Deep Reinforcement Learning for de-novo Drug Design Mariya Popova, Alexandr Isayev, Alexander Tropsha propose a novel computational strategy for de-novo design of molecules with desired properties called ReLeaSE (Reinforcement Learning for Structural Evolution). Based on deep and reinforcement learning, ReLeaSE integrates two deep neural networks – generative and predictive – trained separately but used together to generate novel chemical libraries. ReLeaSE uses only SMILES strings to represent molecules. Generative models are trained with stack-augmented memory networks to produce chemically feasible SMILES strings, while predictive models forecast desired properties of de novo compounds. In the first phase, models are trained separately with supervised learning, and in the second phase, they are trained jointly with reinforcement learning to bias generation towards desired properties. The method was tested to design libraries with bias toward structural complexity or specific physical properties, and to develop novel JAK2 inhibitors. The approach can generate targeted chemical libraries optimized for single or multiple properties. The method combines deep reinforcement learning with neural networks to design chemical compounds with desired physical, chemical, or bioactivity properties. Reinforcement learning is used to optimize target properties by training generative and predictive models. The generative model acts as an agent, while the predictive model acts as a critic, assigning rewards to generated molecules. The reward is based on the predictive model's forecast of properties. The method uses SMILES strings for molecule representation and integrates both phases into a single workflow with a reinforcement learning module. The generative model is a stack-augmented recurrent neural network (Stack-RNN) that outputs molecules in SMILES notation. It learns hidden rules of forming sequences of letters that correspond to valid SMILES strings. The predictive model is a deep neural network that estimates molecular properties. The method was tested to generate over one million compounds, with 95% being chemically valid. The model's synthetic accessibility score (SAS) was used to assess synthetic feasibility, with over 99.5% of generated molecules having SAS values below 6, indicating synthetic accessibility. The method was tested for property prediction, showing high accuracy for LogP and melting temperature. It was also used to design libraries with specific target properties, such as melting temperature, LogP, and JAK2 inhibition. The method was able to generate molecules with desired properties, including those with low or high activity. The method was also used to design libraries enriched with certain substructures, showing the ability to bias generation toward desired chemical features. The method was evaluated against traditional QSAR models, showing superior performance in property prediction and generation of novel compounds. The method's ability to generate chemically valid and synthetically accessible molecules with desired properties makes it a promising approach for de-novo drug design. The method's integration of deep reinforcement learning and neural networks allows for efficient and effective generation of targeted chemical libraries.Deep Reinforcement Learning for de-novo Drug Design Mariya Popova, Alexandr Isayev, Alexander Tropsha propose a novel computational strategy for de-novo design of molecules with desired properties called ReLeaSE (Reinforcement Learning for Structural Evolution). Based on deep and reinforcement learning, ReLeaSE integrates two deep neural networks – generative and predictive – trained separately but used together to generate novel chemical libraries. ReLeaSE uses only SMILES strings to represent molecules. Generative models are trained with stack-augmented memory networks to produce chemically feasible SMILES strings, while predictive models forecast desired properties of de novo compounds. In the first phase, models are trained separately with supervised learning, and in the second phase, they are trained jointly with reinforcement learning to bias generation towards desired properties. The method was tested to design libraries with bias toward structural complexity or specific physical properties, and to develop novel JAK2 inhibitors. The approach can generate targeted chemical libraries optimized for single or multiple properties. The method combines deep reinforcement learning with neural networks to design chemical compounds with desired physical, chemical, or bioactivity properties. Reinforcement learning is used to optimize target properties by training generative and predictive models. The generative model acts as an agent, while the predictive model acts as a critic, assigning rewards to generated molecules. The reward is based on the predictive model's forecast of properties. The method uses SMILES strings for molecule representation and integrates both phases into a single workflow with a reinforcement learning module. The generative model is a stack-augmented recurrent neural network (Stack-RNN) that outputs molecules in SMILES notation. It learns hidden rules of forming sequences of letters that correspond to valid SMILES strings. The predictive model is a deep neural network that estimates molecular properties. The method was tested to generate over one million compounds, with 95% being chemically valid. The model's synthetic accessibility score (SAS) was used to assess synthetic feasibility, with over 99.5% of generated molecules having SAS values below 6, indicating synthetic accessibility. The method was tested for property prediction, showing high accuracy for LogP and melting temperature. It was also used to design libraries with specific target properties, such as melting temperature, LogP, and JAK2 inhibition. The method was able to generate molecules with desired properties, including those with low or high activity. The method was also used to design libraries enriched with certain substructures, showing the ability to bias generation toward desired chemical features. The method was evaluated against traditional QSAR models, showing superior performance in property prediction and generation of novel compounds. The method's ability to generate chemically valid and synthetically accessible molecules with desired properties makes it a promising approach for de-novo drug design. The method's integration of deep reinforcement learning and neural networks allows for efficient and effective generation of targeted chemical libraries.