Insight into glycogen synthase kinase-3β inhibitory activity of phyto-constituents from Melissa officinalis: in silico studies

Abstract

Over activity of Glycogen synthase kinase-3β (GSK-3β), a serine/threonine-protein kinase has been implicated in a number of diseases including stroke, type II diabetes and Alzheimer disease (AD). This study aimed to find novel inhibitors of GSK-3β from phyto-constituents of Melissa officinalis with the aid of computational analysis. Molecular docking, induced-fit docking (IFD), calculation of binding free energy via the MM-GBSA approach and Lipinski’s rule of five (RO5) were employed to filter the compounds and determine their druggability. Most importantly, the compounds pIC₅₀ were predicted by machine learning-based model generated by AutoQSAR algorithm. The generated model was validated to affirm its predictive model. The best model obtained was Model kpls_desc_38 (R² = 0.8467 and Q² = 0.8069), and this external validated model was utilized to predict the bioactivities of the lead compounds. While a number of characterized compounds from Melissa officinalis showed better docking score, binding free energy alongside adherence to RO5 than co-cystallized ligand, only three compounds (salvianolic acid C, ellagic acid and naringenin) showed more satisfactory pIC₅₀. The results obtained in this study can be useful to design potent inhibitors of GSK-3β.

Introduction

Glycogen synthase kinase-3β (GSK-3β) is a serine/threonine-protein kinase, primarily located in cytosol demonstrates important roles in different disease molecular pathophysiology (Mancinelli et al. 2017). Owing to its role in type 2 diabetes and obesity as determined by in vitro and in vivo studies, GSK-3β has gained popularity as a possible drug target (Gum et al. 2003; Ring et al. 2003). It’s also related to Alzheimer's disease (AD) and mood disorders (Hsiung et al. 2003), osteoporosis (Smith and Frenkel 2005), arteriosclerosis (Robertson et al. 2006), and cancer (Inoki et al. 2006). Unlike other protein kinases, GSK-3β under normal conditions is constitutively active and undergoes a rapid and temporary inhibition in response to a variety of external signals (Dorm 2005).

GSK-3β has been explored as a therapeutic target for a range of human diseases including cancer due to its diverse cellular functions (Yuki and Chikashi 2015). It is thus regarded as an obvious target for disease drug development, including neurodegenerative diseases such as Alzheimer's diseases, diabetes mellitus, and cancer (Takahashi and Sasaguri 2009; Gao et al. 2013). This target is unique in that it is constitutively active in cells and its inhibition is liable for cell signalling (Inoki et al. 2006). GSK-3β plays significant roles in numerous signalling pathways that regulate a variety of cellular processes (Xu et al. 2009; Cheng et al. 2011).

Melissa officinalis L. also known as lemon balm, a perennial herb in the family Lamiaceae (Fig. 1) occurs naturally in the Mediterranean and West Asia but is widely cultivated in Europe and North America (Moradkhani et al. 2010). Its leaf contains several phyto-compounds, such as flavonoids, polyphenolic compounds, monoterpenoid aldehyde, tannins, monoterpene glycosides, triterpenesesquiterpenes and essential oils (Sofowora et al. 2013). The usage of M. Officinalis as a supplement ingredient and functional food has increased over time due to its many medicinal properties including sedative, carminative and antispasmodic effects (Ożarowski et al. 2016). Lemon balm leaf, plant, and essential oil are used in herbal medicine (Senderski 2009).

Fig. 1.

Aerial parts of Melissa officnialis (Zarei et al. 2015)

A number of researches have been geared towards the discovery and design of selective GSK-3β inhibitors. Some identified GSK-3 inhibitors include small molecules isolated from organic and aquatic sources or obtained from chemical synthesis. They can function through numerous mechanisms, including competitive or non-competitive ATP inhibition. Attempts have been made to evolve and develop novel GSK-3β inhibitors in academia and industry (Xie et al. 2017). Many chemical families are known to emerge as GSK-3 inhibitors, with great structural diversity. Therefore GSK-3 is regarded as an ideal target for new drug discovery.

