Fig. 1

The overview of model architecture. Chemical structure, gene expression and genetic mutation serve as input of the model. Graph neural network encoded the drugs and convolutional neural network extracted gene expression and genetic mutation features simultaneously. Through self-attention, we incorporated the chemical similarity into the input to train gene weight layer \(W'\). Gene weight layer \(W'\) combined gene expression, genetic mutation and drug similarity. Then the model concatenated drug vector and genetic vector to predict the \(IC_{50}\)