Random Forest and XGBoost-based ensemble models for colorectal cancer exome variant classification and web application deployment for early prediction

Colorectal cancer (CRC) remains a significant global health challenge, necessitating advanced tools for improved diagnosis and personalized treatment strategies. The increasing availability of exome datasets presents an opportunity to leverage machine learning (ML) for identifying potential biomarkers and developing classification models for CRC. This study aims to harness these resources to create a user-friendly web application for the CRC community. Objectives: This research focused on developing a comprehensive AI-driven classification models for CRC by integrating mutation data and implementing a web application to assist clinicians, researchers, and diagnosticians. Key objectives included streamlining analysis protocols and creating an automated framework applicable to other genetic diseases. Publicly available CRC exome datasets from NCBI SRA were analyzed using a custom-built automated NGS pipeline. Feature engineering was performed to select relevant genomic variants. ML algorithms, including Random Forest and XGBoost, were employed for model building, training, and validation. Model performance was evaluated using F1-scores, ROC curves, and precision-recall curves. The best-performing models were then integrated into a Streamlit web application. Initial models (SVM and DNN) showed low accuracy and were not pursued further. Random Forest and XGBoost models demonstrated strong performance, achieving overall F1-scores of 0.93 and 0.92, respectively, in classifying CRC subtypes. Confusion matrices indicated minimal misclassifications, and ROC curves showed excellent ability to distinguish between classes. This study successfully developed a high-performing AI-driven classification models for CRC, leveraging exome datasets and machine learning. The resulting web application deployed on Posit Connect Cloud through Shiny Python provides a valuable resource for the CRC community, facilitating streamlined analysis, biomarker identification, and improved decision-making. The framework developed can be adapted for automated analysis of other genetic diseases, contributing to the advancement of precision diagnostics and personalized treatment strategies.