Molecular docking is a very fitting and low-cost method to understand the reaction mechanism of proteins or enzymes with ligands with high accuracy for rational drug design and discovery by analyzing the conformation and orientation of molecules into a molecular target binding site (Liu et al. 2018; Kitchen et al. 2004). As a result of the development of the first algorithms in the 1980s, molecular docking is now an important tool in drug discovery (Meng et al. 2011), and the most frequently used technique to predict the binding orientation of potential drugs against protein targets (Kapetanovic 2008). Hence, this study aimed at the use of molecular docking to predict the most potent Glycogen synthase 3-beta (GSK-3β) inhibitors with drug-like properties from M. officinalis.

Materials and methods

Preparation of crystal protein

The crystal structure of GSK-3β (PDB ID-1UV5) was downloaded via the protein preparation wizard of maestro v11.8. The protocol described in our previous studies was used to prepare crystal structure of the protein (Iwaloye et al. 2020a, b). The protein was preprocessed by creating zero bonds to metals, deleting waters from 5.0 Å of het groups, adjusting bond orders and setting the het states at pH 7.0 ± 2.0 (Schrödinger Suite 2012). The protein was refined by optimizing the H-bond network using PROPKA and removing water molecules with less than 3 H-bonds to non-waters (Olsson et al. 2011). The retrained minimization was carried out using the OPLS3 force field to avoid steric clashes that may exist in the structure. The minimization was terminated while the RMSD of non-hydrogen atoms reached 0.30 Å.

Preparation of the phyto-compounds

The library of compounds was built by drawing characterized compounds of Melissa officinalis using Marvinsketch (version 19.26) as documented in a different literature review (Triantaphyllou et al. 2001; Patora and Klimek 2002; Heitz et al. 2000; Tagashira and Ohtake 1998). The compounds were prepared using Ligprep. The Ligprep panel enables the conversion of structures; generate variations of structures and elimination of unwanted structures. After stereoisomer computation was left to generate at most 32 per ligand and the output format was left as maestro, the OPLS3 force field was left at pH 7.0 ± 2.0 using epik (Schrödinger Suite 2012).

Receptor grid generation

The receptor grid file was generated using a receptor grid generation panel, which represents the active sites of the receptor for glide ligand docking jobs. The ligand-binding site was defined by picking the co-crystallized ligand of the protein structure on the workspace. The van der Waals radii of the receptor atoms with partial atomic charge was set scaling factor of 1.0 and partial cutoff of 0.25 to soften the potential for non-polar parts of the receptor. The receptor grid box resolution was centered at coordinates 93.91, 68.0 and 9.8 in respect to x, y, and z-axis.

Glide extra precision docking

The prepared library of compounds was docked into the active site of the protein crystal using extra precision with the ligand sampling set generated as flexible. The choice of the best-docked structure for each ligand was made using model energy score (emodel) that combines glide score, the non-bonded interaction energy and the excess internal energy of the generated ligand conformation.

Induced fit docking (flexible docking)

To accurately predict the binding affinity of the novel inhibitors to the prepared protein crystal, Induced fit docking (IFD) was implemented. IFD is an in silico approach that uses Glide and the Refinement module in Prime that accurately predicts ligand binding modes and concomitant structural changes in the receptor (Sherman et al. 2006).

Calculation of binding free energy

The Prime MM-GBSA panel was used to calculate binding free energy for the ligand-receptor complex using the MM-GBSA technology available with Prime (2018). MMGBSA quantifies the difference in energy between the free and the complex state of both the ligand and the protein after energy minimization. In the prime MM-GBA panel, the OPLS3 force field was selected and VSGB was used as the continuum solvent model. Other options were set to default.

The equations for calculating binding energy are as follows.

$$\mathrm{\Delta }\text{G bind}=\mathrm{\Delta }\text{E}+\mathrm{\Delta }\text{Gsolv}+\mathrm{\Delta }\text{GSA}$$ 1

$$\mathrm{\Delta }\text{E}=\text{Ecomplex}-\text{E protein}-\text{Eligand}$$ 2

where Ecomplex, Eprotein, and Eligand indicate the minimized energies for protein-inhibitor complex, protein, and inhibitor, respectively.

$$\mathrm{\Delta }\text{Gsolv}=\mathrm{\Delta }\text{Gsolv}\left(\text{complex}\right)-\mathrm{\Delta }\text{Gsolv}\left(\text{protein}\right)-\mathrm{\Delta }\text{Gsolv}\left(\text{ligand}\right)$$ 3

$$\mathrm{\Delta }\text{GSA}=\mathrm{\Delta }\text{GSA}\left(\text{complex}\right)-\mathrm{\Delta }\text{GSA}\left(\text{protein}\right)-\mathrm{\Delta }\text{GSA}\left(\text{ligand}\right)$$ 4

where ∆GSA is the non-polar contribution to the solvation energy due to the surface area. GSA (complex), GSA (protein) and GSA (ligand) are the surface energies of complex, protein and ligand respectively.

