Fig. 5

Classifier architecture. a The LSTM model architecture. The protein sequences were converted into fixed-size vectors and subsequently passed through an embedding layer with a length of 128. This was followed by a 1D convolutional layer comprising 64 filters and a subsequent 1D max pooling layer. Next, an LSTM layer with 100 units was implemented, followed by a final classification layer that employed a sigmoid function. b The attention model architecture. The attention model replaced the LSTM layer of the LSTM model with an attention layer, while the remaining modules remained unchanged. c The ProtT5 model architecture. The protein sequences were initially fed into the ProtT5 Layer, followed by an MLP Layer comprised of two fully connected layers with a hidden size of 100. Just like the LSTM and attention method, the final step used a sigmoid function for classification