An integrated oral-gut microbiome and clinical feature model predicts myelosuppression risk after neoadjuvant chemoradiotherapy in colorectal cancer

Abstract Background Colorectal cancer (CRC) is a leading global malignancy, with neoadjuvant chemoradiotherapy (nCRT) as standard care for locally advanced rectal cancer. Although gut bacteria are linked to CRC progression and nCRT response, current studies predominantly focus on genus-level bacterial profiling (via 16S rRNA sequencing), neglecting non-bacterial microbes and oral microbiota. Emerging evidence implicates oral microbiota in radiotherapy toxicity. While existing models rely on unimodal gut microbiome data, integrating oral/gut microbiomes and clinical indicators through artificial intelligence (AI) may enhance myelosuppression risk prediction. We aimed to develop a multimodal AI model to predict post-nCRT myelosuppression in CRC patients for personalized intervention. Methods Using Python 3.9.19 and Anaconda 4.14.0, we engineered classifiers for myelosuppression prediction. After 5-fold cross-validation, near 18,000 oral/gut microbial features were filtered via LASSO and Random Forest, with redundancy reduced by RFECV. Seven algorithms underwent grid-search hyperparameter tuning across 35 train-test splits (7 repeats of 5-fold CV). Features consistently selected (≥50% splits) were retained. The final model was optimized by area under curve, accuracy, precision, recall, and F1-score on an independent validation cohort. Results The integrated model demonstrated robust performance: Multiple classifiers achieved AUC=1.00 (internal validation). Six classifiers attained AUC>0.60, with the Gradient Boosting Machine (GBM) model showing superior myelosuppression prediction. Conclusion We established an AI model integrating multimodal biomarkers to accurately predict myelosuppression risk post-nCRT in CRC, providing a clinical tool for toxicity mitigation. (Acknowledgements: This study was supported by National Natural Science Foundation of China, Grant No. 32300085 to Zhenhui Chen, Grant No. 32370139 to Hongying Fan; Guangdong Basic and Applied Basic Research Foundation, Grant No. 2025A1515010567 to Zhenhui Chen)