Fig. 7: Performance of the 10MG model for distinguishing thyroid cancer differentiation states.

10 MG ResNet classifier accurately stratifies thyroid tumors across five differentiation states with clear SHAP-based biomarker insights. a Schematic representation of ResNet model development and validation for distinguishing the differentiation states of thyroid cancer. This model demonstrated high accuracy and robustness in distinguishing thyroid cancer differentiation states. b Pie chart depicting the composition of the training and test sets integrated from the FUSCC and public GSE cohorts, with transcriptomic data processed for batch effect removal. The cohort included 453 subjects: normal controls (n = 139), PTC patients (n = 179), FTC patients (n = 27), PDTC patients (n = 44), and ATC patients (n = 64). Sample origins were annotated for the PDTC and ATC subgroups. c Bar chart showing the average SHAP values for the top 10 metabolic genes contributing to the ResNet model’s predictions. The most influential genes included GPX3, TPO, CYP1B1, and PDE8B, reflecting their significant roles in differentiating thyroid cancer states. d Embedding plot displaying sample clustering on the basis of expression levels of the top two SHAP-ranked features. e Bee schematic illustrating the contribution levels of 10MG genes across differentiation states, quantified through SHAP value analysis. f ROC curve for the test cohort. The AUROC ranged from 0.87 to 0.96, demonstrating the model’s robustness in distinguishing differentiation states. g Bar plots demonstrating the contribution levels of 10MG genes across differentiation states, quantified by SHAP values.