Changes in the microbiome reflect tumor stage

We performed whole-exome sequencing on a set of 88 samples, comprised of 44 pairs of tumor (adenocarcinomas) and normal colon tissue sample from the same patient, with previously characterized tissue-associated microbiomes[2]. The mutations in each of the tumors’ protein-coding regions were identified relative to the patient-matched normal sample and annotated as either synonymous, non-synonymous, or loss-of-function (LoF) mutations (S1 and S2 Tables, S1 and S2 Figs). The ranges of mutations per tumor in our data set were between 19–6678 (total), 10–3646 (missense) and 0–208 (LoF). The mutations were collapsed by gene as well as by pathways using both Kyoto Encyclopedia of Genes and Genomes (KEGG) and pathway interaction database (PID) annotations[31–34] (see Materials and methods). We performed quality control of the data and stringent filtering at every step, with the goal of reducing false positives in mutation calling and statistical prediction (e.g., requiring 30x coverage at a site in both the tumor and matched normal sample to call a mutation; see Materials and methods). While these requirements are likely to increase the frequency of false negatives (true mutations that simply do not meet our criteria), this rigorous strategy is appropriate as a means of increasing the biological relevance of our findings. Of note, when comparing the common LoF mutations found in our dataset to those found in colorectal tumors sampled as part of The Cancer Genome Atlas (TCGA) project, we find several commonalities, including a high frequency of LoF mutations in APC, as well as numerous missense mutations in KRAS, NRAS, and TP53, as expected (S1 Table)[35]. In general, the range of mutations across our sample set were also in line with those identified as part of TCGA and other CRC exome sequencing studies (see S15 and S16 Tables for comparisons)[35–39].

We first investigated the relationship between microbial communities and tumor stage (Fig 1). We hypothesize that the structure and composition of the associated microbiome can be affected by relevant physiological and anatomical differences between the tumors at different stages that would provide different microenvironmental niches for microbes. We identified the changes in the microbial communities surrounding each tumor as a function of stage by grouping the tumors into low stage (stages 1–2) and high stage (stages 3–4) classes, due to the low number of total tumors with available stage information, and applied linear discriminant analysis (LDA) effect size (LEfSe) to the raw operational taxonomic unit (OTU) tables corresponding to these tumors (S3, S5 and S6 Tables)[40]. The set of taxon abundances was transformed to generate a single value representing a risk index classifier for membership in the low-stage or high-stage group (Fig 1A; see Methods). To ascertain the error associated with these risk indices, a leave-one-out (LOO) cross-validation approach was applied. We also used the LOO results to generate receiver operating characteristic (ROC) curves and to calculate the area under the curve (AUC; see Fig 1B). In addition, we performed a permutation test to assess the method’s robustness (S5 Table). Using this approach, we demonstrate that the changes in abundances of 31 microbial taxa can be used to generate a classifier that distinguishes between low-stage and high-stage tumors at a fixed specificity of 80% and an accuracy of 77.5% (P = 0.02 by Mann-Whitney U test, and P = 0.007 by a permutation test; S5 Table). The resulting changes seen in our analysis of the microbial communities that vary by tumor stage were similar to those found in previous studies, including one using a Chinese patient cohort[4,41]. In both cases, there were significant changes among several taxa within the phylum Bacteroidetes, including Porphyromonadaceae, and Cyclobacteriaceae (Fig 1 and S5 Table). We applied the model generated from our data to two independent datasets that were analyses of the CRC microbiome and that also reported tumor staging[10,42] (S3 Fig). In the case of Flemer et al. the same approach successfully separated the low stage from high stage tumors, whereas the trend was the same in the case of the Yoon, et al. data, it is likely that the n-value was too low to achieve significance (S3 Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Correlation between the microbial community at a tumor that differentiates between tumor stage.

