Skip to main content
. 2022 Nov 29;13:1067562. doi: 10.3389/fgene.2022.1067562

FIGURE 3.

FIGURE 3

The detailed model architecture of PromBERT for promoter prediction. DNA sequences of fixed length are tokenized using a custom 1bp tokenizer and fed into 144 attention modules. Based on the final tensors in the pooling layer, the classifier calculates the probabilities of promoter and non-promoter using the sigmoid function. Backpropagation was conducted using binary cross entropy loss.