Fecal Short-Chain Fatty Acids Are Not Predictive of Colonic Tumor Status and Cannot Be Predicted Based on Bacterial Community Structure

Considering that colorectal cancer is the third leading cancer-related cause of death within the United States, it is important to detect colorectal tumors early and to prevent the formation of tumors. Short-chain fatty acids (SCFAs) are often used as a surrogate for measuring gut health and for being anticarcinogenic because of their anti-inflammatory properties. We evaluated the fecal SCFA concentrations of a cohort of individuals with different colonic tumor burdens who were previously analyzed to identify microbiome-based biomarkers of tumors. We were unable to find an association between SCFA concentration and tumor burden or use SCFAs to improve our microbiome-based models of classifying people based on their tumor status. Furthermore, we were unable to find an association between the fecal community structure and SCFA concentrations. Our results indicate that the association between fecal SCFAs, the gut microbiome, and tumor burden is weak.

C olorectal cancer is the third leading cancer-related cause of death within the United States (1). Less than 10% of cases can be attributed to genetic risk factors (2). This leaves a significant role for environmental, behavioral, and dietary factors (3,4). Colorectal cancer is thought to be initiated by a series of mutations that accumulate as the mutated cells begin to proliferate, leading to adenomatous lesions, which are succeeded by carcinomas (2). Throughout this progression, there are ample opportunities for bacterial populations to have a role, as some bacteria are known to cause mutations, induce inflammation, and accelerate tumorigenesis (5)(6)(7). Additional crosssectional studies of humans have identified microbiome-based biomarkers of disease (8). These studies suggest that in some cases, it is the loss of bacterial populations that produce short-chain fatty acids (SCFAs) that results in increased inflammation and tumorigenesis.
Many microbiome studies use the concentrations of SCFAs and the presence of 16S rRNA gene sequences from organisms and the genes involved in producing them as a biomarker of a healthy microbiota (9,10). Depending on the concentrations, SCFAs can have proliferative activities at low concentrations or antiproliferative activities at higher concentrations; they can also have anti-inflammatory activities (11). Direct supplementation of SCFAs or feeding of fiber caused an overall reduction in the tumor burden in mouse models of colorectal cancer (12). These results suggest that supplementation with fiber, which many colonic bacteria ferment to produce SCFAs, may confer beneficial effects against colorectal cancer. Regardless, there is a lack of consistent evidence that increasing SCFA concentrations can protect against colorectal cancer in humans. Case-control studies that have investigated possible associations between SCFAs and colon tumor status have been plagued by relatively small numbers of subjects but have reported increased total and relative fecal acetate levels and decreased relative fecal butyrate concentrations in subjects with colonic lesions (13). In randomized controlled trials, fiber supplementation has been inconsistently associated with protection against tumor formation and recurrence (14,15). Such studies are confounded by difficulties ensuring subjects took the proper dose and using subjects with prior polyp history who may be beyond a point of benefiting from fiber supplementation. Together, these findings temper enthusiasm for treatments that target the production of SCFAs or for using them as biomarkers for protection against tumorigenesis.
Fecal SCFA concentrations did not vary with diagnosis or treatment. To test for a significant association between colorectal cancer and SCFAs, we quantified the concentration of acetate, propionate, and butyrate in feces of previously characterized individuals with healthy colons (n ϭ 172) and those with colonic adenomas (n ϭ 198) or carcinomas (n ϭ 120) (16). We were unable to detect a significant difference in any SCFA concentration across the diagnosis groups (all P Ͼ 0.15; Fig. 1A). Among the individuals with adenomas and carcinomas, a subset (n adenoma ϭ 41 and n carcinoma ϭ 26) were treated and sampled a year later (17). None of the individuals showed signs of recurrence, and yet none of the SCFAs exhibited a significant change with treatment (all P Ͼ 0.058; Fig. 1B). For both the pretreatment cross-sectional data and the pre/ posttreatment data, we also failed to detect any significant differences in the relative concentrations of any SCFAs (P Ͼ 0.16). Finally, we pooled the SCFA concentrations on a total and per molecule carbon basis and again failed to observe any significant differences (P Ͼ 0.077). Although some of the P values from our analyses were close to 0.05, the effect sizes were all relatively small and inconsistent given the disease progression (Fig. 1). These results demonstrated that there were no significant associations between fecal SCFA concentration and diagnosis or treatment.
