liveBook · Manning

Dual-stream architecture

As explained in Build AI Drug Discovery Pipelines

A dual-stream architecture is a model design that processes two related inputs through separate parallel computational streams before combining their representations.

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

How Build AI Drug Discovery Pipelines covers it

How deep it goes
Explains inputs, encoders, pooling, concatenation, and regression, but defers detailed GNN learning mechanisms.
Examples
Uses paired drug and protein graphs, separate graph embeddings, concatenation, and an MLP affinity output as its concrete workflow.
What it assumes
Assumes graph representations and GNN basics while relying on molecular and protein feature engineering for effective inputs.
What you can do after
Build it from scratch by implementing two graph encoders, pooling their outputs, and training a fusion MLP for affinity regression.
How it connects
Connects graph neural networks, message passing, global pooling, multimodal learning, transfer learning, and binding affinity.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Build AI Drug Discovery Pipelines. The section links point to the parts of the book it draws on.