What is gne in networking?
Any node can serve as a Gateway Network Element (GNE). A node becomes a GNE when a TL1 user connects to it and enters a command destined for another node. The node receiving the TL1 message from another node for processing is called a End-Point Network Element (ENE).
Although a wide variety of machine learning models have been developed to analyze high-throughput datasets for GI prediction , there are still some major challenges, such as efficient analysis of large heterogeneous datasets, integration of biological information, and effective feature engineering. To address these challenges, we propose a novel deep learning framework to integrate diverse biological information for GI network inference.
Our proposed method frames GI network inference as a problem of network embedding. In particular, we represent gene interactions as a network of genes and their interactions and create a deep learning framework to automatically learn an informative representation which integrates both the topological property and the gene expression property. A key insight behind our gene network embedding method is the “guilt by association” assumption , that is, genes that are co-localized or have similar topological roles in the interaction network are likely to be functionally correlated. This insight not only allows us to discover similar genes and proteins but also to infer the properties of unknown ones. Our network embedding generates a lower-dimensional vector representation of the gene topological characteristics. The relationships between genes including higher-order topological properties are captured by the distances between genes in the embedding space. The new low-dimensional representation of a GI network can be used for various downstream tasks, such as gene function prediction, gene interaction prediction, and gene ontology reconstruction .
Furthermore, since the network embedding method can only preserve the topological properties of a GI network, and fails to generalize for genes with no interaction information, our scalable deep learning method also integrates heterogeneous gene information, such as expression data from high throughput technologies, into the GI network inference. Our method projects genes with similar attributes closer to each other in the embedding space, even if they may not have similar topological properties. The results show that by integrating additional gene information in the network embedding process, the prediction performance is improved significantly.
GI prediction is a long-standing problem. The proposed machine learning methods include statistical correlation, mutual information , dimensionality reduction , and network-based methods (e.g. common neighborhood, network embedding) . Among these methods, some methods such as statistical correlation and mutual information consider only gene expression whereas other methods use only topological properties to predict GIs.
Network-based methods have been proposed to leverage topological properties of GI networks . Neighborhood-based methods quantify the proximity between genes, based on common neighbors in GI network . The proximity scores assigned to a pair of genes rely on the number of neighbors that the pair has in common. Adjacency matrix, representing the interaction network, or proximity matrix, obtained from neighborhood-based methods, are processed with network embedding methods to learn embeddings that preserve the structural properties of the network. Structure-preserving network embedding methods such as Isomap are proposed as a dimensionality reduction technique. Since the goal of these methods is solely for graph reconstruction, the embedding space may not be suitable for GI network inference. Besides, these methods construct the graphs from the data features where proximity between genes is well defined in the original feature space . On the other hand, in GI networks, gene proximities are not explicitly defined, and they depend on the specific analytic tasks and application scenarios.