(A) Low-stage (stages 1–2) and high-stage (stages 3–4) tumors can be differentiated using a risk index classifier generated from microbial abundance data (y-axis). The central black bar indicates the median, and the thin black bars represent the 25th and 75th percentiles. (B) A receiver operating characteristic (ROC) curve was generated using a 10-fold cross-validation (blue dotted lines). The average of the 10-fold cross-validation curves is represented as a thick black line. (C) Differences in the mean abundances of a subset of the taxa predicted to interact differentially with high-stage and low-stage tumors. This subset represents those taxa that had a mean difference in abundance of greater than 0.1%, proportionally.

https://doi.org/10.1371/journal.pgen.1007376.g001

The comparisons of interest here are between tumor samples at high stage and tumor samples at low stage. This study utilized patient-matched normal samples, but these are not true normals, as they came from the same cancer patients who themselves have low or high stage cancer. To address the question of how the CRC samples, grouped by stage, compare to independent normal samples from healthy individuals, we obtained unpublished colonic mucosal microbiome data from a separate, but otherwise methodologically similar, study being performed at the University of Minnesota that included the tissue-associated microbiome from individuals without cancer (n = 12) undergoing routine colonoscopy. The OTU tables from the normal samples as well as the low-stage and high-stage samples were merged and assessed using the aforementioned LEfSe LOO approach (see Methods) to generate a model that is relevant to the normal and cancer samples (S4 Fig). The high-stage and low-stage samples were still able to be separated, as were the differences between the normal and the low-stage. Interestingly, the high-stage samples were not able to be statistically separated from the normal samples using this model.

Tumor mutations correlate with consistent changes in the proximal microbiome

Next, we attempted to use a similar approach to classify tumors based on mutational profiles. We initially focused on individual genes that harbor loss-of-function (LoF) mutations, as those, we predicted, would be the most likely to have a physiologically relevant interaction with the surrounding microbiome. We applied a prevalence filter to include only those mutations that were present in at least 10 or more patients at the gene level, including 11 genes in the analysis. The purpose of the filter was to limit the number of statistical tests while maintaining a reasonable number of possibly cancer-driving mutations; see S7 Table for how changes in this cutoff affect the number of genes included in the analysis. The raw OTU table was collapsed to the level of genus for the analysis. A visualization of the correlations between gene-level mutational status and the associated microbial abundances revealed differing patterns of abundances that suggests an interaction between the 11 most prevalent LoF tumor mutations and the microbiome (Fig 2 and S5 Fig). We hypothesized that the presence of mutation-specific patterns of microbial abundances could be statistically described by prediction of tumor LoF mutations in individual genes using the microbiome. For each of eleven genes that passed prevalence filtering cutoff, we identified the associated microbial taxa (Fig 2A, S6 and S8 Tables), generated risk indices for each patient (Fig 2B and 2C), and plotted the mean differences in abundances for a subset of microbial taxa interacting with each mutation (Fig 2D). We found that we are able to use microbiome composition profiles to predict the existence of tumor LoF mutations in the human genes APC, ANKRD36C, CTBP2, KMT2C, and ZNF717 (Q-value = 0.0011, 0.0011, 0.019, 0.019, and 0.055, respectively, by permutation test after False Discovery Rate (FDR) correction for multiple tests; Fig 2). The risk indices for each mutation were generated using sets of microbial taxa that ranges from 22 (ZNF717) to 53 (ANKRD36C) taxa (S6 Table). The taxa that showed the most dramatic differences in abundance when comparing tumors with and without mutations are shown in Fig 2D. For example, the abundance of Christensenellaceae is relatively lower in tumors with APC mutations, but relatively higher in tumors with ZNF717 mutations.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Commonly mutated genes show a predicted interaction with changes in the abundances of several microbial taxa.

