(C) PLOS One
This story was originally published by PLOS One and is unaltered.
. . . . . . . . . .



MTMol-GPT: De novo multi-target molecular generation with transformer-based generative adversarial imitation learning [1]

['Chengwei Ai', 'School Of Computer Science', 'Engineering', 'Central South University', 'Changsha', 'Hongpeng Yang', 'Department Of Computer Science', 'University Of South Carolina', 'Columbia', 'South Carolina']

Date: 2024-08

To authenticate the potential drug-likeness of the synthesized compounds, we undertook two rigorous examinations. Initially, we employed molecular docking to explore the affinity of our constructed drug-like compounds with DRD2 and HTR1A targets. Subsequently, through the utilization of pharmacophore mapping, we initiated an exhaustive evaluation to determine if the molecular architectures conceived by MTMol-GPT could harmonize with the essential pharmacophores of DRD2 and HTR1A targets, thereby affirming their potential therapeutic effectiveness.

Molecular docking.

We performed molecular docking to calculate the binding affinity of our generated molecules to validate whether they can target multiple proteins of interest as potential ligands. The structures of our target proteins DRD2 (PDB ID: 6LUQ) and HTR1A (PDB ID: 7E2Y) were obtained from the Protein Data Bank (PDB) [51]. We generated 1, 000 molecules with MTMol-GPT, and extracted 1000 inactive molecules for DRD2 and HTR1A from ExCAPE-DB, respectively. Next, we randomly selected the same amount of active molecules from datasets for DRD2 and HTR1A.

After collecting the data, AutoDock Vina was used to perform molecular docking (details refer to Section AutoDock Vina), and the results of Docking scores (Ds) were shown in Fig 3. As for DRD2, the root mean square deviation (RMSD) value of ligand haloperidol between the experimental conformation in complex and our redocked pose is 0.38Å. And the RMSD for serotonin in HTR1A is 0.68Å. Both RMSD values are smaller than 2Å, which means our docking settings are reliable for these two receptors (see Fig F in S1 Text). The results indicate that molecules generated by both MTMol-GPT and SF-MTMol-GPT have outstanding performance on both DRD2 and HTR1A targets. Comparing the results for the two targets, the median docking score of generated molecules is smaller than that of inactive molecules. And the distribution of scores is also similar to that of active molecules. This indicates that the docking performance of the generated molecules on the two targets respectively is consistent with that of the existing active molecules.

In addition, molecules generated by MTMol-GPT are capable of dual-targeting DRD2 and HTR1A, which was validated through the comparison with dense regions (Fig 4A). Therefore, we used the generated 1, 000 molecules with MTMol-GPT and compared them with the molecules from DRD2 (Fig 4B) and HTR1A (Fig 4C) testing datasets. To enhance the visualization of molecular distribution, we generated density maps with t-SNE and dimensionality-reduced coordinates for both DRD2 and HTR1A target molecule datasets.

PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 4. Visualization of molecule distributions from generated, DRD2, and HTR1A datasets using t-SNE. A) Molecules from generated, DRD2 and HTR1A testing datasets. B) Molecules from generated and DRD2 testing datasets. C) Molecules from generated and HTR1A testing datasets. The density plot shows DRD2 and HTR1A target distributions: red for DRD2, yellow for HTR1A. Generated molecules are colored by their docking values to both targets, with darker blue indicating stronger binding (lower values). https://doi.org/10.1371/journal.pcbi.1012229.g004

In Fig 4A, while most generated molecules reside in the shared area of the two targets, a few also gather in distinct regions with higher color intensity, and lower docking values. Specifically, Fig 4B and 4C illustrates that the generated molecules primarily cluster within the regions of known DRD2- and HTR1A-targeting molecules. By overlapping these two targets’ separate distributions, we found that our generated molecules lay in the overlap regions, indicating the similarity between the generated molecules and known ligands.

To gain an insight into diverse molecules, we defined six clustering regions in total and examined molecules from these specific areas, including the overlapping zones (regions 1 and 2) of DRD2 and HTR1A, areas unique to DRD2 (regions 3 and 4), areas unique to HTR1A (region 5), and zones outside the central distribution of both targets (region 6). Molecules in undefined regions were excluded due to indistinct clustering. After a pre-selection process that involved filtering out molecules beyond the QED, logP, and SA value ranges defined by the DRD2 and HTR1A training sets, we selected one molecule with the lowest average docking values of two targets from each region (details in Table C in S1 Text).

Then we displayed the best docking conformation of these six representative molecules towards both the DRD2 and HTR1A targets in Fig 5A and 5B. Compared with their original ligands haloperidol and serotonin −9.68 and −6.06, respectively, our generated molecules can reach lower docking scores and better binding potential. Moreover, the generated molecules displayed various 3D conformations, further indicating that our model is capable of generating diverse and novel molecules that target two proteins simultaneously. Meanwhile, we analyzed the non-covalent interactions between protein receptors and docked molecules by PLIP [52]. DRD2 crystal structures with different ligands own the hydrophobic interaction with Asp114 and pi-stacking with Trp386 (see 6LUQ and 6CM4 Fig Ga in S1 Text). As for our molecules generated by MTMol-GPT, we can also find some of these key interactions from the best docking poses(see Fig Gb in S1 Text). For example, Mol_975 also forms the hydrophobic interactions with Leu94, Phe110, and Asp114, while Mol_237 contributes to Leu94, Phe110, and Val115. Other representative molecules rely on similar interactions to bind with DRD2. Likewise, similar interactive residues in HTR1A are captured to form hydrophobic interactions when binding with distinct molecules, including Tyr96, Lys119, and Asn386 (see Fig H in S1 Text).

[END]
---
[1] Url: https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1012229

Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/

You are viewing proxied material from magical.fish. The copyright of proxied material belongs to its original authors. Any comments or complaints in relation to proxied material should be directed to the original authors of the content concerned. Please see the disclaimer for more details.