Drug-likeness test of the phytochemicals

Lipinski's rule of five was used to determine the drug-likeness of the compounds and this parameter was predicted by Canvas (Duan et al. 2010a, b).

Validation of molecular docking results

The protocol for docking in this study was validated by docking the prepared inhibitors of GSK-3β downloaded from database server of the CHEMBL. Extra precision (XP) docking score of selected compounds were plot against their pChEMBL value to obtain r² spearman correlation. The plotted graph is illustrated in Fig. 2. The docking protocol was further validated by docking native ligand (co-crystal ligand) with the prepared crystal structure of GSK-3β to determine the root mean square deviation (RMSD). A RMSD value of 0.39 Å (Fig. 3) showed the docking procedure is reproducible (Elekofehinti et al. 2018).

Fig. 2.

The correlation coefficient graph between the experimentally determined pIC₅₀ of GS3K-3β their docked score with R² of 0.88 indicating that the docking experiment can replicate the experimentally determined values of the inhibitors

Fig. 3.

Superposition of the co-crystal ligand with its docked pose

Machine learning principles using automated QSAR

Generation and preparation of dataset

The experimental dataset containing GSK-3β inhibitors were retrieved from CHEMBL database online server (www.ebi.ac.uk/chembl/), by blasting the FASTA sequence of the GSK-3β with online sever (CHEMBL262). Bioactivities of 116 inhibitors of GSK-3β were retrieved with their respective pIC50 were compiled from literatures Olesen et al. (2003), Terrence et al. (2006), Saitoh et al. (2009), Luo et al. (2016) and Cociova et al. (2017). Compounds without pIC₅₀ were deleted from excel sheet before conversion to sdf (structure data file) format using data-warrior package (v.2) (Sander et al. 2005). The sdf format was exported to the workspace of maestro for preparation by ligprep (Schrödinger Suite 2012). The prepared compounds were eventually exported to Canvas cheminformatics (Duan et al. 2010a, b). Canvas clusters the inhibitors based on their Tanimoto similarity between sets of Hashed linear binary fingerprint descriptors, to determine the structural diversity among the inhibitor, and to select representatives from each resulting cluster. This computational study generated a total of 49 clusters and one representative was selected from each cluster to develop our QSAR model.

Principle of autoQSAR

AutoQSAR is a machine-learning algorithm provided by Schrödinger suite that builds and applies QSAR models through automation (Dixon et al. 2016). In order to build a predictive model, AutoQSAR takes the 1D, 2D and 3D structural data of a molecule along with a property (eg: IC₅₀) to be modeled, as an input. It will then compute the fingerprints and descriptors using machine-learning statistical methods for creating a predictive QSAR model. The predictive accuracy of the model is evaluated using various parameters such as ranking score, Root Mean Square Error (RMSE), Standard Deviation (SD), Q² and R² values (de Oliveira and Katekawa 2017).

Validation methods for QSAR models

External validation

The predictive power of a QSAR model can be estimated by the following statistical characteristics of the test set which was recommended by Golbraikh and Tropsha (2002):

• i.Correlation coefficient R between the predicted and observed activities
• ii.Coefficients of determination (R²) (predicted vs. observed activities ${\text{r}}_2^0$, and observed vs. predicted activities ${\text{r}}_2^{0\prime }$) (Sachs 1984).
• iii.Slopes k and k′ of the regression lines through the origin.