(A) A heatmap of the scaled abundances values (cells) for a subset of taxa chosen as they were identified as discriminatory in each leave-one-out iteration (columns) that were found significantly associated with prevalent LoF mutations (rows). Scaled abundances are from the patients with the indicated mutations. (B) LoF mutations in each of the indicated genes can be predicted using a risk index as a classifier (y-axis). The central black bar indicates the median, and the thin black bars represent the 25th and 75th percentiles. (C) ROC curves were generated for each of the indicated mutations using a 10-fold cross-validation (blue dotted lines). The average of the 10-fold cross-validation curves is represented as a thick black line. (D) Differences in the mean abundances of a subset of the taxa predicted to interact differentially with tumors with a LoF mutation relative to those without the indicated mutation. This subset represents those taxa that had a mean difference in abundance of greater than 0.1%, proportionally.

https://doi.org/10.1371/journal.pgen.1007376.g002

To determine if our ability to find associations between specific mutations and microbial signatures was simply the result of being confounded by stage (e.g. perhaps high-stage tumors all have mutations in APC, meaning that the ability to separate tumors by mutations in this gene are potentially just a function of stage), we used mutation presence/absence information and tumor stage to look for confounding correlations between the two using Fisher’s exact test. The results show that there is no confounding associations between mutation status and tumor stage (S9 Table).

Next, we applied our interaction prediction approach, as described above, to the pathway-level mutational data (Fig 3; see Methods). Following visualization of the pathway level abundances (S6 and S7 Figs), we found that LoF mutations in each of the 21 KEGG pathways passing prevalence filter can be significantly predicted with a fixed specificity of 80% and an accuracy up to 86% (Q-values < 0.02 by permutation test after FDR correction; Fig 3A–3D, S10 Table, S8 and S9 Figs). Similarly, microbiome composition significantly predicted LoF mutations in 15 of the 19 tested PID pathways (Q-values < 0.04 by permutation test after FDR correction) (Fig 3E–3H, S10 Table, S10 and S11 Figs). The full sets of taxon abundances that were specifically associated with each of the LoF mutations in the genes and pathways can be found in S11, S12, S13 and S14 Tables and S12 Fig. As an additional control for our analysis pipeline, we also took the same sets of genes, KEGG, and PID pathways and tested them using an alternative analysis tool, MaAsLin (see Methods) [43]. When comparing these two approaches, we found a correspondence among the LoF mutations in genes and pathways with significant associations in the microbiome (genes: 2/5, KEGG: 16/21, and PID: 15/15) (S10 Table). In general, the number of taxa within each of the sets used to generate the risk indices was lower than that used for the gene-level analyses, with an average of 37 taxa per gene compared to 7 taxa per pathway. When comparing results using the gene-level interactions and the pathway level interactions, for instance looking at mutations in APC (Fig 2) and comparing them to mutations in the KEGG-defined Wnt signaling pathway and the PID-defined Canonical Wnt signaling pathway (Fig 3), the interactions at the pathway level are more statistically significant (AUC for APC = 0.81, KEGG = 0.88, PID = 0.90). This trend is consistent and can be visualized as a density histogram of interaction prediction accuracies (S13 Fig), indicating a stronger signal of interaction with the microbiome when using pathway-level mutational data.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Pathways harboring prevalent LoF mutations correlate with changes in the abundances of sets of microbial taxa.

(A) A heatmap of the scaled abundances values (cells) for a subset of taxa (columns) that are found significantly associated with KEGG pathways harboring LoF mutations (rows). Scaled abundances are from the patients with mutations in the indicated pathways. (B) LoF mutations in each of the indicated pathways can be predicted using a risk index as a classifier (y-axis). The central black bar indicates the median, and the thin black bars represent the 2nd and 4th quartiles. (C) ROC curves were generated for each of the indicated pathways using a 10-fold cross-validation (blue dotted lines). The average of the 10-fold cross-validation curves is represented as a thick black line. (D) Differences in the mean abundances of a subset of the taxa predicted to interact differentially with tumors harboring mutations in the indicated pathways relative to those without a mutation. This subset represents those taxa that had a mean difference in abundance of greater than 0.1%, proportionally. (E—F) Identically structured visualizations as in (A—D), but for PID pathway data rather than the KEGG pathways.

