Abstract <p>Modern machine learning models leveraging multi-omics data face significant privacy challenges due to the sensitive nature of patient information. Communication overhead and missing features in each institution can lead to a substantial decline in federated learning performance. In response to these concerns, we propose VFLING—Federated Learning for Multi-Omics Data Integration with Graphs—a secure one-shot communication federated learning framework. We note that medical data reflects disease characteristics from different omics, while the relationships between samples exhibit relative stability across these omics. To minimize data transmission while maximizing the utilization of each participant’s feature information we develop a strategy that transmits not only the local features but also the relationships or topology in one-shot communication. By fusing the omics based on the locally learned graph structure instead of features, VFLING enables improved performance even when some features are missing from individual parties. Extensive experiments demonstrate that VFLING outperforms existing frameworks, paving the way for applications in the medical field. Local features and graph topology are shared to the trainable server in a single communication, enhancing model accuracy through integrated data. This approach improves robustness despite incomplete information. Local Parties and Server Integration: Local parties learn and transmit both local features and graph topology to the server in a single communication, maximizing effective information transfer. The trainable server then integrates data across parties using graph relationships, enhancing model robustness and accuracy despite incomplete feature sets.</p> Graphical Abstract <p></p>