Machine learning integration of tissue-specific metagenomic signatures for colorectal cancer diagnosis.

Abstract

Colorectal cancer (CRC) represents a significant global health burden. Leveraging machine learning (ML) with metagenomic and tissue-specific data presents new opportunities for improving diagnostic accuracy and understanding the microbiome's role in CRC. This study was conducted to enhance diagnostic efficiency and identify crucial bacterial biomarkers in CRC using various ML models applied to metagenomic data. A total of 33 samples were analyzed, comprising 20 healthy controls and 13 CRC patients. Each sample included demographic data (age, gender) and bacterial information (Bacteroides, Enterococcus, Faecalibacterium, Proteobacteria, Gammaproteobacteria, Firmicutes, Enterobacteriaceae, Clostridia). Six models: Logistic Regression, Naive Bayes, Decision Tree, Support Vector Machine (SVM) with both linear and polynomial kernels and Multilayer Perceptron (MLP) were employed. Performance was evaluated using leave-one-out cross-validation (LOOCV). To address the class imbalance, F1-score was utilized as the primary metric for feature selection. A consensus-based feature elimination strategy, where bacterial features were iteratively removed only if their exclusion improved or maintained the F1-score across the majority of the models was implemented. For the MLP, a grid search was integrated into each iteration to optimize hidden layer architectures and solvers, thereby ensuring that robust performance was achieved for each feature subset. The analysis was conducted using a 10-feature initial set consisting of 2 demographic and 8 microbial features. Model performances were optimized through a consensus-based feature elimination strategy, and it was determined that diagnostic success increased with the exclusion of the Faecalibacterium, Age, and Enterobacteriaceae features during the process. The highest performance was achieved with the SVM model with Linear kernel when Bacteroides was excluded from the 9-feature subset (Table 4), reaching an accuracy of 87.88% and an F1-score of 83.33%. Within the final biomarker set, Enterococcus and Firmicutes were identified as the most critical predictive features due to the sharpest declines in F1-score observed in their absence. This study demonstrates that the systematic elimination of initial clinical and metagenomic features maximizes CRC diagnostic accuracy and model stability. The process, initiated with a 10-feature baseline set was subsequently refined to establish a high-precision diagnostic mechanism with an F1-score of 83.33%. The identified final microbial signatures, consisting of 5-6 taxa, provide a clinically applicable, non-invasive diagnostic foundation with low input requirements.