Multiple linear regression
After calculating molecular descriptors for 46 derivatives (Table 2), a reliable model was built using five descriptors: S-B, NHBD, SOVD, TD, and TVC. Molecules 12, 14, 23, 29, 30, 33, 44, 45, and 46 were chosen for the test set, while the remaining molecules were included in the training set. The QSAR model produced by the MLR technique is shown in Eq.(5)
$${\mathbf{pIC}}_{50} = 12.46558 + 1.07186 \times {\mathbf{S}} - {\mathbf{B}} + 1.47232 \times {\mathbf{NHBD}} + 0.03583 \times {\mathbf{SOVD}} - 0.67679 \times {\mathbf{TD}} - 717635 \times {\mathbf{TVC}}$$
(5)
$${\text{N }} = { 37 };{\text{ R}} = \, 0.{88 };{\text{ R}}^{{2}} = 0.{78 };{\text{R}}_{{{\text{Ajus}}}}^{2} = \, 0.{74 };{\text{ RMSE}} = 0.{286}; {\text{F}} = {21}.{384} ;{\text{ Pr }} < \, 0.000{1};R_{cv}^{2} = 0.{64}$$
N denotes the number of compounds in the training set, while RMSE represents the root mean square error.
As a result, Fig.2 demonstrates a strong linear correlation between the five descriptors selected and the biological activity values (pIC50). The criteria used to evaluate the QSAR model were as follows: R2, RMSE, F, P-value and \({R}_{cv}^{2}\). The higher R2 value (0.78), lower RMSE value of 0.286, and high statistical confidence level (F = 21.384) indicate that the QSAR model is statistically acceptable. The P-value (Pr < 0.0001) shows the model’s statistical significance at a level greater than 95%. S-B, NHBD, and SOVD have a positive influence, while TVC and TD have a negative influence, as shown in Fig.3. The cross-validation correlation coefficient (\({R}_{cv}^{2}\) = 0.64) indicates the QSAR model’s correctness. The MLR-based QSAR model shows a significant correlation between observed and predicted pIC50 values. To improve the correlation, another QSAR model will be developed using MNLR and ANN techniques.
Comparison of predicted and observed activity values using the MLR model.
Analysis of molecular descriptor contributions in an MLR model.
Multiple nonlinear regression
A nonlinear QSAR model was constructed using the MNLR technique, as shown in Eq.6
$$\begin{aligned} & {\mathbf{pIC}}_{{50}} = 14.06127 - 0.02836 \times {\mathbf{S}} - {\mathbf{B}} + 0.48261{{*}}{\mathbf{NHBD}} - 1.00626 \times {\mathbf{SOVD}} + 5.06009 \times {\mathbf{TD}} - 2229421 \times {\mathbf{TVC}} \\ & \quad + 0.60322 \times \left( {{\mathbf{S}} - {\mathbf{B}}} \right)^{2} + 0.00495 \times \left( {\varvec{SOVD}} \right)^{2} - 0.15036 \times \left( {\varvec{TD}} \right)^{2} + 4387 \times 10^{8} \times \left( {\varvec{TVC}} \right)^{2} \\ \end{aligned}$$
(6)
N = 37; R = 0.93; R2 = 0.87; RMSE = 0.274.
The evaluation metrics of the non-linear QSAR model (R2 = 0.87, RMSE = 0.274) demonstrate the statistical validity of this model with a p-value less than 0.05. Table 3 presents the observed and estimated biological activities for both the training and test sets using both linear and non-linear models.
MLR-QSAR modeling efficiency
Cross-validation
Figure4 illustrates the results of cross-validation using the Leave-One-Out (LOO) approach. The results of this cross-validation (R2 = 0.65 and RMSE = 0.31) show that cross-validation has no significant effect on the QSAR model. These clear findings demonstrate the proposed QSAR model’s stability and robustness. However, it is important to note that cross-validation alone is insufficient to fully assess the capabilities of QSAR models.
