Automated community detection of biomedical composite networks across network embedding and dynamic optimization

The integration of multi-source, heterogeneous biomedical data - such as cross-hospital electronic health records (EHRs) and multi-omics databases - is pivotal for understanding disease mechanisms and analyzing patient populations. However, current methods face considerable challenges in addressing the heterogeneity across different network entities, the ambiguity of community boundaries, and the strong reliance on manual intervention. These limitations hinder the comprehensive analysis of complex biomedical relationships. To overcome these issues, we propose the LesNet framework, which introduces the first fully automated community detection method tailored for biomedical composite networks, with a focus on improving both efficiency and accuracy in data integration and analysis.

LesNet addresses the problem of data heterogeneity through a cross-network dynamic alignment technique. This method bridges different types of entities (patient IDs), and jointly embeds both topological structures (patient-medication interaction networks) and semantic features (clinical text-medication function associations). The framework employs a self-supervised reinforcement learning approach for community detection. It uses known biomedical modules as seed inputs and generates discriminative embeddings through contrastive learning. The dynamic optimization of community boundaries is achieved through reinforcement learning, simulating expert decision-making processes. This approach allows for the integration of phenotype-associated patients while removing non-relative entity, ensuring that the identified patient communities are biologically relevant.

The LesNet framework was validated using simulated cross-hospital multi-omics data, demonstrating significant improvements in both accuracy and efficiency. The framework achieved a 15% increase in the F1 score for disease-related module detection compared to baseline methods. Moreover, its automated process exceeded manual workflows in filtering efficiency by 80%. These results highlight the effectiveness of LesNet in reducing manual efforts while maintaining high performance. The proposed method offers a scalable and automated tool for multi-source data analysis in precision medicine, with potential applications in cancer subtype classification, drug group recommendation, and other areas of biomedical research. Ultimately, LesNet presents a promising approach for advancing precision medicine by addressing key challenges in multi-source biomedical data analysis.