[4C3-GS-10] AI application: prediction 2

[4C3-GS-10-02] Interaction Prediction between Glycoproteins and Drugs with Encoding of Glycan Using Pre-trained Language Model

〇Eiji Shinkawa¹, Kouichi Nagatsuka¹, Yuki Murata¹, Tamiko Ono¹, Masae Hosoda¹, Kiyoko Kinoshita¹, Masayasu Atsumi¹ (1. Soka University)

Keywords:Neural network, Glycan, Interaction prediction, Transformer, Pre-trained

In a previous study (Mutual Attention Neural Network), the authors confirmed that the prediction accuracy of protein-drug interaction was improved by using the frequency vector of glycans that modify proteins, but the problem remained that the structural information of glycans was not utilized for interaction prediction. In this study, we propose a new encoding method for glycan structure series data using a pre-training language model to demonstrate the usefulness of glycan structure information and pre-training in predicting glycoprotein-drug interactions. A mutual attention neural network incorporating this new glycan encoder is developed and compared with the previous study's model. As a result, we confirmed the improvement of prediction accuracy compared with the previous research model, and showed that the use of glycan structure information and prior learning is useful for predicting the interaction between glycoproteins and drugs.