could predict the efficacy of anti-PD-1 immunotherapy in patients with GI cancer,[74]^8 and researches on other tumor types also implicate the role of gut microbiome in modulating host response of various immunotherapies.[75]^9^,[76]^10^,[77]^11^,[78]^12 Metabolites of the gut microbiome, such as short-chain fatty acids (SCFAs)[79]^13^,[80]^14^,[81]^15^,[82]^16 and inosine,[83]^17 are shown to influence host immunity—synergistically modulating anti-tumor effects. Two independent investigations demonstrated that responder-derived fecal microbiota transplantation (FMT) could restore anti-PD-1 responses in patients with PD-1-refractory melanoma.[84]^18^,[85]^19 In spite of these encouraging findings, microbiota signatures related to efficacy vary from different cohorts,[86]^20 and this phenomenon may be attributed to the heterogeneity between tumor types and even within the same cancer type. To a certain extent, MSI-H/dMMR patients present highly consistent molecular characteristics, and thus research in this subtype is more likely to find microbiota biomarkers that can be generalized and applied in multiple tumor types. Moreover, previous studies have mainly focused on primary resistance, which further restricts their clinical translation, considering a rather large proportion of patients usually develop acquired resistance to immunotherapy treatment. We present here the findings of our investigation into the gut microbiome profiles of 77 patients with advanced MSI-H/dMMR GI receiving anti-PD-1/PD-L1 therapy and further validated the findings on another independent cohort (N = 39). To examine and identify interplay between gut microbiota, their metabolic products, and host immune responses germane to primary and acquired drug resistance, we applied integrated multi-omics analysis to paired blood and fecal samples at baseline and during treatments, then we further developed an effective predictive machine learning model based on microbiome signatures. A summary of the microbial, metabolomic, and immunologic signatures identified in this study as relating to primary and/or acquired anti-PD-1/PD-L1 resistance ensues. Results Advanced MSI-H/dMMR GI cancer cohort A total of 98 patients with advanced GI cancer with MSI-H/dMMR were consented and enrolled at the Beijing Cancer Hospital from February 2018 to October 2020, including 30 patients with gastric cancer (GC) and 68 patients with colorectal cancer (CRC). Fifteen patients were excluded because anti-PD-1/PD-L1 was combined with chemo- or targeted therapies (N = 12) or used as adjuvant therapy (N = 3). Six patients did not provide sufficient fecal samples for metagenomic sequencing and therefore were removed from further analysis ([87]Figure S1). The current cohort included 77 patients with 18 patients with GC and 59 patients with CRC (including 3 patients with rectal cancer and 56 patients with colon cancer). All participants were at stage III/IV and undergoing treatment with anti-PD-1/PD-L1, in some instances combined with anti-CTLA-4 immunotherapy. All patients were identified as MSI-H or dMMR through corresponding analytic measurements ([88]Table 1). We defined a patient as a responder (R) if the patient achieved an objective response (complete response [CR]/partial response [PR]/stable disease [SD]) lasting at least 6 months upon treatment start or as a non-R (NR) (progressive disease [PD] observed within 6 months of treatment start). Response rates (∼66% response rate) across our cohort of 77 patients with MSI-H/dMMR GI cancer were comparable to other cohorts.[89]^7 No significant association was observed between clinical benefits and metadata such as age, gender, and BMI ([90]Table 1). Baseline is defined as pre-treatment or no more than 3 weeks after immunotherapy treatment. A total of 290 fecal samples were collected at baseline and along the treatment period ([91]Figure 1). Meanwhile, baseline blood samples (plasma) from a subset of patients (15 GC, 55 CRC) were also collected for metabolome and cytokine/chemokine panel measurement. For the validation cohort, 39 MSI-H/dMMR patients on anti-PD-1/PD-L1 treatment were enrolled, with 19 patients with GC and 20 patients with CRC ([92]Table S2). From the 39 patients, a total of 85 fecal samples were collected for downstream metagenomics analysis, and baseline fecal samples were collected from 31 patients. Table 1. Cohort characteristics Clinical factor Total (N = 77) (%) R (N = 51) (%) NR (N = 26) (%) p value Gender 0.0822 Female 29 (37.66) 23 (45.10) 6 (23.08) – Male 48 (62.34) 28 (54.90) 20 (76.92) – Cancer 1 Colorectal cancer 59 (76.62) 39 (76.47) 20 (76.92) – Gastric cancer 18 (23.38) 12 (23.53) 6 (23.08) – BMI 0.1014 Lean (BMI < 18.5) 13 (16.88) 9 (17.65) 4 (15.38) – Normal (18.5 ≤ BMI < 25) 48 (62.34) 28 (54.90) 20 (76.92) – Obese (≥25) 16 (20.78) 14 (27.45) 2 (7.7) – Age 0.2295 <50 38 (49.35) 28 (54.90) 10 (38.46) – ≥50 39 (50.65) 23 (45.10) 16 (61.54) – Lynch syndrome[93]^a 1 Yes 14 (18.18) – – – No 43 (55.84) – – – N/A 20 (25.97) – – – RAS/RAF status (colon cancer N = 59) 0.1936 NRAS mutation 1 (1.69) 0 (0.00) 1 (5.00) – KRAS mutation 25 (42.37) 16 (41.03) 9 (45.00) – BRAF V600E MT 4 (6.78) 4 (10.26) 0 (0.00) – BRAF other mutation[94]^b 2 (3.39) 1 (2.56) 1 (5.00) – Wild type 13 (22.03) 6 (15.38) 7 (35.00) – NA 14 (23.73) 12 (30.77) 2 (10.00) – Left_or_Right (colon cancer N = 59) 0.384 Left 20 (33.90) 11 (28.21) 9 (45.00) – Right 36 (61.02) 25 (64.10) 11 (55.00) – N/A 3 (5.08) 3 (7.69) 0 (0.00) – Helicobacter pylori (gastric cancer N = 18) 0.061 Neg 6 (33.33) 6 (50.00) 0 (0.00) – Pos 5 (27.785) 2 (16.67) 3 (50.00) – N/A 7 (38.89) 4 (33.33) 3 (50.00) – [95]Open in a new tab BMI, body mass index; N/A, not available; Neg, negative; Pos, positive. ^a Lynch syndrome was defined as patients with pathogenic germline mutation. ^b BRAF other mutation contains D594G and p.P403fs. Figure 1. [96]Figure 1 [97]Open in a new tab Study design and clinical sample collection (A) 18 patients with gastric cancer (GC) and 59 patients with colorectal cancer (CRC) with MSI-H/dMMR were recruited. Patients were classified as being responders (Rs) or non-Rs (NR). Acquired