This paper introduces a method for molecular de novo design using deep reinforcement learning (RL). The approach involves training a sequence-based generative model to learn how to generate molecular structures with specific desirable properties by using augmented episodic likelihood. The model is demonstrated to perform tasks such as generating analogues to a query structure and generating compounds predicted to be active against a biological target. As a proof of principle, the model is first trained to generate molecules without sulphur. It is then trained to generate analogues to the drug Celecoxib, and finally, it is tuned to generate compounds predicted to be active against the dopamine receptor type 2 (DRD2), achieving over 95% predicted activity. The method uses a recurrent neural network (RNN) trained on the ChEMBL database. The RNN is fine-tuned using a policy-based RL approach, where the goal is to maximize the expected return by learning an augmented episodic likelihood that combines prior likelihood and a user-defined scoring function. This approach allows the model to generate structures that are more likely to be active, including experimentally confirmed actives not included in the training data. The study compares the performance of the model with traditional policy gradient methods and shows that the RL approach achieves significant improvements. The model is also tested on tasks such as generating structures similar to a query structure and generating compounds with predicted biological activity. The results indicate that the model can generate a high fraction of predicted actives, including structures that are not part of the training data. The paper highlights the potential of using RL for molecular de novo design, demonstrating that the approach can generate structures with desired properties and improve the efficiency of drug discovery. The method is shown to be robust to hyperparameters such as σ and learning rate, and can be extended to incorporate multiple parameters for more complex drug discovery tasks. The study provides a framework for using RL in molecular design and shows that it can be a valuable tool for drug discovery.This paper introduces a method for molecular de novo design using deep reinforcement learning (RL). The approach involves training a sequence-based generative model to learn how to generate molecular structures with specific desirable properties by using augmented episodic likelihood. The model is demonstrated to perform tasks such as generating analogues to a query structure and generating compounds predicted to be active against a biological target. As a proof of principle, the model is first trained to generate molecules without sulphur. It is then trained to generate analogues to the drug Celecoxib, and finally, it is tuned to generate compounds predicted to be active against the dopamine receptor type 2 (DRD2), achieving over 95% predicted activity. The method uses a recurrent neural network (RNN) trained on the ChEMBL database. The RNN is fine-tuned using a policy-based RL approach, where the goal is to maximize the expected return by learning an augmented episodic likelihood that combines prior likelihood and a user-defined scoring function. This approach allows the model to generate structures that are more likely to be active, including experimentally confirmed actives not included in the training data. The study compares the performance of the model with traditional policy gradient methods and shows that the RL approach achieves significant improvements. The model is also tested on tasks such as generating structures similar to a query structure and generating compounds with predicted biological activity. The results indicate that the model can generate a high fraction of predicted actives, including structures that are not part of the training data. The paper highlights the potential of using RL for molecular de novo design, demonstrating that the approach can generate structures with desired properties and improve the efficiency of drug discovery. The method is shown to be robust to hyperparameters such as σ and learning rate, and can be extended to incorporate multiple parameters for more complex drug discovery tasks. The study provides a framework for using RL in molecular design and shows that it can be a valuable tool for drug discovery.