https://doi.org/10.1371/journal.pgen.1007376.g003

Predicted microbiome interaction network affected by tumor mutational profile

Lastly, we assessed the correlations between taxa among tumors with and without LoF mutations (Fig 4; see Methods). We found striking differences in the structure of the network comparing tumors with and without a Lof mutation in APC the correlations between taxa (Fig 4A). For example, in tumors with mutations in APC, the abundance of Christensenellaceae is positively correlated with Rhodocyclaceae and negatively correlated with Pedobacter. In tumors lacking LoF mutations in APC, these correlations are lost and Christensenellaceae is instead negatively correlated with Saprospiraceae and Gemm 1. We also assessed the network of correlations across tumors with mutations in PID pathways (Fig 4B). This analysis highlighted that some pathway-level mutations show a shared set of correlations between taxa, while others appear independent. It is interesting to note that there are several PID pathways that biologically linked (e.g. Degradation of beta-catenin, Regulation of nuclear b-catenin signaling, Presenilin action in Notch and Wnt Signaling, and Canonical Wnt signaling pathway). These pathways do indeed share core elements, but as defined by the PID authors, are comprised of distinct combinations of genes (S14 Fig)[34]. Several of the taxa that can be used to predict LoF mutations in p75(NTR) signaling share correlations among each other as well as with taxa associated with mutations in PDGFR-beta signaling and direct p53 effectors.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Interaction networks among bacteria are defined by host tumor mutations.

(A) SparCC analysis of the microbial abundances of the taxa identified by LEfSe for tumors with LoF mutations in APC (left) and without mutation (right) produce distinct patterns of correlations (edges) between a common set of taxa (nodes). Direct correlations are indicated as red edges and inverse correlations as blue edges (SparCC R > = 0.25, P < = 0.05 for displayed edges). (B) SparCC analysis was run simultaneously for all taxa identified by LEfSe when predicting interactions with mutations in PID pathways. There are interactions (dashed edges) between the taxa (grey nodes) associated with mutations across sets of PID pathways (green nodes). The solid edges indicate SparCC R-values (red for direct and blue for inverse correlations). The grey taxon nodes are scaled to the average abundance of the taxa in the associated tumor set. Edge color indicates the direction of the interaction, red for negative and blue for positive. Note that while several of the pathways (green nodes) have closely related general functions (e.g. “Canonical Wnt signaling pathway” and “Degradation of beta-catenin”), the underlying gene sets that comprise these pathways are distinct and result in independent correlations with microbial taxa.

https://doi.org/10.1371/journal.pgen.1007376.g004

Ethics statement

All research conformed to the Helsinki Declaration and was approved by the University of Minnesota Institutional Review Board, protocol 1310E44403.

Patient inclusion and DNA extraction

88 tissue samples from 44 individuals were used, with one tumor and one normal sample from each individual. These de-identified samples were obtained from the University of Minnesota Biological Materials Procurement Network (Bionet), a facility that archives research samples from patients who have provided written, informed consent. These samples were previously utilized and are described in detail in a previous study[75]. The patient information provided for this retrospective cohort did not include information related to immune status or inflammation. To reiterate these points, all research conformed to the Helsinki Declaration and was approved by the University of Minnesota Institutional Review Board, protocol 1310E44403. Tissue pairs were resected concurrently, rinsed with sterile water, flash frozen in liquid nitrogen, and characterized by staff pathologists. The criteria for selection were limited to the availability of patient-matched normal and tumor tissue specimens. In all cases, normal tissue was confirmed by the pathologist to be tumor free. The tumor stages were classified using a modified stage grouping based on the TNM scale (collapsing the letter-scales into single numbers—eg. IIIA, IIIB, and IIIC are reported as stage 3. Additional patient metadata are provided in S4 Table and in the indicated work[75]. Microbial community data from 16S rRNA gene sequencing of an independent cohort of healthy patients (n = 12) who underwent routine colonoscopies was acquired from a separate, unpublished dataset and used in the comparison of low-stage, high-stage, and normal microbial communities.

