A dual-stream architecture processes two related inputs through separate parallel neural-network streams, then fuses their representations for one prediction. In drug-target affinity prediction, one stream specializes in the drug graph and the other in the protein graph. section 11.5 section 11.5.1
A typical dual-stream graph pipeline contains these stages:
- Accept a drug graph and a protein graph as a paired input. section 11.5.1
- Pass each graph through its own stack of GCN, GAT, or GIN convolutional layers. section 11.5.1
- Apply global mean or max pooling to convert each variable-sized graph into a fixed-size embedding. section 11.5.1
- Concatenate the drug and protein embeddings and use an MLP to regress a scalar binding-affinity prediction. section 11.5.1
Separate processing lets each encoder learn features suited to its modality: molecular patterns for drugs and sequence or structural patterns for proteins. It also supports modifying one stream for new information, such as 3D drug conformations or experimental protein structures, while inspecting intermediate representations for interpretation. section 11.5.1
