Purpose:

The gut microbiome is involved in antitumor immunotherapy and chemotherapy responses; however, evidence-based research on the role of gut microbiome in predicting response to neoadjuvant chemoradiotherapy (nCRT) in patients with locally advanced rectal cancer (LARC) remains scarce. This prospective, longitudinal study aimed to evaluate the feasibility of the gut microbiome in predicting nCRT responses.

Experimental Design:

We collected 167 fecal samples from 84 patients with LARC before and after nCRT and 31 specimens from healthy individuals for 16S rRNA sequencing. Patients were divided into responders and nonresponders according to pathologic response to nCRT. After identifying microbial biomarkers related to nCRT responses, we constructed a random forest classifier for nCRT response prediction of a training cohort of baseline samples from 37 patients and validated the classifier in another cohort of 47 patients.

Results:

We observed significant microbiome alterations represented by a decrease in LARC-related pathogens and an increase in Lactobacillus and Streptococcus during nCRT. Furthermore, a prominent microbiota difference between responders and nonresponders was noticed in the baseline samples. Microbes related with butyrate production, including Roseburia, Dorea, and Anaerostipes, were overrepresented in responders, whereas Coriobacteriaceae and Fusobacterium were overrepresented in nonresponders. Ten biomarkers were selected for the response-prediction classifier, including Dorea, Anaerostipes, and Streptococcus, which yielded an area under the curve value of 93.57% [95% confidence interval (CI), 85.76%–100%] in the training cohort and 73.53% (95% CI, 58.96%–88.11%) in the validation cohort.

Conclusions:

The gut microbiome offers novel potential biomarkers for predicting nCRT responses, which has important manifestations in the clinical management of these patients.

Translational Relevance

A reliable tool to predict response to neoadjuvant chemoradiotherapy (nCRT) in patients with locally advanced rectal cancer (LARC) is a critical need. Given that both chemotherapy and radiotherapy were reported to impact tumor immune microenvironment, we hypothesized that a correlation between the gut microbiome and nCRT responses may exist. We report the largest prospective, longitudinal study to date illustrating the predictive value of gut microbiome in nCRT. We identified potential microbial biomarkers, highlighting the significance of bacterial metabolites (butyrate) and microbes (e.g., butyrate-producing bacteria, Fusobacterium) in clinical treatment. These novel biomarkers may not only provide manipulation targets for future research, but also form the basis of establishing a response-prediction tool. To translate our findings into a clinical application, a response-prediction random forest classifier was established and validated to robustly discriminate responders and nonresponders to nCRT, which provided a noninvasive and repeatable tool for predicting response to nCRT prior to treatment initiation.

Colorectal cancer is one of the most common cancers and accounts for approximately 10% of cancer-related deaths worldwide (1). Approximately one third of colorectal cancers occur in the rectum (2), and at least 25% of cases are in an advanced stage at initial diagnosis. Considering the high risk of locoregional recurrence, neoadjuvant chemoradiotherapy (nCRT) has been established as a standard treatment for patients with locally advanced rectal cancer (LARC; ref. 3). Good responses to nCRT improve local control and can even shift clinical strategies from radical surgeries to “watch and wait” (4). However, nCRT responses vary among patients; in addition, nearly 20% of patients undergoing nCRT suffer from treatment-related toxicities, such as fatigue, radiation proctitis, and myelosuppression. In this context, a reliable tool that can allow prediction of therapeutic response to nCRT is a critical need, but is not yet available.

Radiotherapy and chemotherapy are the two components of nCRT, both of which have been reported to impact the tumor immune microenvironment (5). Intriguingly, preclinical and clinical studies have suggested that the gut microbiome actively participates in various biochemical and pathophysiologic reactions in the human body, especially those related to inflammation and immunomodulation (6–8). Therefore, it is rational to speculate that the gut microbiome may affect nCRT responses. Microbes correlating with immunotherapy or chemotherapy responses have been reported previously (9–15); with regard to radiotherapy, however, limited research has focused on the dynamic feature (16, 17) of the gut microbiome in patients upon irradiation and its association with radiation-related toxicities (18–20). A pilot study with 45 fecal samples from patients with LARC receiving nCRT provided some evidence on this issue (21), which merits further validation.

To this end, we initiated this prospective, longitudinal trial in 84 patients with LARC treated with nCRT. After evaluating the feasibility of the gut microbiome in predicting response to nCRT, we established a random forest classifier for its clinical application. In addition, our research shed light on how the gut microbiome effects radiotherapy responses, which could be clinically transformative.

Study design and study participants

This prospective, longitudinal study was performed at Fudan University Shanghai Cancer Center (FUSCC, Shanghai, P.R. China), in accordance with the Declaration of Helsinki of 1975. Ethical approval was obtained from the Institutional Review Board of FUSCC (Shanghai, P.R. China, 2007220-13), and written informed consent was provided by all subjects before sampling.

We recruited 85 patients with LARC who underwent nCRT from June 2018 to April 2019. Patients were excluded if having exposure to any antibiotics, prebiotics, probiotics, steroids, or immunosuppressants within 4 weeks prior to fecal sampling. More details on the inclusion and exclusion criteria are provided in the Supplementary Materials and Methods. Upon critical review, 84 patients were eligible, and 1 patient with liver metastases was excluded. In addition, 31 healthy individuals free of malignancies, as assessed by the investigators, were enrolled as controls. The healthy individuals were cohabitants of the enrolled patients with LARC, so as to control for confounders from lifestyle or diet. The nCRT consisting of conventional radiotherapy at 45–50 Gy, with the daily fraction being 1.8–2 Gy, combined with concurrent capecitabine-based chemotherapy was conducted according to our institutional protocol. Specially, in a previous phase III randomized clinical trial (CinClare, NCT02605265; ref. 22) conducted at our institution, the pathologic complete response rate in patients receiving capecitabine plus irinotecan guided by the UGT1A1*28 genotype status was found to be significantly higher than that in patients receiving capecitabine alone during nCRT (30% vs. 15%; P = 0.001), and the toxicities were acceptable. Motivated by the promising results and for ethical considerations, UGT1A1-guided prescription of irinotecan during nCRT has entered into our clinical practice, even though irinotecan is still not routinely recommended in guidelines.

We prospectively collected the demographic and clinicopathologic data of the patients. We used the American Joint Committee on Cancer (AJCC, eighth edition) four-tier tumor regression grade (TRG) system to evaluate the response to nCRT and accordingly divided patients into two groups. Among the 84 patients with LARC, 45 patients with TRG scores of 0–1 were grouped as responders, and 39 patients with TRG scores of 2–3 were grouped as nonresponders. The study design is summarized in the flow diagram (Supplementary Fig. S1A).

Fecal sampling and DNA extraction

Fecal samples were collected with reference to the method provided in the Manual of Procedures for Human Microbiome Project Protocol (#07-001). For each of the enrolled patients with LARC, two fecal samples were sequentially collected at the initial day (pre-nCRT, n = 84) and within 3 days upon completion of (post-nCRT, n = 83) nCRT treatment. From each of the healthy individuals, one fecal sample was collected at the baseline period (control, n = 31). All specimens were stored at −80°C in the laboratory before subsequent processing. DNAs were extracted from samples using the E.Z.N.A Stool DNA Kit (D4015, Omega, Inc.) according to the manufacturer's instructions.

PCR amplification and 16S rRNA sequencing

To amplify the hypervariable V3-V4 region of the 16S rRNA gene, DNA extracted from fecal samples was amplified with modified versions of primers 341F and 805R. PCR was conducted under the following conditions: initial denaturation at 98°C for 30 seconds; 32 cycles of denaturation at 98°C for 10 seconds, annealing at 54°C for 30 seconds, and extension at 72°C for 45 seconds; and final extension at 72°C for 10 minutes. The PCR products were purified with AMPure XT Beads (Beckman Coulter Genomics) and quantified using Qubit (Invitrogen). Sequencing was performed on an Illumina MiSeq Platform by a sequencing provider (LC-Bio).

The raw sequence data have been deposited in the Genome Sequence Archive in National Genomics Data Center, under accession number CRA002850, that are accessible at https://bigd.big.ac.cn/gsa.

Sequence processing and bioinformatic analysis

Paired-end raw sequences were merged using FLASH (v1.2.8; ref. 23), and clean sequences were obtained after quality checking with fqtrim (v0.9.4). The removal of chimeras, generation of representative sequences, and operation taxonomy units (OTU) feature table were completed by Vsearch (v2.11.1; ref. 24). The representative sequences were aligned to the Ribosomal Database Project classifier for taxonomic annotation. Normalization to the smallest sample size was conducted prior to diversity calculation, and the alpha and beta diversities were generated by using Usearch (v10.0.240; ref. 25). Four alpha indices were chosen to depict taxa diversity within samples, including richness, Chao1, and Shannon index, which measure taxa abundance, as well as Simpson index, which represents both richness and evenness. Four beta distances were used to measure intergroup dissimilarities, among which UniFrac provides phylogenetic information (26), whereas Jaccard distance, Bray Curtis distance, and Euclidean distance only provide information on the abundance or presence of taxa.

