Kidney renal clear cell carcinoma (KIRC) accounts for approximately 80% of all renal cell carcinoma cases. Despite surgery remaining the primary treatment, the 30% postoperative recurrence rate and poor long-term outcomes highlight the urgent need for better prognostic biomarkers beyond current imaging-based methods.
Cuproptosis is a newly discovered form of programmed cell death triggered by copper ion overload. Unlike apoptosis, necrosis, or ferroptosis, cuproptosis acts through copper-dependent toxic oxidative stress in the tricarboxylic acid cycle. High expression of the copper transporter CTR1 has been independently linked to reduced overall survival in KIRC, establishing cuproptosis-related genes as candidate prognostic markers.
Traditional prognostic models for KIRC rely on Cox proportional hazards regression, which captures only linear relationships between biomarkers and outcomes. Neural networks, by contrast, can model nonlinear interactions in high-dimensional gene expression data, potentially offering superior predictive performance for complex diseases like KIRC.
The researchers began with 12 cuproptosis-related genes identified from the literature and performed Wilcoxon differential expression testing on TCGA-KIRC data (532 tumor samples), selecting 10 significantly differentially expressed genes as inputs to their prognostic model. External validation used ICGC-KIRC data (518 samples).
The deep neural network (DNN) comprised one input layer, three hidden layers, and one output layer, trained in TensorFlow with a Tanh activation function and stochastic gradient descent optimizer. Hybrid L1 and L2 regularization with dropout prevented overfitting, and Bayesian hyperparameter optimization tuned the learning rate, regularization weights, and dropout rate.
For drug-target prediction, the researchers screened 251 anticancer drugs by comparing estimated IC50 values between high-risk and low-risk KIRC patient groups. Four sensitive drugs were selected and connected as new nodes to a KIRC drug-target graph (1,250 nodes, 405,226 edges) built from DrugBank data. GraphSAGE, an inductive graph neural network, was chosen to predict which targets each drug could interact with, even for drug nodes not seen during training.
The DNN model achieved an AUC of 0.739 on the TCGA training set and 0.707 on the independent ICGC validation set, outperforming the traditional Cox proportional hazards model at 1-, 3-, and 5-year prediction time points in both datasets. High-risk and low-risk groups defined by the optimal cutoff showed highly significant survival differences (p less than 0.001).
Risk grouping correlated significantly with multiple clinical features including tumor stage, grade, T stage, and M stage (all p less than 0.001), and the DNN model retained independent prognostic ability in both univariate and multivariate Cox analyses. The model also performed well regardless of whether patients had early-stage (I-II) or late-stage (III-IV) disease.
A combined analysis of DNN risk score and tumor mutation burden (TMB) showed that patients classified as both high-risk and high-TMB had the shortest survival, while the risk score contributed greater prognostic weight than TMB alone. A nomogram model integrating DNN risk score with age, stage, grade, T stage, and M stage provided individualized survival probability estimates with good calibration across the first, third, and fifth years.
Sensitivity analysis of 251 anticancer drugs identified four agents with significantly lower estimated IC50 values in high-risk KIRC patients compared to low-risk patients: 5-fluorouracil, enzastaurin, bexarotene, and A-443654. Cell viability and colony formation assays on the 786-O and Caki-1 KIRC cell lines confirmed that all four drugs inhibited proliferation and clonogenicity in vitro, with 5-fluorouracil showing the strongest inhibitory effect.
GraphSAGE predicted drug-target interactions for all four drugs against the KIRC target network. The top-ranked predicted targets included Cytochrome P450 2D6, UDP-glucuronosyltransferase 1A, and Proto-oncogene tyrosine-protein kinase receptor Ret. All seven target types appearing in the top 15 predicted interaction nodes were directly linked to KIRC, supporting the biological plausibility of the predictions.
Four specific predicted drug-target interactions were corroborated by published literature, including the inhibitory effect of 5-fluorouracil on Cytochrome P450 2D6 and UDP-glucuronosyltransferase 1A, and bexarotene's co-regulation of the RET oncogene, providing external validation that GraphSAGE was capturing real pharmacological relationships.
GraphSAGE (Graph Sample and Aggregate) was chosen over transductive graph neural networks like GCN because it learns aggregation functions rather than fixed node embeddings, enabling it to generate predictions for entirely new nodes (the four newly identified drugs) not seen during training. This inductive capability is essential for drug discovery applications where new compounds must be evaluated without retraining the entire model.
The GraphSAGE model aggregated neighbor node features using element-wise averaging across three graph convolutional layers, building progressively larger neighborhoods to capture structural context across the drug-target interaction graph. Training for 300 epochs yielded the best performance, with an average test set accuracy of 0.817 plus or minus 0.013, outperforming both GAT (Graph Attention Network) and GTN (Graph Transformer Network) comparison models.
The KIRC drug-target graph was constructed from 15 drugs and 341 targets (11 known KIRC drugs and 4 newly identified, plus 48 known KIRC targets supplemented with 293 structurally similar targets), resulting in 1,250 nodes connected by over 405,000 edges. This dense graph provided sufficient neighborhood context for GraphSAGE to learn meaningful drug-target interaction patterns.
The combined DNN prognosis model and GraphSAGE drug-target predictor form a two-stage decision support tool: the DNN identifies which patients are at high risk of poor outcomes, and GraphSAGE predicts which molecular targets those patients' tumors are most vulnerable to given the drugs found to be most effective in high-risk patients.
The cuproptosis gene framework is biologically grounded, as copper metabolism is genuinely dysregulated in KIRC. High expression of CTR1, one of the cuproptosis-related transporters, independently predicts poor survival, suggesting that the 10-gene signature captures a real metabolic vulnerability in these tumors rather than an incidental statistical correlation.
Current clinical practice still relies primarily on TNM staging for treatment decisions. The nomogram model developed in this study provides a more granular survival estimate by integrating molecular risk with clinical features, offering a practical tool for postoperative counseling and for stratifying patients into surveillance versus treatment intensification protocols.
A central limitation is that the drug sensitivity predictions were based on estimated IC50 values derived from bioinformatics tools rather than direct drug response measurements in KIRC patient-derived models. The cell line experiments confirmed antiproliferative effects but did not test all predicted drug-target interactions in KIRC-specific cellular contexts.
The model was validated on TCGA and ICGC datasets, which are large but may not fully represent the clinical diversity of KIRC patients. The two datasets required separate normalization due to batch effects from different sequencing platforms, and prospective clinical validation in an independent patient cohort with standardized treatment data would be necessary before clinical adoption.
The drug-target predictions from GraphSAGE for most of the four drugs remain novel and unconfirmed, and the most interesting predicted targets, those not yet documented in prior literature, represent hypotheses requiring experimental validation through biochemical binding assays and mechanistic in vivo studies in kidney cancer models.