A model is considered robust if it meets the following criteria (Golbraikh and Tropsha 2002):

$$\begin{array}{c}{R}_{\text{pred}}^2>0.6,\text{r2}-{\text{r}}_0^2/\text{r2}<0.1,\\ {\text{r}}^2-{\text{r}}_{0\prime }^2/{\text{r}}^2<0.\text{1 and}\\ 0.85<\text{k}<1.\text{15 or}0.85<{\text{k}}^{\prime }<1.15\\ \end{array}$$

Results and discussion

Molecular docking studies

In the context of therapeutic application, GSK-3β has become an interesting drug target due to its exclusive and central role in the pathogenesis of varieties of disease. Several studies have documented that over-expression of Glycogen synthase kinase-3β accounts for memory impairment/increased β-amyloid production, diabetic type II, stroke, cancer and chronic inflammatory disease (Eldar-Finkelman 2002; Beurel 2011; Liu 2014).

An extensive literature survey was done to identify and select phyto-compounds against GSk-3β based on their medicinal properties for drug designing using molecular docking studies. Melissa officinalis became the ideal plant due to its medicinal strength (Kamdem et al. 2013; Ammon et al. 2006) and the small molecular weight of its photo-constituents. The present study utilized glide XP docking and induced fit docking to painstakingly and accurately predict the binding affinity and docking score of the compounds with GSk-3β, thereby denoting compounds with favourable interaction. Table 1 presented the molecular docking results and the interacting residues of the respective ligand–protein complex. The docking score of the hit compounds ranged from − 7.289 to − 17.284 kcal/mol. Luteoin 3′-O-β-d-glucoronopyranoside(VI) attained the highest binding affinity with a score of − 17.284 kcal/mol. The next ranked compounds are luteoin 7-O-β-d-glucoronopyranoside(V), chlorogenic acid and salvianolic acid F with docking scores of − 17.199 kcal/mol, − 15.650 kcal/mol and − 14.285 kcal/mol respectively. However, the co-crystallized, our ligand choice of comparison attained the least docking score (− 4.638 kcal/mol). This suggests that the hit compounds derived from Melissa officnialis are promising agents as GSk-3β inhibitors. Induced fit docking (IFD) offered a more accurate prediction of binding affinity by allowing the protein to undergo rotation on binding to a ligand. The results of the IFD score of the hit compounds did not follow the same trend with the docking score. While luteoin 3′-O-b-d-glucoronopyranoside(VI) still retained the most favourable interaction with the IFD score of − 716.889 kcal/mol, Naringenin and apigenin have more favourable interaction than the other compounds by recording IFD score of − 716.819 kcal/mol and − 715.304 kcal/mol respectively.

Table 1.

Molecular docking results and Interacting residues of the respective ligand–protein complex

Compounds	Docking score (kcal/mol)	Induced fit docking score (kcal/mol)	H-bond	H-bonding distance (Å)	Predicted pIC50
Luteoin 3′-O-β-d-glucoronopyranoside(VI)	− 17.284	− 716.889	VAL135, ASN64, 2[LYS183], GLY68	2.50, 1.68 1.78, 2.91 2.72	6.452
Luteoin 7-O-β-d-glucoronopyranoside(V)	− 17.199	− 713.219	PRO136, ILE62 2[LYS183] ASN64, ASN186	2.21, 2.02 4.42, 4.51 1.95, 1.75	6.415
Chlorogenic acid	− 15.650	− 709.342	ASP200, LYS183, ASN64 2[VAL135]	2.02, 4.15, 2.29 1.82, 2.28	6.643
Salvianolic acid F	− 14.285	− 708.102	PHE67, VAL135	1.77, 2.22	6.453
Salvianolic acid C	− 14.140	− 711.207	2[LYS183], ASN186, 2[VAL135] ARG141	1.66, 3.76, 2.23 2.01, 2.09 2.41	7.770
Salvianolic acid A	− 13.213	− 709.096	ILE62, PRO136	2.09, 2.16	6.744
Melitric acid A	− 12.445	− 710.219	VAL135, GLN185 2[LYS183] 2[LYS85], ASN64	2.08, 2.99, 3.64 2.02, 2.65 4.85	6.427
Yunnaneic acid	− 12.173	− 709.805	ASN64, LYS183 2[GLN185] LYS85	1.77, 2.95 1.95, 1.86 3.22	6.190
Ellagic acid	− 10.439	− 705.418	LYS85, VAL135	2.27, 1.83	8.169
Naringenin	− 8.182	− 716.819	ASP133, PRO136	1.99, 1.93	7.850
Decadienal	− 7.289	− 705.512	VAL135	1.94	4.178
Rosmarinic acid	− 9.465	− 712.618	ASN64, VAL135, ASP133	2.92, 2.61, 2.22	6.855
Apigenin	− 9.671	− 715.304	ASP133, PRO136	1.88, 1.99	6.811
Co-cyrstallized ligand	− 4.638	− 704.432	ASN64, LYS183, ASN186, GLN185	2.24, 5.38, 1.80, 1.84	7.487

