Methods, systems and non-transitory computer readable media for automated design of molecules with desired properties using artificial intelligence
The subject matter described herein includes computational methods, systems and non-transitory computer readable media for de-novo drug discovery, which is based on deep learning and reinforcement learning techniques. The subject matter described herein allows generating chemical compounds with desired properties. Two deep neural networks-generative and predictive, represent the general workflow. The process of training consists of two stages. During the first stage, both models are trained separately with supervised learning algorithms, and during the second stage, models are trained jointly with reinforcement learning approach. In this study, we conduct a computational experiment, which demonstrates the efficiency of proposed strategy to maximize, minimize or impose a desired range to a property. We also thoroughly evaluate our models with quantitative approaches and provide visualization and interpretation of internal representation vectors for both predictive and generative models.
1. A system for automated design of molecules using artificial intelligence, the system comprising:
a computing platform including at least one processor and a memory;
a generative model implemented by the at least one processor for utilizing a first artificial neural network trained in a first training stage to generate representations of valid molecules in a predetermined notation, wherein the first artificial neural network comprises a stack-augmented recurrent neural network trained to model valences of atoms and sequence dependencies including ring openings, ring closures, and bracket sequences to generate the representations of the valid molecules;
a predictive model implemented by the at least one processor and trained in the first training stage separately from the training of the generative model in the first training stage for predicting, from the representations of the valid molecules, numerical properties of the valid molecules; and
a reward function implemented by the at least one processor for generating reward values for the valid molecules, where the reward values are functions of the predicted numerical properties generated by the trained predictive model and a user-specified property, wherein the generative model and the predictive model are trained jointly in a second training stage utilizing the reward values to teach the first artificial neural network to output representations of the valid molecules having the user-specified property,
wherein, in the second training stage, the generative and predictive models are combined into a single reinforcement learning system in which the generative model operates as an agent whose action space is represented by a simplified molecular input line entry system (SMILES) notation alphabet and state space is represented by all possible strings in the SMILES notation alphabet and the predictive model operates as a critic which estimates behavior of the agent by assigning a reward value to each representation of one of the valid molecules generated by the generative model, and
wherein the predictive model includes an embedding layer, a long short term memory layer, and two dense layers.
2. The system of claim 1 wherein the first artificial neural network utilizes a deep learning method.
3. The system of claim 1 wherein the predictive model utilizes a second neural network designed to predict the properties of the valid molecules from the representations of the valid molecules in the predetermined notation.
4. The system of claim 1 wherein at least some of the valid molecules whose representations are generated by the generative model are novel chemical entities.
5. The system of claim 1 wherein the representations of valid molecules generated by the generative model represent molecules that can be synthesized.
6. The system of claim 1 wherein the generative model is configured to generate libraries of the representations of the valid molecules with desired profiles of properties.
7. The system of claim 1 wherein the user-specified property comprises a chemical property, a physical property, or a biological property.
8. The system of claim 7 wherein the user-specified property comprises a biological activity.
9. A method for automated design of molecules using artificial intelligence, the method comprising:
training, in a first training stage, a generative model comprising a first artificial neural network implemented by at least one processor to output representations of valid molecules in a predetermined notation, wherein the first artificial neural network comprises a stack-augmented recurrent neural network trained to model valences of atoms and sequence dependencies including ring openings, ring closures, and bracket sequences to generate the representations of the valid molecules;
training, in the first training stage and separately from the training of the generative model in the first training stage, a predictive model comprising a second neural network to predict properties of the valid molecules in the predetermined notation;
jointly training, in a second training stage, the generative model and the predictive model using reward values generated by a reward function which generates a reward value that is a function of a numerical property predicted by the trained predicted model to teach the first artificial neural network to output representations of the valid molecules having a user-specified property;
utilizing the trained generative model including the first artificial neural network to generate representations of valid molecules in the predetermined notation; and
utilizing the trained predictive model to predict properties of the valid molecules whose representations are output by the trained generative model,
wherein, in the second training stage, the generative and predictive models are combined into a single reinforcement learning system in which the generative model operates as an agent whose action space is represented by a simplified molecular input line entry system (SMILES) notation alphabet and state space is represented by all possible strings in the SMILES notation alphabet and the predictive model operates as a critic which estimates behavior of the agent by assigning a reward value to each representation of one of the valid molecules generated by the generative model, and
wherein the predictive model includes an embedding layer, a long short term memory layer, and two dense layers.
10. The method of claim 9 wherein the first artificial neural network utilizes a deep learning method.
11. The method of claim 9 wherein the predictive model utilizes a second neural network designed to predict the properties of the valid molecules from the representations of the valid molecules in the predetermined notation.
12. The method of claim 9 wherein at least some of the valid molecules whose representations are generated by the generative model are novel chemical entities.
13. The method of claim 9 wherein the user-specified property comprises a chemical property, a physical property, or a biological property.
14. The method of claim 13 wherein the user-specified property comprises a biological activity.
15. The method of claim 9 wherein the representations of the valid molecules generated by the generative model represent molecules that can be synthesized.
16. A non-transitory computer readable medium having stored thereon executable instructions that when executed by a processor of a computer control the computer to perform steps comprising:
training, in a first training stage, a generative model comprising a first artificial neural network to output valid molecules in a predetermined notation, wherein the first artificial neural network comprises a stack-augmented recurrent neural network trained to model valences of atoms and sequence dependencies including ring openings, ring closures, and bracket sequences to generate the representations of the valid molecules;
training, in the first training stage and separately from the training of the generative model in the first training stage, a predictive model comprising a second neural network to predict properties of molecules in the predetermined notation;
jointly training, in a second training stage, the generative model and the predictive model using reward values generated by a reward function which generates a reward value that is a function of a numerical property predicted by the trained predicted model to teach the first artificial neural network to output representations of the valid molecules having a user-specified property;
utilizing the trained generative model including the first artificial neural network to generate valid molecules in the predetermined notation; and
utilizing the trained predictive model to predict properties of the molecules output by the trained generative model,
wherein, in the second training stage, the generative and predictive models are combined into a single reinforcement learning system in which the generative model operates as an agent whose action space is represented by a simplified molecular input line entry system (SMILES) notation alphabet and state space is represented by all possible strings in the SMILES notation alphabet and the predictive model operates as a critic which estimates behavior of the agent by assigning a reward value to each representation of one of the valid molecules generated by the generative model, and
wherein the predictive model includes an embedding layer, a long short term memory layer, and two dense layers.