-
We have implemented a novel computational strategy for design of drug molecules with the desired properties.The model implements a deep learning and reinforcement learning approaches.
-
The model integrates two different deep neural networks - generator and discriminator with reinforcement learning.
-
These two models are trained seperately but are used jointly to generate novel targeted chemical libraries.
-
Our model uses string representation of molecules by their simplified molecule-input line entry-system(SMILES).
-
In this system, the generative model is used to produce novel chemically feasible molecules, that is, it plays a role of an agent, whereas the discriminative model (that predicts the properties of novel compounds) plays the role of a critic, which estimates the agent’s behavior by assigning a numerical reward (or penalty) to every generated molecule.
-
The reward is a function of the numerical property generated by the predictive model, and the generative model is trained to maximize the expected reward.
-
In the first phase of the method, generative and discriminator models are trained separately with a supervised learning algorithm.
Yellow colored vioilin plot determines druglikeliness of the base dataset whereas the other determines the same for the generated data by GAN.
-
In the second phase, both models are trained jointly with the RL approach to bias the generation of new chemical structures toward those with the desired physical and/or biological properties.
-
In a nutshell we have used GAN model integrated with RL to design chemical molecules with specific range of physical properties, such as solubility, novelty, synthesizability.