Skip to main content
Download PDF

Identifying important microbial and genomic biomarkers for differentiating right- versus left-sided colorectal cancer using random forest models

BMC Cancer volume 23, Article number: 647 (2023) Cite this article

Abstract

Background

Colorectal cancer (CRC) is a heterogeneous disease, with subtypes that have different clinical behaviours and subsequent prognoses. There is a growing body of evidence suggesting that right-sided colorectal cancer (RCC) and left-sided colorectal cancer (LCC) also differ in treatment success and patient outcomes. Biomarkers that differentiate between RCC and LCC are not well-established. Here, we apply random forest (RF) machine learning methods to identify genomic or microbial biomarkers that differentiate RCC and LCC.

Methods

RNA-seq expression data for 58,677 coding and non-coding human genes and count data for 28,557 human unmapped reads were obtained from 308 patient CRC tumour samples. We created three RF models for datasets of human genes-only, microbes-only, and genes-and-microbes combined. We used a permutation test to identify features of significant importance. Finally, we used differential expression (DE) and paired Wilcoxon-rank sum tests to associate features with a particular side.

Results

RF model accuracy scores were 90%, 70%, and 87% with area under curve (AUC) of 0.9, 0.76, and 0.89 for the human genomic, microbial, and combined feature sets, respectively. 15 features were identified as significant in the model of genes-only, 54 microbes in the model of microbes-only, and 28 genes and 18 microbes in the model with genes-and-microbes combined. PRAC1 expression was the most important feature for differentiating RCC and LCC in the genes-only model, with HOXB13, SPAG16, HOXC4, and RNLS also playing a role. Ruminococcus gnavus and Clostridium acetireducens were the most important in the microbial-only model. MYOM3, HOXC4, Coprococcus eutactus, PRAC1, lncRNA AC012531.25, Ruminococcus gnavus, RNLS, HOXC6, SPAG16 and Fusobacterium nucleatum were most important in the combined model.

Conclusions

Many of the identified genes and microbes among all models have previously established associations with CRC. However, the ability of RF models to account for inter-feature relationships within the underlying decision trees may yield a more sensitive and biologically interconnected set of genomic and microbial biomarkers.

Peer Review reports

Background

Despite being part of the same organ, colorectal cancer tumours can have different pathogenicity, histology and patient outcomes depending on subtype [1] and which side of the splenic flexure they occur [2]. Left-sided colorectal cancer (LCC, or distal colorectal cancer) affects the rectum, sigmoid colon, descending colon, and distal one-third of the transverse colon. It is generally more common in men, diagnosed at an earlier stage, more responsive to treatment, and patients exhibit a higher rate of survival [3]. Right-sided colorectal cancer (RCC, or proximal colorectal cancer) affects the proximal two-thirds of the transverse colon, ascending colon, and caecum [3]. It is generally more common in women, less responsive to existing treatments, and has poorer outcomes [2]. Numerous studies have reported vast differences between LCC and RCC in terms of diagnostics, prognostics, histology, epidemiology, pathology, treatment response, and survival [4,5,6]. Among other things, these differences suggest that LCC and RCC should be distinguished when developing new treatment regimens and therapeutic drugs [7, 8].

Gut microbiota has been shown to play an influential role in CRC carcinogenesis and progression. However, the mechanisms by which this occurs largely remains unknown [9]. In addition to cancer progression, it has also been postulated that the gut microbiome may affect gene expression and downstream patient treatment responses [10]. To test these hypotheses, there is a need for studies that explore the influence of the gut microbiome on the genomic expression inside colorectal cancer tumor cells.

Machine learning (ML) methods are frequently applied for classification in tasks that rely on high-dimensional genomic data. Here, to query the relationships between the expression of genomic features in CRC and microbial content, we use Random Forest (RF) classification [11]. We selected RF as it can account for interactions and correlations among large numbers of features [12]. Furthermore, RF models do not require normalization or scaling, which makes it possible to combine completely different types of data, for example, microbial count data and RNA-seq datasets.

Here, we explore the ability of RF models to predict CRC sidedness using three different datasets: human genomic feature expression level (RNA-seq), microbial count data (from unmapped human reads), and a combined genomic feature and microbial count dataset. We subsequently use differential expression (DE) analysis on the most important features of the RF model (i.e., biomarkers) to find differential genomic and microbial features and relationships between RCC and LCC. Finally, we discuss the possible biological mechanisms driving differences in these biomarkers.

