Figure 2.

TSignal architecture consists of ProtBERT encoder and multi-head attention-based transformer decoder. Initial embedding involves a dense representation and a linear positional encoding. ProtBERT embeddings are concatenated with organism group and positional representations prior to using them as key and value vectors in the decoder, together with query vector from dense representations of the sequence labels. See Section 2.1 for an in-depth description of the TSignal model.