Rational design AI assisted of novel Glutaminase allosteric inhibitors

Motivation Cancer cells frequently exhibit uncoupling of the glycolytic pathway from the tricarboxylic acid cycle and as a result, often become dependent on their ability to increase glutamine catabolism. GLS1 exists as two isozymes: kidney glutaminase (KGA) and glutaminase C (GAC), which control the hydrolysis of glutamine to glutamate, resolving the ‘glutamine addiction’ of cancer cells. Therefore, they play a central role in supporting cancer growth and proliferation. Several medicinal chemistry projects are focused on developing new and effective GLS1 inhibitors. Several compounds have been identified as active site or allosteric inhibitors. This latter group mainly includes a large number of bis-2-(5-phenylacetamido-1,3,4-thiadiazol-2-yl)ethyl sulphide derivatives, called BPTES. Despite their pre-clinical effectiveness, only a few inhibitors have advanced to clinical trials, CB-839 and IPN60090 [1]. In recent years, AI methods, such as graph neural networks, reinforcement learning, and generated models, along with structure-based methods (i.e., molecular dynamics simulations and molecular docking) are contributing to drug discovery and analysis of drug responses. Thus, with the aim of proposing efficient GAC allosteric inhibitors as hits for a drug discovery project, the present contribution presents a preliminary study to find novel compounds, with generated models integrated with structure-based methods. Methods Crystallographic data available in PDB, roughly 23 coordinates acquisitions, were carefully evaluated to select 5HL1 as the most appropriate 3D structure to be exploited for docking studies. An in-house dataset of allosteric inhibitors was created from literature data, including patents. For each molecule, the isomeric SMILES string and 3D-structure were generated and the experimental activity as GAC allosteric modulator was annotated. Several generative Deep Neural models based on autoencoder architectures were trained on the dataset to generate new molecules that share the same chemical characteristics of the inhibitor dataset. Ligand structures were generated using RDKit[2] and XTB[3] softwares. The new molecules were then manually checked and docked against the receptor. Docking studies were performed using Autodock-Vina with the Vinardo scoring function. The Vinardo scoring function appeared to best reproduce the interactions between ligands and GAC as known from experiments. Results The generative Deep Neural model was trained on the dataset of 860 molecules and used to produce 118 new ones that are potentially interesting as allosteric modulators. In order to filter these compounds, selecting the most promising, docking studies were performed on the CB-839 inhibitor currently under clinical trials, chosen as the reference and on the first 118 newly designed molecules. The compounds with binding affinity better than CB-839 generally displayed the same interactions with the key residues Lys320 of the two dimers and conserved the H-bond network of the reference, thus maintaining some of the key determinants for the inhibitory activity. AdMEtox and drug-likeness evaluations on these molecules, highlighted for some of them a more promising profile when compared with CB-839. Molecular dynamics simulations and MMPBSA analysis on CB-839, to refine its molecular interactions with the target, are running. The most promising ligands identified within this project will be synthesized and the biological activity towards GAC will be determined. References 1. Katt W P et al. “A tale of two glutaminases: homologous enzymes with distinct roles in tumorigenesis.” Future Med. Chem.2 (2017): 223-243. 2. RDKit: Open-source cheminformatics. https://www.rdkit.org 3. Bannwarth C, et al. “Extended tight-binding quantum chemistry methods.” WIREs Comput. Mol. Sci. (2020): e01493 DISSEMINATION MATERIAL Summary Artificial Intelligence AI is gaining interest in drug discovery, especially if applied to innovative therapeutical targets, where the development of specific drugs is mandatory. In this context, its application for identifying new compounds potentially acting as allosteric inhibitors of GLS1, an enzyme involved in several solid tumor types. Currently, there are only two compounds under clinical trials, but there are no specific drugs available in therapy. Thus, the aim of this project is to develop new derivatives through AI and computational approaches, to be subsequently synthesized and tested for their efficacy. Motivation Cancer cells frequently exhibit uncoupling of the glycolytic pathway from the tricarboxylic acid cycle and as a result, often become dependent on their ability to increase glutamine catabolism. GLS1 exists as two isozymes: kidney glutaminase (KGA) and glutaminase C (GAC), which control the hydrolysis of glutamine to glutamate, resolving the ‘glutamine addiction’ of cancer cells. Therefore, they play a central role in supporting cancer growth and proliferation. Several medicinal chemistry projects are focused on developing new and effective GLS1 inhibitors. Several compounds have been identified as active site or allosteric inhibitors. This latter group mainly includes a large number of bis-2-(5-phenylacetamido-1,3,4-thiadiazol-2-yl)ethyl sulphide derivatives, called BPTES. Despite their pre-clinical effectiveness, only a few inhibitors have advanced to clinical trials, CB-839 and IPN60090 [1]. In recent years, AI methods, such as graph neural networks, reinforcement learning, and generated models, along with structure-based methods (i.e., molecular dynamics simulations and molecular docking) are contributing to drug discovery and analysis of drug responses. Thus, with the aim of proposing efficient GAC allosteric inhibitors as hits for a drug discovery project, the present contribution presents a preliminary study to find novel compounds, with generated models integrated with structure-based methods. Methods Crystallographic data available in PDB, roughly 23 coordinates acquisitions, were carefully evaluated to select 5HL1 as the most appropriate 3D structure to be exploited for docking studies. An in-house dataset of allosteric inhibitors was created from literature data, including patents. For each molecule, the isomeric SMILES string and 3D-structure were generated and the experimental activity as GAC allosteric modulator was annotated. Several generative Deep Neural models based on autoencoder architectures were trained on the dataset to generate new molecules that share the same chemical characteristics of the inhibitor dataset. Ligand structures were generated using RDKit[2] and XTB[3] softwares. The new molecules were then manually checked and docked against the receptor. Docking studies were performed using Autodock-Vina with the Vinardo scoring function. The Vinardo scoring function appeared to best reproduce the interactions between ligands and GAC as known from experiments. Results The generative Deep Neural model was trained on the dataset of 860 molecules and used to produce 118 new ones that are potentially interesting as allosteric modulators. In order to filter these compounds, selecting the most promising, docking studies were performed on the CB-839 inhibitor currently under clinical trials, chosen as the reference and on the first 118 newly designed molecules. The compounds with binding affinity better than CB-839 generally displayed the same interactions with the key residues Lys320 of the two dimers and conserved the H-bond network of the reference, thus maintaining some of the key determinants for the inhibitory activity. AdMEtox and drug-likeness evaluations on these molecules, highlighted for some of them a more promising profile when compared with CB-839. Molecular dynamics simulations and MMPBSA analysis on CB-839, to refine its molecular interactions with the target, are running. The most promising ligands identified within this project will be synthesized and the biological activity towards GAC will be determined. References 1. Katt W P et al. “A tale of two glutaminases: homologous enzymes with distinct roles in tumorigenesis.” Future Med. Chem.2 (2017): 223-243. 2. RDKit: Open-source cheminformatics. https://www.rdkit.org 3. Bannwarth C, et al. “Extended tight-binding quantum chemistry methods.” WIREs Comput. Mol. Sci. (2020): e01493 DISSEMINATION MATERIAL Summary Artificial Intelligence AI is gaining interest in drug discovery, especially if applied to innovative therapeutical targets, where the development of specific drugs is mandatory. In this context, its application for identifying new compounds potentially acting as allosteric inhibitors of GLS1, an enzyme involved in several solid tumor types. Currently, there are only two compounds under clinical trials, but there are no specific drugs available in therapy. Thus, the aim of this project is to develop new derivatives through AI and computational approaches, to be subsequently synthesized and tested for their efficacy.