Correlation of observed and predicted activities calculated using LOO-CV.
Artificial neural networks
The ANN approach is employed to construct a QSAR model with a 5-3-1 architecture and a ρ value of 1.68. The ρ value between 1 and 3 indicates that the number 3 in the hidden layer is proportional to the number of descriptors 5 in the input layer, predicting the pIC50 values represented by the output layer 1.
The QSAR model developed using the ANN approach has a very high coefficient of determination R2 of 0.89, indicating an excellent fit to the data. Moreover, the RMSE is low (0.17), suggesting that the model predictions are highly accurate. The results demonstrate that the QSAR model is statistically reliable in predicting inhibitory activity as balanced multifunctional agents against Alzheimer’s disease.
Consequently, the five descriptors (S-B, NHBD, SOVD, TVC and TD) are valuable for predicting pIC50 values. These descriptors were selected for their relevance to this study.
Figure5 illustrates that the predicted pIC50 values are evenly distributed across the training set, ensuring that the ANN model predictions closely correspond to the experimentally observed pIC50 values.
Correlation between the observed and the predicted activities calculated by ANN.
Y-randomization test
Randomization is a common strategy to ensure the reliability of QSAR models. Once a regression model is selected, Y-randomization is used to validate it. This involves shuffling the activity values of the compounds and rebuilding the model to assess its sensitivity to random variations in the data. This process is repeated 100 times to assess the model’s overall robustness and identify any potential overfitting. As shown in Eq.(5), these randomized models have an average correlation coefficient of 0.36, \({R}^{2}\) value of 0.14 and \({{Q}_{cv}}^{2}\) of 0.63.
Random target values yielded significantly lower average \({R}^{2}\) and \({{Q}_{cv}}^{2}\) values compared to the model values. This suggests that the relationships between descriptors and activities in Model 1 are not coincidental, confirming the model’s robustness. This randomization test validates the model’s predictive power and reliability.
External validation
External validation was conducted using the Golbraikh-Tropsha criteria68, with the aim of evaluating the ability of the QSAR models. We predicted pIC50 activity values for compounds in the test set, as shown in Table 4. This assessment includes calculating the correlation coefficient (R2), which is a crucial criterion for determining how effectively externally validated models can predict the activities of molecules that were not included in the model development process, as shown in Fig.6.
Correlations between the observed/predicted activity by the MLR model for the test set.
The R2 test provided a value of 0.68 with a root mean square deviation of 0.282, and all values were within the acceptable range, indicating that the Golbraikh and Tropsha criterion was successfully validated. These findings suggest that the developed QSAR model is quite successful. Furthermore, external validation of the QSAR models indicates their high capacity to reliably predict pIC50 values for the experimental inhibitory activity of balanced multifunctional drugs against Alzheimer’s disease.
Applicability domain (AD)
The William plot for the AD of the model is shown in Fig.7.
The Williams graph of the model presented by Eq.(5).
The AD of the QSAR model was established through leverage analysis, as demonstrated by the Williams plot (Fig.7). In the William plots. the results indicate that the leverage values of all the compounds in the training and test sets were lower than the warning leverage (h* = 0.50) except for compound 1, which was greater than the warning lever effect. This compound belongs to the training set.
The accurate prediction of the QSAR model for the test sets can be attributed to their lack of outliers. Consequently, all tested compounds fall within the AD, signifying the credibility of their predicted activity values.
