Developing a thyroid cancer differentiation state classification system using deep residual networks and metabolic signature profiling.

We developed a deep residual network (ResNet) framework to classify thyroid cancer differentiation states by integrating multiomic data and interpretability analysis. Our framework incorporated untargeted metabolomic, whole-exome sequencing, and transcriptomic data from 158 thyroid tumors and 57 matched normal tissues, encompassing well-differentiated, poorly differentiated, and anaplastic thyroid cancers. We further examined single-cell RNA sequencing datasets from the Gene Expression Omnibus (GEO) to map key metabolic reprogramming pathways in dedifferentiated thyroid cancer. We systematically integrated transcriptomic data from follicular epithelial-derived thyroid carcinomas across all GEO cohorts, establishing a pan-pathological classification model based on a 10-gene metabolic signature. To complement this approach, a 10-metabolite model was developed via the ResNet architecture, capitalizing on the direct pathophysiological responsiveness of metabolites to tumor progression states. By employing Shapley additive explanations, we highlighted critical metabolic signatures driving differentiation states. Our findings reveal how metabolic shifts underpin thyroid cancer progression, and based on these findings, we propose an accurate, interpretable model that may facilitate early diagnosis and inform clinical decision-making.