Methods & materials

Patients, samples and processing

308 colorectal cancer tumour samples were obtained from patients via surgical resection (partial colectomy). Patients with inherited CRC and those who had received preoperative chemotherapy or radiotherapy were excluded. Patients were over the age of 18 and gave written informed consent. Tumour tissue was obtained between January 2002 and January 2016, with a median tumour tissue date of August 2006. The study was approved by the University of Otago, New Zealand, Human Research Ethics Committee (approval number: H16/037). Patient and clinical data, including anatomical site of tumour, in addition to genomic and microbial data profiles were available for all patients. Samples were snap frozen in liquid nitrogen at time of surgery and stored at -80 °C and transitioned for RNA Extraction using RNAlater®-ICE. RNA was then extracted using the QIAGEN RNAEasy mini kit and sequenced using Illumina HiSeq machines (2 × 125 bp PE v4 sequencing). The samples were machine-randomized to limit any machine-specific noise or calibration bias. Raw Sequence Reads are available at SRA Accession: PRJNA788974.

Sequence reads were first mapped to the human genome (GRCh38) using STAR (v2.73a). The remaining unmapped reads were classified using Kaiju (v1.6.2) to obtain microbial abundances [13, 14]. Raw genomic reads were TPM (transcripts per kilobase million) normalised prior to data analysis to remove gene length and sequencing depth biases. Microbial abundances were CPM (Counts per million) normalized.

Random forest model generation

The RF models were built on the following training datasets: the first contained 58,677 TPM normalized genomic features, the second contained CPM normalized microbial counts for 28,557 taxa, and the third contained a combination of both. A separate validation cohort of 30 samples (15 RCC, 15 LCC) was held out from model generation, leaving 278 patient samples for model development. Genomic and microbial data was available for all 308 patients.

The RF models were parametrized in parallel on high-powered cluster computing nodes with 8,136 cores in 226 × Broadwell nodes, and a total system memory of 31 TB.

Table 1 Patient Demographics & Cancer Characteristics
Full size table

RF models were generated using the Python-based scikit-learn random forest module [15, 16]. Model hyperparameters were optimized independently for all three training datasets using a grid search with 5-fold cross validation (scikit-learn package GridSearchCV) [15]. To narrow down suitable hyperparameter sets, the influence of 8 hyperparameters on F1 scores (the weighted average of precision and recall) were each independently observed across a typical range of values for each, while holding the other hyperparameters to their default values (Supplementary Figs. 1–3). GridSearchCV with 5-fold cross validation was then used on the smaller set of hyperparameter combinations on the training datasets. A final set of model parameters was selected for each dataset based on highest performing receiver-operator characteristic area-under the curve score (AUROC score), accuracy (total correctly classified cases), and F1 score. For cases in which the performance scores were identical, the model with the fewest features was selected.

Each model was trained using the finalized set of parameters using a 75% train, 25% test split on the dataset of tumor samples from the 278 different patients. The model metrics of accuracy, out-of-bag score, f1 score, ROC AUC score, recall, and precision for each of the three models are reported in Table 2. Overfitting was assessed by comparing model accuracy with out-of-bag score (number of correct predictions in the out-of-bag sample) and accuracy of the validation cohort. If the model accuracy differed from the out-of-bag score by 0.1 or more, we inferred that there was a strong likelihood overfitting had occurred. A threshold analysis was also performed for each model, but we found that all optimised thresholds were within 10% of the default value, so we used the default threshold value (0.5). The model was then validated on our validation cohort of 30 samples, this is sometimes referred to as the testing set, and is independent from the testing data used in model training. ROC curves were generated for all three models using the Python package seaborn [17].

Feature importance and retention

The feature importance scores (Gini impurity values) were extracted from each of our three RF models. Given that the models have large numbers of features (greater than 50), it is prudent to perform feature reduction such that only features with high importance (weight) and a high degree of statistical evidence are retained. Using the R package Rf2pval [18], we implemented a rank-based permutation approach to obtain distributions of feature-importance scores at each rank under a null hypothesis where none of the features are associated with the target variable, and assign p-values to the features. We generated 100 randomized datasets in which the target variable (‘side’) was permuted, retrained the RF models on each, and obtained feature importance scores and scoring metrics for each permuted model. We retained only features with p-values less than 0.05 (Fig. 2a-c). A threshold for feature reduction was identified using the overlap of feature importance scores from the true model with the mean of the permuted feature importance scores (Tables 1, 2 and 3). Family-wise error rate via resampling was used for measuring the probability of making one or more false discoveries during multiple-testing and was calculated using the Rf2pval package for all three models to be FWER < 0.05, or less than 5% chance of our features listed above our threshold being incorrectly identified.