Combining SCFA and microbiome data does not improve the ability to diagnose individuals as having adenomas or carcinomas using a random forest model. We previously found that binning 16S rRNA gene sequence data into operational taxonomic units (OTUs) based on 97% similarity or into genera enabled us to classify individuals as having adenomas or carcinomas using random forest machine learning models (8,16). We repeated that analysis but added the concentrations of the SCFAs as possible features to train the models (see Fig. S1 in the supplemental material). Models trained using SCFAs to classify individuals as having adenomas or carcinomas rather than healthy colons had median areas under the receiver operator characteristic curve (AUROC) that were significantly greater than 0.5 (P adenoma Ͻ 0.001 and P carcinoma Ͻ 0.001). However, the AUROC values to detect the presence of adenomas or carcinomas were only 0.54 and 0.55, respectively, indicating that SCFAs had poor predictive power on their own ( Fig. 2A). When we trained the models with the SCFA concentrations and OTU or genus-level relative abundances, the AUROC values were not significantly different from the same models trained without the SCFA concentrations (P Ͼ 0.15; Fig. 2A). These data demonstrate that knowledge of the SCFA profile from a subject's fecal sample did not improve the ability to diagnose a colonic lesion.
Knowledge of microbial community structure does not predict SCFA concentrations using a random forest model. We next asked whether the fecal community structure was predictive of fecal SCFA concentrations, regardless of a person's diagnosis. We trained random forest regression models using 16S rRNA gene sequence data binned into OTUs and genera to predict the concentrations of the SCFAs and Colorectal Cancer ® SCFAs (Fig. S2). The largest R 2 between the observed SCFA concentrations and the modeled concentrations was 0.14, which was observed when using genus data to predict butyrate concentrations (Fig. 2B). We also used a smaller data set of shotgun metagenomic sequencing data generated from a subset of our cohort (n normal ϭ 27, n adenoma ϭ 25, and n cancer ϭ 26) (18). We binned genes extracted from the assembled metagenomes into operational protein families (OPFs) or KEGG categories and trained random forest regression models using metagenomic sequence data to predict the concentrations of the SCFAs (Fig. S2). Similar to the analysis using 16S rRNA gene sequence data, the metagenomic data were not predictive of SCFA concentration. The largest R 2 was 0.055, which was observed when using KEGG data to predict propionate concentrations (Fig. 2B). Because of the limited number of samples that we were able to generate metagenomic sequence data from, we used our 16S rRNA gene sequence data to impute metagenomes that were binned into metabolic pathways or KEGG categories using PICRUSt (Fig. S2). SCFA concentrations could not be predicted based on the imputed metagenomic data. The largest R 2 was 0.085, which was observed when using KEGG data to predict propionate concentrations (Fig. 2B). The The median AUROC for diagnosing individuals as having adenomas or carcinomas using SCFAs was slightly better than than chance (depicted by the horizontal line at 0.50) but did not improve performance of the models generated using 16S rRNA gene sequence data. (B) Regression models that were trained using 16S rRNA gene sequence, metagenomic, and PICRUSt data to predict the concentrations of SCFAs performed poorly (all median R 2 values of Ͻ0.14). Regression models generated using 16S rRNA gene sequence and PICRUSt data included data from 490 samples, and those generated using metagenomic data included data from 78 samples. Sze et al. ® inability to model SCFA concentrations from microbiome data indicates that the knowledge of the abundance of organisms and their genes was insufficient to predict fecal SCFA concentrations.
Conclusion. Our data indicate that fecal SCFA concentrations are not associated with the presence of adenomas or carcinomas and that they provide weak predictive power to improve the ability to diagnose someone with one of these lesions. Furthermore, knowledge of the taxonomic and genetic structure of gut microbiota was not meaningfully predictive of SCFA concentrations. These results complement existing literature that suggest that fiber consumption and the production of SCFAs are unable to prevent the risk of developing colonic tumors. It is important to note that our analysis was based on characterizations of SCFA and microbiome profiles using fecal samples at a single time point. Furthermore, observations along the mucosa near the site of lesions may provide a stronger association. This may be a cautionary result to temper enthusiasm for SCFAs as a biomarker of gut health more generally. Going forward, it is critical to develop additional hypotheses for how the microbiome and host interact to drive tumorigenesis so that we can better understand tumorigenesis and identify biomarkers that will allow early detection of lesions.
Methods. (i) Study design and sampling. The overall study design and the resulting sequence data have been previously described (16,17). In brief, fecal samples were obtained from 172 individuals with healthy colons, 198 individuals with colonic adenomas, and 120 individuals with carcinomas. Of the individuals diagnosed as having adenomas or carcinomas, a subset (n adenoma ϭ 41 and n carcinoma ϭ26) were sampled after treatment of the lesion (median ϭ 255 days between sampling, interquartile range [IQR] ϭ 233 to 334 days). Tumor diagnosis was made by colonoscopic examination and histopathological review of the biopsy specimens (16). The University of Michigan Institutional Review Board approved the studies that generated the samples, and informed consent was obtained from all participants in accordance with the guidelines set out by the Helsinki Declaration.
