A membrane protein function prediction method fusing spatial structure weighting and hierarchical constraint

Abstract

The application discloses a membrane protein function prediction method fusing spatial structure weighting and hierarchical constraint, and relates to the technical field of biological information, and comprises the following steps: obtaining an amino acid sequence and a corresponding weighted spatial contact graph of a membrane protein to be predicted; a pre-trained protein language model extracts sequence features of the amino acid sequence; a structure perception graph attention module performs structure enhancement on the sequence features based on the weighted spatial contact graph, so as to obtain structure enhanced features; the sequence features and the structure enhanced features are fused, and protein sample representation is generated. Through fusion of the weighted spatial contact graph and the structure perception graph attention module, the representation capability of the membrane protein structure information is enhanced; through cross-attention and a gating mechanism, semantic alignment of protein residue features and GO terms is realized; through a hierarchical consistency loss function, the hierarchical relationship of prediction probability is constrained, and the consistency of the prediction result and the GO ontology logic is improved.

Images