3D2SMILES: Translating Physical Molecular Models into Digital DeepSMILES Notations Using Deep Learning

Physical molecular models are widely used in educational settings for teaching organic and other branches of chemistry, offering an intuitive way of understanding molecular structures. Conversely, virtual models, while less intuitive, provide additional functionalities such as the ability to retrieve molecular names and other properties. Currently, to the best of our knowledge, there is a gap between 3D molecular models and their digital counterparts. This paper introduces a computer vision model designed to bridge this gap by converting images of physical molecular models into their digital DeepSMILES representations. This conversion facilitates further information retrieval, enhancing educational utility. We developed both synthetic and real datasets to train our model and evaluated its performance across various dataset combinations, model architectures, and dataset sizes. Additionally, we attempted to improve the model's accuracy by multi-image input and beam search. We achieved 62.0\% top-1 accuracy and 80.3\% top-3 accuracy with beam search and multi-image input on our validation set. We also explored the model's characteristics, such as explainability by saliency maps, error analysis, and examined its calibration. We also discussed the model's limitations and directions for future research.