PTMFusionNet: A Deep Learning Approach for Predicting Disease Related Post-translational Modification and Classifying Disease Subtypes

With the advancement of technologies such as mass spectrometry, it has become possible to simultaneously perform large-scale detection of protein intensity and corresponding post-translational modification (PTM) information, thereby facilitating clinical diagnosis and treatment. However, existing PTM information is insufficient to fully integrate with protein expression data. We propose a deep learning method called PTMFusionNet, which predicts potential disease-related PTMs and integrates them with protein expression data to classify disease subtypes. PTMFusionNet includes two Graph Convolutional Network (GCN) models: the Layer-Attention Graph Convolutional Network (LAGCN) and the Feature Weighting Graph Convolutional Network (FWGCN). LAGCN is used to predict PTM potentiality scores, while FWGCN integrates these scores with protein expression data for disease subtype classification. Experimental results across three datasets (KIPAN, COADREAD, and THCA) demonstrate that PTMFusionNet outperforms benchmark algorithms in accuracy, F1 score, and AUC, highlighting its robustness in identifying critical PTM biomarkers and advancing disease subtyping.