Abstract
In this work, we introduce QuanGAT, a hybrid framework integrating quantum neural networks (QNNs), graph attention networks (GATs), and federated learning to tackle DNA mutation prediction in biomedical graphs in a privacy-preserving, noise-aware setting. Motivated by quantum biology's potential to describe mutation-level mechanisms, our approach is able to simulate decentralized genomic environments in order to protect sensitive data, while capturing complex biological variability. QuanGAT employs a QNN encoder based on parameterized quantum circuits, incorporating quantum noise, via the so-called depolarizing and amplitude damping channels. Furthermore, in order to enforce differential privacy, QuanGAT utilizes an attention-based architecture enhanced with Laplace noise. We evaluated the model on mutation-enriched protein-protein interaction networks, partitioned across simulated clients under centralized, federated, and noisy scenarios. In particular, we employed QuanGAT in three different datasets, namely PPI, STRING, and OBGN-Proteins. Especially in federated learning settings, QuanGAT consistently outperformed certain state-of-the-art graph neural networks up to 4.5 % in terms of accuracy and up to 6.3 % in terms of macro F1 score. Our findings demonstrate that integrating QNN encodings with attention-based graph learning may improve DNA mutation prediction in decentralized, privacy-aware settings.