The paper "Augmented Memory: Sample-Efficient Generative Molecular Design with Reinforcement Learning" by Jeff Guo and Philippe Schwaller addresses the challenge of sample efficiency in de novo molecular design. The authors propose a novel algorithm called Augmented Memory, which combines data augmentation with experience replay to improve the performance of molecular generative models. The key contributions of the paper include: 1. **Augmented Memory Algorithm**: This algorithm significantly enhances sample efficiency by updating the model multiple times using scores obtained from oracle calls. It combines experience replay with SMILES augmentation to bias the model towards high-rewarding molecules. 2. **Selective Memory Purge**: This heuristic is introduced to prevent mode collapse by removing entries in the replay buffer that are discouraged based on chemical scaffolds. 3. **Performance Evaluation**: The method is evaluated on the Practical Molecular Optimization (PMO) benchmark, drug discovery case studies, and materials design case studies. Augmented Memory outperforms existing algorithms in terms of sample efficiency and diversity. 4. **Drug Discovery Case Study**: The method is applied to generate potential dopamine type 2 receptor (DRD2) inhibitors, demonstrating its ability to perform multiparameter optimization (MPO) and generate molecules with improved docking scores. 5. **Materials Design Case Study**: Augmented Memory is extended to optimize quantum-mechanical properties, showing its versatility in generating molecules with desired properties. The paper highlights the importance of experience replay in policy-based reinforcement learning for molecular design and provides a robust framework for sample-efficient de novo molecular design. The code for the method is available at <https://github.com/schwaller-group/augmented_memory>.The paper "Augmented Memory: Sample-Efficient Generative Molecular Design with Reinforcement Learning" by Jeff Guo and Philippe Schwaller addresses the challenge of sample efficiency in de novo molecular design. The authors propose a novel algorithm called Augmented Memory, which combines data augmentation with experience replay to improve the performance of molecular generative models. The key contributions of the paper include: 1. **Augmented Memory Algorithm**: This algorithm significantly enhances sample efficiency by updating the model multiple times using scores obtained from oracle calls. It combines experience replay with SMILES augmentation to bias the model towards high-rewarding molecules. 2. **Selective Memory Purge**: This heuristic is introduced to prevent mode collapse by removing entries in the replay buffer that are discouraged based on chemical scaffolds. 3. **Performance Evaluation**: The method is evaluated on the Practical Molecular Optimization (PMO) benchmark, drug discovery case studies, and materials design case studies. Augmented Memory outperforms existing algorithms in terms of sample efficiency and diversity. 4. **Drug Discovery Case Study**: The method is applied to generate potential dopamine type 2 receptor (DRD2) inhibitors, demonstrating its ability to perform multiparameter optimization (MPO) and generate molecules with improved docking scores. 5. **Materials Design Case Study**: Augmented Memory is extended to optimize quantum-mechanical properties, showing its versatility in generating molecules with desired properties. The paper highlights the importance of experience replay in policy-based reinforcement learning for molecular design and provides a robust framework for sample-efficient de novo molecular design. The code for the method is available at <https://github.com/schwaller-group/augmented_memory>.