Constrained Multi-objective Optimization-Based Temporal Network Observability for Biomarker Identification of Individual Patients

Identifying the biomarkers from the personalized gene interaction network of individual patients is important for disease diagnosis. However, existing methods not only ignore the prior biomarkers for practical use but also ignore the observability of the entire system. Therefore, this paper proposes a new constrained multi-objective optimization-based temporal network observability model (CMTNO) to identify biomarkers, which not only requires minimizing the number of selected nodes including ordinary nodes and prior nodes (the first optimization objective) but also maximizing the number of selected prior nodes (the second optimization objective) on the premise of ensuring network observability (the constraint condition). Considering the temporal feature of cancer (patients belong to different stages and each patient contains one task), an experience learning-based constrained multi-objective evolutionary algorithm is designed to solve the CMTNO problems. The selected probabilities of ordinary nodes and prior nodes are treated as experience, stored in two separate archives and updated by the optimal solutions on each task. Experience utilization refers to using two archives to generate new initial populations for new patients, in order to improve the optimization efficiency of the algorithm. Besides, a two-step neighbor-based connectivity method is proposed to distinguish different nodes with similar connectivity to further improve the effectiveness of archives. The proposed model and algorithm are evaluated on three kinds of cancer patients' data under two kinds of network models, and results show their effectiveness in identifying effective biomarkers.