Src kinase app: valid inhibitor generation and prediction with explanation using predictive model and selfies

Dealing with a small Experimental dataset using a generative model produces a model with underfitting and reduces its ability to generate a new valid compound. Even in the presence of free available chemical databases SMILES string has to use a complex and computationally intensive model to solve validation problems. SELFIES solve all validation problems but further activity optimization is needed with the absence of an app that records molecules generated. In this study, the author uses a predictive model to provide a dataset by a virtual screen of 3 million compounds from a chemical online database in addition to experimental active dataset. Data feed to a different model of one layer Recurrent Neural Network model using both SELFIES and SMILES for about 2-4 epochs. Structure-based drug design was used and Src Kinase as a target to validate both the predictive model and compounds produced by Recurrent Neural Network and further filtration happens using Molecular Dynamics Simulation. SELFIES outperform SMILES in producing valid molecules in all types of Recurrent Neural Network simple structures. Recurrent Neural Network can produce active compounds using the GRU layer without any activity optimization from just 4 runs 100 molecules each. The novelty of the result can be compared to the result coming from predictive model virtual screen data. Recurrent Neural Network can produce novel compounds with key interaction residue with the target protein. All Predictive Models were deployed and ExplainableAI is used to guide generated molecules. MERN stack app SaveMol is used to save molecules produced with substructure research ability and apps links provide here(https://github.com/phalem/Src).