Drug screening to inhibit serine/threonineproteinkinase mTOR using LSTM: A descriptor free QSARapproach

Intorduction: Cancer is a widespread disease that has significantmortality rate all over the world. Mammalian targetof rapamycin (mTOR) is a crucial signaling node that is oftendysregulated in human cancer. Therefore, we decided to focuson predicting the inhibitory activity of molecules againstmTOR kinase by artificial intelligence (AI), which could aidin the development of cancer treatments. AI has transformedmany industries including pharmaceuticals. Moreover, Machinelearning and deep learning have been game-changers indrug discovery, allowing researchers to analyze large data sets and detect patterns. This has led to the discovery of new compounds.AI has also made industries more efficient and productiveby automating processes and analyzing data. It is expectedthat AI will continue to revolutionize the way we work as technologyadvances.Materials and Method: We used SMILES, a one-dimensionalrepresentation of chemical structures, and obtained the mTORkinase activity data from the BindingDB database. After cleaningup the data, we vectorized the SMILES strings using onehotencoding and split the dataset into training, validation, andtest sets. we designed the model architecture to consist of anLSTM layer, two fully connected layers.Results: After the training, the resulting deep learning modelcan predict the IC۵۰ value for a candidate molecule to inhibitmTOR with agreeable accuracy.Conclusion: Due to the model's black-box nature, the abstractfeatures it creates as descriptors are not easily interpretable byhumans. Additionally, using SMILES instead of ۳D-structurefiles of molecules lacks spatial information, which could limitthe model's accuracy.Overall, this research demonstrates a promising method for developinga descriptor-free QSAR model that is simple, accurate,and fast to predict molecular activity.

Keywords:

QSAR , Deep-Learning , mTOR , Descriptor-free , LSTM