From SMILES to graphs: molecule representation in deep learning to predict chemical reactions

The study of molecular reactions is a critical aspect of chemistry and drug design, as it allows for the prediction of the interactions between molecules and their potential outcomes. With advancements in deep learning, there has been a growing effort to leverage these techniques to predict molecular reactions and accelerate the discovery process. However, while many models exist, they often utilize SMILES strings to represent molecules, leading to a loss of important structural information. In this master's thesis, we explore the use of graph representation of molecules as an alternative to SMILES strings in the context of deep learning models for reaction prediction. By adapting an existing model to work with graph representation, we aim to determine if this approach can lead to improved results compared to those obtained with SMILES strings. Overall, the goal of this study is to advance the understanding of how different molecular representations can impact the performance of deep learning models in reaction prediction and contribute to the ongoing efforts in accelerating the molecular discovery process. Through rigorous experimentation and analysis, we hope to draw conclusions about the benefits of using graph representation for molecular reaction prediction and provide insights for future research in this field. Unfortunately, our experiments did not meet our expectations due to the model's poor performance. Our findings suggest that our model was not capable of correctly predicting the entire product molecules. Additionally, the lack of time and computational resources hindered proper model selection and training. Despite the unsatisfactory results, this study provides insights into how existing models can be adapted to work with graph representations in deep learning models for reaction prediction, emphasizing the importance of time and computational resources when dealing with these data structures.