In exploration of microbiome data, unconstrained principal coordinates analysis (PCoA) was performed using abovementioned beta distances. The community types of LARC specimens were analyzed by unsupervised clustering, including Ward minimum variance hierarchical clustering, K-means partitioning, and partitioning around medoids (PAM) clustering, using normalized OTU abundance. In taxonomic comparison analysis, abundance tables were prefiltered to include only taxa with average relative abundances >0.01% and then analyzed by Welch t test provided by the Statistical Analysis of Metagenomics Profile (STAMP, v2.1.3). Corrected P values (q value) less than 0.05 after multiple comparisons correction using Storey FDR method were considered significant (27). In assessment of discriminatory taxa between groups, linear discriminant analysis (LDA) effect size (LEfSe; ref. 28) was performed to conduct high-dimensional nonparametrical comparisons, with the LDA score set at 2.0. In network analysis, the microbial cooccurrence networks were inferred by using Spearman rank correlations and visualized in R software (v3.6.2; ref. 29), with a threshold of correlation magnitude and P value equal to 0.7 and 0.01, respectively. The microbial phenotypes were predicted by using the Bugbase (https://www.biorxiv.org/content/biorxiv/early/2017/05/02/133462.full.pdf).

Statistical analysis

According to data feature, continuous data were analyzed with Student t test or Mann–Whitney test, and categorical data were analyzed with Pearson χ2 test or Fisher exact test. Permutational multivariate ANOVA (PERMANOVA) was used for statistical significance testing of beta dissimilarities between selected groups. To adjust for potential covariates, we utilized the microbiome regression-based kernel association test (MiRKAT), which tests the association between microbiome composition and outcomes of interest, based on both phylogenetic and nonphylogenetic distances. Aside from good control of type I error rate and easy covariate adjustment, MiRKAT also allows for simultaneous consideration of several types of beta distances, avoiding the problem of cherry picking the best result (30). Missing clinical data were imputed using the random forest method provided by “mice” package in R software. All statistical analyses were conducted in R software and a two-sided P < 0.05 was considered statistically significant.

Establishment of response-prediction classifier

The random forest analysis was conducted for selecting discriminatory biomarkers and constructing a response-prediction classifier. The training cohort and validation cohort were randomly selected from pre-nCRT samples at a ratio of 1:1. According to published methods (31, 32), we first constructed a classifier I incorporating microbial variables only, and the set of input features consisted of relative abundance of 40 genera that were found to be potential discriminatory taxa in STAMP and LEfSe analysis. Following five repeats of 10-fold cross-validation, a cut-off point on the averaged error curve was defined as the minimum error plus the SD at that corresponding point. Among all the sets of variables, with a corresponding error rate less than the cutoff, the set with the fewest number of variables was selected as the optimal set. With the variables in the optimal set, a response-prediction classifier was constructed and tested in the training and validation cohorts, respectively. The probability of response (POR) was calculated as the percentage of samples predicted as “responders” in all samples entering the classifier. Classifier II was constructed in a similar way, with the set of input features consisting of not only the 40 genera, but also 26 clinical variables, including patient age, gender, body mass index (BMI), tumor location, tumor size, tumor stage, etc. More details on the model construction can be found in the Supplementary Materials and Methods.

The area under the receiver-operating characteristic curve (AUC) was used to assess the performance of the classifier. For evaluating the clinical translational value of the classifier, other performance metrics, including sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), were also summarized at the optimal cut-off value of 0.5 defined by the ROC curve.

The gut microbiome in patients with LARC significantly differed from that in healthy individuals

From the 198 samples, a total of 4,897,877 sequences (mean length, 392 bp) were mapped to 3,267 OTUs. Rarefaction curves of OTU richness reached a plateau (Supplementary Fig. S2A), indicating that our sequencing depth was sufficient.

To obtain an overview of the microbiome profiles in patients with LARC, we first performed comparison analysis between patients with LARC and healthy individuals. No significant differences in alpha indices were observed between patients with LARC and healthy individuals (Supplementary Fig. S2B). In PCoA, no matter calculated with which beta distance, patients with LARC and healthy individuals clustered distinctly (Fig. 1A; Supplementary Fig. S2C). Notably, compared with the relatively homogeneous specimens from healthy individuals, samples from patients with LARC seemed to self-segregate into two clusters. To analyze the LARC community types, we applied unsupervised clustering, including Ward hierarchical clustering (Fig. 1A), K-means (Fig. 3C), and PAM (Supplementary Fig. S2D), on the baseline specimens from patients with LARC. In every method of clustering, the specimens were prominently separated into two clusters, which were labeled as LARC type1 and LARC type2, respectively. Using these labels as grouping variables, we revisualized the PCoA in all baseline samples. Consistent findings were observed in results from every method of clustering: LARC type1 was significantly distinct from healthy individuals (P < 0.001 in all three methods) and from LARC type2 (P < 0.001 in all three methods), whereas no significant differences were found between LARC type2 and healthy individuals (P = 0.14, 0.09, and 0.05 in Ward hierarchical clustering, K-means, and PAM respectively).

Figure 1.

The gut microbiome in patients with LARC (n = 84) significantly differed from that in healthy individuals (n = 31). A, PCoA calculated using Euclidean distance at the OTU level separated patients with LARC from controls significantly (PERMANOVA; P < 0.001); PC1 and PC2 explained 8.815% and 6.885% of variance, respectively (left, top). PCoA calculated using Euclidean distance separated LARC type1 (defined by Ward hierarchical clustering) from controls significantly (P < 0.001), whereas LARC type2 resembled the controls (P = 0.14; left, bottom). Dendrogram of Ward minimum variance hierarchical clustering using Bray Curtis distance at the OTU level, accompanied by the stacked bar plot depicting the common taxa at the family level in the two LARC community types (right). Association between the LARC community types with therapeutic response status was significant (P = 0.008). B, Stacked bar plot showing composition of common bacteria at the phylum level in samples from patients with LARC before nCRT (n = 84) and after nCRT (n = 83) and from controls (n = 31), respectively. C, Differential relative abundances of 15 LARC-enriched genera by the STAMP, the average relative abundance of each identified taxa in their respective groups and the 95% CI (left and right, respectively). D, Microbial cooccurrence network deduced using Spearman rank correlations, based on samples from healthy individuals. Only statistically significant (P < 0.01) connection with magnitude >0.7 (positive correlation, red edges) or <−0.7 (negative correlation, blue edges) is shown. Each node represents OTU and the size of node is proportional to corresponding abundances. Each node is labeled at the genus level and colored according to affiliated phylum. E, Microbial cooccurrence network based on samples from patients with LARC. Control, healthy individuals; NR, nonresponder; R, responder.

Figure 1.

The gut microbiome in patients with LARC (n = 84) significantly differed from that in healthy individuals (n = 31). A, PCoA calculated using Euclidean distance at the OTU level separated patients with LARC from controls significantly (PERMANOVA; P < 0.001); PC1 and PC2 explained 8.815% and 6.885% of variance, respectively (left, top). PCoA calculated using Euclidean distance separated LARC type1 (defined by Ward hierarchical clustering) from controls significantly (P < 0.001), whereas LARC type2 resembled the controls (P = 0.14; left, bottom). Dendrogram of Ward minimum variance hierarchical clustering using Bray Curtis distance at the OTU level, accompanied by the stacked bar plot depicting the common taxa at the family level in the two LARC community types (right). Association between the LARC community types with therapeutic response status was significant (P = 0.008). B, Stacked bar plot showing composition of common bacteria at the phylum level in samples from patients with LARC before nCRT (n = 84) and after nCRT (n = 83) and from controls (n = 31), respectively. C, Differential relative abundances of 15 LARC-enriched genera by the STAMP, the average relative abundance of each identified taxa in their respective groups and the 95% CI (left and right, respectively). D, Microbial cooccurrence network deduced using Spearman rank correlations, based on samples from healthy individuals. Only statistically significant (P < 0.01) connection with magnitude >0.7 (positive correlation, red edges) or <−0.7 (negative correlation, blue edges) is shown. Each node represents OTU and the size of node is proportional to corresponding abundances. Each node is labeled at the genus level and colored according to affiliated phylum. E, Microbial cooccurrence network based on samples from patients with LARC. Control, healthy individuals; NR, nonresponder; R, responder.