resistance (AR) patients and long Rs (LRs) were further differentiated in Rs. (B) In total, 290 stool and 70 blood plasma samples were collected. CR, complete response; PD, progressive disease; PR, partial response; SD, stable disease. See also [98]Figure S1. Characteristics of gut microbiome with primary resistance to immunotherapy To explore the global signature associated with primary resistance to immunotherapy, we performed multi-omics analysis of the baseline samples. We first analyzed the metagenomics data collected from the fecal samples. Consistent with the findings of our previous study involving 74 patients with GI cancer,[99]^8 no significant differences were observed in alpha diversity between Rs and NRs based on Shannon indices ([100]Figure S2A). Beta diversity (Principal Coordinates Analysis, PCoA) depicting the overall gut microbial structure showed no significant difference between Rs and NRs (permutational multivariate analysis of variance [PERMANOVA] p = 0.5) ([101]Figure 2A). To find the common characteristics of MSI-H patients and avoid the influence of cancer type, we performed a combined analysis and blocked this factor with the MaAsLin2 (Microbiome Multivariable Association with Linear Models) analysis. Figure 2. [102]Figure 2 [103]Open in a new tab Gut microbiome composition, functional pathways, and blood metabolites associated with drug response (A) PCoA plot of fecal samples arranged by response and diagnosis using Bray-Curtis dissimilarity. The x and y axes show the first and second principal coordinates, along with the percentage of variance explained on each dimension. NR, non-responders; R, responders. (B) A volcano plot of taxa differentially abundant at baseline between Rs (n = 26) and NRs (n = 22) based on MaasLin2 results. Taxa that differed significantly between groups (p < 0.05) were named accordingly. (C) A volcano plot of metabolome differentially abundant between Rs (n = 46) and NRs (n = 24) based on MaasLin2 results. Metabolome that differed significantly between groups (p < 0.05) were named accordingly. The horizontal axis represents the coefficient value. If the coefficient value is greater than zero, then the metabolite is enriched in the Rs. The vertical axis represents the q values (the adjusted p value). (D) Differences in the relative abundance of different short-chain fatty acids between Rs (n = 46) and NRs (n = 24) based on Wilcoxon test. (E) Top 25 functional pathways enriched in Rs based on the untargeted metabolome data, sorted by p value. See also [104]Figures S2–S7 and [105]Tables S1, [106]S2, [107]S3, [108]S4, and [109]S5. We explored the significantly altered species between R and NR patients ([110]Figure S2B) and found that Bacteroides caccae, Porphyromonadaceae, Parabacteroides, Acidaminococcaceae, Alistipes finegoldii, Paraprevotella clara, Bacteroides massiliensis, and Alistipes putredinis were enriched in Rs, while Veillonella parvula, Veillonella atypica, Peptostreptococcaceae, Streptococcus thermophilus, and Micrococcaceae were enriched in the NR group ([111]Figure 2B; [112]Table S1). Time series analysis including the longitudinal samples revealed that 11 species were consistently differential between R and NR patients, in that V. parvula and V. atypica were consistently in higher abundance in NR samples, while Coprobacter, A. putredinis, B. caccae, and B. clarus were consistently higher in R samples ([113]Figures S3 and [114]S4). Our analysis on the independent validation cohort with 31 baseline samples ([115]Table S2) showed that B. caccae, Porphyromonadaceae, and Parabacteroides were also enriched in Rs and that Micrococcaceae were enriched in the NR group ([116]Figure S5), which further validates the reliability and stability of our findings. Characteristics of fecal metabolome with primary resistance to immunotherapy To render an unbiased view of metabolic profiles of patients with different treatment responsiveness, we performed global metabolomic profiling using liquid chromatography-mass spectrometry (LC-MS) on baseline blood plasma samples. We detected and putatively identified 241 and 300 distinct spectral features by LC/MS (electrospray ionization [ESI]+) and LC/MS (ESI−) modes, respectively ([117]Table S3). Tight clustering of the quality control (QC) samples in the principal-component analysis (PCA) plot was observed in both LC/MS (ESI+) and LC/MS (ESI−) modes ([118]Figure S6), indicating good instrument stability and reproducibility of the obtained data. We compared the metabolite profiles of Rs and NRs to obtain potential pathways associated with primary drug resistance. In total, 143 metabolites were identified as differential between Rs and NRs ([119]Figure 2C; [120]Table S4). Of these, two types of metabolites are noteworthy, including ones involved in arginine metabolism, such as arginine, ornithine, glutamine, asymmetric dimethylarginine, asparagine, serine, citrulline, and proline. We also explored the relationship between arginine and gut microbes and found several microbes significantly associated with arginine quantities ([121]Table S5), including Paraprevotella clara (p < 0.05, correlation coeffficient = 0.4) and V. atypica (p < 0.05, correlation coeffficient = −0.4), which were previously found to be differential features between Rs and NRs. The other type of metabolites is associated with SCFA metabolism, such as 4-aminobutanoate, capric acid, 2-ketohexanoic acid, 4-acetamidobutanoic, and 3-hydroxybutyric acid. Targeted metabolomics on SCFAs was also applied, and we found that butyrate, caproic acid, and iso-butyrate were significantly enriched in Rs ([122]Figure 2D). Pathway enrichment analysis between R and NR groups also showed significance in propanoate metabolism, which is related to SCFA metabolism, along with arginine and proline metabolism ([123]Figures 2E and [124]S7). Clinical benefits transferred by FMT in mice To validate the role of the gut microbiome in mediating patients’ responses to immunotherapy, fecal samples from 6 patients (3 Rs, 3 NRs) were transplanted into an MC38 mouse model pre-treated with broad-spectrum antibiotics. While