Src kinase app: valid inhibitor generation and prediction with explanation using predictive model and selfies
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Generative models trained on small datasets often underperform, limiting their ability to explore chemical space effectively. While SELFIES (Self-Referencing Embedded Strings) guarantee syntactic validity, their integration with activity prediction and practical deployment remains underdeveloped. Here, the author presents an accessible workflow that combines predictive modeling, large-scale virtual screening, and recurrent neural network (RNN) training with both SMILES and SELFIES representations. Using Src kinase as a model target, the author demonstrates that GRU-SELFIES architectures consistently outperform SMILES-based counterparts, producing valid, drug-like molecules with favorable docking scores in as few as four sampling runs. The generated compounds showed a higher docking score distribution compared to the training set. This result was obtained by applying the predictive model to a small dataset, where it was able to identify compounds with higher docking scores. Several generated compounds exhibited stable interactions with critical Src kinase residues during molecular dynamics simulations for 200ns, supporting their potential as inhibitors. To ensure reproducibility and accessibility, all predictive and generative models were deployed as interactive web applications with integrated explainable AI (XAI). Additionally, the author introduces SaveMol , a MERN-based platform for storing and searching generated molecules. Finally, over 196k dataset protein-ligand complex, came from training data and small benchmark dataset as an engine for diffusion model deposited here: https://doi.org/10.5281/zenodo.17931255 . All links available at: https://github.com/phalem/Src
Fields of Study & Subjects
Computational Drug Discovery Methods
Machine Learning in Materials Science
Synthesis and Characterization of Heterocyclic Compounds
Publication details
- Journal / source
- ChemRxiv
- Publication date
- 30 March 2026
- Language
- EN
- Document type
- Preprint