Hydrogen bonding interaction

The binding site of GSk-3β is known to contain a group of polar residues such as LYS85, ASP200, and GLU51 that play a leading role in the ligand-ATP recognition; ASP200 specifically interacts with the phosphate group of ATP (De Bondt et al. 1993). The interacting residues of the protein with lead compounds were listed in Table 1. Our results revealed that few of the inhibitors were involved in hydrogen bond interaction with LYS85 and ASP200. Yunnaneic acid, ellagic acid, and melitric acid A formed hydrogen interaction with LYS85. Consequently, chlorogenic acid appeared to form interaction with ASP200. A similar interaction with potential inhibitors was also reported by Padavala et al. (2010). Several studies have recognized VAL135 and ASP133 as key residues for H-bond interaction with a diverse range of GSk-3β inhibitors (Smith et al. 2001; Padavala et al. 2010), in addition to GLN185, LYS183, ILE62, ASN186 and ARG141 (Witherington et al. 2003; Bertrand et al. 2003; Buescher and Phiel 2010). Interestingly all the lead compounds showed interaction with most of these amino acid residues. Figures 4, 5, 6 depicted the docking conformation of the compounds with the three most favorable interactions with GSk-3β. The residues present in the active site of GSk-3β interacted with luteoin 3′-O-β-d-glucoronopyranoside(VI) were VAL135, ASN64, 2[LYS183], and GLY68, with the hydrogen bonding distance between the compound and GSk-3β were found to be 2.50 Å, 1.68 Å, 1.78 Å, 2.91 Å and 2.72 Å. Naringenin showed interaction with ASP133 and PRO136 using the hydroxyl group attached to its phenyl rings. The hydrogen-bonding distance between naringenin and GSk-3β is estimated to be 1.99 Å and 1.93 Å.

Fig. 4.

Binding pose of luteoin 3′-O-β-d-glucoronopyranoside(VI) with GS3K-3β revealing interacting amino acid residues within the active site of GS3K-3β in 2D

Fig. 5.

Binding pose of Naringenin with GS3K-3β revealing interacting amino acid residues within the active site of GS3K-3β in 2D

Fig. 6.

Binding pose of Apigenin with GS3K-3β revealing interacting amino acid residues within the active site of GS3K-3β in 2D

Binding energy assessment

To validate the docking scores, the free energy of binding was calculated via the MMGBSA post docking program, which predicts binding free energies for compounds/ligands by utilizing the combination of molecular mechanics calculations and salvation models. It has been demonstrated in many studies that the MMGBSA post docking method is the most reliable for rating the affinity of a ligand on binding to its protein target (Maffucci et al. 2018; Sun et al. 2014) since results obtained through MMGBSA for binding energies calculations were found to be highly reproducible (Genheden and Ryde 2015). The accuracy of the docking was affirmed by examining the lowest energy poses predicted by the scoring function. The evaluation of binding free energy is listed in Table 2. Interestingly, compounds with the good docking score showed favorable binding energy. In terms of binding free energy, the major energy contributors were identified as van der Waals (∆Gvdw), Coulomb interaction (∆GColulomb), Hydrogen bond (ΔGHbond) and lipophilic energy (∆GsolLipo) that enhance the binding affinity of the compounds towards the binding pocket of the protein.

Table 2.