evaluating potential synergistic effects of FMT alongside anti-PD1 treatment ([125]Figure 3A), we observed a significant reduction in tumor size in mice transplanted with Rs’ fecal material (R group) compared to those receiving NRs’ stool (NR group) or anti-PD1 alone (negative control [NC] group) ([126]Figure 3B). Using multiplex immunochemistry (mIHC) to appraise T cell populations in the tumor microenvironments (TMEs), we observed increased fractions of helper T cells (CD3^+CD4^+CD8^−), cytotoxic T cells (CD3^+CD4^−CD8^+), and regulatory T (Treg) cells (CD4^+CD8^−FoxP3^+) in mice receiving Rs’ stool ([127]Figures 3C and 3D). These results suggest that FMT of the Rs’ fecal samples can alter the immune cell landscape in the tumor microenvironment (TME) and enhance immunotherapy. Figure 3. [128]Figure 3 [129]Open in a new tab Enhanced anti-tumor effects in mice following fecal microbiota transplantation (FMT) with feces from Rs (A) Experimental design. (B) Tumor growth curve. Graphic representation of the results of a one-way ANOVA followed by Tukey’s multiple comparisons test. Fecal material was obtained from 3 Rs and 3 NRs, with each gavaged to 3 mice. (n = 9 for R/NR group; n = 3 for control group). ∗∗∗ represents p values < 0.01. (C) Helper T cells (CD3^+CD4^+CD8^−), cytotoxic T cells (CD3^+CD4^−CD8^+), and Treg (CD4^+CD8^−FoxP3^+) T cell population in the tumor microenvironment, quantified through mIHC. Graphic representation of the results of a one-way ANOVA followed by Tukey’s multiple comparisons test. ∗ represents p values < 0.05. (D) Representative images of mIHC of tumor tissue samples. NC, negative control; R, responders; NR, non-responders. Yellow: interferon γ (IFN-γ), green: Foxp3, magenta: CD8, red: CD3, and cyan: CD4. (E) Heatmap showing the differential metabolites identified from blood metabolome. (F) Dot plot showing significantly enriched KEGG pathways based on mouse tumor gene expression data. (G) Network analysis of tumor size, blood metabolites, and T cell population in the tumor microenvironment. Spearman correlations were calculated between each feature comparison, and those yielding an adjusted p value (p.adjust-value) <0.05 were retained. Node size indicates feature degree. See also [130]Tables S6 and [131]S7. We further performed a multi-omics analysis to dissect the differences between the R and NR groups, including metabolomics profiling of mouse blood samples and RNA sequencing (RNA-seq) analysis of mouse tumor tissues. A total of 97 differential metabolites were identified, including arginine, which was previously found in higher quantities in R patients ([132]Figure 3E; [133]Table S6). RNA-seq pathway enrichment analysis also pinpointed arginine biosynthesis as significant ([134]Figure 3F), in which the expression of four genes involved in the urea cycle (nitric oxide synthase 2 [NOS2]/gm5424/arg1/arg2) was up-regulated in the R group ([135]Table S7). These results indicate that an active ammonia metabolism and arginine synthesis may impact the immune checkpoint inhibitor (ICI) response. Furthermore, we explored co-occurrence relationships between blood metabolites and tumor immune cell populations and found that asymmetric dimethylarginine, a metabolite related to arginine metabolism, was significantly associated with Treg cells, indicating the role of arginine metabolism in modulating the immune function ([136]Figure 3G). Gut-microbiome-based machine learning model well predicted the efficacy of immunotherapy We sought to find out whether it is feasible to predict the response probability prior to ICI treatment and identify biomarkers associated with clinical outcomes using gut microbiome data. To achieve this goal, we built a LightGBM machine learning model with microbial species data ([137]Figure 4A). The model was trained on data from patients that had baseline samples, and 10-fold cross-validation was conducted to assess the model’s performance. To evaluate the robustness of the model, we collected another independent validation cohort comprising 31 baseline stool samples, as detailed in [138]Table S2. This trained model demonstrated excellent performance in predicting efficacy of ICI treatment using microbial species data. Specifically, we achieved area under the curve (AUC) of 0.9 in the training dataset, 0.88 in the testing set, and 0.79 in the independent validation set ([139]Figure 4A). By analyzing the contributions of features to our model’s prediction of response status, we identified that B. caccae, V. parvula, V. atypica, and Clostridiales bacterium were the top 4 indicative features to ICI response ([140]Figure 4B). It is noteworthy that the first three have been previously identified as significantly different. These findings highlight the potential of gut microbiome as a predictor to identify patients who are likely to benefit from immunotherapy, and the selected microbial species may be potential biomarkers to predict response of ICI treatment. Figure 4. [141]Figure 4 [142]Open in a new tab Machine learning model based on microbial biosignature (A) Receiver operating characteristic (ROC) curves of machine learning models based on species (AUC = 0.76). (B) Feature contribution represented by the SHAP value in the three above-mentioned machine learning models. Biosignatures of patients with acquired resistance to immunotherapy During clinical practice, a considerable number of patients developed acquired resistance (AR) to immunotherapy, which was defined as patients who fail to respond to ICI following objective response or prolonged SD for more than 6 months.