The Local Tumor Microbiome Is Associated with Survival in Late-Stage Colorectal Cancer Patients

ABSTRACT The gut microbiome is associated with survival in colorectal cancer. Single organisms have been identified as markers of poor prognosis. However, in situ imaging of tumors demonstrate a polymicrobial tumor-associated community. To understand the role of these polymicrobial communities in survival, we conducted a nested case-control study in late-stage cancer patients undergoing resection for primary adenocarcinoma. The microbiome of paired tumor and adjacent normal tissue samples was profiled using 16S rRNA sequencing. We found a consistent difference in the microbiome between paired tumor and adjacent tissue, despite strong individual microbial identities. Furthermore, a larger difference between normal and tumor tissue was associated with prognosis: patients with shorter survival had a larger difference between normal and tumor tissue. Within the tumor tissue, we identified a 39-member community statistic associated with survival; for every log2-fold increase in this value, an individual’s odds of survival increased by 20% (odds ratio survival 1.20; 95% confidence interval = 1.04 to 1.33). Our results suggest that a polymicrobial tumor-specific microbiome is associated with survival in late-stage colorectal cancer patients. IMPORTANCE Microbiome studies in colorectal cancer (CRC) have primarily focused on the role of single organisms in cancer progression. Recent work has identified specific organisms throughout the intestinal tract, which may affect survival; however, the results are inconsistent. We found differences between the tumor microbiome and the microbiome of the rest of the intestine in patients, and the magnitude of this difference was associated with survival, or, the more like a healthy gut a tumor looked, the better a patient’s prognosis. Our results suggest that future microbiome-based interventions to affect survival in CRC will need to target the tumor community.


Supplemental tables
. Patient characteristics in the cohort Table S2. Per-Subject Predictors of the Microbiome in CTF ordination space Table S3. ASVs separating tumor and normal tissue based on differential ranking Table S4. Effect of tissue type on beta diversity in an unpaired analysis Table S5. Dissimilarity between normal and tumor tissue and survival Table S6. Feature Ranks for tumor vs normal tissue, based on long survival Table S7. Feature Ranks for interaction term (short survival in tumor tissue) Table S8. Tumor rPCA PC 1 features-associated ASVS for the ALR calculation Table S9. Tumor rPCA PC 2 associated features used in the ALR calculation Table S10. STORMS check list Figure S1. A family level view of the microbiome Figure S2. There is a strong individual effect on the microbiome Figure S3. Metadata predictors of the microbiome in Compositional Tensor Factorization. Figure S4. Subject aware techniques detect a consistent difference between tissue types and replicate associated taxa Figure S4. Survival separates individuals by tumor-associated rPCA quadrant

ASV Cumulative
Tumor Tissue

Antibiotics Usage
List what is known about antibiotics usage before or during sample collection.

STORMS
If participants were excluded due to current or recent antibiotics usage, state this here.
Other factors (e.g. proton pump inhibitors, probiotics, etc.) that may influence the microbiome should also be described as well. For example, hypothesized confounders may be controlled for by multivariable adjustment. Consider using a directed acyclic graph (DAG) to describe your causal model and justify any variables controlled for. DAGs can be made using www.dagitty.net.
Yes Table S1, Table S2, Figure S3 6.1 Selection bias Discuss potential for selection or survival bias.

STORMS
Selection bias can occur when some members of the target study population are more likely to be included in the study/final analytic sample than others. Some examples include survival bias (where part of the target study population is more likely to die before they can Lines 40-46; be studied), convenience sampling (where members of the target study population are not selected at random), and loss to follow-up (when probability of dropping out is related to one of the things being studied).

Bioinformatic and Statistical Methods
Describe any transformations to quantitative variables used in analyses (e.g. use of percentages instead of counts, normalization, rarefaction, categorization).

STROBE
Define or clarify any subjective terms such as "dominant," "dysbiosis," and similar words used in interpretation of results.
When interpreting the findings, consider how the interpretation of the findings may be summarized or quoted for the general public such as in press releases or news articles.
If causal language is used in the interpretation (such as "alters," "affects," "results in," "causes," or "impacts"), assumptions made for causal inference should be Yes explicitly stated as part of 6.0 and 13.0.
Distinguish between function potential (ie inferred from metagenomics) and observed activity (     The use of a subject-aware CTF ordination shows a clear difference in the microbiome between paired tissue samples from the same individual along all three PCs. (D) A common, directional change can be seen along PC 1 and PC2 when the difference between normal tissue and tumor tissue is plotted as a vector. The difference between normal-and tumor tissue can also be observed along individual components: (E) PC 1, (F) PC 2, and (G) PC 3. Ticks and dashed zero-lines along PC 1 (E) and PC 2 (F) match the two-dimensional axes in (D). All boxplots are shown with a Cohen's d effect size statistic for a one-sided t-test and p-values from a permutative one sample t-test, 999 permutations. (H) There is a change in the tissue associated ALR in our replication cohort between tissue types. The replication data was fit to reference sequences from the original cohort; the relationship between the tissue type and ALR was modeled using a linear mixed effects model treating the individual as a random intercept. The data is labeled with the estimated coefficient and parametric p-value.

Figure S4. Survival separates individuals by tumor-associated rPCA quadrant
Left shows the quadrants in the rPCA space. Middle plot shows the log odds ratio (OR) for pairwise quadrants with 95% CI. Negative log OR indicates long survival, positive values indicate short survival. Red line indicates 0. Right shows the OR for the crude (light gray) and adjusted (dark gray) of short survival. The Adjusted OR is adjusted for age, sex, ASA score, tumor location, surgery year, TNM stage, and grade of differentiation.