Calculation of binding free energy

Compounds	ΔGBinda	ΔGBindbcoulomb	ΔGBindcvdw	ΔGBinddlipo	ΔGBindeHbond
Luteoin 3′-O-β-d-glucoronopyranoside(VI)	− 36.58	− 169.83	− 28.09	− 12.62	− 3.64
Luteoin 7-O-β-d-glucoronopyranoside(V)	− 55.08	− 193.71	− 46.00	− 14.57	− 3.26
Chlorogenic acid	− 59.06	− 184.44	− 21.64	− 14.12	− 4.40
Salvianolic acid F	− 42.75	− 156.21	− 30.74	− 15.54	− 2.19
Salvianolic acid C	− 40.79	− 317.24	− 48.20	− 14.76	− 3.57
Salvianolic acid A	− 58.63	− 181.40	− 42.68	− 13.74	− 2.12
Melitric acid A	− 31.41	− 157.46	− 46.40	− 14.60	− 3.97
Yunnaneic acid	− 38.46	− 209.02	− 47.16	− 14.92	− 2.78
Ellagic acid	− 51.04	− 23.18	− 35.39	− 15.55	− 2.45
Naringenin	− 45.00	− 26.40	− 37.05	− 9.82	− 1.12
Decadienal	− 37.36	− 11.50	− 20.55	− 16.15	− 0.97
Rosmarinic acid	− 7.04	− 107.38	− 32.52	− 16.55	− 1.91
Apigenin	− 44.98	− 26.40	− 37.05	− 9.82	− 1.12
Co-cyrstallized Ligand	− 31.11	− 66.33	− 31.46	− 11.73	− 1.72

^aMM-GBSA free energy (kcal/mol) of binding

^bContribution to the MM-GBSA free energy of binding (kcal/mol) from the Coulomb energy

^cContribution to the MMGBSA free energy of binding (kcal/mol) from the van der Waals energy

^dContribution to the MM-GBSA free energy of binding (kcal/mol) from lipophilic binding

^eContribution to the MM-GBSA free energy of binding (kcal/mol) from hydrogen bonding

ADME studies

The predicted ADME properties (Table 3) include a number of rotatable bonds, the molecular weight of the molecule, number of hydrogen bond acceptors, prediction of binding to human serum albumin, number of hydrogen bond donors, predicted octanol–water partition coefficient and number of violations of Lipinski's rule of five (RO5). Lipinski's RO5 helps to evaluate the drug-likeness, and determine the prospect of small molecules in becoming an orally active drug for humans. The rule permits a molecular weight < 500 Da, octanol–water partition coefficient < 5, hydrogen bond donor ≤ 5 and hydrogen bond acceptor ≤ 10 (Lipinski et al. 2001). Prospective drug candidate that obey the Ro5 tend to have lower attrition rates at the stage of clinical trials and for this reason, it has an increased chance of becoming and staying marketable (Gombar et al. 2003). Compounds of Melissa officinalis have shown excellent results and are in accordance with this rule. In view of this, they can be developed as a promising lead in the design of GSk-3β inhibitors.

Table 3.

Prediction of ADME properties

Compounds name	aROF	bMW	cAlogP	dHBA	eHBD	fRB	gPSA
Luteoin 3′-O-β-d-glucoronopyranoside(VI)	0	460.344	− 0.359	9	5	4	213.010
Luteoin 7-O-β-d-glucoronopyranoside(V)	1	461.353	0.318	10	6	4	210.180
Chlorogenic acid	0	353.301	− 1.200	7	5	5	167.580
Salvianolic acid F	0	329.324	1.172	4	4	5	121.050
Salvianolic acid C	0	476.389	2.552	7	4	7	183.550
Salvianolic acid A	0	493.439	2.704	7	5	9	190.640
Melitric acid A	1	536.441	2.187	8	5	11	216.940
Yunnaneic acid	0	359.307	1.795	6	4	7	147.350
Ellagic acid	0	301.185	0.744	8	4	0	141.340
Naringenin	0	272.253	1.620	4	2	1	89.820
Decadienal	0	150.218	3.243	1	1	1	20.230
rosmarinic acid	0	359.307	1.795	6	4	7	147.350
Apigenin	0	271.245	1.620	4	2	1	89.820
Co-cyrstallized ligand	0	376.332	− 1.167	2	4	1	85.710