[143]^21 On the other hand, long Rs (LRs) were defined as having a sustained benefit over 1 year and without disease progression at the time of analysis. Thus, to understand the role of gut microbiome in the process of AR, we further divided Rs into AR and LR groups. Baseline multi-omics analysis identified a few differential features (microbes, metabolites, or cytokines/chemokines) between AR and LR samples, including significantly increased interleukin-5 (IL-5), hydrocinnamic acid, 2-hexadecenal, and biliverdin in the AR group as well as increased salicyluric acid and glycoursodeoxycholic acid in the LR group ([144]Table S8). To get a systematic view of gut microbiome’s role during AR, we further analyzed patients’ gut microbiome before and after they acquired ICI resistance. A total of 18 species associated with drug resistance were identified ([145]Figure 5A). Notably, two Alistipes species (A. onderdonkii and A. putredinis), which were found to be lower in AR samples, were also decreased in the primary resistant group (NRs), indicating their potential role in ICI resistance. Species such as Ruminococcus bromii, Bifidobacterium adolescentis, and Bifidobacterium pseudocatenulatum were significantly decreased during the drug resistance process, and they were also found to be increased for a long time in LR patients ([146]Figure 5B). Figure 5. [147]Figure 5 [148]Open in a new tab Gut microbiome associated with AR (A) Differential microbes identified during the progression of drug resistance. (B) Microbial dynamics in representative LR patients, focusing on the differential microbes found during the progression of AR. See also [149]Figures S8–S11 and [150]Table S8. Additionally, the differences between NR and AR microbial composition at baseline ([151]Figures S8 and [152]S9) and dynamics during treatment ([153]Figures S10 and [154]S11) were evaluated. Bacteroidales, Barnesiella, and Parabacteroides, which were higher in AR baseline samples, were still in higher quantities during the treatment period. Granulicatella and Rothia, which showed no significant difference between AR and NR baseline samples, were increased in AR patients as treatment prolonged. Peptostreptococcaceae, however, was consistently enriched in NR patients along the treatment frame. Discussion The role of gut microbiome in mediating cancer immunotherapy has attracted increased attention in recent years.[155]^22^,[156]^23 Several studies have identified the correlation between gut microbiota and the efficacy of immunotherapy in melanoma, non-small cell lung cancer, etc.[157]^9^,[158]^11^,[159]^17^,[160]^18^,[161]^19^,[162]^24^,[163]^25 ^,[164]^26 However, there have been variations in microbial signatures across different studies, leading to some inconsistencies. Several recent clinical trials have revealed the potential roles and applications of FMT in promoting response[165]^10^,[166]^19 and overcoming resistance to immunotherapy in melanoma.[167]^18^,[168]^27 Nevertheless, achieving precise manipulation remains challenging. The complex and diverse nature of gut microbiome makes it difficult to establish consistent associations between specific microbiota and treatment response. Particularly, the interactions between gut microbiome and immunotherapy in GI cancer have been scarcely investigated. Here, we investigated the roles of the gut microbial ecosystem in primary resistance and AR to immunotherapy in patients with advanced MSI-H/dMMR GI cancer, which focused on the rare but clinically significant subpopulation with MSI-H/dMMR. Among the various biomarkers for immuno-oncology treatment, MSI and MMR status plays an important role in GI cancer. By focusing on a cohort with MSI-H/dMMR characteristics, we expect to elucidate the interplay between gut microbiome, host immune status, and drug resistance. As the occurrence of MSI-H/dMMR in patients with GI cancer is rather low (3%–5% in metastatic/recurrent GI cancers), we have screened a large cohort of almost 1,000 patients with GI cancer and were fortunate enough to gather a cohort of 59 patients with CRC and 18 patients with GC who met the strict inclusion and exclusion criteria presented in the study. We analyzed gut microbiome, blood metabolome, and cytokines/chemokines of 77 patients at baseline and during ICI treatment. Several key microbes and metabolites were identified to be associated with response, which was validated in another independent cohort comprising 39 patients. We further validated our findings using FMT and developed a machine learning model to predict response with high accuracy. Overall, our findings indicated that gut microbiome plays an essential role in drug resistance and could be used as a tool for preventative diagnosis. Moreover, we focused on the gut-microbiome-related signatures specific to the MSI-H/dMMR molecular subtype in GI cancer in this study. Although a cancer-specific microbiome might exist, here we aim to identify a more general pattern related to MSI-H/dMMR to help understand whether primary resistance and AR within this subpopulation can increase the overall response rate to ICI in patients with GI cancer with MSI-H/dMMR. Thus, during the analysis, we conducted a combined analysis and blocked the confounder factors to identify gut microbiome signatures related to MSI status. The results indicate that there were certain shared microbiome-based biosignatures, which related to patients’ MSI status rather than cancer types. We identified gut bacteria, blood metabolites, and corresponding functional pathways associated with primary resistance. Notably, arginine was enriched in Rs of our cohort as well as the R group in the animal trial. Arginine is a conditionally essential amino acid, and its metabolism is regulated by gut microbiota[169]^28. Arginine is shown to promote immunotherapy by regulating T cell metabolism through three transcription regulators (baz1b, Psip1, and TSN).[170]^29 Multi-stage cooperative nanodrugs based on arginine and cisplatin CpG have been shown to promote CD8^+ T cell infiltration and exert a synergistic effect with anti-PD-L1 treatment both in vitro and in vivo.[171]^30 Engineered probiotic Escherichia coli Nissle 1917, which converts ammonia to L-arginine, could significantly increase T cell infiltration in the TME and enhance the efficacy of anti-PD-L1 immunotherapy.[172]^31 Meanwhile, we observed that asymmetric dimethylarginine (ADMA), another metabolite related to arginine metabolism, was positively associated with Treg cells in the FMT animal trial and suspected that ADMA may have a different mode of action than arginine in regulating the immune response. For example, Chen et al. found that ADMA impaired proliferation and phagocytosis of tumor cells for macrophages and even caused death in cultures incubated without arginine.