Combined Deep Learning and Molecular Modeling Techniques on the Virtual Screening of New mTOR Inhibitors from the Thai Mushroom Database.

Abstract

The mammalian target of rapamycin (mTOR) is a protein kinase of the PI3K/Akt signaling pathway that regulates cell growth and division and is an attractive target for cancer therapy. Many reports on finding alternative mTOR inhibitors available in a database contain a mixture of active compound data with different mechanisms, which results in an increased complexity for training the machine learning models based on the chemical features of active compounds. In this study, a deep learning model supported by principal component analysis (PCA) and structural methods was used to search for an alternative mTOR inhibitor from mushrooms. The mTORC1 active compound data set from the PubChem database was first filtered for only the compounds resided near the first-generation inhibitors (rapalogs) within the first two PCA coordinates of chemical features. A deep learning model trained by the filtered data set captured the main characteristics of rapalogs and displayed the importance of steroid cores. After that, another layer of virtual screening by molecular docking calculations was performed on ternary complexes of FKBP12-FRB domains and six compound candidates with high "active" probability scores predicted by the deep learning models. Finally, all-atom molecular dynamics simulations and MMPBSA binding energy analysis were performed on two selected candidates in comparison to rapamycin, which confirmed the importance of ring groups and steroid cores for interaction networks. Trihydroxysterol from Lentinus polychrous Lev. was predicted as an interesting candidate due to the small but effective interaction network that facilitated FKBP12-FRB interactions and further stabilized the ternary complex.