Fig. 7

Overview of the proposed iDNA-ABF. A shows the DNA methylation dataset collection where different datasets belonging to three main DNA methylation types are reorganized into their training datasets and independent datasets. The overall architecture of our iDNA-ABF is presented in B–E. B Multi-scale information processing module, exploiting two scales (3-mer and 6-mer) of tokenizers separately to process the input sequence and adaptively obtain corresponding embeddings. C BERT encode module, using BERT encoders to extract high-latent feature representations. D Multi-scale extraction module, generating final output feature representations based on multi-scale embeddings. E Classification module, integrating binary classification probability values to make prediction. F The workflow of the interpretable analysis. In brief, our model uses attention mechanisms to extract and learn sequential motifs from query sequences