Fig. 1

The framework of PMF-CPI. For a compound-protein pair, the protein sequence is transformed into context embedding via pretrained TAPE and then fed into LSTM. And a compound SMILES is turned into a molecular graph by RDKit and encoded with graph neural networks. After that, two encoding vectors are merged through the Kronecker product, and MLP provides a final output. PMF-CPI can process two task types, regression for CPI binding affinities or classification for interactions