[173]^32 These findings suggest that although both arginine and ADMA are metabolites of arginine metabolism, they may have different biological functions and can have opposing effects on the immune function. In this study, we also found the increase of certain SCFAs in Rs, including butyrate. As the most studied microbiota-derived SCFA, butyrate uncoupled the tricarboxylic acid cycle from glycolytic input in CD8^+ T cells and then enhanced memory potential of activated CD8^+ T cells.[174]^33 He et al. demonstrated that butyrate treatment directly boosted the anti-tumor cytotoxic CD8^+ T cell responses both in vitro and in vivo in an ID2-dependent manner through the IL-12 signaling pathway.[175]^15 Besides butyrate, capric acid and isobutyrate were also significantly enriched in Rs from our MSI-H/dMMR cohort. Compared to butyrate, there is very limited research on the mechanisms by which caproic acid and isobutyrate regulate host immunity. Whether caproic acid and isobutyrate exhibit similar mechanisms to butyrate still needs further exploration. Our finding of the R-enriched species and pathways that are capable of producing these SCFAs (Porphyromonadaceae) may provide extra evidence on the pro-immunotherapy effect of commensal bacteria.[176]^34 We further developed machine learning models with the goal to evaluate the usability of the microbiome-related signature as anti-PD-1/PD-L1 immunotherapy biomarkers. Our predictive models exhibited excellent predictive performance in an independent external validation set. Another highlight of our work was the focus on AR. Having hypothesized that gut microflora play an active role in mediating AR, we set out to identify significant differences in the gut microbiome of AR patients before and after resistance. Several microbes possibly contributing to drug resistance were identified, including Bifidobacterium. Among the three Bifidobacterium species we identified (B. adolescentis, B. pseudocatenulatum, and B. breve), the first two were enriched before AR, while the last one was enriched after resistance occurred. Although previous reports have shown that commensal Bifidobacterium could promote immunotherapy efficacy in mouse models,[177]^17^,[178]^25 the effect of different species could vary as shown in our findings, considering the vast differences in the Bifidobacterium genus regarding host preference and metabolic functions.[179]^35^,[180]^36^,[181]^37 Further, alterations in microbial composition between NR and AR have been observed at baseline and during the treatment. Specific taxa such as Barnesiella and Parabacteroides were enriched at baseline in AR and remained higher during the treatment, which was reported to be associated with immunomodulation in tumor.[182]^38^,[183]^39 Overall, these findings suggest the modulation roles of gut microbiome in immunotherapy, and further studies are still needed to fully elucidate these complex dynamics in relation to clinical outcomes. Our study has some limitations that, in turn, may serve as future avenues of investigation and exploration. For the FMT animal models, we observed some but not statistically significant tumor inhibition effects of NR samples with ICI compared to the NC group. This may be due to the fact that FMT of exogenous microbial community itself could to some extent stimulate immune system and thus benefit ICI efficacy regardless of stool sample origin (R or NR), which warrants further investigation. Secondly, although we collected MSI-H samples from over 1,000 patients and validated our findings in an independent validation cohort, we acknowledge the limited patient number in this study, which is due to the low occurrence of MSI-H/dMMR in GI cancer. Additionally, due to the lack of similar reports or data online, we cannot validate our findings in other GI tumor types at present. We will consistently gather samples and stay updated on the latest research developments. Once we accumulate an ample sample size, we plan to extend our findings to larger patient groups and other GI tumor categories. Conclusion By interrogating a rare cohort of patients with MSI-H/dMMR GI cancer with a suite of integrated multi-omics analyses, we identified biosignatures whose differential presence/absence or relative abundance correlated significantly with primary or acquired modes of immunotherapeutic drug resistance. Especially, microbes, pathways, and/or metabolites involved in arginine metabolism and SCFA metabolism indicate certain common molecular mechanisms related to gut microbes and hosts’ response to immunotherapy. Robust machine learning models were also built to predict patient response to immunotherapy in addition to the “golden” MSI status, which could greatly improve the accuracy of patient stratification to maximize treatment benefits. We believe that these exciting findings will help to guide the clinical practice in cancer immunotherapy and further explore the feasibility and efficacy of FMT in reversing the immunotherapy resistance in patients with GI cancer. Limitations of the study We acknowledge the limited patient number in this study due to the rare nature of MSI-H/dMMR subtypes. Although the findings were validated in an independent cohort, caution should be taken when extrapolating to a different population. Moreover, in vivo FMT trials confirmed the role of gut microbiome in enhancing ICI efficacy, but the roles of arginine metabolism and related metabolites are unclear and warrant further investigations. STAR★Methods Key resources table REAGENT or RESOURCE SOURCE IDENTIFIER Antibodies __________________________________________________________________ Anti-CD3 Invitrogen Cat#MA5-14524; RRID: AB_10982026 Anti-CD8 Cell Signaling Cat#98941; RRID: AB_2756376 Anti-CD4 Cell Signaling Cat#25229; RRID: AB_2798898 Anti-Foxp3 Cell Signaling Cat#12653; RRID: AB_2797979 Anti-IFN-γ abcam Cat#ab216642; RRID: AB_3076637 __________________________________________________________________ Biological samples __________________________________________________________________ Stools from advanced gastric or colorectal cancer patients with MSI-H/dMMR This paper N/A Blood samples from advanced gastric or colorectal cancer patients with MSI-H/dMMR This paper N/A Tumor tissues from MC38 bearing mice This paper N/A __________________________________________________________________ Chemicals, peptides, and recombinant proteins __________________________________________________________________ Ampicillin MedChemExpress Cat#HY-B0522 Neomycin MedChemExpress Cat#HY-150520 Metronidazole MedChemExpress Cat#HY-B0318 Vancomycin MedChemExpress Cat#HY-B0671 __________________________________________________________________ Critical commercial assays __________________________________________________________________ V-PLEX Plus Kits Meso Scale Diagnostics, Rockville, MD, USA Cat#K15198G Wehealthgene® Fecal Microlution TM Collection kit Wehealthgene Cat#ML-001A QIAamp PowerFecal DNA Kit Qiagen Cat#12830-50 __________________________________________________________________ Deposited data __________________________________________________________________ Metagenomics sequencing This paper Accession number SRA: PRJNA1042925 Mass spectrum metabolomics data This paper Metabolomics Workbench repository: MTBLS9021; [184]https://www.ebi.ac.uk/metabolights RNA-seq data of tumor tissues from MC38 bearing mice This paper Accession number SRA: PRJNA1044251 __________________________________________________________________ Experimental models: Cell lines __________________________________________________________________ MC38 cell line National Infrastructure of Cell-Line Resource, Peking Union Medical College, Beijing, China Cat#1101MOU-PUMC000523 __________________________________________________________________ Experimental models: Organisms/strains __________________________________________________________________ C57/BL6J mice Beijing Huafukang Bioscience [185]http://www.hfkbio.com/ __________________________________________________________________ Software and algorithms __________________________________________________________________ R v3.6.3 R Core Team [186]https://cran.r-project.org/ Trimmomatic Bolger et al.[187]^40 [188]http://www.usadellab.org/cms/?page=trimmomatic Bowtie2 Langmead et al.[189]^41 [190]https://bowtie-bio.sourceforge.net/bowtie2/index.shtml MetaPhlAn2 Truong et al.[191]^42 [192]https://huttenhower.sph.harvard.edu/metaphlan2/ HUMAnN2 Franzosa et al.[193]^43 [194]https://huttenhower.sph.harvard.edu/humann2/ MaAsLin2 Mallick et al.[195]^44 [196]https://huttenhower.sph.harvard.edu/maaslin/ [197]Open in a new tab Resource availability Lead contact Further information and request for resources should be directed to and will be fulfilled by the lead contact, Zhi Peng zhipeng@bjmu.edu.cn (Z.P.). Materials availability This study did not generate new unique reagents. Data and code availability Metagenomics and RNA-seq data generated in this study have been deposited in the NCBI database under BioProject accession number SRA: PRJNA 1042925 and SRA: PRJNA1044251. The raw LC/MS data as well as the processed metabolic profiles and corresponding metadata for the human (deidentified) and animal samples is publicly available on the Metabolomics Workbench repository: MTBLS9021([198]https://www.ebi.ac.uk/metabolights). Other data supporting the findings of this study are available from the [199]lead contact, Zhi Peng zhipeng@bjmu.edu.cn (Z.P.), upon request. Experimental model and study participant details Cohort recruitment and ethical approval Advanced gastric or colorectal cancer patients with MSI-H/dMMR were enrolled for participation at Beijing Cancer Hospital following the criteria described below. MSI and MMR are related to DNA repair mechanisms and measured using different methods. MSI is measured by analyzing microsatellite repeats in the patient’s DNA using PCR or next-generation sequencing, while MMR is measured by assessing the expression of proteins involved in the mismatch repair pathway by immunohistochemistry (IHC). In addition to the discovery cohort, we also recruited a validation cohort from the same center following the same criteria. Inclusion and exclusion criteria for participants have been established to ensure patient safety and minimize the risk of confounding factors that could interfere with the study results. The Inclusion Criteria included: 1) Patients aged 18 years or older with gastrointestinal tumors who are expected to receive immune checkpoint inhibitor therapy. 2) Patients who have a confirmed diagnosis of mismatch repair deficiency by immunohistochemistry (IHC), or microsatellite instability-high by polymerase chain reaction (PCR), next-generation sequencing (NGS) testing. 3) Patients who provide written informed consent to participate in the trial. The Exclusion Criteria included: 1) Patients receiving chemotherapy or targeted therapy simultaneously when receiving the immune checkpoint inhibitor therapy. 2) Patients who have had a fever or serious infection within one month prior to starting immune checkpoint inhibitor therapy. 3) Patients who have received broad-spectrum antibiotics orally or intravenously within one month prior to or during the immune checkpoint inhibitor therapy. 4) Patients who have undergone gastrointestinal resection or diversion surgery within the past six months. 