Table 2 Random Forest Model Results
Full size table
Table 3 Top ranking features from the RF model trained on the genes-only dataset (Left). Top ranking features with p-values less than 0.05 and their importance scores discovered by our genes model (Left). Side-paired differential expression (fold change) analysis results of TPM values for the same features (Right) Wilcoxon-rank sum test was used to calculate p-values and FDR (Benjamini & Hochberg)
Full size table

Differential expression, feature side-assignment and heatmap generation

DE analysis was performed on each of the feature lists from the three models using edgeR [18]. Wilcoxon-rank sum tests were used to calculate p-values to test for DE of each model’s features. Heatmaps were generated for assessing feature clustering compared with the clinical labels of cancer stage, metastasis, subtype, gender site and side using the function heatmap.2 in the R package gplots [19]. The heatmap implements row-scaled z-scores of the transcripts per kilobase million (TPM) read counts, with hierarchical clustering using Pearson distance correlation, and average-linkage distance.

Results

Random forest model performance

We found that the random forest models from all three datasets clearly differentiated between LCC and RCC. Model accuracy on the validation sets ranged from 0.7 to 0.9, with genomic features having an accuracy of 0.94 and 0.9 on the training and validations sets, respectively microbial counts having an accuracy of 0.76 and 0.7, and genomic features with microbial counts having an accuracy of 0.8 and 0.87 (Fig. 1). Out-of-bag scores were 0.73, 0.74 and 0.74 for the three datasets. The strongest predictors between the LCC and RCC were genomic features, although classifications based on microbial count differences were also consistent. We found 15 statistically significant features in the genes-only model, 54 in the microbes-only model, and 46 in the genes-and-microbes model (Fig. 2A-C; Tables 3, 4 and 5).

Fig. 1
figure 1

Receiver Operating Characteristic Curves (ROC) as calculated on the held-out validation set. a ROC curve of the genes-only model. b ROC curve of the microbial-only model. c ROC curve of the genes-and-microbes model

Full size image
Fig. 2
figure 2

Feature importance plots showing rank-based feature importance scores of the permuted data and the scores of the real (unpermuted) data. The cutoff for features reported as significant was determined based on an alpha threshold of 0.05, and are to the left of the vertical red line. a genes-only model. b microbes-only model. c shows the genes-and-microbes model

Full size image
Table 4 As per Table 3, with microbes-only model data. Top ranking features with p-values less than 0.05 and their importance scores discovered by our microbes-only model (Left). Side-paired differential expression (fold change) analysis results of CPM values for the same features (Right) Wilcoxon-rank sum test was used to calculate p-values and FDR (Benjamini & Hochberg)
Full size table

Significant model features

Of the 15 significant features in our gene-only dataset, (Fig. 2A; Table 3) the highest importance score was for the Prostate Cancer Susceptibility Candidate 1 (PRAC1) gene, which has higher expression in LCC. Other significantly important features included those in the HOX family of genes, HOXB13, HOXC4, HOXC6 and HOXC8.

Table 3. Top ranking features from the RF model trained on the genes-only dataset (Left). Top ranking features with p-values less than 0.05 and their importance scores discovered by our genes model (Left). Side-paired differential expression (fold change) analysis results of TPM values for the same features (Right) Wilcoxon-rank sum test was used to calculate p-values and FDR (Benjamini & Hochberg).

In the microbes-only dataset, 54 features were identified by Rf2pval as significantly important (Fig. 2B; Table 4). The taxon with the highest importance score was Ruminococcus gnavus, which shows higher counts in RCC (Table 4). Clostridium acetireducens ranked second and was more abundant in RCC.

Table 5 As per Tables 3 and 4, with genes-and-microbes model data. Top ranking features with p-values less than 0.05 and their importance scores discovered by our genes-and-microbes model (Left). Side-paired differential expression (fold change) analysis results of TPM and CPM values for the same features (Right) Wilcoxon-rank sum test was used to calculate p-values and FDR (Benjamini & Hochberg)
Full size table

Table 4. As per Table 3, with microbes-only model data. Top ranking features with p-values less than 0.05 and their importance scores discovered by our microbes-only model (Left). Side-paired differential expression (fold change) analysis results of CPM values for the same features (Right) Wilcoxon-rank sum test was used to calculate p-values and FDR (Benjamini & Hochberg).

46 features were deemed significant in the genes-and-microbes model (Fig. 2C; Table 5): 28 genomic features and 18 microbes. Notable features include MYOM3, HOXC4, Coprococcus eutatus, PRAC1, lncRNA AC012531.3, Ruminococcus gnavus, RNLS, HOXC6, SPAG16, and Fusobacterium nucleatum.

Table 5. As per Tables 3 and 4, with genes-and-microbes model data. Top ranking features with p-values less than 0.05 and their importance scores discovered by our genes-and-microbes model (Left). Side-paired differential expression (fold change) analysis results of TPM and CPM values for the same features (Right) Wilcoxon-rank sum test was used to calculate p-values and FDR (Benjamini & Hochberg).