Microbiome characterization

The microbiome data used in the study was generated previously and is described exhaustively in[75]. Briefly, microbial DNA was extracted from patient-matched normal and tumor tissue samples using sonication for lysis and the AllPrep nucleic acid extraction kit (Qiagen, Valencia, CA). The V5-V6 regions of the 16S rRNA gene were PCR amplified with the addition of barcodes for multiplexing using the forward and reverse primer sets V5F and V6R from Cai, et al.[76]. The barcoded amplicons were pooled and Illumina adapters were ligated to the reads. A single lane on an Illumina MiSeq instrument was used (250 cycles, paired-end) to generate 16S rRNA gene sequences. The sequencing resulted in approximately 10.7 million total reads passing quality filtering in total, with a mean value of 121,470 quality reads per sample. The forward and reverse read pairs were merged using the USEARCH v7 program ‘fastq_mergepairs’, allowing stagger, with no mismatches allowed[77]. OTUs were picked using the closed-reference picking script in QIIME v1.7.0 using the Greengenes database (August 2013 release)[78–80]. The similarity threshold was set at 97%, reverse-read matching was enabled, and reference-based chimera calling was disabled.

Exome sequence data generation

Genomic DNA samples were quantified using a fluorometric assay, the Quant-iT PicoGreen dsDNA Assay Kit (Life Technologies, Grand Island, NY). Samples were considered passing quality control (QC) if they contained greater than 300 nanograms (ng) of DNA and display an A260:280 ratio above 1.7. Full workflow details for library preparation are outlined in the Nextera Rapid Capture Enrichment Protocol Guide (Illumina, Inc., San Diego, CA). In brief, libraries for Illumina next-generation sequencing were generated using Nextera library creation reagents (Illumina, Inc., San Diego, CA). A total of 50 ng of genomic DNA per sample were used as input for the library preparation. The DNA was tagmented (simultaneously tagged and fragmented) using Nextera transposome based fragmentation and transposition as part of the Nextera Rapid Capture Enrichment kit (Illumina, Inc., San Diego, CA). This process added Nextera adapters with complementarity to PCR primers containing sequences that allow addition of Illumina flow cell adapters and dual-indexed barcodes. The tagmented DNA was amplified using dual indexed barcoded primers. The amplified and indexed samples were pooled (8 samples per pool) and quantified to ensure appropriate DNA concentrations and fragment sizes using the fluorometric PicoGreen assay and the Bioanalyzer High-Sensitivity DNA Chip (Agilent Technologies, Santa Clara, CA). Libraries were considered to pass QC as long as they contained more than 500 ng of DNA and had an average peak size between 200–1000 base pairs. For hybridization and sequence capture, 500 nanograms of amplified library was hybridized to biotinylated oligonucleotide probes complementary to regions of interest at 58°C for 24 hours. Library-probe hybrids were captured using streptavidin-coated magnetic beads and subjected to multiple washing steps to remove non-specifically bound material. The washed and eluted library was subjected to a second hybridization and capture to further enrich target sequences. The captured material was then amplified using 12 cycles of PCR. The captured, amplified libraries underwent QC using a Bioanalyzer, and fluorometric PicoGreen assay. Libraries were considered to pass QC as long as they contained a DNA concentration greater than 10 nM and had an average size between 300–400 base pairs. Libraries were hybridized to a paired end flow cell at a concentration of 10 pM and individual fragments were clonally amplified by bridge amplification on the Illumina cBot (Illumina, Inc., San Diego, CA). Eleven lanes on an Illumina HiSeq 2000 (Illumina, Inc., San Diego, CA) were required to generate the desired sequences. Once clustering was complete, the flow cell was loaded on the HiSeq 2000 and sequenced using Illumina’s SBS chemistry at 100 bp per read. Upon completion of read 1, base pair index reads were performed to uniquely identify clustered libraries. Finally, the library fragments were resynthesized in the reverse direction and sequenced from the opposite end of the read 1 fragment, thus producing the paired end read 2. Full workflow details are outlined in Illumina’s cBot User Guide and HiSeq 2000 User Guides. Base call (.bcl) files for each cycle of sequencing were generated by Illumina Real Time Analysis (RTA) software. The base call files and run folders were then exported to servers maintained at the Minnesota Supercomputing Institute. Primary analysis and de-multiplexing was performed using Illumina’s CASAVA software 1.8.2. The end result of the CASAVA workflow was de-multiplexed FASTQ files that were utilized in subsequent analysis for read QC, mapping, and mutation calling.