^aLipinski rule of five

^bMolecular weight

^cPredictedoctanol/water partition coefficient

^dHydrogen bond acceptor

^eHydrogen bon donor

^fRotatable bond

^gPolar surface area

Automated QSAR analysis

The screening process further continued with the aid of machine learning-based predictive model (pIC₅₀ Calculation) generated by AutoQSAR panel of Schrodinger. Given a learning set of chemical structures and an activity property from CHEMBL database, a total of 497 physicochemical and topological descriptors are computed, together with a variety of Canvas fingerprints (Dixon et al. 2016), giving out a large pool of independent variables from which to build models. The automated module split the dataset randomly into 80% training set, and 20% test set. Models are built on each training set from all possible combination of machine learning method, and sets of independent variables that are supported by each machine learning methods. The observed activities and predicted activities of training set and test set in negative logarithm of inhibitor concentration(pIC₅₀) was represented in Table 4. The algorithm generated 10 best models and the results of the top 5 models are shown in Table 5. The best model Model kpls_desc_38 recorded a standard deviation (S.D) of 0.5505, R² of 0.8467, root mean square error (RMSE) 0.5366 and Q² of 0.8069. The best model was computed from kernel based partial least square regression (KPLS), which supports the use of descriptors and fingerprints as independent variables, and fingerprint desc_38, a model code generated from a combination of the machine learning method (MLM). The scatter plot depicting predicted pIC₅₀ versus experimental pIC₅₀ for best generated model is shown in Fig. 7. The Predicted pIC₅₀ using the best model for the data set of the lead compounds and co-ligand are tabulated in Table 1. It is worth noting that three of the compounds are observed to have better predicted pIC₅₀ than co-ligand.

Table 4.

Details of AutoQSAR predicted activities compared with the observed activities

S/n	CHEMBL CID	Set	Observed pIC50	Predicted pIC50	Residue error
1	CHEMBL68397	Train	7.4000	6.6281	− 0.7719
2	CHEMBL67799	Train	7.8000	7.6181	− 0.1819
3	CHEMBL76427	Train	4.7500	4.3437	− 0.4063
4	CHEMBL76427	Train	4.7500	4.3437	− 0.4063
5	CHEMBL250714	Train	6.0600	5.9896	− 0.0704
6	CHEMBL109944	Train	7.3500	7.7096	0.3596
7	CHEMBL111620	Train	7.5800	7.5083	− 0.0717
8	CHEMBL112564	Train	5.1600	5.9676	0.8076
9	CHEMBL1957099	Train	7.2000	6.2389	− 0.9611
10	CHEMBL1957078	Train	6.5000	6.1836	− 0.3164
11	CHEMBL1957942	Train	5.5500	6.0766	0.5266
12	CHEMBL326208	Train	7.8500	7.6141	− 0.2359
13	CHEMBL497398	Train	4.6000	5.0189	0.4189
14	CHEMBL188938	Train	8.3000	7.3258	− 0.9742
15	CHEMBL576540	Train	7.0900	7.2598	0.1698
16	CHEMBL183171	Train	6.2600	6.6905	0.4305
17	CHEMBL1834116	Train	5.4700	5.6292	0.1592
18	CHEMBL1834122	Train	4.7700	4.8932	0.1232
19	CHEMBL1957096	Train	4.5000	5.5342	1.0342
20	CHEMBL1957089	Test	5.6000	5.9919	0.3919
21	CHEMBL1940983	Train	6.7500	7.0294	0.2794
22	CHEMBL1801622	Train	5.8200	6.2720	0.4520
23	CHEMBL1801637	Train	8.3000	8.5850	0.2850
24	CHEMBL399078	Test	5.3400	6.1300	0.7900
25	CHEMBL398670	Train	6.0400	5.5198	− 0.5202
26	CHEMBL249141	Train	5.4500	5.6634	0.2134
27	CHEMBL426587	Train	5.3200	5.8617	0.5417
28	CHEMBL461139	Train	6.5200	6.6372	0.1172
29	CHEMBL1801625	Train	5.7000	5.6008	− 0.0992
30	CHEMBL251886	Test	6.5000	5.7403	− 0.7597
31	CHEMBL179725	Test	7.8900	7.3075	− 0.5825
32	CHEMBL458210	Train	6.2700	6.2182	− 0.0518
33	CHEMBL270265	Train	9.1500	10.1193	0.9693
34	CHEMBL156987	Train	7.1200	7.2798	0.1598
35	CHEMBL408564	Train	8.1500	6.9164	− 1.2336
36	CHEMBL270473	Train	8.8200	8.0888	− 0.7312
37	CHEMBL576165	Train	7.1600	6.6797	− 0.4803
38	CHEMBL573933	Test	6.8000	6.4804	− 0.3196
39	CHEMBL334084	Train	6.6200	6.8859	0.2659
40	CHEMBL1081616	Test	7.5100	7.1326	− 0.3774
41	CHEMBL471430	Test	8.3100	7.7477	− 0.5623
42	CHEMBL470566	Train	7.0300	6.7009	− 0.3291
43	CHEMBL513576	Test	6.1700	6.5270	0.3570
44	CHEMBL26241	Train	5.1500	5.6899	0.5399
45	CHEMBL26101	Train	4.7000	4.5452	− 0.1548
46	CHEMBL142327	Train	4.1200	4.0730	− 0.0470
47	CHEMBL359135	Test	4.3000	3.8430	− 0.4570
48	CHEMBL596364	Train	4.1200	4.7567	0.6367
49	CHEMBL405759	Train	7.5200	7.0731	− 0.4469

