Document resource
Background The efficacy of radiotherapy for rectal cancer (RC) depends heavily on tumor radiosensitivity, yet reliable predictive tools remain lacking in clinical practice. This study integrates metabolomic and radiomic data to develop a multimodal intelligent model, aiming to accurately predict radiosensitivity in RC patients and provide robust decision support for individualized treatment.Methods Patients with pathologically confirmed colorectal cancer (CRC) were retrospectively enrolled. The discovery cohort consisted of 186 serum samples and 131 radiomic datasets. Two independent validation cohorts included 91 and 62 patients, respectively. Untargeted metabolomics was performed using UPLC-MS/MS. Spearman correlation analysis was applied to eliminate features with high collinearity. Minimum redundancy maximum relevance (mRMR) and least absolute shrinkage and selection operator (LASSO) regression were utilized for dimensionality reduction and feature selection to identify the core radiomic signatures.Results A total of 1192 metabolites were detected. Comparisons among pathological response high-sensitivity (prHS), medium-sensitivity (prMS), and low-sensitivity (prLS) groups identified 34, 47, and 40 differential metabolites, respectively (P < 0.05) ( IDDF2026-ABS-0155 Figure 1. OPLS-DA plots and heatmaps illustrating the significantly differential metabolites among the prHS, prMS, and prLS groups). Pathway analysis revealed significant enrichment in pyrimidine and bile acid metabolism (IDDF2026-ABS-0155 Figure 2. Volcano plots of the top altered metabolites and pathway enrichment analysis for the distinct sensitivity groups). A 9-metabolite panel demonstrated robust diagnostic performance in the validation cohorts, yielding area under the curve (AUC) values of 77.20% and 74.90% for distinguishing prHS+prMS from prLS (IDDF2026-ABS-0155 Figure 3. ROC curves demonstrating the predictive performance of the metabolite model in the Guangzhou and Fujian validation cohorts). Furthermore, the constructed radiomic model achieved AUCs of 70.2% and 70.5% when differentiating prHS+prMS vs. prLS, and prMS vs. prLS, respectively (IDDF2026-ABS-0155 Figure 4. ROC curves evaluating the predictive efficacy of the clinical base model and the radiomic model).Conclusions Integrating metabolomic and radiomic features enables the construction of a multimodal model with strong predictive efficacy for RC radiosensitivity. This model holds significant clinical potential for guiding personalized radiotherapy strategies.Abstract IDDF2026-ABS-0155 Figure 1Abstract IDDF2026-ABS-0155 Figure 2Abstract IDDF2026-ABS-0155 Figure 3Abstract IDDF2026-ABS-0155 Figure 4