Molecular docking simulations
The results shown in Fig.8 confirm that most active ligands, C25 and C27, share common intermolecular interactions, including Trp279 amino acid residue as the active site of anti-Alzheimer’s drug (1EVE.pdb), more than Asp72 and Tyr334 amino acids residues. These interactions were all detected with the lowest possible binding energies in kcal/mol, -7.26 and -6.1 kcal/mol, respectively. The novel-designed compound, labeled Pred12, which was predicted in the current study with the highest acetylcholinesterase inhibitory activity based on the validated QSAR model, was equally docked to the same targeted receptor (1EVE.pdb). The docking results revealed similar intermolecular interactions, such as Trp84 and Trp279 amino acid residues as two active sites of the targeted protein, in addition to other common chemical bonds produced towards Tyr334 and Phe330 amino acid residues. One additional hydrogen bond was detected towards the Phe288 amino acid residue, making the novel-designed compound highly stable in terms of energy order (binding energy of -7.94 kcal/mol)73. Moreover, the molecular docking protocol has been successfully validated29,70 as the candidate ligands labeled C25 and C27 more than most active molecules from the newly designed compounds marked Pred12 were effectively docked to two active sites of the three-dimensional structure of the anti-Alzheimer drug complexed with its target acetylcholinesterase co-crystallized ligand as shown in Fig.9.
Intermolecular contacts in two and three dimensions between 1EVE.pdb protein and C25. C27. and Pred12 with binding energies of − 7.26. − 6.1. and − 7.94 kcal/mol. respectively.
Active sites of the targeted receptor encoded in PDB basis by 1EVE.pdb.
MD simulation
The complexes of the human AChE receptor complexed with ligands 61, 726, and 794 were selected for MD simulation to evaluate their thermodynamic stability over a 100 ns timescale. A stable drug-receptor complex is essential for exerting a therapeutic response. The macromolecular complex underwent a 100 ns MD simulation using Schrodinger’s Desmond software version 2022.4. The target AChE receptor’s monomeric chain consists of 534 amino acids, comprising 4254 heavy atoms out of a total of 8369 atoms. To assess thermodynamic stability, structural changes and root-mean-square deviation (RMSD) analysis of the macromolecular backbone were performed throughout the 100 ns simulation. Ligand 61, with thirteen flexible bonds and thirty-four heavy atoms out of seventy-two total, displayed high stability in complex with human AChE throughout the simulation. The RMSD value of the receptor’s backbone fluctuated between 1.2 and 2.4 Å, while the bound ligand 61 exhibited minor RMSD fluctuations within the receptor cavity. ranging from 3.0 to 5.0 Å.
RMSF analysis, which measures the deviation of atoms from their initial positions, is a crucial parameter for assessing the flexibility and dynamic behavior of the macromolecular complex. To assess protein dynamics and stability, RMSF analysis was conducted on the human AChE receptor complexed with ligand 61. The MD simulation results showed that the RMSF values for the Cα backbone remained within a stable range of 0.5–2.5 Å, except for a few terminal residues, while for ligand 61, it was found to be ranging from 1.0 to 2.0 Å. RMSF of human AChE receptor’s backbone and complexed ligand 61, observed during MD analysis, is depicted in Fig.10.
RMSF for Cα chain of human AChE complexed with ligand 61 detected while executing 100ns MD simulation.
The stability of the receptor-ligand complexes was evaluated by monitoring hydrophobic contacts, ionic interactions and hydrogen bonds formed during MD simulations. Ligand 61 interacted with the human AChE receptor via hydrophobic contacts with Tyr70, Trp84, Tyr121, Trp279, Leu282, Phe330 and Phe331. Additionally, hydrogen bonds formed with Tyr121, Phe284, Asp285, Ser286, Phe288 and Arg289 while Trp279 and Ser286 interacted through water bridges.
