EPI-DynFusion: Enhancer-Promoter Interaction Prediction Model Based on Sequence Features and Dynamic Fusion Mechanisms

Ao Zhang¹Jianhua Jia^1*Mingwei Sun¹Xin Wei²

• ¹School of Information Engineering, Jingdezhen Ceramic Institute, Jingdezhen, China
• ²Business School, Jiangxi Institute of Fashion Technology, Nanchang, 330044, China, Nanchang, Jiangxi Province, China

Enhancer-promoter interactions (EPIs) are critical for regulating gene expression. While traditional wet-lab techniques for detecting EPI provide valuable experimental data, they are often limited by cost-related challenges. Thus, the development of more efficient computational methods is essential for enhancing our understanding of EPI mechanisms. Although various computational approaches have been introduced to tackle this issue, many existing deep learning and machine learning methods rely on simplistic feature averaging or concatenation for feature fusion. These methods fail to adequately capture the complex relationships among different features, leading to inflexibility and an inability to adaptively prioritize varying levels of importance among features. Consequently, this can result in suboptimal model performance, particularly in complex and dynamic data scenarios. To mitigate the issue of inflexible feature fusion in current methods, we propose a deep learning model named EPI-DynFusion. This model first encodes sequences using pre-trained DNA vectors and extracts local features through convolutional neural networks (CNN). It then innovatively integrates Transformer and Bidirectional Gated Recurrent Unit (BiGRU) architectures with dynamic feature fusion to adaptively capture deep feature associations. Additionally, the model incorporates the Convolutional Block Attention Module (CBAM) to enhance feature extraction accuracy. Based on this architecture, we developed EPI-DynFusion-gen with improved generalization capability and EPI-DynFusion-best with superior performance through dataset repartitioning and model optimization.Based on the analysis of experimental results obtained from six benchmark cell lines, we found that the average area under the curve (AUROC) for the specific model, the generic model, and the best model was 94.8%, 95%, and 96.2%, respectively, while the average exact recall (AUPR) of these models exhibited variable performance, with values of 81.2% for the specific model, 71.1% for the generic model, and 83.2% for the best model. In conclusion, our proposed EPI-DynFusion method, which predicts enhancer-promoter interactions based solely on DNA sequences, demonstrates state-of-the-art performance in this field.