The application discloses a targeting
superantigen fusion protein based on an improved SE(3)-
Transformer and an implementation method. In an offline stage,
amino acid sequences are first converted into one-hot encoding or
language model embedding (such as ESM-2), and are spliced with
multiple sequence alignment (MSA) features for geometric initialization. A neural network (improved SE(3)-
Transformer) combined with
multiple sequence alignment (MSA) and an attention mechanism is constructed to predict the coordinates of C alpha, C, N and O atoms for main chain prediction, and the neural network is trained through a
gradient descent method based on a physical
heuristic potential item. In a
verification stage, the improved SE(3)-
Transformer after training is used to generate a predicted structure, conformational stability is verified through a simplified force field, and
fine tuning is performed based on a
confidence score. The application can accurately predict the structure of a
target antigen epitope and an
antibody variable region, optimize a
superantigen functional domain in combination with a graph neural network, and dynamically design a flexible connecting
peptide to realize modular fusion.