SSCGO

This model takes protein sequences as input, retrieves functional description texts from databases, and uses AlphaFold2 to predict protein structures. It generates position-specific scoring matrices (PSSMs) via PSI-BLAST, encodes them, and uses them as node features in a cross-domain long-range topological network to enable the propagation of spatial constraint features. It encodes protein secondary structures and original sequences and maps them to a two-dimensional space, then inputs them into the Omni-Scale Spatial Topological Module (OSTM) to capture long-range interactions. Functional description text is encoded using BioBERT, and local and long-range features are captured using a cascaded multi-scale perception and bidirectional extended short-term memory network (BiLSTM). Finally, a weighted coupling self-evolution strategy is used to deeply aggregate multi-modal features.