DMPNN-Bert: a deep learning architecture for molecular property prediction

Abstract

Abstract: Molecular property prediction is a fundamental research problem in the fields of drug discovery, chemical synthesis prediction. To establish a universal molecular property prediction model, this study proposed six molecular properties prediction models. For capture molecular features, this study combines the representational ability of molecular graphs and the advantage of attention mechanism. Based on three different molecular graph representation of MPNN, DMPNN, dyMPN, to combine two different kinds of deep learning algorithm with the attention mechanism of Transformer and Bert. The results were compared with MPNN and DMPNN. The evaluation indexes of ROC-AUC, RMSE and MAE are applied in this paper. Ten benchmark datasets were used to test the performance of eight models. The results based on the proposed DMPNN combine Bert (DMPNN-Bert) achieves in seven of ten benchmark datasets, which illustrate that the prediction performance of the proposed model.