GATAEPR: A Graph Attention Autoencoder for Predicting Disease Progression Relationships in Cancer Patients: Predicting the Disease Progression in Cancer Patients

The genomic-level heterogeneity of cancer patients implies that the relationship of cancer progression may be not just a simple linear developmental pathway, but more likely to have a complex developmental network. However, few studies have focused on the prediction for cancer patients' disease progression. Therefore, we propose a novel framework based on graph neural networks to predict the disease progression relationships in cancer patients. Considering the genomic-level heterogeneity of cancer patients, we identified cancer-related gene sets to control noise and found that many of these genes were reported to be related with cancer. We identified similar cases by the k-nearest neighbors, and determined coarse-grained disease progression through available clinical cancer stages, treating the patient's neighbors as possible source or downstream of cancer development. To learn the complex disease progression relationships among cancer patients, we developed a link prediction model called GATAEPR, which includes an encoder with graph attention mechanism and a gravity-informed decoder. We evaluated the performance of GATAEPR from various evaluation metrics and compared GATAEPR with other existing methods, and found that GATAEP performed best for trend prediction of cancer progression in similar cases. We proposed to predict disease progression in similar cases, while non-similar cases have limited reference value. We are the first to predict disease progression for cancer patients based on deep graph neural networks, which is an innovative practice for deep learning in the field of cancer patient disease progression research. GATAEPR can improve the temporal granularity of clinical stages, and is expected to be a common data processing tool before constructing cancer disease development dynamics and bring inspiration to the research of complex disease development mechanism.