Close modal

In terms of microbial composition, Firmicutes, Actinobacteria, and Bacteroidetes were the predominant phyla in all samples (Fig. 1B). To reveal the general microbiome features in patients with LARC compared with healthy individuals, taxonomic comparisons were made by treating patients with LARC as a whole group, despite the heterogeneity discovered in unsupervised clustering. STAMP revealed that 15 genera were significantly enriched in patients with LARC compared with healthy individuals (Fig. 1C), including Prevotella, Porphyromonas, Fusobacterium, Parvimonas, and Peptostreptococcus. High-dimensional comparison from LEfSe further indicated an overrepresentation of Bacteroides fragilis (B. fragilis), Phascolarctobacterium faecium, and Ruminococcus torques in patients with LARC, while Akkermansia muciniphila (A. muciniphila), Roseburia inulinivorans (R. inulinivorans), Bifidobacterium kashlwanohense, and Streptococcus salivarius subsp. salivarius were significantly more abundant in healthy individuals (Supplementary Fig. S2E and S2F).

Aside from specific microbes, alterations in the microbial community were also highlighted to understand the microbial contributions to colorectal cancer progression (33). Therefore, we examined the microbiota network and bacterial phenotype prediction. Drastic differences were observed between the microbiota networks of patients with LARC and healthy individuals (Fig. 1D and E). Contrary to healthy individuals, in whom the gut microbiota interactions occurred mostly within genera, the gut microbiota in patients with LARC formed two densely connected modules composed of diverse genera. One module was centered on butyrate-producing bacteria, including Faecalibacterium prausnitzii, Roseburia, and Eubacterium. Butyrate is a preferred energy source for normal colonocytes and is associated with gut health by regulating cellular functions, including gene expression, inflammation, differentiation, and apoptosis (34, 35), and butyrate-producing bacteria are reported to inhibit tumor growth (34, 36). Hence, we considered this module to consist of “beneficial microbiota” driven by butyrate-producing microbes. The other module was centered on Bilophila wadsworthia and Bacteroides ovatus, both of which are reported to be related to colorectal cancer (37, 38). Specifically, Bilophila wadsworthia is a sulphite-reducing bacterium that is correspondingly elevated in saturated fat–enriched diets (39) and stimulates a Th type (Th1) immune response leading to colitis (39, 40). Thus, we considered this module to consist of “pathogenic microbiota.” Regarding bacterial phenotype prediction, the potentially pathogenic bacteria were increased in patients with LARC versus healthy individuals (Supplementary Fig. S2G). The two groups were similar with respect to other bacterial phenotypes, such as biofilm-forming ability and stress-tolerant elements. In summary, our results demonstrated that the gut microbiome in patients with LARC significantly differed from that in healthy individuals, regarding composition, interactions, and phenotypes.

Gut microbiome variations over the course of nCRT associated with therapeutic response

After identifying the signature of the gut microbiome in patients with LARC, we then explored what impacts nCRT would impose on the gut microbiome, and whether or not these perturbations on microbiome would correlate with therapeutic response. Repeated fecal sampling, before and following nCRT, was conducted to determine how the gut microbiome interacts with nCRT treatment and outcome in a time-series manner.

In our analysis, the pre-nCRT samples demonstrated significantly higher diversity and a trend toward higher unevenness than the post-nCRT samples (Fig. 2A); this variation trend was also true in the subgroup analysis stratified by response status. Intriguingly, the variation in the responder subgroup was more prominent than that in the nonresponder subgroup, with the decline in richness and Chao1 index, both reaching statistical significance in responder subgroup but not in the nonresponder subgroup (Supplementary Fig. S3A and S3B). PCoA based on Euclidean distance (Fig. 2B) and UniFrac distance (Supplementary Fig. S3C) showed pronounced clustering effects by sampling timepoint (P = 0.01 and P = 0.002, respectively).

Figure 2.

Gut microbiome variations over the course of nCRT associated with therapeutic response. A, Microbial richness and Chao1 diversity were significantly reduced in patients with LARC following nCRT (P = 0.025 and P = 0.028, respectively), and Simpson index showed a trend toward higher evenness following nCRT (P = 0.38). B, PCoA using Euclidean distance at the OTU level separated pretreatment samples from posttreatment samples significantly (P = 0.01). C, Differentially abundant taxa between pretreatment and posttreatment samples analyzed by LEfSe are projected as histogram (left) and cladogram (right). All listed taxa were significantly (Kruskal–Wallis test, P < 0.05; LDA score > 2) enriched for their respective groups (pre-nCRT, green and post-nCRT, red). D, Box plot showing the relative abundances of the differentially abundant taxa in pretreatment (green) and posttreatment (red) groups, identified using LEfSe. *, P < 0.05; **, P < 0.01; ***, P < 0.001. E, Microbial cooccurrence network constructed on the basis of posttreatment samples of the responder group. F, Microbial cooccurrence network constructed on the basis of posttreatment samples of the nonresponder group. Control, healthy individuals; postNR, posttreatment samples from nCRT nonresponders; postR, posttreatment samples from nCRT responders.

Figure 2.

Gut microbiome variations over the course of nCRT associated with therapeutic response. A, Microbial richness and Chao1 diversity were significantly reduced in patients with LARC following nCRT (P = 0.025 and P = 0.028, respectively), and Simpson index showed a trend toward higher evenness following nCRT (P = 0.38). B, PCoA using Euclidean distance at the OTU level separated pretreatment samples from posttreatment samples significantly (P = 0.01). C, Differentially abundant taxa between pretreatment and posttreatment samples analyzed by LEfSe are projected as histogram (left) and cladogram (right). All listed taxa were significantly (Kruskal–Wallis test, P < 0.05; LDA score > 2) enriched for their respective groups (pre-nCRT, green and post-nCRT, red). D, Box plot showing the relative abundances of the differentially abundant taxa in pretreatment (green) and posttreatment (red) groups, identified using LEfSe. *, P < 0.05; **, P < 0.01; ***, P < 0.001. E, Microbial cooccurrence network constructed on the basis of posttreatment samples of the responder group. F, Microbial cooccurrence network constructed on the basis of posttreatment samples of the nonresponder group. Control, healthy individuals; postNR, posttreatment samples from nCRT nonresponders; postR, posttreatment samples from nCRT responders.

Close modal

In taxonomic analyses, LEfSe results (Fig. 2C and D) revealed that Fusobacterium was significantly decreased following nCRT, in addition to Peptostreptococcus, Parvimonas, and Porphyromonas, and two genera in Lactobacillales, that is, Lactobacillus and Streptococcus, were significantly increased following nCRT. In particular, the increase in Streptococcus, belonging to the order Lactobacillales, was exclusively found in the responder subgroup (Supplementary Fig. S3D and S3E). This finding indicated that an increase of the commensal bacteria, Lactobacillales, may correlate with nCRT responses. Of note, even though the gut microbiome in patients with LARC was vastly altered by nCRT, it was still significantly different from that in healthy individuals at the genus level (Supplementary Fig. S3F).

The significance of Streptococcus was also represented in the microbiota network (Fig. 2E and F). In post-nCRT samples, the “pathogenic microbiota” interacted with Streptococcus in the surroundings, regardless of therapeutic responses. Specifically, in post-nCRT samples of the responder group, the “pathogenic microbiota” and the “beneficial microbiota” were connected via Streptococcus, which was not observed in post-nCRT samples of the nonresponder group. Furthermore, our results showed that most interactions between genera were positive or symbiotic; nevertheless, except in healthy individuals, negative or antagonistic interactions were observed in post-nCRT samples of the responder group. Specifically, two bacteria related with butyrate production, that is, Roseburia and Dorea, competed with Ralstonia, which was enriched in the nonresponder group, as we will mention in the next part. Another butyrate-producing bacterium, Anaerostipes, which was enriched in the responder group, as mentioned in the next part, together with Streptococcus, competed with the colorectal cancer–related Bilophila.

In line with the above results that nCRT decreased the abundance of LARC-related bacteria, Bugbase showed that potentially pathogenic bacteria were significantly reduced in patients with LARC after nCRT (Supplementary Fig. S3G). Altogether, our results suggested that nCRT significantly lowered the microbial richness in patients with LARC, the extent of which may correlate with nCRT responses. Furthermore, nCRT reshaped the gut microbiome by decreasing gut pathogens and increasing favorable commensals, the latter of which was exclusively observed in the responder subgroup.

