Data Sheet 2_EPI-DynFusion: enhancer-promoter interaction prediction model based on sequence features and dynamic fusion mechanisms.docx

IntroductionEnhancer–promoter interactions (EPIs) play a vital role in the regulation of gene expression. Although traditional wet-lab methods provide valuable insights into EPIs, they are often constrained by high costs and limited scalability. As a result, the development of efficient computational models has become essential. However, many current deep learning and machine learning approaches utilize simplistic feature fusion strategies, such as direct averaging or concatenation, which fail to effectively model complex relationships and dynamic importance across features. This often results in suboptimal performance in challenging biological contexts.
MethodsTo address these limitations, we propose a deep learning model named EPI-DynFusion. This model begins by encoding DNA sequences using pre-trained DNA embeddings and extracting local features through convolutional neural networks (CNNs). It then integrates a Transformer and Bidirectional Gated Recurrent Unit (BiGRU) architecture with a Dynamic Feature Fusion mechanism to adaptively learn deep dependencies among features. Furthermore, we incorporate the Convolutional Block Attention Module (CBAM) to enhance the model’s ability to focus on informative regions. Based on this core architecture, we develop two variants: EPI-DynFusion-gen, a general model, and EPI-DynFusion-best, a fine-tuned version for cell line–specific data.
ResultsWe evaluated the performance of our models across six benchmark cell lines. The average area under the receiver operating characteristic curve (AUROC) scores achieved by the specific, generic, and best models were 94.8%, 95.0%, and 96.2%, respectively. The average area under the precision-recall curve (AUPR) scores were 81.2%, 71.1%, and 83.3%, respectively, demonstrating the superior performance of the fine-tuned model in the precision-recall space. These results confirm that the proposed fusion strategies and attention mechanisms contribute to significant improvements in performance.
DiscussionIn conclusion, EPI-DynFusion presents a robust and scalable framework for predicting enhancer–promoter interactions solely based on DNA sequence information. By addressing the limitations of conventional fusion techniques and incorporating attention mechanisms alongside sequence modeling, our method achieves state-of-the-art performance while enhancing the interpretability and generalizability of enhancer–promoter interaction prediction tasks.