Abstract
Background: Predicting protein-DNA binding sites in vivo is a challenging but urgent task in many fields such as drug design and development. Most promoters contain many transcription factor (TF) binding sites, but only a small number of sites have been identified by time-consuming biochemical experiments. To address this challenge, numerous computational approaches have been proposed to predict TF binding sites from DNA sequences. However, current deep learning methods often face issues such as gradient vanishing as the model depth increases, leading to suboptimal feature extraction.
Results: We propose a model called CRA-KAN (where C stands for convolutional neural network, R stands for recurrent neural network, and A stands for attention mechanism) to predict transcription factor binding sites. This hybrid deep neural network incorporates the KAN network to replace the traditional multi-layer perceptron, combines convolutional neural networks with bidirectional long short-term memory (BiLSTM) networks, and utilizes an attention mechanism to focus on DNA sequence regions with transcription factor binding motifs. Residual connections are introduced to facilitate optimization by learning residuals between network layers. Testing on 50 common ChIP-seq benchmark datasets shows that CRA-KAN outperforms other state-of-the-art methods like DeepBind, DanQ, DeepD2V, and DeepSEA in predicting TF binding sites.
Conclusions: The CRA-KAN model significantly improves prediction accuracy for transcription factor binding sites by effectively integrating multiple neural network architectures and mechanisms. This approach not only enhances feature extraction but also stabilizes training and boosts generalization capabilities. The promising results on multiple key performance indicators demonstrate the potential of CRA-KAN in bioinformatics applications.