A merged molecular representation learning for molecular properties prediction with a web-based service

Deep learning has brought a dramatic development in molecular property prediction which is crucial in the field of drug discovery. Various methods such as fingerprint, SMILES, graphs have been proposed for representing molecules. Recently, unlabeled molecule data is used to improve performance for various pre-training methods. The main challenge of molecular properties predictions is designing a data representation and model that can show good performance for various datasets. However, performance deviation due to scarcity of dataset exists in constructing the model. We propose a new self-supervised method to learn the characteristics and structures of molecules by integrating existing methods. The key of our model is learning structures with matrix embedding and learning logics that can infer descriptors via QED prediction. As a result, our method improves the generalization of the data and achieves the best average performance by benchmarking downstream tasks. Moreover, we develop a web-based fine-tuning service to utilize our model on various tasks.