Hidden Fingerprints: AI-Powered Comparative Microbiome Analysis Reveals Shared and Cancer-Specific Bacterial Signatures in Colorectal, Breast, and Prostate Cancer

Early detection of colorectal, breast, and prostate cancer remains challenging due to non-specific early symptoms and limited access to affordable screenings in diverse communities. This project explores a machine learning approach to identify hidden microbial signatures across cancer microenvironments using public genomic datasets from TCMA and NCBI BioProjects. CLR-transformed microbial abundance profiles were used to train models under two evaluation designs: binary intra-organ models comparing cancer versus healthy tissue, and multi-class models comparing all three cancer types. The binary comparisons, which hold sample type constant and isolate the effect of cancer status alone, are the more clinically meaningful test of the hypothesis. They revealed a divergent picture: breast cancer was distinguished from healthy adjacent tissue at a modest AUC of 0.6262, while colorectal cancer showed no detectable signal (AUC 0.487, chance level), aligning with statistical testing showing zero significantly differentially abundant genera in colorectal cancer versus 22 in breast cancer. By contrast, multi-class models, using SVM RBF, Random Forest, XGBoost, and L1-Logistic Regression with SMOTE applied strictly within training folds, achieved striking performance, with SVM reaching 99.27% cross-validation accuracy and a macro-AUC of 1.0000. SHAP values mirrored Random Forest Gini rankings, confirming Mahella, Blautia, [Ruminococcus], Dorea, Intestinibacillus, and Lachnospira as top predictive genera. However, this near-perfect performance is driven by baseline sample-type differences across stool, tissue, and urine rather than cancer-specific dysbiosis, and should not be mistaken for a universal cancer-microbiome signature. This highlights both the potential and limitations of microbiome-based screening.