Construction and Validation of Multi-Omics Predictive Models for Colorectal Cancer Using Machine-Learning Approaches

Area of Science:

• Oncology
• Bioinformatics
• Machine Learning in Medicine

Background:

• Colorectal cancer (CRC) management requires accurate prediction of key biomarkers.
• Tumor mutational burden (TMB), microsatellite instability (MSI), and specific gene mutations (NTRK/PIK3CA) are critical for treatment selection.
• Predicting these markers using only routine clinicopathological variables remains a challenge.

Purpose of the Study:

• To develop and validate a multi-omics nomogram for predicting TMB, MSI, NTRK/PIK3CA mutation status, and overall survival (OS) in CRC.
• To utilize routine clinicopathological variables for accurate and accessible biomarker prediction.
• To compare the performance of different machine learning algorithms for predictive modeling.

Main Methods:

• Utilized TCGA data (n=398) for training and a prospective cohort (n=120) for external validation.
• Employed feature selection techniques including LASSO regression to identify optimal clinicopathological variables.
• Compared four machine learning algorithms (LR, SVM, DT, RF) using AUC, F1 score, and decision-curve analysis, with the best model validated and calibrated.

Main Results:

• The Random Forest (RF) method achieved the highest predictive power for TMB (AUC=0.9597) and MSI (AUC=0.8225).
• RF also demonstrated strong performance in predicting NTRK and PIK3CA gene status using TMB and MSI indicators.
• The developed models showed high accuracy in predicting key molecular features of colorectal cancer.

Conclusions:

• A robust predictive model using routine clinicopathological variables for TMB, MSI, and gene mutation status in CRC was successfully constructed and validated.
• This model can assist clinicians in identifying high-risk patients and informing treatment strategies.
• Further research with larger sample sizes is recommended to optimize and confirm the model's clinical utility.