Baseline gut microbiome correlated with therapeutic response to nCRT in patients with LARC

On the basis of the above results, we speculated that the gut microbiome may provide potential biomarkers for predicting response to nCRT in patients with LARC. To test this hypothesis, we conducted cross-sectional comparison analysis between the responder and nonresponder groups. Baseline demographics and clinical characteristics of the patients in the two groups are summarized in Table 1.

Table 1.

Baseline characteristics of the patients with LARC enrolled in this study.

Patients with LARC (N = 84)
Baseline characteristicsNR (n = 39)R (n = 45)P
Age (year) 55.46 ± 9.47 56.64 ± 10.43 0.59 
Gender 
 Female 11 (28.21%) 15 (33.33%) 0.61 
 Male 28 (71.79%) 30 (66.67%)  
BMI (kg/m223.35 ± 2.82 23.27 ± 2.67 0.89 
Tumor location 
 Low 19 (48.72%) 29 (64.44%) 0.15 
 Middle-high 20 (51.28%) 16 (35.56%)  
Tumor size (cm) 5.55 ± 1.69 4.72 ± 1.85 0.03 
Clinical T stage 
 cT2 1 (2.56%) 8 (17.78%) 0.04 
 cT3 29 (74.36%) 32 (71.11%)  
 cT4 9 (23.08%) 5 (11.11%)  
Clinical N stage 
 cN+ 37 (94.87%) 41 (91.11%) 0.68 
 cN0 2 (5.13%) 4 (8.89%)  
Clinical staging before nCRT 
 I 0 (0) 4 (8.89%) 0.12 
 IIA 1 (2.56%) 0 (0)  
 IIIA 1 (2.56%) 4 (8.89%)  
 IIIB 26 (66.67%) 29 (64.44%)  
 IIIC 11 (28.21%) 8 (17.78%)  
Mesorectal fascia invasion 16 (41.03%) 14 (31.11%) 0.34 
Extramural vascular invasion 13 (33.33%) 15 (33.33%) 
Chemotherapy before nCRT 
 Yes 5 (12.82%) 2 (4.44%) 0.24 
 No 34 (87.18%) 43 (95.56%)  
Concurrent chemotherapy regimen 
 Capecitabine 5 (12.82%) 5 (11.11%) 
 Capecitabine + irinotecan 33 (84.62%) 39 (86.67%)  
 Capecitabine + oxaliplatin 1 (2.56%) 1 (2.22%)  
Cycles of irinotecan infusion 
 1–2 11 (33.33%) 5 (12.82%) 0.06 
 3 5 (15.15%) 13 (33.33%)  
 4–5 17 (51.52%) 21 (53.85%)  
Total dose of irinotecan (mg) 430.12 ± 180.06 464.62 ± 160.54 0.40 
Dose of capecitabine (tablets per day) 2.99 ± 1.12 3.36 ± 1.12 0.18 
Chemotherapy following nCRT 
 Yes 34 (87.18%) 40 (88.89%) 
 No 5 (12.82%) 5 (11.11%)  
Days of radiotherapy 36.00 ± 3.90 36.38 ± 5.85 0.39 
Dose of radiotherapy 
 45 Gy/25 Fx 0 (0) 2 (4.44%) 0.5 
 50 Gy/25 Fx 39 (100.00%) 43 (95.56%)  
Weight loss during nCRT (kg) −1.08 ± 2.38 −0.80 ± 2.54 0.72 
Use of antibiotics during nCRT 
 Yes 13 (33.33%) 8 (17.78%) 0.1 
 No 26 (66.67%) 37 (82.22%)  
Patients with LARC (N = 84)
Baseline characteristicsNR (n = 39)R (n = 45)P
Age (year) 55.46 ± 9.47 56.64 ± 10.43 0.59 
Gender 
 Female 11 (28.21%) 15 (33.33%) 0.61 
 Male 28 (71.79%) 30 (66.67%)  
BMI (kg/m223.35 ± 2.82 23.27 ± 2.67 0.89 
Tumor location 
 Low 19 (48.72%) 29 (64.44%) 0.15 
 Middle-high 20 (51.28%) 16 (35.56%)  
Tumor size (cm) 5.55 ± 1.69 4.72 ± 1.85 0.03 
Clinical T stage 
 cT2 1 (2.56%) 8 (17.78%) 0.04 
 cT3 29 (74.36%) 32 (71.11%)  
 cT4 9 (23.08%) 5 (11.11%)  
Clinical N stage 
 cN+ 37 (94.87%) 41 (91.11%) 0.68 
 cN0 2 (5.13%) 4 (8.89%)  
Clinical staging before nCRT 
 I 0 (0) 4 (8.89%) 0.12 
 IIA 1 (2.56%) 0 (0)  
 IIIA 1 (2.56%) 4 (8.89%)  
 IIIB 26 (66.67%) 29 (64.44%)  
 IIIC 11 (28.21%) 8 (17.78%)  
Mesorectal fascia invasion 16 (41.03%) 14 (31.11%) 0.34 
Extramural vascular invasion 13 (33.33%) 15 (33.33%) 
Chemotherapy before nCRT 
 Yes 5 (12.82%) 2 (4.44%) 0.24 
 No 34 (87.18%) 43 (95.56%)  
Concurrent chemotherapy regimen 
 Capecitabine 5 (12.82%) 5 (11.11%) 
 Capecitabine + irinotecan 33 (84.62%) 39 (86.67%)  
 Capecitabine + oxaliplatin 1 (2.56%) 1 (2.22%)  
Cycles of irinotecan infusion 
 1–2 11 (33.33%) 5 (12.82%) 0.06 
 3 5 (15.15%) 13 (33.33%)  
 4–5 17 (51.52%) 21 (53.85%)  
Total dose of irinotecan (mg) 430.12 ± 180.06 464.62 ± 160.54 0.40 
Dose of capecitabine (tablets per day) 2.99 ± 1.12 3.36 ± 1.12 0.18 
Chemotherapy following nCRT 
 Yes 34 (87.18%) 40 (88.89%) 
 No 5 (12.82%) 5 (11.11%)  
Days of radiotherapy 36.00 ± 3.90 36.38 ± 5.85 0.39 
Dose of radiotherapy 
 45 Gy/25 Fx 0 (0) 2 (4.44%) 0.5 
 50 Gy/25 Fx 39 (100.00%) 43 (95.56%)  
Weight loss during nCRT (kg) −1.08 ± 2.38 −0.80 ± 2.54 0.72 
Use of antibiotics during nCRT 
 Yes 13 (33.33%) 8 (17.78%) 0.1 
 No 26 (66.67%) 37 (82.22%)  

Note: Continuous data are expressed as mean ± SD and categorical data are expressed as counts (percentage). Clinical staging before nCRT was performed according to the AJCC (eighth edition). P values < 0.05 are in bold.

Abbreviations: Fx, Fractions; NR, nonresponders; R, responders.

In our research, no significant differences in alpha diversity were noted between the responder and nonresponder groups, regardless of the sampling timepoint (Supplementary Fig. S4A and S4B). In PCoA, however, a notable clustering effect by response status was revealed. The results were significant in both pre-nCRT and post-nCRT samples based on the Euclidean distance (Fig. 3A and B) and in post-nCRT samples based on the Jaccard distance or Bray Curtis distance (Supplementary Fig. S4C and S4D). Considering that patients in the responder group had a significantly lower tumor stage (P = 0.04) and smaller tumor size (P = 0.03), compared with patients in the nonresponder group (as summarized in Table 1), we adjusted for the two variables using MiRAKT. MiRAKT is a kernel regression approach for testing the association between microbiome composition and outcome of interest, featured by fast computation, good control of type I error rate, and easy covariate adjustment (30). Using MiRAKT under individual beta distances of weighted UniFrac distance, unweighted UniFrac distance, Bray Curtis distance, and Euclidean distance, P values were found to be 0.22, 0.01, 0.08, and 0.03, respectively. The optimal MiRAKT, which allowed for simultaneous consideration of several beta distances so as to avoid cherry picking the best result, generated a P value of 0.02. Hence, the association between microbiome profiles and nCRT therapeutic response remained statistically significant after the potential confounders were controlled for.

Figure 3.

