Fig. 1

Diagram of the model for feature extraction from sequence information. (1) The encoder model transforms amino acid sequences into sequences of vector embeddings. (2) The similarity prediction module utilizes pairs of proteins represented by their vector embedding sequences to predict their shared structural classification of proteins (SCOP) level. Sequence alignment is performed based on the L1 distance between their vector embeddings, employing the sequence-structure alignment (SSA)method. Subsequently, a similarity score is computed from the alignment and linked to shared SCOP levels through ordinal regression. (3) The contact prediction module leverages the vector embedding sequence to predict contacts between amino acid positions within each protein. The contact loss is determined by comparing these predictions with observed contacts in the protein’s 3D structure. The parameters of the encoder are adjusted by utilizing error signals from both tasks