Attention-based Feature Fusion for Reconstructing Gene-Regulatory Interactions

Reconstructing gene regulatory networks (GRNs) from expression data is vital for understanding gene transcrip- tion. Although increasingly advanced algorithms, particularly deep learning models, are proposed to mine potential gene regulatory interactions, insufficient effort has been invested in improving the reliability of features in the presence of biological variability between expression samples. In this research, we propose a robust feature fusion method inspired by emerging attention-based techniques in computer vision. Our method can capture the functional asymmetry between transcription factors (TF) and target genes with an important adaptation using differentiated attention heads. Based on three different gene expression datasets: in silico, E.coli, and S.cerevisiae, we demonstrate that our method is superior to other state-of-the- art competitors. The overall GRN reconstruction performance of our method yields a gain of 3%-14% in the AUC score over the competing models.