Baseline gut microbiome correlated with therapeutic response to nCRT in patients with LARC. A, PCoA using Euclidean distance at the OTU level showed a significant clustering effect by response status in baseline (pre-nCRT) samples (n = 84; P = 0.03). B, PCoA using Euclidean distance at the OTU level showed a significant clustering effect by response status in post-nCRT samples (n = 83; P = 0.01). C, K-means clustering classified the baseline samples into two clusters, and a significantly higher response rate was seen with LARC type2 (19/26 = 73.1%) compared with LARC type1 (26/58 = 44.8%; P = 0.03). D, Differentially abundant taxa between the responder (green) and nonresponder (red) groups analyzed by LEfSe are projected as cladogram (middle) and histogram (right). All listed taxa were significantly (Kruskal–Wallis test, P < 0.05; LDA score > 2) enriched for their respective groups. Box plot (left) showing the relative abundances of the differentially abundant taxa in their respective groups. *, P < 0.05; **, P < 0.01. E, Microbial cooccurrence network constructed on the basis of pre-nCRT fecal samples of the responder group. F, Microbial cooccurrence network constructed on the basis of pre-nCRT fecal samples of the nonresponder group. preNR, baseline samples from nCRT nonresponders; preR, baseline samples from nCRT responders.

Figure 3.

Baseline gut microbiome correlated with therapeutic response to nCRT in patients with LARC. A, PCoA using Euclidean distance at the OTU level showed a significant clustering effect by response status in baseline (pre-nCRT) samples (n = 84; P = 0.03). B, PCoA using Euclidean distance at the OTU level showed a significant clustering effect by response status in post-nCRT samples (n = 83; P = 0.01). C, K-means clustering classified the baseline samples into two clusters, and a significantly higher response rate was seen with LARC type2 (19/26 = 73.1%) compared with LARC type1 (26/58 = 44.8%; P = 0.03). D, Differentially abundant taxa between the responder (green) and nonresponder (red) groups analyzed by LEfSe are projected as cladogram (middle) and histogram (right). All listed taxa were significantly (Kruskal–Wallis test, P < 0.05; LDA score > 2) enriched for their respective groups. Box plot (left) showing the relative abundances of the differentially abundant taxa in their respective groups. *, P < 0.05; **, P < 0.01. E, Microbial cooccurrence network constructed on the basis of pre-nCRT fecal samples of the responder group. F, Microbial cooccurrence network constructed on the basis of pre-nCRT fecal samples of the nonresponder group. preNR, baseline samples from nCRT nonresponders; preR, baseline samples from nCRT responders.

Close modal

Of interest, a significantly higher response rate was seen with the LARC type2, which was found to resemble healthy individuals in the abovementioned analysis, compared with LARC type1, and the finding was consistent in every method of clustering (Figs. 1A and 3C; Supplementary Fig. S2D).

Considering that complicated treatment factors may bias the results for the prediction of response to nCRT, we then focused on the baseline specimens to explore the discriminatory taxa as biomarkers. As shown by LEfSe, the genus Clostridium XVIII and species R. inulinivorans, Dorea formicigenerans (D. formicigenerans), and Anaerostipes hadrus were significantly enriched in the responder group, and that the family Coriobacteriaceae, genera Fusobacterium and Granulicatella, as well as species Ralstonia pickettii (R. pickettii) and Eisenbergiella tayi (E. tayi) were significantly enriched in the nonresponder group (Fig. 3D).

Differences in terms of response to nCRT were also reflected in microbiota networks within the pre-nCRT samples (Fig. 3E and F). The “beneficial microbiota” was more tightly clustered in the responder group than that in the nonresponder group, whereas the “pathogenic microbiota” was more tightly clustered in the nonresponder group than that in the responder group. Notably, interactions with Bacteroides and Alistipes, two bacteria reported to be related to colorectal cancer (32, 37), were noticed in the surroundings of the “beneficial microbiota” only in the nonresponder group. In the surroundings of the “pathogenic microbiota,” interactions with favorable microbes, including Akkermansia, Bifidobacterium, and various butyrate-producing bacteria, were shown in the responder group, whereas interactions with Collinsella, which is a genus belonging to the nonresponder-related Coriobacteriaceae, were noted in the nonresponder group. Combining taxonomic comparison analysis results and microbiota network features, we inferred that Roseburia, Dorea, and Anaerostipes were associated with nCRT response, whereas Coriobacteriaceae, Fusobacterium, Granulicatella, R. pickettii, and E. tayi may correlate with nCRT resistance.

Establishment of a predictive classifier for response to nCRT based on the gut microbiome profiles

To translate our findings into a clinical application, the random forest algorithm was utilized to construct a response-prediction classifier in patients with LARC. We randomly divided the 84 patients with pre-nCRT samples into two cohorts: the training cohort, consisting of 19 responders and 18 nonresponders, was analyzed for the construction of the classifier, and the validation cohort consisted of the remaining 26 responders and 21 nonresponders. In constructing classifier I, which incorporated microbial variables only, we selected 40 potential discriminatory genera to form the set of input features, on the basis of the abovementioned taxonomic comparison analysis. Following five repeats of 10-fold cross-validation, 10 microbial variables were selected as the optimal set to construct the classifier. The AUC was 93.57% [95% confidence interval (CI), 85.76%–100%] in the training cohort and 73.53% (95% CI, 58.96%–88.11%) in the validation cohort (Fig. 4). With the optimal cut-off value set at 0.5 by the ROC curve, the specificity and PPV in both cohorts reached above 80% (Table 2), indicating that our microbiome-based classifier performed robustly in predicting response to nCRT. Among the 10 genera variables, Dorea, Anaerostipes, and Clostridium XVIII were biomarkers of nCRT responders, and Eisenbergiella, Granulicatella, and Ralstonia were biomarkers of nCRT nonresponders; Ruminococcus2 and Phascolarctobacterium were found to be significantly enriched in patients with LARC and Akkermansia was enriched in healthy individuals; and Streptococcus was significantly elevated following nCRT.

Figure 4.

Establishment of a predictive classifier for response to nCRT based on the gut microbiome profiles. A, The ROC curve for the training cohort based on the random forest classifier constructed by microbial variables only. The AUC was 93.57% (95% CI, 85.76%–100%). B, Box plot for the POR in the training cohort (P < 0.001) according to the random forest classifier constructed by microbial variables only. C, The ROC curve for the validation cohort based on the random forest classifier constructed by microbial variables only; AUC was 73.53% (95% CI, 58.96%–88.11%). D, Box plot of POR in the validation cohort (P = 0.003) according to the random forest classifier constructed by microbial variables only.

Figure 4.

Establishment of a predictive classifier for response to nCRT based on the gut microbiome profiles. A, The ROC curve for the training cohort based on the random forest classifier constructed by microbial variables only. The AUC was 93.57% (95% CI, 85.76%–100%). B, Box plot for the POR in the training cohort (P < 0.001) according to the random forest classifier constructed by microbial variables only. C, The ROC curve for the validation cohort based on the random forest classifier constructed by microbial variables only; AUC was 73.53% (95% CI, 58.96%–88.11%). D, Box plot of POR in the validation cohort (P = 0.003) according to the random forest classifier constructed by microbial variables only.

Close modal
Table 2.

Performance of the random forest response-prediction classifiers for patients with LARC.

Value in classifier Ia (95% CI)Value in classifier IIb (95% CI)
Training cohortValidation cohortTraining cohortValidation cohort
Variables (n = 37) (n = 47) (n = 37) (n = 47) 
AUC, % 93.57 (85.76–100) 73.53 (58.96–88.11) 86.26 (74.03–98.49) 73.26 (58.66–87.86) 
Cutoff 0.5 0.5 0.5 0.5 
ACC, % 89.19 (88.68–89.70) 72.34 (71.51–73.17) 81.08 (80.27–81.89) 70.21 (69.34–71.08) 
SE, % 89.47 (75.67–100) 65.38 (47.10–83.67) 78.94 (60.62–97.28) 61.54 (42.84–80.24) 
SP, % 88.89 (74.37–100) 80.95 (64.16–97.75) 83.33 (66.12–100) 80.95 (64.16–97.75) 
PPV, % 89.47 (75.67–100) 80.95 (64.16–97.75) 83.33 (66.12–100) 80.00 (62.47–97.53) 
NPV, % 88.89 (74.37–100) 65.38 (47.10–83.67) 78.94 (60.62–97.28) 62.96 (44.75–81.18) 
Pc 0.08 0.95 0.08 0.95 
Value in classifier Ia (95% CI)Value in classifier IIb (95% CI)
Training cohortValidation cohortTraining cohortValidation cohort
Variables (n = 37) (n = 47) (n = 37) (n = 47) 
AUC, % 93.57 (85.76–100) 73.53 (58.96–88.11) 86.26 (74.03–98.49) 73.26 (58.66–87.86) 
Cutoff 0.5 0.5 0.5 0.5 
ACC, % 89.19 (88.68–89.70) 72.34 (71.51–73.17) 81.08 (80.27–81.89) 70.21 (69.34–71.08) 
SE, % 89.47 (75.67–100) 65.38 (47.10–83.67) 78.94 (60.62–97.28) 61.54 (42.84–80.24) 
SP, % 88.89 (74.37–100) 80.95 (64.16–97.75) 83.33 (66.12–100) 80.95 (64.16–97.75) 
PPV, % 89.47 (75.67–100) 80.95 (64.16–97.75) 83.33 (66.12–100) 80.00 (62.47–97.53) 
NPV, % 88.89 (74.37–100) 65.38 (47.10–83.67) 78.94 (60.62–97.28) 62.96 (44.75–81.18) 
Pc 0.08 0.95 0.08 0.95 

Abbreviations: ACC, accuracy; SE, sensitivity; SP, specificity.

aClassifier I was constructed on the basis of microbial variables only.

bClassifier II was constructed on the basis of both microbial and clinical variables.

cP value of Delong test for ROC curves in each cohort using the two classifiers.