Table 5.

Parameters corresponds to five best model generated by AutoQSAR

Model code	Score	SD	R2	RMSE	Q2
Kpls_desc_38	0.8224	0.5505	0.8467	0.5366	0.8069
Pls_38	0.7846	0.6662	0.7756	0.5444	0.8013
Kpls_radial_31	0.7652	0.6530	0.7631	0.6276	0.7458
Kpls_dendritic	0.7444	0.6250	0.7830	0.6427	0.7334
Kpla_linear_31	0.7058	0.6033	0.7978	0.6639	0.7156

Fig. 7.

Scatter plot analysis of best model predicted from AutoQSAR

External validation of QSAR model

Validation methods are required to affirm the robustness of a model on unseen data. The method of root mean-squared error (RMSE) is one of the internal methods of validating a model (Wold and Ericksson 1998; Yasri and Hartsough 2001).

Strategies for external validation are important and it is of great interest to adopt all available validation strategies to check robustness of the model. All the parameters for external validation of structure based pharmacophore model are presented in Table 6. Cross validation (Q²) value of 0.8069, the correlation coefficient (R²) value of 0.8467. The slopes of regression lines through origin (K and K₀ value) and substantial values of correlation coefficients (${\text{R}}_0^2$ and ${\text{R}}_{0\prime }^2$ ) were obtained from Observed pIC₅₀ and Predicted pIC₅₀ activity of the dataset. The predictive ability of the selected model was also confirmed by external. A value of $r_{\text{pred}}^2$ is greater than 0.6 may be taken as an indicator of good external predictability. All these values met the necessary criteria for a robustness of QSAR model.

Table 6.

External validation parameters for QutoQSAR

External validation parameters	Model kpls_desc_38	Limitations
Q2a	0.8069	Q^2 > 0.5
R2b	0.8467	R^2 close to 1
K valuec	0.9444	0.85 ≤ k ≤ 1.15
R02d	0.8412	Close to R^2
K’ valuee	1.0040	0.85 ≤ k ≤ 1.15
R0′2f	0.8412	Close to R^2
Rm(Loo)2 g	0.7763	Rm(Loo)2 > 0.5
rpred2h	0.6979	${\text{r}}_{\text{pred}}^2$ > 0.5

^aCross-validated coefficient

^bCorrelation coefficient between actual and predicted values

^cSlope values of regression lines

^dCoefficient for regression through origin values

^eSlope values of regression lines

^fCoefficient for regression through origin values g. modified squared correlation coefficient using LOO method

^hPredictive correlation coefficient value

Conclusion

Glycogen synthase kinase-3β is a drug target for Alzheimer's, type II diabetes and other diseases. This study shows the binding ability of a library of compounds generated from Melissa officinalisas potential inhibitors of GSk-3β. It is important to emphasize that three compounds which are salvianolic acid C, ellagic acid and naringenin are found to have better docking score, binding free energy and predicted pIC₅₀ values alongside satisfactory drug likeness than co-crystallized native compounds. Therefore, we recommend data provided in this study should further be validated by in vivo and in vitro studies.