In another simulation, the complexed ligand 726 complexed with the human AChE enzyme, revealed that it constitutes fifteen flexible bonds, comprising thirty-six heavy atoms of seventy-two atoms in total. The human AChE-ligand 726 conjugate has displayed high stability throughout the simulation. RMSD (Root Mean Square Deviation) analysis revealed that the receptor’s backbone exhibited fluctuations between 1.2 and 2.1 Å during the simulation. The bound ligand 726 also showed minor fluctuations within the receptor cavity. ranging from 3.2 to 7.0 Å. Molecular dynamics simulations of the human AChE receptor complexed with ligand 726 revealed the RMSF (Root Mean Square Fluctuation) for the Cα backbone remained within 0.5–2.0 Å, except for some terminal residues, while for ligand 726, it was found to be ranging from 1.0 to 4.0 Å. The RMSF of human AChE receptor’s backbone and complexed ligand 726, observed during MD analysis, is depicted in Fig.11. Throughout the simulation, ligand 726 was found to be interacting with human AChE receptor via the formation of hydrophobic bonds with the amino acids Tyr70, Phe75, Trp279, Phe290, Phe331, Tyr334, Pro337, Leu358 whereas Tyr70, Val71, Asp72, Trp84, Tyr121, Phe288, Arg289 and Tyr334 interacted via hydrogen bonds. Additionally, the amino acid Tyr70, Trp84, Asp285 and Tyr334 formed water bridges.
RMSF for Cα chain of human AChE complexed with ligand 726 detected while executing 100ns MD simulation.
MD simulation of the ligand complex 794, complexed with the human AChE enzyme, revealed that it constitutes nine flexible bonds, comprising thirty-four heavy atoms of fifty-eight atoms in total. The human AChE-ligand 794 conjugate has displayed high stability throughout the simulation. The RMSD value of the receptor’s backbone was found to fluctuate between 1.0 and 1.75 Å, whereas the bound ligand 794 exhibited minor fluctuations in its RMSD value within the receptor cavity, ranging from 3.6 to 5.0 Å. Figure12 illustrates the root-mean-square deviation (RMSD) profiles for human AChE receptor complexes with ligands 61, 726 and 794 respectively.
RMSD for Cα chain of human AChE complexed with ligand 61 (a), ligand 726 (b) and ligand 794 (c) detected while executing 100ns MD simulation.
Molecular dynamics (MD) simulations of the AChE-ligand 794 complex revealed that the RMSF for the Cα backbone remained within 0.5–2.0 Å for most residues, except for a few terminal residues, while for ligand 794, it was found to be ranging from 1.0 to 3.0 Å. The RMSF of the human AChE receptor’s backbone and complexed ligand 794, observed during MD analysis, is depicted in Fig.13.
RMSF for Cα chain of human AChE complexed with ligand 794 detected while executing 100ns MD simulation.
Throughout the simulation, ligand 794 was found to be interacting with the human AChE receptor via formation of hydrophobic bonds with the amino acids Met83, Trp84, Tyr121, Trp279, Phe288, Phe290, Phe330, Phe331, Tyr334 and His440. Additionally, hydrogen bonds were formed with Tyr70, Glu73, Gln74, Tyr121 and Ser122 while the amino acid Gln69, Tyr70, Val71, Asp72, Gln74, Tyr121, Phe288, Arg289 and Tyr334 interact via water bridges. Figure14 shows the interactions of ligand 61 (a), ligand 726 (b) and ligand 794 (c) with human AChE receptor residues.
Protein–Ligand Contacts: Protein–ligand interactions identified between the human AChE receptor and ligand 61 (a), ligand 726 (b) and ligand 794 (c) respectively. The interactions were visualized using different colored bars with green representing h.
Design of new compounds
Table 5 presents the predicted values of molecular descriptors and pIC50 activities, calculated using the multiple linear regression study, which was used to develop a quantitative structure–activity relationship model for predicting the inhibitory activity of pyridazin-3(2H)-one derivatives against Alzheimer’s disease. The model was built using a dataset of compounds with known inhibitory activity, and the selected descriptors were calculated using molecular modeling software. The MLR model showed good statistical performance, with a high coefficient of determination (R2) and low root mean square error (RMSE) values.
The primary goal of this research is to develop novel Alzheimer’s disease inhibitors derived from pyridazin-3(2H)-one, using the suggestions obtained from the 2D-QSAR investigations. In this work, thirteen pyridazin-3(2H)-one derivatives (Pred1–Pred13) were designed to improve the inhibitory efficacy against Alzheimer’s disease (Table 6). To evaluate the potential of the newly proposed compounds, we used the same molecular descriptors as those used for the existing series of molecules. We then applied the MLR model (a regression analysis method) to predict their activity. The results presented in Table 8 indicate that these novel compounds exhibit comparable or superior inhibitory activity to the most active compounds in the series. This suggests their potential as promising candidates for Alzheimer’s disease treatment.