In addition to the microbial variables, 26 clinical variables were also incorporated for variable selection in constructing classifier II, including patient age, gender, BMI, and parameters concerning tumor biology and treatment details. Following five repeats of 10-fold cross-validation, 11 variables were selected from the 66 variables, among which the dose of capecitabine was the only clinical variable. No significant differences were noticed in the performance of the two classifiers (Table 2; Supplementary Fig. S5).

Construction of a classifier for surveillance and response prediction based on the gut microbiome merits investigation because fecal sampling is noninvasive and repeatable. To the best of our knowledge, random forest clustering analysis is one of the most commonly used classifiers in gut microbiome analysis (41) and has been established for the early diagnosis of several cancer types, including colorectal cancer (31, 32). This study is the first to report the construction of a microbiome-based random forest classifier for predicting response to nCRT. With high AUC, specificity, and PPV values, our classifier performed robustly in screening out responders, which may facilitate personalized treatment decisions.

Our results are consistent with previous reports of how the gut microbiome engages in colorectal cancer tumorigenesis. Prevotella, Porphyromonas, Fusobacterium, Parvimonas, Peptostreptococcus, and B. fragilis have been widely reported to be colorectal cancer–related pathogenic bacteria (8, 37, 38, 42). These microbes were all found to be significantly abundant in our patients with LARC versus healthy individuals. Contrary to the typical tumorigenic flora in patients with LARC, the enriched microbes in healthy individuals, including A. muciniphila, R. inulinivorans, Bifidobacterium kashlwanohense, and Streptococcus salivarius subspecies, are known as favorable commensals associated with host immune homeostasis, mucus production, and barrier function maintenance (35).

Variations in the human or murine gut microbiome after irradiation exposure have been reported previously (16, 17, 20), but have not been discussed in the setting of nCRT, let alone their potential association with response to nCRT. To some extent, our results of microbial alterations during nCRT agreed with reports utilizing radiation only, such as the decrease in bacterial richness (17, 20) and the increase in Lactobacillus following irradiation (16, 17). Lactobacillus produces lactate and inhibits gut pathogens, thus the increase in Lactobacillus indicated that nCRT, especially the radiotherapy part, may have imposed a positive impact on the gut microbiome. More importantly, we demonstrated for the first time that the aforementioned LARC-related microbes, Fusobacterium, Peptostreptococcus, Parvimonas, and Porphyromonas, were all significantly decreased following nCRT, reinforcing the benefits from nCRT on the gut microbiome in patients with LARC. Of interest, Streptococcus parasanguinis was reported to be enriched in responders to immunotherapy (12). On the basis of our results, the genus Streptococcus was found to be significantly increased exclusively in responder subgroup following nCRT and presented a vital role in post-nCRT networks. What these observations on Streptococcus mean remains to be elucidated. In this regard, we may have provided additional knowledge on how the gut microbiome interacts with radiotherapy responses.

It is noteworthy to mention that potential microbial biomarkers for predicting response to nCRT were discovered in our baseline specimens. Dorea and Anaerostipes were overrepresented in responders, and selected as discriminatory variables in our response-prediction random forest classifier, whereas Coriobacteriaceae and Fusobacterium were overrepresented in nonresponders. In the literature, the role of Fusobacterium in chemoresistance has been widely discussed (6, 43). Furthermore, Anaerostipes was reported to be a butyrate producer, and Dorea was reported to produce acetate and lactate, which may serve as substrates for butyrate production (44). Of interest, butyrate-producing bacteria were reported to be significantly enriched in responders to immunotherapy (9, 12, 13), and the genus Collinsella belonging to the family Coriobacteriaceae was also reported to be significantly overrepresented in nonresponders to immunotherapy (12). Although in line with some of the findings from immunotherapy and chemotherapy research, enrichment of the microbes in responders or nonresponders in our study does not indicate a direct association. Fortunately, published experimental studies have provided some mechanistic insights into this issue.

Fusobacterium has been proposed to predict recurrence and poor prognosis in colorectal cancer (45). Evidence suggests that Fusobacterium nucleatum (F. nucleatum) tracks with tumor stage, correlates with a higher degree of microsatellite instability (45) and recruitment of myeloid-derived suppressor cells into the tumor microenvironment (46), and leads to chemoresistance of 5-fluorouracil and oxaliplatin (14, 47). Potential mechanisms include: (i) F. nucleatum activates the β-catenin pathway by expressing FadAc, which binds E-cadherin, thereby regulating tumor cell growth (48); (ii) F. nucleatum inhibits apoptosis in favor of autophagy by downregulating miRNA18a via the TLR-4/MYD88 pathway (14); (iii) F. nucleatum upregulates expression of the antiapoptotic protein, BIRC3, via the TLR4/NK-κB pathway (49); and (iv) F. nucleatum suppresses cytotoxicity of natural killer cells via Fap2-TIGIT (50). We conjecture that the mechanism of Fusobacterium in association with resistance to nCRT may involve both direct regulation on tumor cell cycle, apoptosis, or autophagy and immunomodulation on tumor microenvironment, and further experimental research is warranted.

As mentioned previously, the two microbes enriched in responders are both related with butyrate production. Butyrate is a type of short-chain fatty acid (SCFA) generated from dietary fiber by several gut microbes in the phylum Firmicutes (8, 51). Butyrate promotes regulatory T-cell accumulation and triggers anti-inflammatory responses (9) by binding G-protein–coupled receptor 109a expressed on myeloid cells and colonocytes. Both positive and negative influences on colorectal cancer tumorigenesis by butyrate-producing microbes have been reported, and the influence of butyrate is believed to correlate with butyrate concentration and host genotype (34, 51). Likewise, Faecalibacterium prausnitzii (9, 12, 13) and D. formicigenerans (12) were reported to be associated with immunotherapy response, whereas R. intestinalis (52) was associated with immunotherapy resistance. What type of immunomodulation of the butyrate-producing microbes occurred during nCRT and how it would affect therapeutic response remain unknown. An interesting phenomenon is that vancomycin, an antibiotic acting mainly on Gram-positive bacteria, including butyrate-producing bacteria, has a synergistic effect with radiotherapy (53). Furthermore, butyrate produced by vancomycin-depleted gut microbes abrogated this synergistic effect by impacting dendritic cell functions (54). These contradictory findings indicate that further investigation on the role of butyrate and butyrate-producing microbes in nCRT is definitely needed.

Evidence-based research on the role of the gut microbiome in chemoradiotherapy responses is scarce. A recent study with 45 baseline fecal samples from patients with LARC undergoing nCRT reported that Bacteroides was enriched in nonresponders, and Duodenibacillus massiliensis was associated with nCRT responses (21). However, these previous results may be biased by hidden confounders considering that their enrollment lasted for 6 years and the sample size was relatively small. Moreover, the cross-sectional sampling design of the previous study is believed to be less statistically powerful than a repeated sampling design, as we did in this longitudinal study. The limitations of our research lie in two aspects. The first is that the study population was from our institution only. To improve the performance and reliability of the predictive classifier, more patients from our institution and other institutions are needed for internal and external validation. Second, we used genus-level variables in constructing the classifier to assure steadiness of the results. With the help of meta-genomics analysis, species-level biomarkers and more information on bacterial functions may be provided.

In conclusion, we have performed systematic and comprehensive research on the gut microbiome from the perspective of predicting response to nCRT with a considerable sample size. The taxonomic profile in patients with LARC in this study reflected the typical tumorigenic flora of colorectal cancer, as reported before, and microbial alterations over the course of nCRT were found to be associated with therapeutic response. We successfully identified several microbial biomarkers of nCRT responses, and some mechanistic explanations for these microbial biomarkers were provided in the literature, especially Fusobacterium and butyrate-producing bacteria. With robust discrimination and predictive power by the response-prediction random forest classifier, our results highlight the possibility of bridging microbiome research with therapeutic management.

No disclosures were reported.