5) Patients with active autoimmune diseases. Response to treatments was evaluated every 6 weeks and confirmed no less than 4 weeks from the date first recorded. In accordance with guidelines put forth in Response Evaluation Criteria in Solid Tumors version 1.1 (RECIST 1.1), patients in this study deemed to be either responders or non-responders. Patients classified as responders were further divided into acquired resistance [AR; where the best response is stable disease (SD), partial response (PR) or complete response (CR) and progressive disease (PD) is acquired after 6 months] or long responders [LR, where SD or partial response lasts more than 1 year without any evidence of disease progression at the time of recruitment]. Patients were classified as non-responders if the best response was PD or SD/PR/CR with acquired PD in 6 months. The definition of acquired resistance in ICI treatment is still evolving, and our approach was based on the current expert consensus from Society for Immunotherapy of Cancer (SITC), which provide a framework for defining and studying resistance to ICI therapy. Clinical laboratory examination and antibiotic (ATB) usage within 30 days were documented for all patients. This study was approved by the Ethics Committee of Peking University Cancer Hospital in 2018 (approval number: 2018KT66). Informed content was obtained from all patients enrolled for the collection of clinical information and samples, and all tests and procedures were conducted in accordance with the Declaration of Helsinki. Animal experiment This animal study was approved by the Animal Care and Use Committee at Beijing cancer hospital (Approval Number EAEC2021-02). We constructed the MC38 cell line transplant tumor model of C57BL/6J mice (female, aged 6 weeks, purchased from the Beijing Huafukang Bioscience [200]http://www.hfkbio.com/). Animals were kept in the SPF condition. Antibiotics (ANVM, 1 g/L ampicillin, 1 g/L neomycin, 1 g/L metronidazole, 0.5 g/L vancomycin) were administered through drinking water at 2 weeks prior to stool gavage (FMT). MC38 murine colon adenocarcinoma cells were cultured in DMEM medium (Gibco) with 10% FBS and 1% penicillin-streptomycin (Thermo Fisher Scientific) at 37°C with 5% CO2. MC38 cells were subcutaneously injected into the right flank of mice at a dosage of 5 × 105 one week after the initiation of FMT. Once the tumor reached ∼50 mm3, mice were intraperitoneally injected with 100ug PD-1 blockade (RMP1-14, BioXcell) every 3 days for 12 days (4 injections). Tumor volume was calculated by measuring length (a) and width (b) via vernier caliper and applying the equation v = ab2/2. To prepare the stool for gavage, one gram of stool sample was suspended in 5 mL of phosphate-buffered saline (PBS). Mice were orally gavaged with 200 μL of processed donor stool samples or PBS (as the negative control, NC) every three days. Fecal material was obtained from 3 responders and 3 non-responders, with each gavaged to 3 mice. The 3 responders included 2 gastric cancer (GC) patients and 1 colorectal cancer (CRC) patient while the 3 non-responders included 2 CRC patients and 1 GC patient. Total 21 mice were used in this study, including the NC group. All mice were sacrificed 4 weeks after tumor transplantation and tumor tissues were collected for downstream analysis. Method details Fecal microbiome sample preparation Stool samples were collected from patients at multiple timepoints throughout their immunotherapy regimen as previously described8 ([201]Figure 1). Briefly, fecal samples were collected from each patient at baseline and along the treatment period. We assessed treatment efficacy at each patient visit, which occurred every two weeks, and stool samples were collected at the same time. We continued to collect samples until the patient became the progression of disease (PD) state. The collected fecal samples were stored using the Wehealthgene Fecal Microlution TM Collection kit (catalog No. ML-001A, Wehealthgene). Then the samples were transported to the laboratory within 3 days. DNA was extracted using QIAamp PowerFecal DNA Kit (catalog No. 12830-50, Qiagen), followed by library construction and sequencing on an Illumina NovoSeq 6000 platform (Novo Gene). Blood metabolome sample preparation Baseline blood samples were drawn from 70 patients, and plasma was prepared by centrifugation at 3,000 rpm for 10 min at 4°C and stored at −80°C. In preparation for metabolomics analysis, plasma samples were thawed and centrifuged at 14,000 × g for 20 min in a cold room (4°C–8°C), and supernatants were transferred to sterile 1.5 mL microfuge tubes. A volume of 400 μL methanol (pre-chilled to −80°C) was added to each 100 μL supernatant. The final 80% (v/v) methanol solution was shaken and incubated at −80°C for 2h. Following centrifugation (14,000 × g, 10 min, 4°C), supernatants were transferred into autosampler vials and subjected to direct UHPLC-MS analysis. Ultimate 3000 UHPLC (Dionex) coupled with Orbitrap mass spectrometry (Thermo Fisher) was used to perform LC separation. In the negative mode, the BEH C18 column (2.1 × 100 mm, Waters) was applied for analysis at 0.25 mL/min. Mobile phase A was crafted by mixing 1L of HPLC-grade water containing 0.3953 g of Ammonium bicarbonate (pH ∼8). Mobile phase B was 100% ACN. A gradient was established as follows: 0∼3min, 1% B; 10–17 min, 99% B; 17.1–20.0 min, 1% B. In positive mode, samples were passed over a Waters BEH amide column (100 ∗2.1 mm, 1.7 μm) heated to 40°C using a gradient of 1% B to 99% B over 12.5 min (solvent A: 95% ACN +5% H2O + 10mM NH4FA; solvent B: 50% IPA +50% ACN +10mM NH4FA). Data with masses ranging from m/z 80–1200 to m/z 70–1050 were acquired at both the positive and negative ion modes with data dependent MSMS acquisition. The full scan and fragment spectra were collected at a resolution of 60,000 and 15,000, respectively. Detailed mass spectrometer parameters were as follows: spray voltage at 2.8 kV for negative and 3.2 kV for positive; capillary temperature set at 320°C; heater temperature set at 300°C; sheath gas flow rate set to 35; auxiliary gas flow rate set to 10. Metabolite identification was achieved based on Tracefinder (Thermofisher, CA) searches using a home-built database. Multiplex panel immunofluorescence (mIHC) staining of mice tumor tissues Multiplex panel immunofluorescence (mIHC) staining of tumor tissues was performed by Crown Bioscience Inc. Briefly, FFPE blocks were sectioned with a manual rotary microtome, 4 μm thickness/section. The following antibodies were used with the Bond RX autostainer: anti-CD3 (dilution 1:500, Invitrogen), anti-CD8 (dilution 1:400, Cell Signaling), anti-CD4 (dilution 1:100, Cell Signaling), anti-Foxp3 (dilution 1:400, Cell Signaling), and anti-IFN-γ (dilution 1:100, Abcam). All stained sections were scanned with Vectra multiplexed imaging systems at 20× magnification. High resolution