Fig. 2: Embedding chemical space for generative discovery of polypharmacology drugs.
From: De novo generation of multi-target compounds using deep generative chemistry
a Use of a variational autoencoder (VAE) to create an embedded representation of chemical structure (middle), where input ChEMBL compounds (left) are encoded and decoded with separate deep neural networks. Here the embedding is approximated in two dimensions (Methods). Example embedded compounds are shown (points), along with their Murcko scaffold classification (colors). b MEK1 or mTOR-targeting compounds (blue or purple points) in the chemical embedding space, approximated in two dimensions as for (a). c Reinforcement learning strategy for de novo generation of compounds recognizing two targets. Compounds are sampled from the chemical embedding (top) and scored by the predicted inhibitory concentration against each target (pIC50) alongside a panel of metrics to assess synthesizability and drug likeness (middle). Top-scoring compounds are used to refocus the chemical embedding for progressive epochs of compound sampling (descending arrows, bottom). Source data are provided as a Source Data file.
