WaveDNA:Wavelet-Based Encoding Enables Transfer Learning from Vision Models for Transcription Factor Binding Site Prediction

Transcription factors (TFs) regulate gene expression by binding to short, specific DNA sequences known as transcription factor binding sites (TFBSs). Accurate identification of TFBSs is fundamental for understanding transcriptional regulation. Deep learning (DL) has become a strong choice for modeling TF–DNA binding thanks to its ability to capture complex non-linear patterns and hierarchical dependencies. Here, we introduce a simple yet powerful technique that encodes DNA sequences as two-dimensional wavelet images, facilitating direct transfer of pretrained image-based convolutional neural networks. We call this approach WaveDNA, a lightweight and interpretable DL framework that enables efficient transfer learning without relying on large-scale genomic pretraining. Across diverse ENCODE ChIP-seq datasets spanning multiple TFs, we show that WaveDNA enables pretrained vision models to achieve competitive perofmance on TFBS prediction tasks while using roughly fivefold fewer parameters and substantially less compute. In addition, representing sequences as images makes it straightforward to apply established computer-vision interpretability tools to visualize learned binding patterns. Overall, WaveDNA provides a practical and interpretable way to model TF–DNA interactions with good accuracy and low computational cost.