We used hierarchical clustering to ascertain connections between gene expression profiles and clinical characteristics and consensus subtyping scores (Fig. 3). As expected, of the six clinical characteristics that we considered (cancer stage, post-operative metastasis, consensus molecular subtype (CMS), gender, and site), side is most closely linked to the gene expression levels of our top genomic features. There is a clear cluster of left-sided CRC samples that show higher expression levels of PRAC1 and HOXB13 (left side of heatmap). There is also a subset of RCC that show higher expression of HOXC4, HOXC6, and HOXC8 (middle of heatmap), although not all RCC exhibit this pattern. Heatmap for microbes-only model is shown in Supplementary Fig. 4, and the heatmap for genes-and-microbes model is shown in Supplementary Fig. 5.

Fig. 3
figure 3

A heatmap of scaled gene expression values of the top-scoring genomic features discovered by the genes-only RF model and clinical characteristics. Hierarchical clustering of both genes and patients is via Pearson correlation, based on average linkage distance. The colors indicate row-scaled z-scores of TPM RNA-seq gene expression ratios

Full size image

In total, our models discovered 107 unique genomic and microbial features which played a significant role in the differentiating between CRC anatomical sides. Only six genomic features were common to both the genomic and genomic-plus-microbial models: PRAC1, SPAG16, HOXC4, RNLS, HOXC6 and PRKAA2; and six microbes which were common to both our microbes-only, and genes-and-microbes models: Ruminococcus gnavus, Ruminococcus sp. TF10-6, Selenomonas sp. CM52, Verrucomicrobia bacterium, Anaerostipes caccae and Turicibacter sanguinis.

Discussion

Many of the previous studies on RCC vs. LCC and gene expression have used the publicly available TCGA data. Here, we used a novel dataset of 308 patients, with microbial data from human unmapped reads, which adds to the growing body of evidence of the genomic and microbial differences between the sites [20, 21].

One difficulty in characterizing the roles of the microbiome and the genome in RCC vs. LCC is that there is genomic and microbial heterogeneity both between and within the two anatomical locations [22]. A primary reason for this heterogeneity is that the proximal and distal areas of the colon have different embryonic origins and physiological functions: the right-side of the colon is derived from the embryonic midgut and is involved in digestion, and the left side of the colon is derived from the embryonic hindgut and is involved primarily in the storage of fecal matter and water absorption. Despite these different functions, the microbial content is similar in these two parts of the colon because they are attached, and peristaltic movement allows stool matter to pass both forwards and backwards [23]. Numerous studies have shown a strong correlation between gut dysbiosis and CRC, but less is known about the microbial taxa that differentiate RCC and LCC and, perhaps, play a role in carcinogenesis [24]. Fusobacterium, Prevotella, Clostridium, Akkermansia, and Ruminococcus are among the most frequently reported bacteria in studies on CRC-related microbial dysbiosis [24, 25]. All were deemed significantly important microbial taxa in the RF models presented here.

All three RF models showed strong predictive accuracy. The microbes-only model showed the poorest predictive capability while the genes-only model was the highest performing. It is perhaps surprising that the combined model was not the most predictive. We postulate that this may be due to the fact that while microbes and genes may both affect CRC, microbial taxa are in fact indirect players, with effects that are reflected as altered genomic expression within the tumour, leading to cancer growth.

Finally, there were a number of highly important genes that differed between the genomic features only and combined RF models. One other point of interest is that there are some different top genes in the genes-only model when compared with the genes-and-microbes model. This may suggest that these genes and microbes act in consort and our genes-and-microbes RF model may have identified some underlying biological interactions.

Patterns in RCC

The RF models showed that increased expression of the HOX family of genes was characteristic of RCC. Specifically, we observed an upregulation of HOXC4, HOXC6, HOXC8, and HOX-related lncRNA AC012531.3, and a downregulation of HOXB13 (Tables 3 and 5). The HOX (homeobox) gene family is most well-known for guiding embryonic development [26]. HOX mutations that cause either increased or decreased expression have been associated with several types of cancer [27] as tumor suppressors and proto-oncogenes. However, their role in CRC is not well understood [27].

The top microbes identified by the microbes-only model include Ruminococcus gnavus, Clostridium acetereducens, Lachnospiraceae, and, Ruminococcus sp. TF10-6 (Table 4). R. gnavus causes inflammation in Crohn’s disease models, and influences immunotherapy responses in CRC [28, 29]. C. acetireducens is an anaerobic bacterium that has no previously known associations to CRC. However, it is known to oxidize alanine to produce butyrate, and butyrate has been associated with CRC tumorigenesis [30]. Lachnospiraceae spp are also known to produce short-chain fatty acids which are known to have increased abundance in CRC patients [31]; Ruminococcus sp. TF10-6, also more abundant in RCC; and Akkermansia muciniphila, more abundant in LCC. There is some evidence that the largely uncharacterized lncRNA AC012531.3 which is located in one of the HOX gene loci, plays a role in colorectal cancer carcinogenesis [32].

