Protein-targeted drug compound identification

A system for identifying drug compounds for targeting proteins in tissue cells, the system comprising:

• a neural network model which comprises an attention-based protein encoder, pretrained in an autoencoder architecture to encode an input protein sequence into an output vector in a latent space representing proteins, and a molecular decoder which is pretrained in an autoencoder architecture to generate compound data, defining a compound molecule, from an input vector in a latent space representing molecules, wherein the protein encoder and molecular decoder are coupled such that the input vector of the molecular decoder is dependent on the output vector of the protein encoder for an input protein sequence;
• and
• control logic adapted to train the model in a reinforcement learning architecture in which reward values dependent on affinity of compound molecules, defined by compound data generated by the molecular decoder for respective protein sequences encoded by the protein encoder, to proteins corresponding to those sequences are used to progressively train the model to optimize the reward value for compound data generated thereby;
• wherein the control logic is further adapted, in response to receipt of a protein sequence for a target protein after training the model, to supply the protein sequence to the model to generate compound data corresponding to a set of drug compounds for targeting that protein, to match one or more existing drug compounds to a given patient using the set of drug compounds, to select a treatment for the given patient based on the matching and to facilitate the treatment of the given patient using the selected treatment.

A computer program product for identifying drug compounds for targeting proteins in tissue cells, said computer program product comprising a non-transitory computer readable storage medium having program instructions embodied therein, the program instructions being executable by a computing system to cause the computing system to:

• implement a neural network model which comprises an attention-based protein encoder, pretrained in an autoencoder architecture to encode an input protein sequence into an output vector in a latent space representing proteins, and a molecular decoder which is pretrained in an autoencoder architecture to generate compound data, defining a compound molecule, from an input vector in a latent space representing molecules, wherein the protein encoder and molecular decoder are coupled such that the input vector of the molecular decoder is dependent on the output vector of the protein encoder for an input protein sequence;
• train the model in a reinforcement learning architecture in which reward values dependent on affinity of compound molecules, defined by compound data generated by the molecular decoder for respective protein sequences encoded by the protein encoder, to proteins corresponding to those sequences are used to progressively train the model to optimize the reward value for compound data generated thereby;
• after training the model, supply a protein sequence for a target protein to the model to generate compound data corresponding to a set of drug compounds for targeting that protein;
• match one or more existing drug compounds to a given patient using the set of drug compounds;
• select a treatment for the given patient based on the matching;
• and
• facilitate treatment of the given patient using the selected treatment.