Hybrid Attentive Graph Neural Networks for time series gene expression clustering

Understanding gene expression patterns over time is crucial for unraveling the complexities of biological processes. An essential aspect of analyzing time series gene expression data involves the clustering of genes with similar responses, where the expression levels of genes are measured at specific time points. However, the main challenge in working with these time series data lies in uncovering concealed information embedded within their temporal structure. In this paper, we propose a novel Hybrid model that incorporates Attentive Graph Neural Networks, namely HAGNET, designed for clustering of time series gene expressions. In particular, HAGNET integrates both global patterns and temporal dependencies through a hybrid architecture of Graph Neural Networks (GNNs). Moreover, to account for the dynamic relationships of time series, HAGNET employs a custom temporal attention mechanism that filters each subsequent graph during the training process. The effectiveness of HAGNET is demonstrated on both synthetic and real-world biological datasets, outperforming several state-of-the-art methods. Finally, for reproduction purposes, we make the implementation of HAGNET publicly available at https://github.com/egiovanoudi/HAGNET . • HAGNET clusters genes based on time series gene expressions. • A hybrid architecture of GNNs combines global patterns and temporal dependencies. • A connectivity design with a custom temporal attention mechanism is proposed. • HAGNET outperforms state-of-the-art methods.