For the genes-and-microbes model the top features include Coprococcus eutactus, Ruminococcus gnavus, Fusobacterium nucleatum, Lachnospira sp., and Prevotella copri. Coprococcus eutactus has a very high feature importance score and is the microbe with the highest association to RCC in our genes-and-microbes model (Table 5). C. eutactus has previously been associated with longer cancer progression-free survival, and was not found in the microbes-only model, which could hint at a genomic-microbial interaction between C. eutactus and CRC side [33]. Ruminococcus gnavus, Ruminococcus sp. and Lachnospira were previously identified as being important to CRC and associated with structural segregation of the mucosa [34]. Fusobacterium nucleatum was found to be important in the genes-and-microbes model, but was not discovered by the microbes-only model. F. nucleatum is one of the most commonly associated species with CRC, and it is believed to act as a pathobiont [35, 36]. It is also known to cause periodontal disease and is currently being explored as a biomarker for high-risk CRC. Given that F. nucleatum was only significant in the genes-and-microbes model, and that it is known to be only situationally pathogenic, this suggests this taxon may become pathogenic under specific gene co-activation [35, 36]. Prevotella copri, also identified uniquely by our genes-and-microbes model, has been shown to be significantly enriched in the gut microbiome of CRC patients compared with normal patients [37].

Patterns in LCC

One recurring pattern in LCC is the expression of genes known to be associated with the prostate or prostate cancer. This is of interest given the heightened prevalence of LCC in men, and the left-sided colon’s close proximity to the prostate [38]. Prostate cancers and LCCs can be challenging to distinguish from biopsy samples, due to similarities in morphology and immunohistochemistry [39]. Genes that are of high importance in our RF models that are associated with both LCC and prostate cancer include PRAC1, HOXB13, SPAG16 (Tables 3 and 5) [26]. PRAC1 has been previously associated with LCC as well as prostate cancer [21, 40]. HOXB13 has a protective effect against tumor proliferation in RCC [41], and a reduction in the expression of HOXB13 via hypermethylation of the DNMT3B-HOXB13-C-myc signaling axis has been associated with tumor proliferation and metastasis in RCC [41]. Our results indicated that HOXB13 is under-expressed in RCC relative to LCC, which adds evidence to the hypothesis that decreased HOXB13 expression is specifically associated with RCC [41]. Elevated HOXC6 has been linked to poor overall survival in LCC patients, but not RCC patients [41]. MYOM3 was the top-ranking feature in our genes-and-microbes model and has a higher expression in LCC as determined using differential gene expression analysis. MYOM3 has not been studied in CRC but it has been linked to clinical outcomes in renal and lung cancer [42, 43].

While our microbes-only model mostly identified microbes associated with RCC, Akkermansia muciphila was more common in LCC (Table 4). A. muciniphila degrades mucin in the gut, and has previously been shown to exacerbate colitis-associated CRC development in mice [44], and is associated with total pathological response in treatment of non-small cell lung cancer [45]. Akkermansia has been noted as one of three microbes most likely to have a causal association with differential CRC treatment effectiveness [10].

The genes-and-microbes model also identified microbes that for the most part were enriched in RCC (Table 5). Only two microbes in this model are present at higher levels in LCC, namely, Verrucomicrobia bacterium and Fimbriiglobus ruber. Verrucomicrobia has been studied as a biomarker for the early detection of CRC [46]. However, Fimbriiglobus is largely uncharacterized.

Conclusions

Understanding microbial-genomic interactions may be important for informing treatment regimens in colorectal cancer. This study uses machine learning random forest (RF) models and differential gene expression (DE) to discover and associate genetic and microbial biomarkers with LCC and RCC. Three RF models with accuracy scores of 0.9, 0.7, and 0.87 were created and these yielded 15, 54 and 46 significantly important features. DE analysis was used to quantify changes in expression between CRC side. Our genes-and-microbes model identified microbes that did not appear in our microbes-only model, including C. eutactus, F. nucleatum and P. copri, and this may indicate that the random forest model is uncovering interactive effects between genes and microbes. RCC was most associated with the HOX family of genes, including HOX-associated lncRNA AC012531.25. LCC was highly associated with prostate cancer related genes, which is of interest as LCC is more common in men. The future of CRC research lies in personalized genomics, and the biomarkers identified by these three classification models may play an important role in the observed variability in clinicopathological and treatment outcomes of CRC patients.

Data availability

Raw Sequence Reads available at SRA Accession: PRJNA788974. Code is available from: https://github.com/tkolisnik/Kolisnik-Identifying-Important-CRC-Biomarkers-With-RF.

The R package Rf2pval used in these analyses is available at: www.github.com/tkolisnik/Rf2pval.

SciKit Learn can be imported from: https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.RandomForestClassifier.html.