Lipinski’s rule
A potent drug is a chemical substance that has successfully passed several screening phases. Most drugs fail in preclinical testing because they lack the unique characteristics needed to be considered drug candidates. ADMET evaluation, an essential parameter in drug development, is of great importance in the early stages to reduce the failure rate of drug candidates.
In our study, Thirteen predicted molecules were subjected to in-silico ADMET testing, and all showed a high probability of being absorbed in the gut, as shown in Table 8. Those molecules that penetrate the gut also have a marked tendency to be absorbed by the human organism. An in vivo study confirmed that these 13 predicted molecules could effectively improve cognitive dysfunction in scopolamine-treated individuals by regulating acetylcholine levels and inhibiting oxidative stress. These results highlight the potential of these predicted molecules as novel, balanced candidates for the pharmacotherapy of Alzheimer’s disease.
Next, we assessed the drug-like physicochemical properties of the 13 predicted molecules with the aim of determining their potential for administration as drugs. Using Lipinski’s rule as a benchmark, these compounds met the criteria and fulfilled the topological polar surface area (TPSA) requirements for central nervous system drugs (Table 7)71. In addition, the thirteen predicted molecules were expected to have moderate water solubility and good gastrointestinal absorption capacity. Consequently, we concluded that these 13 predicted molecules possessed favorable physicochemical properties like those of drugs.
ADMET properties
To ensure the potential relevance of molecules designed as drugs, we took into account pharmacokinetic parameters such as ADMET. In silico ADMET properties were predicted using the pkCSM online tool72, and the details are provided in the corresponding table (Table 8).
Absorption levels below 30% indicate poor absorption, while values above 90% indicate acceptable absorption in the human digestive system. It is important to note that a volume of distribution (VDss) greater than 0.45 is considered significant. The predicted compounds demonstrated excellent intestinal absorption, suggesting widespread distribution throughout the body. Enzymatic metabolism, the biochemical transformation of drugs in the body, plays a crucial role in modifying pharmacological molecules. This process can generate various metabolites that influence drug responses to varying degrees73. Given the potential differences in physicochemical and pharmacological properties, a thorough investigation of drug metabolism is essential.
Cytochrome P450 (CYP450) enzymes, particularly those in the CYP1, CYP2, CYP3, and CYP4 families, are responsible for over 90% of phase I metabolism. Among these, CYP3A4 is especially relevant to our research, as the novel compounds we developed act as both substrates and inhibitors of this enzyme75.
Clearance, a measure of how quickly a drug is eliminated from the body, is influenced by various factors. The newly discovered compounds exhibited notably high clearance values, ensuring optimal drug retention. Additionally, assessing the non-toxicity of predicted molecules is a critical step in drug selection. Further investigations are necessary to evaluate the toxicity profiles of these compounds. Fortunately, all the molecules we developed are non-toxic, adding to their potential as drug candidates.
In the pursuit of developing drug candidates for Alzheimer’s disease, the focus has been on inhibiting self-induced Aβ1-42 aggregation and self-induced Aβ1-42 fibril disaggregation. However, as reported in the 2022 study by Yichun et al., certain limitations have been observed in the development of synthetic drugs targeting these processes. These limitations include the inability to cross the blood–brain barrier (BBB), safety profile issues, limited solubility, and undesirable side effects. Therefore, addressing these challenges is essential in the development of novel drug candidates for Alzheimer’s disease.
These findings emphasize the importance of the predicted molecules for the development of dual-binding AChE inhibitors targeting both the PAS and the catalytic active site (CAS) of AChE, which could offer multiple activities against AChE, Aβ and delay the progression of AD.
