MT-GNN: A Multi-Task Graph Neural Network for Robust Survival and Recurrence Prediction in Wilms' Tumor

Accurate prognosis prediction for Wilms' Tumor (WT), the most common pediatric renal malignancy, is crucial for tailoring personalized treatment strategies. However, developing robust computational models for WT remains a significant challenge due to the 'high-dimensional, low-sample-size' nature of pediatric cancer data. Existing data-driven methods often treat genes as independent features, ignoring complex biological interactions, which leads to overfitting when trained on limited samples. To address these challenges, we propose MT-GNN, a novel Multi-Task Graph Neural Network (GNN) framework that integrates biological prior knowledge to disentangle prognostic heterogeneity. Specifically, we construct patient-specific gene graphs guided by Protein-Protein Interaction (PPI) networks, enabling the model to capture non-Euclidean topological features of gene interactions. Furthermore, we introduce a multi-task learning mechanism that jointly optimizes survival analysis (main task) and recurrence prediction (auxiliary task). This mechanism acts as a strong regularizer, encouraging the learning of generalized gene representations that are robust to data scarcity. Our method achieves an AUC of 0.725 in recurrence prediction and a C-Index of 0.679 in survival analysis. Extensive experiments on real-world WT cohorts demonstrate that MT-GNN significantly outperforms state-of-the-art machine learning baselines (including XGBoost) and standard deep learning models (such as CNNs and LSTMs).