Meta-analysis of the gut microbiota in predicting response to cancer immunotherapy in metastatic melanoma

Abstract

BACKGROUNDIdentifying factors conferring responses to therapy in cancer is critical to select the best treatment for patients. For immune checkpoint inhibition (ICI) therapy, mounting evidence suggests that the gut microbiome can determine patient treatment outcomes. However, the extent to which gut microbial features are applicable across different patient cohorts has not been extensively explored.METHODSWe performed a meta-analysis of 4 published shotgun metagenomic studies (Ntot = 130 patients) investigating differential microbiome composition and imputed metabolic function between responders and nonresponders to ICI.RESULTSOur analysis identified both known microbial features enriched in responders, such as Faecalibacterium as the prevailing taxa, as well as additional features, including overrepresentation of Barnesiella intestinihominis and the components of vitamin B metabolism. A classifier designed to predict responders based on these features identified responders in an independent cohort of 27 patients with the area under the receiver operating characteristic curve of 0.625 (95% CI: 0.348-0.899) and was predictive of prognosis (HR = 0.35, P = 0.081).CONCLUSIONThese results suggest the existence of a fecal microbiome signature inherent across responders that may be exploited for diagnostic or therapeutic purposes.FUNDINGThis work was funded by the Knut and Alice Wallenberg Foundation, BioGaia AB, and Cancerfonden.

Keywords: Bioinformatics; Cancer immunotherapy; Melanoma; Microbiology; Oncology.

Figure 1. Meta-analysis workflow.

In brief, fecal MGS data at baseline from 4 studies (N = 130) comparing differences in microbiome composition between R and NR were systematically reanalyzed at the taxonomic, genetic, and functional level. We set aside data from one of the studies (18) (N = 27) in order to validate our findings in an independent cohort. WGS, whole-genome sequencing.

Figure 2. Differences in microbial abundances between responders and nonresponders to ICI therapy.

(A) Multidimensional scaling (MDS) analysis using weighted UniFrac distances of samples (N = 103) pertaining to each study and grouped by response. (B) MDS plot of all samples included in the meta-analysis, grouped by response. Dashed and solid lines indicate 95% normal confidence ellipses for each study and response group, respectively. (C) Hierarchical clustering of log-normalized abundances of differentially abundant taxa between R and NR for only the melanoma subset of patients. Seventeen operational taxonomic unit (OTUs) were identified as differentially abundant with P < 0.05 (Wilcoxon’s rank-sum test).

Figure 3. Alterations in the functional potential of the gut microbiome between responders and nonresponders to ICI therapy.

Log-normalized abundances reported as copies per million (CoPM, analogous to TPM for RNA-Seq data) of differentially abundant MetaCyc pathways present in the fecal microbiome of patients (N = 103). Twenty-nine pathways were identified as differentially abundant with P < 0.05 (unadjusted, Wilcoxon’s rank-sum test). PWY-tags preceding each pathway name are unique IDs associated with each metabolic pathway on MetaCyc. Hierarchical clustering dendrogram for the pathways is omitted due to figure size constraints.

Figure 4. Microbial signatures in patients with melanoma are predictive of PFS in an independent cohort.

(A) Construction of a random forest (RF) classifier for treatment response to ICI, based on fecal sequencing data. Differentially abundant features (k_pool = 49) between R and NR patients (N = 103) were selected as input for training an RF classifier. The trained model was then used to predict treatment responses of patients in an independent cohort (N = 27). (B) Lollipop plot of the top 10 most important variables, evaluated according to the mean decrease in Gini impurity as determined by the RF classifier after model training. Features are color-coded according to species or pathway. (C) Performance characteristics of the RF classifier. Left: Receiver operating characteristic (ROC) curves on the training and test data. AUC, area under the curve. Right: Confusion matrices and prediction scores for the RF classifier on the training and test data. (D) Kaplan-Meier PFS estimates for the R and NR patients, as predicted by the RF classifier (P value using the log-rank test), with number of patients at risk for each 5-month interval and hazard ratio for the R group (calculated using Cox proportional hazards regression).