Fig 1.
Framework for DL-based virtual screening and predictive modeling of natural TNF-α inhibitors.
The model was developed using the ChEMBL dataset, with 2AZ5 as the target protein. After validation, the model was used to screen the Selleckchem Natural Compounds database. Compounds were prioritized based on their predicted pIC50 values, and after additional filtering, four compounds were chosen for MD simulation studies.
Fig 2.
Complete dataset pre-processing pipeline.
Panel A represents the characterization of compounds using PubChem descriptors (881) of both training (ChEMBL) and testing (Selleckchem) datasets and Panel B depicts variance thresholding to retain essential features.
Table 1.
Hyperparameter tuning of neural network architecture.
Fig 3.
Evaluation metrics of the DL regression model.
(a) MSE, (b) MAPE, (c) MAE, and (d) Loss.
Fig 4.
Identification of screened compounds through Selleckchem’s natural product library along with their predicted pIC50 values.
The structures and the pIC50 values of the eligible candidates are included (CID 442977: Imperialine, CID 6070: Veratramine, CID 10098: Jervine, and CID 5390854: Gelsemine).
Fig 5.
Visualization of docking scores and pIC50 values of 68 eligible natural candidates after drug-likeness analysis.
Table 2.
ADME properties of the compounds Imperialine, Veratramine, Jervine, and Gelsemine using ADMETlab2.0 server.
Fig 6.
Docked poses of top hit compounds against target TNF-α.
(a) Imperialine, (b) Veratramine, (c) Jervine, and (d) Gelsemine.
Table 3.
Drug-likeness analysis of final four compounds.
Fig 7.
2D molecular interaction of top hit compounds.
(a) Imperialine, (b) Veratramine, (c) Jervine, and (d) Gelsemine.
Fig 8.
Retrospective validation of our virtual screening protocol.
a) ROC plot based on Tanimoto scores for MF and Layered fingerprint descriptors b) Heat map showing Tanimoto similarity between control (index 1), actives (index 2–21) and decoys (22–50) and c) Box plots with distribution of Tanimoto scores between active-control and decoy-control for both the fingerprint descriptors.
Fig 9.
(a) RMSD, (b) RMSF, (c) Rg values plotted for the native apo-protein -black and apo-protein docked with selected bioactive molecules, (1) Imperialine-red, (2) Veratramine-green, (3) Jervine-yellow, and (4) Gelsemine- purple.
Table 4.
Average structural and dynamic parameters for protein-ligand complex molecular dynamics simulation.
Table 5.
RMSF (nm) of the active side residues for the apoprotein and the docked complexes.
Fig 10.
(a) H-bonds, and (b) SASA values plotted for the native apo-protein (TNF-α in the human body)-black and apo-protein docked with selected bioactive molecules, (1) Imperialine-red, (2) Veratramine-green, (3) Jervine-yellow, and (4) Gelsemine–purple.
Table 6.
The calculation of binding free energy results of 4 selected compounds.
Fig 11.
Principal component analysis of the MD simulation trajectories.
(a) 2AZ5, docked with (b) Imperialine, (c)Veratramine, (d) Jervine, and (e) Gelsemine.
Fig 12.
Dynamic cross-correlation matrix of MD simulation trajectories.
2AZ5 complexed with (a) 2AZ5 apo form, (b) Imperialine, (c) Veratramine, (d) Jervine, and (e) Gelsemine on α-carbon atoms.