Y. Yi: Software, formal analysis, validation, visualization, methodology, writing-original draft. L. Shen: Resources, data curation, formal analysis, funding acquisition, investigation, writing-original draft. W. Shi: Resources, data curation, formal analysis, investigation, writing-original draft. F. Xia: Formal analysis, supervision, project administration. H. Zhang: Data curation, formal analysis, investigation, project administration. Y. Wang: Data curation, formal analysis, investigation, project administration. J. Zhang: Investigation, visualization, methodology. Y. Wang: Investigation, visualization, methodology. X. Sun: Investigation, visualization, methodology. Z. Zhang: Investigation, visualization, methodology. W. Zou: Investigation, visualization, methodology. W. Yang: Investigation, visualization, methodology. L. Zhang: Investigation, visualization, methodology. J. Zhu: Supervision, methodology, project administration. A. Goel: Conceptualization, supervision, writing-review and editing. Y. Ma: Conceptualization, supervision, funding acquisition, writing-review and editing. Z. Zhang: Conceptualization, supervision, writing-review and editing.

This study was supported by grants from the National Natural Science Foundation of China (grant nos., 81773357, 81920108026, and 81871964), Shanghai ACA Foundation (SACACY19B04), Fudan University Shanghai Cancer Center Foundation (YJQN201921), Shanghai Young Top Talents (grant no., QNBJ1701 to Y. Ma), Shanghai Science and Technology Development Foundation (grant no., 19410713300), and the CSCO-Roche Tumor Research Foundation (grant no., Y-2019Roche-079). We thank all the subjects who volunteered to participate in this study, and all the staff who helped to collect fecal samples from the subjects. We thank Dr. Jia'ning Yang for her generous support in sample collection, and we also thank Drs. Xiaodong Yi and Kunyu Qiu for their kind help in bioinformatic analysis.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.

