A multimodal deep learning framework for predicting sunitinib response in advanced clear cell renal cell carcinoma

Abstract Despite sunitinib remaining the cornerstone of systemic therapy for advanced renal cell carcinoma (RCC), reliable predictive biomarkers for therapeutic response remain elusive. We retrospectively analyzed multi‐modal data (complete across four modalities) from 157 clear cell RCC (ccRCC) patients receiving sunitinib monotherapy. The cohort was randomly partitioned into training (70%) and testing (30%) sets. Seven deep learning architectures (DenseNet201, ResNet18, ResNet152, ResNet34, Inception_v3, ShuffleNet_v2, and MobileNet_v3) were deployed to identify imaging biomarkers. Deep transfer learning features derived from ResNet34 were integrated with genomic variants, proteomic profiles, and clinical parameters to develop a multimodal predictive model—termed the CGPR (Clinical, Genomic, Proteomic, and Radiomic) model. Model interpretability was enhanced via SHapley Additive exPlanations (SHAP) to quantify feature contribution weights. Single‐modality‐based models exhibited suboptimal performance, with average area under the receiver operating characteristic curves (AUCs) below 0.8 and significant discrepancies between training and testing sets. Radiomics‐based deep learning models also showed limited efficacy (ResNet34: mean AUC = 0.785). The multimodal CGPR model achieved superior accuracy (mean AUC = 0.946). SHAP analysis identified the deep transfer learning features DL_0 and DL_60, alongside the proteomic marker SRGAP1 (SLIT‐ROBO Rho GTPase Activating Protein 1), as the three most influential predictors. This study demonstrates the transformative potential of multimodal integration for predicting sunitinib response in advanced ccRCC. By harmonizing deep learning signatures (ResNet34‐derived), proteomic biomarkers (SRGAP1), and clinical parameters, the CGPR model surpassed conventional single‐modality approaches in predictive accuracy.