(ii) Measuring specific SCFAs. The measurement of acetate, propionate, isobutyrate, and butyrate used a previously published protocol that used high-performance liquid chromatography (HPLC) (19). Two changes were made to the protocol. First, instead of using fecal samples suspended in DNA Genotek OmniGut tubes, we suspended frozen fecal samples in 1 ml of phosphate-buffered saline (PBS). Second, instead of using the average weight of fecal sample aliquots to normalize SCFA concentrations, we used the actual weight of the fecal samples. These methodological changes did not affect the range of concentrations of these SCFAs between the two studies. The concentrations of isobutyrate were consistently at or below the limit of detection and were not included in our analysis.
(iii) 16S rRNA gene sequence data analysis. Sequence data from Baxter et al. (16) and Sze et al. (17) were obtained from the Sequence Read Archive (studies SRP062005 and SRP096978) and reprocessed using mothur v.1.42 (20). The original studies generated sequence data from the V4 region of the 16S rRNA gene using paired 250nucleotide (nt) reads on an Illumina MiSeq sequencer. The resulting sequence data were assembled into contigs and screened to remove low-quality contigs and chimeras. The curated sequences were then clustered into OTUs at a 97% similarity threshold and assigned to the closest possible genus with an 80% confidence threshold trained on the reference collection from the Ribosomal Database Project (v.16). We used PICRUSt (v.2.1.0-b) with the recommended standard operating protocol to generate imputed metagenomes based on the expected metabolic pathways and KEGG categories (21).
(iv) Metagenomic DNA sequence analysis. A subset of the samples from the samples described by Baxter et al. (16) were used to generate metagenomic sequence data (n normal ϭ 27, n adenoma ϭ 25, and n cancer ϭ 26). These data were generated by Hannigan et al. (18) and deposited into the Sequence Read Archive (study SRP108915). Fecal DNA was subjected to shotgun sequencing on an Illumina HiSeq using 125-bp paired-end reads. The archived sequences were already quality filtered and aligned to the human genome to remove contaminating sequence data. We downloaded the sequences and assembled them into contigs using MEGAHIT (22), which were used to identify open reading frames (ORFs) using Prodigal (23). We determined the abundance of each ORF by mapping the raw reads back to the ORFs using Diamond (24). We clustered the ORFs into operational protein families (OPFs) in which the clustered ORFs were more than 40% identical to each other using mmseq2 (25). We also used mmseq2 to map the ORFs to the KEGG database and clustered the ORFs according to which category the ORFs mapped.
(v) Random forest models. The classification models were built to predict lesion type from microbiome information with or without SCFA concentrations. The regression models were built to predict the SCFA concentrations of acetate, butyrate, and propionate from microbiome information. For classification and regression models, we preprocessed the features by scaling them to vary between zero and one. Features with no variance in the training set were removed from both the training and testing sets. We randomly split the data into training and test sets so that the training set consisted of 80% of the full data set, while the test set was composed of the remaining data. The training set was used for hyperparameter selection and training the model and the test set was used for evaluating prediction performance. For each model, the best performing hyperparameter, mtry, was selected in an internal five-fold cross-validation of the training set with 100 randomizations. The mtry parameter represents the number of features randomly sampled from the available features at a question point in the classification tree (i.e., called splits of nodes) that, when answered, lead to the greatest improvement in classification. Six values of mtry were tested, and the value that provided the largest AUROC or R 2 was selected. We trained the random forest model using the selected mtry value and predicted the held-out test set. The data-split, hyperparameter selection, training, and testing steps were repeated 100 times to get a reliable and robust reading of model prediction performance. We used AUROC and R 2 as the prediction performance metrics for classification and regression models, respectively. We used the randomForest R package (version 4.6-14) as implemented in the caret R package (version 6.0-81) for developing and testing our models.
(vi) Statistical analysis workflow. Data summaries, statistical analysis, and data visualizations were performed using R (v.3.5.1) with the tidyverse package (v.1.2.1). To assess differences in SCFA concentrations between individuals with healthy colons and those with adenomas or carcinomas, we used the Kruskal-Wallis rank sum test. If a test had a P value below 0.05, we then applied a pairwise Wilcoxon rank sum test with a Benjamini-Hochberg correction for multiple comparisons. To assess differences in SCFA concentrations between individual samples before and after treatment, we used paired Wilcoxon rank sum tests to test for significance. To compare the median AUCROC for the held-out data for the model generated using only the SCFAs, we compared the distribution of the data to the expected median of 0.5 using the Wilcoxon rank sum test to test whether the model performed better than would be achieved by randomly assigning the data to each diagnosis. When we compared the random forest models generated without and with SCFA data included, we used Wilcoxon rank sum tests to determine whether the models with the SCFA data included did better.
(vii) Code availability. The code for all sequence curation and analysis steps, including an Rmarkdown version of this article, is available at https://github.com/ SchlossLab/Sze_SCFACRC_mBio_2019/.