P07-12

Natural-Product Screening Toward Discovery of Anti-Aging Glutaminase-1 Inhibitors. An Electronic-Structure Informatics Study

Mio YOKOYAMA *, Mizuki IWASAKI, Yusuke TATEISHI, Manabu SUGIMOTO

Kumamoto University
( * E-mail: 245d8785@st.kumamoto-u.ac.jp )

Accumulation of the senescent cells in aging within the human body can trigger various diseases, leading to death. Recently, the persistence of these senescent cells is linked to Glutaminase-1 (GLS1). Thus, developing GLS1 inhibitors is considered one of the significant challenges in anti-aging [1]. To develop more highly active inhibitors and/or to discover unique compounds that have been unknown so far, in silico screening is expected to play an important role. Aiming to discover unknown scaffolds of GLS1 inhibitors, we herein apply Electronic-Structure Informatics (ESI), which has been suggested by the authors’ group [2]. ESI is an informatics focusing on information obtained through electronic-structure (quantum chemistry) calculations. The molecular descriptors therein have been derived on some theoretical bases. The ESI descriptor set does not directly contain descriptors that reflect structural features. Because they represent features related to electronic structure, they are expected to provide molecules with unexpected scaffolds that differ from known GLS1 inhibitors in the course of inhibitor screening. We explore the natural product (NP) database for screening, anticipating the potential of NP.
We have developed a machine learning model to screen for potential GLS1 inhibitors by calculating ESI descriptors for 260 GLS1 inhibitor molecules obtained from ChEMBL. These descriptors were utilized as explanatory variables in constructing a regression model with the activity value (pIC₅₀) as the objective variable. It was found that the regression model using the extra trees regressor is successful where a coefficient of determination (R²) was 0.777 for the test set of molecules corresponding to 20 % of molecules among the 260 GLS1 inhibitors.
Because the reproducibility of the present model is considered reasonably good, we have constructed the final regression model for compound screening by including all the molecules. In applying this final model, we screened 2647 molecules from a NP database called KampoDB [3] to identify highly active GLS1 inhibitors. The KampoDB is a database for traditional Japanese medicines, so our approach herein corresponds to “drug-repositioning”.
We could have found several potential candidates for GLS1 inhibitors in the KampoDB. For structural modifications (optimization), we took the molecular anatomy-and-remodeling approach: a candidate molecule was decomposed into fragments, and their contributions to pIC₅₀ were analyzed using the regression model, which we call “NP anatomy”. Then, some fragments were replaced with others to enhance the inhibitory activity. We call this latter approach “NP remodeling”. In the presentation, we will show the detailed results of compound screening and our molecular optimization.

[1] Y. Johmura, et al., Science, 371, 265-270 (2021).
[2] M. Sugimoto et al., Chem. Lett., 50, 849-852 (2021).
[3] R. Sawada et al., Sci. Rep., 8, 11216 (2018).