The integration of heterogeneous multi-omics data remains a critical challenge in computational biology, essential for unraveling the complex molecular underpinnings of diseases. Existing methods often struggle to incorporate prior biological knowledge effectively or provide interpretable results. We present BioMGNN v2, an end-to-end, supervised deep learning framework designed to address these challenges by extending our previous BioMGNN architecture. BioMGNN v2 constructs a heterogeneous biomedical knowledge graph that embeds patient-specific multi-omics data within a network of signed, directed biological interactions. It then leverages a heterogeneous graph neural network with relation-specific attention and integrates a complementary patient–patient graph branch, enabling robust representation learning across both biological entities and clinical samples. Self-supervised pretraining with contrastive and masked modeling objectives further enhances generalization, while gated cross-omics fusion and multi-task learning improve classification accuracy and biological relevance. To ensure interpretability, BioMGNN v2 incorporates graph explainers with stability selection to identify reproducible and statistically rigorous biomarker modules, complemented by uncertainty estimation for clinical reliability. We demonstrate the superior performance of our algorithm in patient classification and its ability to uncover biologically plausible and interpretable biomarkers, positioning it as a powerful and trustworthy tool for advancing precision medicine.