Exome data analysis

The exome sequence data contained approximately 4.2 billion reads in total following adapter removal and quality filtering, inclusive of forward and reverse reads, with a mean value of 47.8 million high-quality reads per sample. The raw reads were assessed using FastQC v0.11.2 and the Nextera adapters removed using cutadapt v1.8.1[81,82]. Simultaneously, cutadapt was used to trim reads at bases with quality scores less than 20. Reads shorter than 40 bases were excluded. The trimmed and filtered read pairs were aligned and mapped to the human reference genome (hg19) using bwa v0.7.10 resulting in a bam file for each patient sample[83]. These files were further processed to sort the reads, add read groups, correct the mate-pair information, and mark and remove PCR duplicates using picard tools v1.133 and samtools v0.1.18[84,85]. Tumor-specific mutations were identified using FreeBayes v0.9.14-24-gc292036[86]. Following these steps, 94.0% of the remaining read pairs mapped to the reference genome, hg19. Specifically, SNPs-only were assessed and a minimum coverage at each identified mutation position of more than 30X was required in both the patient normal and tumor samples. These mutations were filtered to only include those that were within protein-coding regions which were then compiled into a single vcf file. We did not remove variants found in dbSNP. This vcf file was assessed using SNPeffect v4.1 K (2015-09-0) in order to predict the potential impact of each of the mutations[87]. Based on these results, the mutations were grouped into three categories: (1) total mutations (2) non-synonymous mutations and (3) loss of function (LoF) mutations. The total mutations group The total mutations group is simply the sum of all the mutations that were found in each tumor. The non-synonymous mutations included all the mutations in the total mutations group that were non-silent. The LoF group only included those mutations that resulted in a premature stop codon, a loss of a stop codon, or a frameshift mutation.

Mutations in genes were collapsed to pathways (PID and KEGG) based on pathway membership as defined by the relevant database authors[31,33,34]. Specifically, we relied upon Uniprot annotations to link genes to pathways with the database file, idmapping.dat.gz, available from Uniprot[88,89]. This, three column database file is used by the web-based mapping interface for cross-referencing between genes and pathway membership datasets, including KEGG and PID[88]. In our case, we annotated each gene in our dataset with pathway membership information by first reducing the size of the idmapping.dat file by including only entries for KEGG and PID. For each LoF mutation in our dataset, we found the corresponding gene in the reduced database file with a matching ENSEMBL Gene ID (S1 Table) and recorded which grouping it was found in (KEGG or PID) and what the name of the defined pathway was (e.g. Canonical Wnt signaling). Many genes had multiple annotations as they are members of more than one of the defined pathways. These annotations were used to identify which tumor samples had mutations in KEGG and PID pathways, defined as having at least one gene harboring a LoF mutation that was a member of the pathway.

