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

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.

pyRforest: A comprehensive R package for genomic data analysis featuring scikit-learn Random Forests in R

“…Specifically, SHAP analysis indicated that PRAC1 and HOXB13 contribute to the predisposition towards left-sided CRC. This observation is consistent with outcomes from other studies, highlighting the role of these biomarkers as being differentially expressed depending on spatial location of the disease [22, 23]. HOXC4 and HOXC6 SHAP directionality deviated from expected values from our previous study: the bimodal feature explanations of these genes indicate the relationship of these genes with outcome may be complex, and that their contribution is context dependent.…”

“…Each feature was assigned a p-value, providing a statistical basis for their inclusion in the downstream analyses (Supplementary Table 4). Notably, the small nuclear protein PRAC1 emerged as the top feature, indicating its potential role in the lateralization of CRC, which is a result consistent with many previous studies [22,23]. HOXB13, and the lncRNA ENSG00000242407 were also identified as important features.…”

“…To demonstrate the utility of pyRforest as an analytical tool, we have presented a case study of a dataset previously analyzed in Kolisnik et al (BMC Cancer, 2023) [22]. We identified 83 significant biomarkers, in contrast with 15 reported in the previous study [22]. This provides a more optimal number of biomarkers for statistical significance calculations in gene-set enrichment studies.…”

“…This is a considerably reduced set compared to the 1008 non-zero features that had non-zero importance in the default scikit-learn `RandomForestClassifier` model without any feature prioritization applied. The pyRforest approach offered a more robust method for feature identification compared to ‘sklearn.inspection’ which was extremely conservative, identifying only a single feature and missing several of the features known from other studies to affect CRC localization [22, 23]. Moreover, ‘sklearn.inspection’ had a significantly higher runtime than pyRforest .…”

“…In this case study, we employed pyRforest to analyze a dataset comprising RNA-seq data from colorectal cancer samples obtained from University of Otago, Christchurch, New Zealand analyzed in our previous study [22]. This dataset includes 308 patient samples, and encompasses a comprehensive range of 58,678 genomic features, including genes and long non-coding RNAs (lncRNAs).…”