The misinformation that is spreading in digital media demands the development of robust automated detection systems that can understand semantic content and contextual metadata. This paper introduces a novel multi-modal graph neural network framework combining BERT-based semantic embeddings with metadata-driven graph construction for improved fake news detection. Our approach builds a heterogeneous graph that incorporates features such as speaker credibility, political affiliation, and topical similarity, besides adopting community detection as a post-hoc analysis technique for exploring misinformation clusters. The proposed method achieves accuracy of 76.9%, F1-score that is 0.744, and recall-0.778 on the PolitiFact dataset, outperforming strong baseline models such as BERT+Linear and TF-IDF+SVM. We adopt a GraphSAGE-based relational multi-modal approach, fusing BERT embeddings with metadata-derived representations. Interpretability is established through community detection and ablation studies that highlight feature contributions. Overall, the proposed framework offers a well-balanced trade-off between performance and transparency, contributing to more reliable mis-information detection.