Dynamic Graph Neural Network Learning for Temporal Omics Data Prediction

Abstract

High-throughput studies of biological systems are rapidly generating a wealth of 'omics'-scale data. Many of these studies are temporal collecting proteomics and genomics data capturing dynamic observations. While temporal omics data are essential to unravel the mechanisms of various diseases, they often include missing (or incomplete) values due to technical and experimental reasons. Data prediction methods, i.e., imputation and forecasting, have been widely used to mitigate these issues. However, existing imputation and forecasting techniques typically address static omics data representing a single time point and perform forecasting on data with complete values. In this paper, we propose a graph-based method for temporal omics data imputation and forecasting that handle omics data containing missing values at multiple time points. The method takes advantage of topological relationships (e.g., protein-protein and gene-gene interactions) among omics data samples and incorporates a graph convolutional network to first infer the missing values at different time points. Then, we combine these inferred values with the original omics data to perform temporal imputation and forecasting using a long short-term memory network. Evaluating the proposed method on two real-world datasets demonstrated a distinct advantage over existing data imputation and forecasting methods. On the omics dataset, the average mean square error of our method improved 11.3% for imputation and 6.4% for forecasting compared to the baseline methods.INDEX TERMS Temporal biological data, missing values, data prediction, graph neural networks

Classification of Cancer Microarray Data Based on Deep Learning: A Review

“…GNNs belong to the realm of deep learning algorithms tailored for the examination and interpretation of structured data encapsulated within graphs. Graphs, comprising interconnected nodes and edges, serve as versatile models to depict relationships and dynamics across diverse domains like social networks, biological systems, citation networks, and recommendation frameworks [23]. The primary objective of GNNs is to acquire nuanced representations of individual nodes within a graph [24]).…”

“…The journey continues with node embeddings, where initial embeddings are assigned to nodes based on their features. GNNs engage in iterative messagepassing steps, allowing nodes to gather information from their neighbors and update their embeddings accordingly; this involves the exchange of neighborhood information and using learnable aggregation functions such as mean, sum, or attention mechanisms [23]. Stacked aggregation layers further refine node embeddings by assimilating information from increasingly expansive neighborhoods to encapsulate local and global graph structures; some employ graph pooling layers, contributing to hierarchical representation [24].…”

A Comprehensive Bibliometric Analysis of Missing Value Imputation

“…To impute and forecast temporal omics data with missing values at various time points, Jing et al proposed a graphbased method that uses topological links across datasets and a graph convolutional network [134]. The integrated system developed by Jena and Dehuri (2022) combines the strengths of rule-based inductive DT and Support Vector Machine (SVM) classifiers, referred to as DT-SVM, to impute missing values and predict the class label of unseen samples [135].…”