The problem of IC50 prediction for ligand-protein pairs using Transformer architecture under limited resources

• Дата: 2026
• Автори: Krysenko, Pavlo, Bektimirov, Alim
• Формат: Стаття
• Мова: Українська
• Опубліковано: Kyiv National University of Construction and Architecture 2026
• Теми: deep learning; ligand; protein; bioinformatics; convolutional neural networks; data compression
• Онлайн доступ: https://es-journal.in.ua/article/view/365080
• Назва журналу: Environmental safety and natural resources

The article discusses approaches to optimizing the training of models for predicting the half-maximal inhibitory concentration (IC50) of ligand-protein pairs under limited computational resources. A method of smart bucketing of data by protein length with a dynamic selection of the number of groups to improve randomization is proposed. To solve the problem of the quadratic complexity of the Transformer architecture, a convolution layer was used to compress the input data. Based on 4 conducted experiments, the relationship between the degree of sequence compression and the obtained root mean square error (RMSE) for lgIC50 was analyzed.