Graph neural networks (GNN) have seen significant growth recently for modeling temporal evolution in dynamic networks. Representation of complex networks in the form of graph data structures has enabled researchers to study how entities within these networks interact with each other. These interactions evolve over time. Developing a generic methodology for modeling this temporal evolution in complex networks for tracking evolving relationships has been a significant challenge. Most of the existing methods fail to extract contextual representations of historical neighborhood interactions for future link prediction. To address these challenges, this paper presents a novel method for modeling temporal evolution in complex networks using GNNs. A Context-Aware Graph Temporal Neural Network (CATGNN) method that uses dynamic neighborhood selection based on common neighbors for a given node is presented. The method uses dynamic neighborhood selection using contextual embeddings extracted from the historical interactions of the down-sampled set of neighbors of a central node based on a common neighborhood. Fixed-sized contextual memory modules are constructed for each node that store the historical interactions of its neighbors and are updated based on the recency and significance of interactions. The proposed method has been evaluated using six real-world datasets and has comparable performance against state-of-the-art methods, both in terms of accuracy and efficiency. It shows an improvement of 7.52 to 0.05% over the baselines in terms of average precision. The results demonstrate that the proposed CATGNN model can capture complex patterns of change that are difficult to identify using traditional techniques by propagating information over the graph structure. The model can be applied in various fields involving complex systems.

Clathrin protein (CP) plays a pivotal role in numerous cellular processes, including endocytosis, signal transduction, and neuronal function. Dysregulation of CP has been associated with a spectrum of diseases. Given its involvement in various cellular functions, CP has garnered significant attention for its potential applications in drug design and medicine, ranging from targeted drug delivery to addressing viral infections, neurological disorders, and cancer. The accurate identification of CP is crucial for unraveling its function and devising novel therapeutic strategies. Computational methods offer a rapid, cost-effective, and less labor-intensive alternative to traditional identification methods, making them especially appealing for high-throughput screening. This paper introduces CL-Pred, a novel computational method for CP identification. CL-Pred leverages three feature descriptors: Dipeptide Deviation from Expected Mean (DDE), Bigram Position Specific Scoring Matrix (BiPSSM), and Position Specific Scoring Matrix-Tetra Slice-Discrete Cosine Transform (PSSM-TS-DCT). The model is trained using three classifiers: Support Vector Machine (SVM), Extremely Randomized Tree (ERT), and Light eXtreme Gradient Boosting (LiXGB). Notably, the LiXGB-based model achieves outstanding performance, demonstrating accuracies of 94.63% and 93.65% on the training and testing datasets, respectively. The proposed CL-Pred method is poised to significantly advance our comprehension of clathrin-mediated endocytosis, cellular physiology, and disease pathogenesis. Furthermore, it holds promise for identifying potential drug targets across a spectrum of diseases.