imagery of whole sections was generated and subjected to quantification analysis with HALOTM software. Quantification and statistical analysis Blood SCFA quantification Short chain fatty acids were analyzed with a 6500plus QTrap mass spectrometer (AB SCIEX, USA) coupled to an ACQUITY UPLC H-Class system (Waters, USA). An ACQUITY UPLC BEH C18 column (2.1 × 100mm, 1.7μm, Waters) was employed with mobile phase A: 100% water and mobile phase B: 100% ACN over a linear gradient of 0–1 min, 2% B; 1–10 min, 50% B; 10–12 min, 98% B; 12–15 min, 2%B. Flow rate was 0.3 mL/min. The column chamber and sample tray were held at 45°C and 10°C, respectively. Data were acquired in multiple reaction monitormode, and ion transitions were optimized using chemical standards. Nebulizer (Gas1), heater (Gas2), and curtain gas pressures were set at 50, 50, and 35 psi, respectively. Source voltage was - 4500 V for negative ion mode, and optimal probe temperature was determined to be 550°C. SCIEX OS 1.6 software was used to identify metabolites and integrate peaks. Blood cytokine/chemokine quantification Concentrations of IL-1β, MIP-1α, PlGF, TARC, VEGF, IL-17A, IL-6, IP-10, MCP-1, IL-16, MIP-1β, IFN-γ, IL-5, TNF-α, IL-13, IL-15, MCP-4, CRP, GM-CSF, ICAM-1, IL-12p70, IL-2, IL-7, IL-8-HA-, Tie-2, VCAM-1, VEGF-D, Flt-1, bFGF, Eotaxin, Eotaxin-3, IL-10, IL-12-IL-23p40, IL-1α, IL-4, IL-8, MDC, SAA, TNF-β, and VEGF-C were measured from patient blood samples collected at baseline using V-PLEX Plus Kits (Catalog NO: K15198G, K15049G, K15190G, K15050G, K15047G; Meso Scale Diagnostics, Rockville, MD, USA), per manufacturer’s instructions. Two technical replicates were measured for each sample and average values were recorded. Assays were performed at AliveX Biotech (Shanghai, China). Fecal microbiome data analysis Raw reads were quality controlled via KneadData (version 0.6.1), which integrated several QC tools [e.g., FastQC ([202]https://www.bioinformatics.babraham.ac.uk/projects/fastqc/) and Trimmomatic.[203]^40 After trimming low-quality bases from reads and omitting reads less than 100 bp in length, bowtie2[204]^41 mapped filtered reads to the human genome (hg19), and host contaminants were removed. Taxonomic profiling was accomplished using MetaPhlAn2 with default parameters,[205]^42 and functional analyses were achieved using HUMAnN2 alongside the Uniref. 90 reference database.[206]^43 Once feature tables were generated (taxa, gene, pathway), alpha community diversity was calculated in accordance with Shannon indices while beta diversity was deduced based on Bray-Curtis distances. Permanova and beta dispersion analyses were applied to identify confounding clinical variables, such as age and gender. Features observed in a differential manner were then selected with MaAsLin2[207]^44 using previously identified confounding factors as fixed effects. When analyzing multiple samples from the same patient, “subject” was also treated as a random effect. All statistical analyses and plotting were performed in Rstudio (R version 3.6.3). Blood metabolome data analysis We applied the standard 80% rule to minimize the effect of missing values and retain only metabolites detectable in 80% or more patients in at least one group. QC samples were prepared by pooling equivalent aliquots of plasma from each of the samples and injected intermittently into the analytical process to monitor the stability of the method. Variables presenting relative standard deviations (RSD) greater than 30% in quality control (QC) samples were removed, replaced with one-half of the minimum value found in that dataset. Total area normalization and logarithmic transformation was performed to stabilize variance across the intensity range. Mean-normalized data from positive and negative ion modes were combined for downstream analysis. PCA, Permanova, and beta dispersion analyses were performed on normalized feature tables to identify potential confounding factors, and pathway enrichment analyses were achieved using the MetaboAnalystR package.[208]^45 Multi-omics analysis of animal experiment Statistical analyses of tumor growth and immune cell infiltration was performed using one-way ANOVA followed by Tukey’s multiple comparisons test (p < 0.05). The method of blood metabolomics is the same as human metabolome analysis. Applying Spearman’s rank correlation coefficient, networks were constructed between blood metabolites and tumor mIHC results. RNA sequencing and data analysis were completed by Novogene Co., Ltd (Beijing, China). Briefly, total amounts and integrity of RNA were assessed using the RNA Nano 6000 Assay Kit of the Bioanalyzer 2100 system (Agilent Technologies, CA, USA) followed by library construction and sequenced by the Illumina NovaSeq 6000 (150bp PE). We used clusterProfiler R package (3.8.1) to test the statistical enrichment of differential expression genes in KEGG pathways ([209]http://www.genome.jp/kegg/). Machine learning model construction and validation To identify potential responders at baseline level, we trained a LightGBM classifier using baseline metagenomic data of MSI-high responders and non-responders. Relative abundance of species was used to build models. Cancer types were transformed into dummy variables to improve model performance. ‘BorutaPy’ was used for feature selection and optimizing the performance in the machine model. It utilizes a randomized algorithm to iteratively compare the importance of each feature with the randomly generated shadow features, which assigns a ranking to each feature. The whole dataset was split into train and test set using ‘train_test_split’ of ‘scikit-learn’ package with parameter ‘test_size = 20%’. We evaluated the performance of LightGBMclassifier using 10-fold cross-validation. GridSearchCV was used for hyperparameter tuning and improving the accuracy of the model. It automates searching through a predefined hyperparameter grid to find the set of hyperparameters that result in the best model performance. Further, the model was validated on another independent cohort. Measurement of feature importance was performed by using ‘SHAP’ package. Acknowledgments