Fig. 1

A summary of the two main versions of the architecture we used. First, the two-branch architecture (A) encodes the explicit feature representations that are available for the compounds and proteins. The resulting embeddings can be aggregated with one of the three strategies (MLP, dot product, tensor product). The second version of the architecture combines both the implicit and explicit information available for the compounds and proteins (B). The explicit features can be encoded in the exact same manner as shown in (A), while the one-hot encoded dummy vectors for the compounds and proteins can be transformed into dense representations using a single fully-connected layer. The intermediate embeddings from the explicit and implicit features are aggregated using the MLP strategy. This output of the MLPs is the compound and protein embeddings that can then be aggregated with any of the three available strategies