1.
Kuipers
EJ
,
Grady
WM
,
Lieberman
D
,
Seufferlein
T
,
Sung
JJ
,
Boelens
PG
, et al
Colorectal cancer
.
Nat Rev Dis Primers
2015
;
1
:
15065
.
2.
Siegel
RL
,
Miller
KD
,
Jemal
A
. 
Cancer statistics, 2020
.
CA Cancer J Clin
2020
;
70
:
7
30
.
3.
Benson
AB
,
Venook
AP
,
Al-Hawary
MM
,
Cederquist
L
,
Chen
YJ
,
Ciombor
KK
, et al
Rectal cancer, version 2.2018, NCCN clinical practice guidelines in oncology
.
J Natl Compr Canc Netw
2018
;
16
:
874
901
.
4.
Dossa
F
,
Chesney
TR
,
Acuna
SA
,
Baxter
NN
. 
A watch-and-wait approach for locally advanced rectal cancer after a clinical complete response following neoadjuvant chemoradiation: a systematic review and meta-analysis
.
Lancet Gastroenterol Hepatol
2017
;
2
:
501
13
.
5.
van den Ende
T
,
van den Boorn
HG
,
Hoonhout
NM
,
van Etten-Jamaludin
FS
,
Meijer
SL
,
Derks
S
, et al
Priming the tumor immune microenvironment with chemo(radio)therapy: a systematic review across tumor types
.
Biochim Biophys Acta Rev Cancer
2020
;
1874
:
188386
.
6.
Ramos
A
,
Hemann
MT
. 
Drugs, bugs, and cancer: fusobacterium nucleatum promotes chemoresistance in colorectal cancer
.
Cell
2017
;
170
:
411
3
.
7.
Hekmatshoar
Y
,
Rahbar Saadat
Y
,
Hosseiniyan Khatibi
SM
,
Ozkan
T
,
Zununi Vahed
F
,
Nariman-Saleh-Fam
Z
, et al
The impact of tumor and gut microbiotas on cancer therapy: beneficial or detrimental?
Life Sci
2019
;
233
:
116680
.
8.
Garrett
WS
. 
Cancer and the microbiota
.
Science
2015
;
348
:
80
6
.
9.
Chaput
N
,
Lepage
P
,
Coutzac
C
,
Soularue
E
,
Roux
KL
,
Monot
C
, et al
Baseline gut microbiota predicts clinical response and colitis in metastatic melanoma patients treated with ipilimumab
.
Ann Oncol
2017
;
28
:
1368
79
.
10.
Routy
B
,
Le Chatelier
E
,
Derosa
L
,
Duong
CPM
,
Alou
MT
,
Daillère
R
, et al
Gut microbiome influences efficacy of PD-1-based immunotherapy against epithelial tumors
.
Science
2018
;
359
:
91
7
.
11.
Vetizou
M
,
Pitt
JM
,
Daillere
R
,
Lepage
P
,
Waldschmitt
N
,
Flament
C
, et al
Anticancer immunotherapy by CTLA-4 blockade relies on the gut microbiota
.
Science
2015
;
350
:
1079
84
.
12.
Frankel
AE
,
Coughlin
LA
,
Kim
J
,
Froehlich
TW
,
Xie
Y
,
Frenkel
EP
, et al
Metagenomic shotgun sequencing and unbiased metabolomic profiling identify specific human gut microbiota and metabolites associated with immune checkpoint therapy efficacy in melanoma patients
.
Neoplasia
2017
;
19
:
848
55
.
13.
Gopalakrishnan
V
,
Spencer
CN
,
Nezi
L
,
Reuben
A
,
Andrews
MC
,
Karpinets
TV
, et al
Gut microbiome modulates response to anti-PD-1 immunotherapy in melanoma patients
.
Science
2018
;
359
:
97
103
.
14.
Yu
TC
,
Guo
F
,
Yu
Y
,
Sun
T
,
Ma
D
,
Han
J
, et al
Fusobacterium nucleatum promotes chemoresistance to colorectal cancer by modulating autophagy
.
Cell
2017
;
170
:
548
63
.
15.
Sivan
A
,
Corrales
L
,
Hubert
N
,
Williams
JB
,
Aquino-Michaels
K
,
Earley
ZM
, et al
Commensal Bifidobacterium promotes antitumor immunity and facilitates anti-PD-L1 efficacy
.
Science
2015
;
350
:
1084
9
.
16.
Lam
V
,
Moulder
JE
,
Salzman
NH
,
Dubinsky
EA
,
Andersen
GL
,
Baker
JE
. 
Intestinal microbiota as novel biomarkers of prior radiation exposure
.
Radiat Res
2012
;
177
:
573
83
.
17.
Nam
YD
,
Kim
HJ
,
Seo
JG
,
Kang
SW
,
Bae
JW
. 
Impact of pelvic radiotherapy on gut microbiota of gynecological cancer patients revealed by massive pyrosequencing
.
PLoS One
2013
;
8
:
e82659
.
18.
Gerassy-Vainberg
S
,
Blatt
A
,
Danin-Poleg
Y
,
Gershovich
K
,
Sabo
E
,
Nevelsky
A
, et al
Radiation induces proinflammatory dysbiosis: transmission of inflammatory susceptibility by host cytokine induction
.
Gut
2018
;
67
:
97
107
.
19.
Wang
A
,
Ling
Z
,
Yang
Z
,
Kiela
PR
,
Wang
T
,
Wang
C
, et al
Gut microbial dysbiosis may predict diarrhea and fatigue in patients undergoing pelvic cancer radiotherapy: a pilot study
.
PLoS One
2015
;
10
:
e0126312
.
20.
Reis Ferreira
M
,
Andreyev
HJN
,
Mohammed
K
,
Truelove
L
,
Gowan
SM
,
Li
J
, et al
Microbiota- and Radiotherapy-induced Gastrointestinal Side-effects (MARS) study: a large pilot study of the microbiome in acute and late-radiation enteropathy
.
Clin Cancer Res
2019
;
25
:
6487
500
.
21.
Jang
BS
,
Chang
JH
,
Chie
EK
,
Kim
K
,
Park
JW
,
Kim
MJ
, et al
Gut microbiome composition is associated with a pathologic response after preoperative chemoradiation in rectal cancer patients
.
Int J Radiat Oncol Biol Phys
2020
;
107
:
736
46
.
22.
Zhang
Z
,
Sun
X
,
Liu
A
,
Zhu
Y
,
Zhu
Y
,
Zhang
T
, et al
A multicenter randomized phase III trial of capecitabine with or without irinotecan driven by UGT1A1 in neoadjuvant chemoradiation of locally advanced rectal cancer (CinClare)
.
J Clin Oncol
2019
;
37
:
3510
.
23.
Magoc
T
,
Salzberg
SL
. 
FLASH: fast length adjustment of short reads to improve genome assemblies
.
Bioinformatics
2011
;
27
:
2957
63
.
24.
Rognes
T
,
Flouri
T
,
Nichols
B
,
Quince
C
,
Mahe
F
. 
VSEARCH: a versatile open source tool for metagenomics
.
PeerJ
2016
;
4
:
e2584
.
25.
Edgar
RC
. 
Search and clustering orders of magnitude faster than BLAST
.
Bioinformatics
2010
;
26
:
2460
1
.
26.
Lozupone
C
,
Knight
R
. 
UniFrac: a new phylogenetic method for comparing microbial communities
.
Appl Environ Microbiol
2005
;
71
:
8228
35
.
27.
Parks
DH
,
Tyson
GW
,
Hugenholtz
P
,
Beiko
RG
. 
STAMP: statistical analysis of taxonomic and functional profiles
.
Bioinformatics
2014
;
30
:
3123
4
.
28.
Segata
N
,
Izard
J
,
Waldron
L
,
Gevers
D
,
Miropolsky
L
,
Garrett
WS
, et al
Metagenomic biomarker discovery and explanation
.
Genome Biol
2011
;
12
:
R60
.
29.
Chen
X
,
Li
P
,
Liu
M
,
Zheng
H
,
He
Y
,
Chen
MX
, et al
Gut dysbiosis induces the development of pre-eclampsia through bacterial translocation
.
Gut
2020
;
69
:
513
22
.
30.
Zhao
N
,
Chen
J
,
Carroll
IM
,
Ringel-Kulka
T
,
Epstein
MP
,
Zhou
H
, et al
Testing in microbiome-profiling studies with MiRKAT, the microbiome regression-based Kernel Association Test
.
Am J Hum Genet
2015
;
96
:
797
807
.
31.
Yu
J
,
Feng
Q
,
Wong
SH
,
Zhang
D
,
Liang
QY
,
Qin
Y
, et al
Metagenomic analysis of faecal microbiome as a tool towards targeted non-invasive biomarkers for colorectal cancer
.
Gut
2017
;
66
:
70
8
.
32.
Feng
Q
,
Liang
S
,
Jia
H
,
Stadlmayr
A
,
Tang
L
,
Lan
Z
, et al
Gut microbiome development along the colorectal adenoma–carcinoma sequence
.
Nat Commun
2015
;
6
:
6528
.
33.
Sears
CL
,
Garrett
WS
. 
Microbes, microbiota, and colon cancer
.
Cell Host Microbe
2014
;
15
:
317
28
.
34.
Donohoe
DR
,
Holley
D
,
Collins
LB
,
Montgomery
SA
,
Whitmore
AC
,
Hillhouse
A
, et al
A gnotobiotic mouse model demonstrates that dietary fiber protects against colorectal tumorigenesis in a microbiota- and butyrate-dependent manner
.
Cancer Discov
2014
;
4
:
1387
97
.
35.
Ren
Z
,
Li
A
,
Jiang
J
,
Zhou
L
,
Yu
Z
,
Lu
H
, et al
Gut microbiome analysis as a tool towards targeted non-invasive biomarkers for early hepatocellular carcinoma
.
Gut
2019
;
68
:
1014
23
.
36.
Deng
X
,
Li
Z
,
Li
G
,
Li
B
,
Jin
X
,
Guoqing
L
. 
Comparison of microbiota in patients treated by surgery or chemotherapy by 16S rRNA sequencing reveals potential biomarkers for colorectal cancer therapy
.
Front Microbiol
2018
;
9
:
1607
.
37.
Almeida
A
,
Mitchell
AL
,
Boland
M
,
Forster
SC
,
Gloor
GB
,
Tarkowska
A
, et al
A new genomic blueprint of the human gut microbiota
.
Nature
2019
;
568
:
499
504
.
38.
Ternes
D
,
Karta
J
,
Tsenkova
M
,
Wilmes
P
,
Haan
S
,
Letellier
E
. 
Microbiome in colorectal cancer: how to get from meta-omics to mechanism?
Trends Microbiol
2020
;
28
:
401
23
.
39.
Devkota
S
,
Wang
Y
,
Musch
MW
,
Leone
V
,
Fehlner-Peach
H
,
Nadimpalli
A
, et al
Dietary-fat-induced taurocholic acid promotes pathobiont expansion and colitis in Il10-/- mice
.
Nature
2012
;
487
:
104
8
.
40.
David
LA
,
Maurice
CF
,
Carmody
RN
,
Gootenberg
DB
,
Button
JE
,
Wolfe
BE
, et al
Diet rapidly and reproducibly alters the human gut microbiome
.
Nature
2014
;
505
:
559
63
.
41.
Zhang
J
,
Zhang
N
,
Liu
YX
,
Zhang
X
,
Hu
B
,
Qin
Y
, et al
Root microbiota shift in rice correlates with resident time in the field and developmental stage
.
Sci China Life Sci
2018
;
61
:
613
21
.
42.
Wargo
JA
. 
The microbiome, cancer, and cancer therapy
.
Nat Med
2019
;
25
:
377
88
.
43.
Yamamura
K
,
Izumi
D
,
Kandimalla
R
,
Sonohara
F
,
Baba
Y
,
Yoshida
N
, et al
Intratumoral fusobacterium nucleatum levels predict therapeutic response to neoadjuvant chemotherapy in esophageal squamous cell carcinoma
.
Clin Cancer Res
2019
;
25
:
6170
9
.
44.
Louis
P
,
Flint
HJ
. 
Formation of propionate and butyrate by the human colonic microbiota
.
Environ Microbiol
2017
;
19
:
29
41
.
45.
Zhang
S
,
Wang
Q
,
Zhou
C
,
Chen
K
,
Chang
H
,
Xiao
W
, et al
Colorectal cancer, radiotherapy and gut microbiota
.
Chin J Cancer Res
2019
;
31
:
212
22
.
46.
Kostic
AD
,
Chun
E
,
Robertson
L
,
Glickman
JN
,
Gallini
CA
,
Michaud
M
, et al
Fusobacterium nucleatum potentiates intestinal tumorigenesis and modulates the tumor-immune microenvironment
.
Cell Host Microbe
2013
;
14
:
207
15
.
47.
Mandip
KC
,
Steer
CJ
. 
Novel mechanisms of chemoresistance by Fusobacterium nucleatum involve not so novel pathways of microRNAs and autophagy
.
Transl Cancer Res
2018
;
7
:
S10
S5
.
48.
Rubinstein
MR
,
Wang
X
,
Liu
W
,
Hao
Y
,
Cai
G
,
Han
YW
. 
Fusobacterium nucleatum promotes colorectal carcinogenesis by modulating E-cadherin/beta-catenin signaling via its FadA adhesin
.
Cell Host Microbe
2013
;
14
:
195
206
.
49.
Zhang
S
,
Yang
Y
,
Weng
W
,
Guo
B
,
Cai
G
,
Ma
Y
, et al
Fusobacterium nucleatum promotes chemoresistance to 5-fluorouracil by upregulation of BIRC3 expression in colorectal cancer
.
J Exp Clin Cancer Res
2019
;
38
:
14
.
50.
Gur
C
,
Ibrahim
Y
,
Isaacson
B
,
Yamin
R
,
Abed
J
,
Gamliel
M
, et al
Binding of the Fap2 protein of Fusobacterium nucleatum to human inhibitory receptor TIGIT protects tumors from immune cell attack
.
Immunity
2015
;
42
:
344
55
.
51.
Belcheva
A
,
Irrazabal
T
,
Robertson
SJ
,
Streutker
C
,
Maughan
H
,
Rubino
S
, et al
Gut microbial metabolism drives transformation of MSH2-deficient colon epithelial cells
.
Cell
2014
;
158
:
288
99
.
52.
Matson
V
,
Fessler
J
,
Bao
R
,
Chongsuwat
T
,
Zha
Y
,
Alegre
ML
, et al
The commensal microbiome is associated with anti-PD-1 efficacy in metastatic melanoma patients
.
Science
2018
;
359
:
104
8
.
53.
Facciabene
A
,
Rafail
S
,
Uribe-Herranz
M
,
Pierini
S
,
Bittinger
K
,
Verginadis
I
, et al
Abstract LB-178: The impact of the gut microbiome on the antitumor effects of radiotherapy
.
Cancer Res
2017
;
77
:
LB
178
.
54.
Uribe-Herranz
M
,
Rafail
S
,
Beghi
S
,
Gil-de-Gomez
L
,
Verginadis
I
,
Bittinger
K
, et al
Gut microbiota modulate dendritic cell antigen presentation and radiotherapy-induced antitumor immune response
.
J Clin Invest
2020
;
130
:
466
79
.