Fig. 2

Summary diagrams of the piRNA target prediction model. a The graph outlines the basic framework of our piRNA targeting prediction model, comprising five essential components. Commencing with the Embedding Layer, it transforms base sequences into numerical representations for subsequent processing. Then, the Feature Extraction Layer harnesses the power of 1D-CNN (One-Dimensional Convolutional Neural Network) to extract potential features from both piRNA and mRNA sequences, capturing their inherent characteristics for the next step. Moving forward, the Interaction Layer simulates the interaction between piRNA and mRNA, incorporating attention mechanisms to dynamically assess and emphasize the binding affinity for each base in the piRNA sequence. Following this, the Global Dominant Base-Pair Matching Extraction Layer leverages attention mechanisms again to identify globally significant base-pair matching patterns that are crucial for piRNA targeting. Finally, in the binary classification layer, the model generates piRNA targeting predictions. Notably, during the pre-training phase, the entire network undergoes concerted training. Conversely, in the fine-tuning stage, the initial three components of the model are strategically frozen, permitting a focused refinement of the Global Dominant Base-Pair Matching Extraction Layer and the classification layer. The pre-training and fine-tuning strategy leverage the shared patterns identified across C. elegans and mice, while subsequently refining the model to better suit the intricate specifics of the mammalian task. b Detailed structure of the model including values of parameters used