References

  1. Fontana E, Eason K, Cervantes A, Salazar R, Sadanandam A. Context matters-consensus molecular subtypes of colorectal cancer as biomarkers for clinical trials. Ann Oncol. 2019;30:520–7.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  2. Yang J, Du XL, Li ST, Wang BY, Wu YY, Chen ZL, et al. Characteristics of differently located colorectal cancers support proximal and distal classification: a Population-Based study of 57,847 patients. PLoS ONE. 2016;11:e0167540.

    Article  PubMed  PubMed Central  Google Scholar 

  3. Stintzing S, Tejpar S, Gibbs P, Thiebach L, Lenz H-J. Understanding the role of primary tumour localisation in colorectal cancer treatment and outcomes. Eur J Cancer. 2017;84:69–80.

    Article  PubMed  PubMed Central  Google Scholar 

  4. Baran B, Mert Ozupek N, Yerli Tetik N, Acar E, Bekcioglu O, Baskin Y. Difference between left-sided and right-sided colorectal Cancer: a focused review of literature. Gastroenterol Res Pract. 2018;11:264–73.

    Article  CAS  Google Scholar 

  5. Bergen ES, Scherleitner P, Ferreira P, Kiesel B, Müller C, Widhalm G, et al. Primary tumor side is associated with prognosis of colorectal cancer patients with brain metastases. ESMO Open. 2021;6:100168.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  6. Zhao Z, Wang D-W, Yan N, Pan S, Li Z-W. Superior survival in right-sided versus left-sided colon signet ring cell carcinoma. Sci Rep. 2020;10:17900.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  7. Narayanan S, Gabriel E, Attwood K, Boland P, Nurkin S. Association of clinicopathologic and molecular markers on stage-specific survival of right Versus left Colon cancer. Clin Colorectal Cancer. 2018;17:e671–8.

    Article  PubMed  Google Scholar 

  8. Nagai Y, Kiyomatsu T, Gohda Y, Otani K, Deguchi K, Yamada K. The primary tumor location in colorectal cancer: a focused review on its impact on surgical management. Glob Health Med. 2021;3:386–93.

    Article  PubMed  PubMed Central  Google Scholar 

  9. Sánchez-Alcoholado L, Ramos-Molina B, Otero A, Laborda-Illanes A, Ordóñez R, Medina JA et al. The Role of the Gut Microbiome in Colorectal Cancer Development and Therapy Response. Cancers. 2020;12.

  10. Cercek A, Lumish M, Sinopoli J, Weiss J, Shia J, Lamendola-Essel M, et al. PD-1 blockade in Mismatch Repair-Deficient, locally advanced rectal Cancer. N Engl J Med. 2022. https://doi.org/10.1056/NEJMoa2201445.

    Article  PubMed  PubMed Central  Google Scholar 

  11. Breiman L. Random forests. Mach Learn. 2001;45:5–32.

    Article  Google Scholar 

  12. Chen X, Ishwaran H. Random forests for genomic data analysis. Genomics. 2012;99:323–9.

    Article  CAS  PubMed  Google Scholar 

  13. Menzel P, Ng KL, Krogh A. Fast and sensitive taxonomic classification for metagenomics with Kaiju. Nat Commun. 2016;9/:7:11257.

    Article  Google Scholar 

  14. Dobin A, Davis CA, Schlesinger F, Drenkow J, Zaleski C, Jha S, et al. STAR: ultrafast universal RNA-seq aligner. Bioinformatics. 2013;29:15–21.

    Article  CAS  PubMed  Google Scholar 

  15. Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O et al. Scikit-learn: Machine Learning in Python. MACHINE LEARNING IN PYTHON.:6.

  16. Mölder F, Jablonski KP, Letcher B, Hall MB, Tomkins-Tinch CH, Sochat V, et al. Sustainable data analysis with Snakemake F1000Res. 2021;10:33.

    PubMed  Google Scholar 

  17. Waskom M. Seaborn: statistical data visualization. J Open Source Softw. 2021;6:3021.

    Article  Google Scholar 

  18. Kolisnik T. Rf2pval: R package for obtaining p-values and cutoffs for features in random forest models.Github.

  19. Warnes B. Bonebakker, Gentleman. gplots: Various R programming tools for plotting data. R package version. 2009.

  20. Liang L, Zeng J-H, Qin X-G, Chen J-Q, Luo D-Z, Chen G. Distinguishable prognostic signatures of Left- and right-sided Colon cancer: a study based on sequencing data. Cell Physiol Biochem. 2018;48:475–90.

