P02-10

Case studies of deep learning-based molecular docking program in medicinal chemistry

Kazuya OSUMI *, Naoya UKEGAWA, Tomohide MASUDA

Pharmaceutical Research Laboratories, Toray Industries, Inc.
( * E-mail: kazuya.osumi.v5@mail.toray )

Molecular docking is a computational procedure in which the non-covalent bonding of molecules, such as a protein receptor and a ligand, is predicted. The scoring function in the docking process is responsible for evaluating the correctness of the pose of the molecule in the binding site and predicting its binding affinity. Several recent efforts, such as GNINA, have demonstrated success in combining 3D grid representations with convolutional neural networks (CNNs). This allows the model to learn its own representation of the protein-ligand interaction in order to determine what constitutes a strong binder.
In this study, we report the examination of the following two cases of utilizing GNINA.

1) Retrospective analysis: Classification of the ligand function
In the research on RORγt inhibitors, a slight modification of the terminal substituent switched the function of the ligand from an antagonist to an agonist. There have been several reports of similar cases, and the function of the ligand has been explained by X-ray co-crystal structure analysis and MD simulations. Therefore, we examined whether it is possible to identify the ligand function more conveniently by docking simulation. We docked each ligand to inactive and active states of the proteins and compared their docking poses and scores. As a result, it was suggested that when using GNINA, the function of the ligand can be identified by the difference in docking scores for both the inactive and active forms of the protein.
2) Prospective prediction: Selection of compounds expected to have high affinity
In the research on kinase inhibitors targeting the CNS, it was inferred from the initial SAR of the hit compound and the known complex structures that there is a space where the substituent can be extended beyond the terminal amino group of the hit compound. Therefore, we constructed a virtual library of approximately 3,000 compounds derived from the combination of the amino group of the hit compound and diverse set of carboxylic acids. After filtering by druglikeness and CNS-MPO scores, we selected 96 compounds expected to have high affinity by docking simulation using GNINA. As a result, 21 compounds were obtained with improved inhibitory activity than the hit compound, suggesting that the selection of compounds by GNINA’s docking score is effective.

[1] Francoeur, P. G., et al., J. Chem. Inf. Model, 2020, 60, 4200-4215.
[2] McNutt, A., et al., J. Cheminformatics, 2021, 13, 43.
[3] Osumi, K., et al., The 39th Medicinal Chemistry Symposium, 2022.
[4] René, O., et al., ACS Med. Chem. Lett. 2015, 6, 276−281.
[5] Yukawa, T., et al., J. Med. Chem. 2019, 62, 1167– 1179.