De Novo Molecular Generation Using Deep Learning for Prioritizing Synthesizability
Inamdar, Rashmi Jakhotiya, K. Jakhotiya · 2023
In this work, we introduce a novel method for identifying new drug candidates through de novo generation of molecules with desired properties. In this research we created a framework to generate specific molecules, which included three deep learning models: two generative Recurrent Neural Networks (RNNs), which learn the patterns in the SMILES syntax and were able to generate canonical SMILES strings; and a deep neural network to predict the effect of generated molecules on the target. One of the generative RNNs is then also taught through a process of Reinforcement Learning (RL) to generate molecules with the desired properties. The novelty of this paper is that we not only evaluate the affinity of the generative molecule against the target, but we also consider the synthetic accessibility of the drug to train the RNN. This allows for our pipeline to generate molecules which have our desired physicochemical properties while remaining simple.