    Article  CAS  PubMed  Google Scholar 

  21. Jiang Y, Yan X, Liu K, Shi Y, Wang C, Hu J, et al. Discovering the molecular differences between right- and left-sided colon cancer using machine learning methods. BMC Cancer. 2020;20:1012.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  22. Liu LU, Holt PR, Krivosheyev V, Moss SF. Human right and left colon differ in epithelial cell apoptosis and in expression of Bak, a pro-apoptotic Bcl-2 homologue. Gut. 1999;45:45–50.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  23. Martin FH, Surgery. Gynecology & obstetrics. Franklin H. Martin Memorial Foundation; 1914.

  24. Liu W, Zhang X, Xu H, Li S, Lau HC-H, Chen Q, et al. Microbial Community Heterogeneity within Colorectal Neoplasia and its correlation with colorectal carcinogenesis. Gastroenterology. 2021;160:2395–408.

    Article  CAS  PubMed  Google Scholar 

  25. Lucas C, Barnich N, Nguyen HTT. Microbiota, Inflammation and Colorectal Cancer. Int J Mol Sci. 2017;18.

  26. Luo Z, Rhie SK, Farnham PJ. The Enigmatic HOX Genes: Can We Crack Their Code? Cancers. 2019;11.

  27. Li B, Huang Q, Wei G-H. The Role of HOX Transcription Factors in Cancer Predisposition and Progression. Cancers. 2019;11.

  28. Rebersek M. Gut microbiome and its role in colorectal cancer. BMC Cancer. 2021;21:1325.

    Article  PubMed  PubMed Central  Google Scholar 

  29. Henke MT, Kenny DJ, Cassilly CD, Vlamakis H, Xavier RJ, Clardy J. Ruminococcus gnavus, a member of the human gut microbiome associated with Crohn’s disease, produces an inflammatory polysaccharide. Proceedings of the National Academy of Sciences. 2019;116:12672–7.

  30. Okumura S, Konishi Y, Narukawa M, Sugiura Y, Yoshimoto S, Arai Y, et al. Gut bacteria identified in colorectal cancer patients promote tumourigenesis via butyrate secretion. Nat Commun. 2021;12:5674.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  31. Yang J, McDowell A, Kim EK, Seo H, Lee WH, Moon C-M, et al. Development of a colorectal cancer diagnostic model and dietary risk assessment through gut microbiome analysis. Exp Mol Med. 2019;51:1–15.

    PubMed  PubMed Central  Google Scholar 

  32. Wang X, Zhou J, Xu M, Yan Y, Huang L, Kuang Y, et al. A 15-lncRNA signature predicts survival and functions as a ceRNA in patients with colorectal cancer. Cancer Manag Res. 2018;10:5799–806.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  33. Peters BA, Wilson M, Moran U, Pavlick A, Izsak A, Wechter T, et al. Relating the gut metagenome and metatranscriptome to immunotherapy responses in melanoma patients. Genome Med. 2019;11:61.

    Article  PubMed  PubMed Central  Google Scholar 

  34. Chen W, Liu F, Ling Z, Tong X, Xiang C. Human intestinal lumen and mucosa-associated microbiota in patients with colorectal cancer. PLoS ONE. 2012;7:e39743.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  35. Wang N, Fang J-Y. Fusobacterium nucleatum, a key pathogenic factor and microbial biomarker for colorectal cancer. Trends Microbiol. 2022. https://doi.org/10.1016/j.tim.2022.08.010.

    Article  PubMed  PubMed Central  Google Scholar 

  36. Han YW. Fusobacterium nucleatum: a commensal-turned pathogen. Curr Opin Microbiol. 2015;23:141–7.

    Article  CAS  PubMed  Google Scholar 

  37. He T, Cheng X, Xing C. The gut microbial diversity of colon cancer patients and the clinical significance. Bioengineered. 2021;12:7046–60.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  38. Lee GH, Malietzis G, Askari A, Bernardo D, Al-Hassi HO, Clark SK. Is right-sided colon cancer different to left-sided colorectal cancer? – a systematic review. Eur J Surg Oncol. 2015;41:300–8.

    Article  CAS  PubMed  Google Scholar 

  39. Owens CL, Epstein JI, Netto GJ. Distinguishing prostatic from colorectal adenocarcinoma on biopsy samples: the role of morphology and immunohistochemistry. Arch Pathol Lab Med. 2007;131:599–603.

    Article  PubMed  Google Scholar 

  40. Hu W, Yang Y, Li X, Huang M, Xu F, Ge W, et al. Multi-omics Approach reveals distinct differences in Left- and right-sided Colon cancer. Mol Cancer Res. 2018;16:476–85.

    Article  CAS  PubMed  Google Scholar 

  41. Xie B, Bai B, Xu Y, Liu Y, Lv Y, Gao X, et al. Tumor-suppressive function and mechanism of HOXB13 in right-sided colon cancer. Signal Transduct Target Ther. 2019;4:51.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  42. Expression of MYOM3 in renal cancer -. The Human Protein Atlas. https://www.proteinatlas.org/ENSG00000142661-MYOM3/pathology/renal+cancer. Accessed 17 Jul 2022.