Joint analysis of microbiome data with tumor stage and mutation status in genes and pathways

In order to identify microbial taxa that were significantly associated with specific characteristics (i.e. tumor stage—Fig 1, LoF mutations in individual genes—Fig 2, and LoF mutations within functional pathways—Fig 3), the OTU table was divided into two groups of tumors, defined by the characteristic of interest (e.g. Low Stage vs. High Stage, Mutation vs. No Mutation). As it would be invalid to generate a model that was built using the actual test sample, we used a “leave-one-out” approach when generating risk indices. To generate a risk index for a tumor, we started with an OTU table that contained all the tumors (43), leaving out the one for which we were calculating the risk index. This 43-tumor OTU table was divided into two groups based on the characteristic of interest, as described above. These data were used as input into LEfSe, as a means of identifying the microbial taxa that were able to discriminate between the two groupings[40]. Taxa were considered significant discriminators if the base 10 logarithm of their LDA score was less than 2, as recommended. The taxa that were significant discriminators were themselves of two categories: those that were more abundant, for instance, at tumors that harbored a mutation and those that were more abundant at tumors that did not harbor a mutation. The relative abundances of these taxa were arcsine root transformed and these transformed values—one for each category—were summed. The difference between these sums was calculated and is the risk index value used in each of the aforementioned figures. The use of the unweighted sum in the risk index, rather than relying entirely on the regression coefficients from LDA, is a simple way to control the degree of flexibility of the model when training on small sample sizes. More detail is described in a previous publication[90]. This procedure was repeated a total of 44 times (once for each tumor, to complete the leave-one-out approach) to obtain a risk index for each of the patients. The significance of the difference in risk indices between the patients found in one group vs. another (e.g. low stage vs. high stage, with LoF mutation, without LoF mutation) was assessed using a Mann-Whitney U test and a permutation test, in which we permuted the labels for a given group 999 times, each time deriving new held-out predictions of the risk indexes for each subject for that gene. Then the observed difference in means between the patients with LoF mutation and patients with LoF mutation risk index predictions using the method on the actual LoF mutation labels to the differences observed in the permutations to obtain an empirical P-value was compared. The resulting P-values were corrected using the false discovery rate (FDR) correction for multiple hypothesis tests.

MaAsLin v0.0.4 was used as a test for method robustness to alternative statistical parameters[43]. Without performing LOO iterations, the same sets of genes, KEGG pathways, and PID pathways that passed mutation prevalence filtering as for the LEfSe-based pipeline were used as metadata input for MaAsLin. Patient age and sex were also included in the model as covariates. The unrarefied OTU table collapsed to the genus level was used as abundance input, with prevalence filters set to only include taxa found at relative abundances of more than 0 in at least 5 tumors and an abundance filter of 0.001, relative abundance. Other parameters in MaAsLin were defaults. The p-values generated by MaAslin were used to identify sets of taxa that had altered abundances in tumors as a function of mutational status. These taxa were used to generate risk indices for each of the samples, as above, but in a non-iterative fashion as the LOO procedure was not used. The RIs for the samples harboring specific LoF mutations were tested against those without a detected LoF mutation, as above. The p-values generated were corrected for multiple testing by FDR. These values are reported in S10 Table.

Receiving Operating Characteristic (ROC) curves were plotted and the area under the curve (AUC) values computed on a dataset containing 10 sets of predictions and corresponding labels obtained from 10-fold cross-validation using ROCR package in R[91]. A risk index threshold was also obtained that best predicts the membership in a stage group or the presence or absence of LoF mutation with a leave-one-out cross-validation on the risk index. Each held-out sample was treated as a new patient on whom the optimal risk index cutoff was tested and subsequently refined to separate patients who had a LoF mutation and patient who did not have a LoF.

Correlation analysis was performed using SparCC on a reduced OTU table containing significant taxa identified using the above prediction methods collapsed to the genus level[92]. Pseudo p-values were calculated using 100 randomized sets. Networks of correlations were visualized using Cytoscape v3.1.0[93]. For easier interpretation, final network visualization was filtered to remove families that have potential for being miscalled.

The patients in this study have associated clinical data as we described previously[75], We used a linear model to determine the extent to which clinical factors may correlate with mutation load. These included patient sex, tumor stage, patient age, patient body mass index (BMI), and microsatellite instability (MSI) status. None of these factors, alone or in combination, were found to significantly impact the mutational data, though it bears noting that MSI status was only available for a subset (13 out of 44) of the patients.