Prediction of COX-2 inhibitory activity using LSTM-network
Maxim N. Kachalkin, Tat'yana Konstantinovna Ryazanova, Irina V. Sokolova, Alexander V. Voronin · 2022
Generative recurrent networks are widely used for de novo drug discovery. We used long-short term memory (LSTM) architecture to obtain a model for predicting the quantitative characteristics of COX-2 inhibitory activity for small molecules. The best accuracy was achieved at epochs 32 and 34 for the training and test data sets (82 and 79%, respectively). To search for new potential candidates from the general data set, the top 10 substances with the lowest IC50values were selected. The value of IC50 is in the range from 0.006 - 0.25 nM. It is noted that the best value of IC50 in indole derivatives. The obtained data were generalized and template structures were built to generate 1000 substances using the RD-kit. Using the developed model, IC50 was calculated for each of them and the best candidates were determined. As a result, 9 compounds were revealed, in which the values were less than 0.2 nM.