  43. Yang J, Feng E, Ren Y, Qiu S, Zhao L, Li X. Long non-coding (lnc)RNA profiling and the role of a key regulator lnc-PNRC2-1 in the transforming growth factor-β1-induced epithelial-mesenchymal transition of CNE1 nasopharyngeal carcinoma cells. J Int Med Res. 2021;49:300060521996515.

    CAS  PubMed  Google Scholar 

  44. Wang F, Cai K, Xiao Q, He L, Xie L, Liu Z. Akkermansia muciniphila administration exacerbated the development of colitis-associated colorectal cancer in mice. J Cancer. 2022;13:124–33.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  45. Cascone T, William WN Jr, Weissferdt A, Leung CH, Lin HY, Pataer A, et al. Neoadjuvant nivolumab or nivolumab plus ipilimumab in operable non-small cell lung cancer: the phase 2 randomized NEOSTAR trial. Nat Med. 2021;27:504–14.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  46. Wu Y, Jiao N, Zhu R, Zhang Y, Wu D, Wang A-J, et al. Identification of microbial markers across populations in early detection of colorectal cancer. Nat Commun. 2021;12:3063.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

Download references

Acknowledgements

The author(s) wish to acknowledge the use of New Zealand eScience Infrastructure (NeSI) high performance computing facilities, consulting support and/or training services as part of this research. New Zealand’s national facilities are provided by NeSI and funded jointly by NeSI’s collaborator institutions and through the Ministry of Business, Innovation & Employment’s Research Infrastructure programme. URL https://www.nesi.org.nz.

Funding

The funding for the data analysis in this work was provided by Massey University School of Natural and Computational Sciences. Funding for data collection was provided by the lab of Dr. Rachel Purcell at University of Otago. Computational resources were provided by New Zealand eScience Infrastructure (NeSI).

Author information

Authors and Affiliations

  1. School of Natural Sciences, Massey University, Auckland, New Zealand

    Tyler Kolisnik, Arielle Kae Sulit, Sebastian Schmeier & Olin Silander

  2. Canada’s Michael Smith Genome Sciences Centre, BC Cancer, Vancouver, BC, Canada

    Tyler Kolisnik

  3. Department of Surgery, University of Otago, Christchurch, New Zealand

    Arielle Kae Sulit, Frank Frizelle & Rachel Purcell

  4. School of Mathematical and Computational Sciences, Massey University, Auckland, New Zealand

    Adam Smith

Authors
  1. Tyler Kolisnik
    View author publications

    You can also search for this author in PubMed Google Scholar

  2. Arielle Kae Sulit
    View author publications

    You can also search for this author in PubMed Google Scholar

  3. Sebastian Schmeier
    View author publications

    You can also search for this author in PubMed Google Scholar

  4. Frank Frizelle
    View author publications

    You can also search for this author in PubMed Google Scholar

  5. Rachel Purcell
    View author publications

    You can also search for this author in PubMed Google Scholar

  6. Adam Smith
    View author publications

    You can also search for this author in PubMed Google Scholar

  7. Olin Silander
    View author publications

    You can also search for this author in PubMed Google Scholar

Contributions

TK conceived the project, and designed and implemented the code for data curation, machine learning, hyperparameter selection, feature reduction, as well as analysis of results and preparation of figures and wrote the first draft and contributed to all versions of the manuscript. AKS mapped the RNA-seq data to the human genome and mapped the human unmapped reads to produce the microbial counts. AS guided the development of statistical methods and contributed to editing the manuscript. OS provided support with idea generation, troubleshooting, statistics, and critical analysis of results, and contributed to editing the manuscript. RP, FF provided sample collection, processing, RNA sequencing and metadata collection, and minor manuscript revisions. SS provided initial insight to the formation of the machine learning pipeline, as well as assistance with conceptualization and data normalization. All authors have read and approved this manuscript.

Corresponding author

Correspondence to Tyler Kolisnik.

Ethics declarations

Ethics approval and consent to participate

The study was approved by the University of Otago, New Zealand, Human Research Ethics Committee (approval number: H16/037). Informed consent was obtained from all subjects and/or their legal guardians. All experiments were performed in accordance with relevant ethics guidelines and the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.

Additional information

Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Supplementary Material 1

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Kolisnik, T., Sulit, A.K., Schmeier, S. et al. Identifying important microbial and genomic biomarkers for differentiating right- versus left-sided colorectal cancer using random forest models. BMC Cancer 23, 647 (2023). https://doi.org/10.1186/s12885-023-10848-9

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://doi.org/10.1186/s12885-023-10848-9

Keywords

BMC Cancer

ISSN: 1471-2407

Contact us