The predicted models of anti-colon cancer and anti-hepatoma activities of substituted 4-anilino coumarin derivatives using quantitative structure-activity relationship (QSAR)

2.1 Hardware and software

The research used a PC with CPU @ 2.00 GHz RAM 4.00 GB Intel® Core™ i3-5005U. The softwares utilized were Hyperchem™ 8.0.10, Gaussian® 09 W, IBM® SPSS® Release 23, and ACD/ChemSketch 14.0. Overall research was being done at Austrian-Indonesian Centre for Computational Chemistry (AIC) Laboratory, Department Chemistry, Faculty of Mathematics and Science, Universitas Gadjah Mada.

2.2 Materials

A data set of 27 compounds (Luo et al., 2017) was used to generate the QSAR equations. The anticancer activity data were set as dependent variables (IC₅₀) to HCT116 and HepG2 (μM) which were modified into logarithmic.

2.3 Method of calculation

Compound 33d as the most-desiring-activity compound was computed using PM6, DFT/BPV86, DFT/PBE, DFT/B3LYP (Van Bay et al., 2021), and Hartree-Fock (HF) method (Hmamouchi et al., 2016). The Basis set applied for HF and DFT calculation was 6-31G. ¹³C NMR chemical shifts were calculated by DFT/B3LYP method and 6-31G basis set. The best method with low correlation error between experimental and calculated chemical shift was selected.

2.4 Geometrical optimization and descriptors calculation

The 27 compounds of substituted 4-anilino coumarin derivatives were depicted employing GaussView 5.0 and optimized by Gaussian 09W application. From the calculations, E_LUMO, E_HOMO, delta E_HOMO-E_LUMO (ΔE), atomic charge (q), and dipole moment (μ) were obtained. The rest descriptors, volume (V), molecular mass (MW), refractivity (MR), partition coefficient (log P), polarisability (α), and surface area (SA) were calculated using Hyperchem™ 8.0.10.

2.5 Model development and validation

QSAR equations were created by utilizing multiple linear regression (MLR) using IBM® SPSS® Release 23. The regression development method was backward elimination with confident interval level 95%. The internal validation should meet the significance test: r²_train > 0.6 (Golbraikh and Tropsha, 2000), SEE < 0.3 (Mishra et al., 2014), test F_cal/F_tab > 1, α(0.05) > sig(0.00) (Pramesti, 2017). The external validation used to eliminate the unqualified equations, r²_test > 0.6, r²_overall > 0.5, and RMSE_test < 0.3 are critical to be calculated (Golbraikh and Tropsha, 2000).

2.6 Drug design

The drug design was done by replacing the R₁, R₂, and R substituents from coumarin derivatives. Hydrophobic, electronic, and steric effects were the main properties to consider. To attain proper hydrophobic character, log P, R₁, R₂, and R of the newly designed molecules were replaced with certain moieties. To avoid unwanted value of Log P, Lipinski’s “rule of five” was implemented.

3.1 Method of calculation

The compound used was 33d that was 4-(1,3-benzodioxol-5-ylamino)-6-methoxy-3-(trifluoroacetyl)–2H-chromen-2-one as it had the pre-eminent anti-cancer activity among the listed derivatives (Luo et al., 2017). In Fig. S1 (see the figure in Supplementary material), HF/6-31G method performs significant inaccuracy to predict the ¹³C NMR spectra than DFT and semiempirical method. High degree of error is proved by its high RMSE calculation which is 15.79 and low linearity between experimental and calculated chemical shifts was showed by r² (0.9785) (Putri et al., 2019). This lack might due to this method’s limitation to carry the electron correlation in calculation process (Young, 2001). Electron correlation is important to improve the accuracy of computed energies and molecular geometries (Piris, 2017).

Semiempirical method (PM6 method) has better result than HF method expressed by higher r² that is 0.9908. But, this method gives higher RMSE value (13.08) compared with several DFT methods indicates the higher error. The sp³ hybridization of nitrogen in the derivatives can be the main cause of false prediction from PM3 or AM1 calculation (Clark, 1993).

From Fig. S1, it can be seen that the three closest theoretical chemical shifts to the experiment are all from DFT methods (BPV86, PBE, and B3LYP). DFT functionals verily rely on organic compounds molecular structure (Van Bay et al., 2021). The determination coefficient of the three DFT functionals are 0.9834 (BPV86) (Putri et al., 2019), 0.9825 (PBE), and 0.9812 (B3LYP) indicating the pleasant coherence between theoretical and experimental chemical shifts. B3LYP RMSE calculation gives higher value comparing with PBE and BPV86. Despite, PBE method provides negligible difference in RMSE calculation (9.52) with BPV86 (9.27) (Putri et al., 2019) but its optimizing process is significantly more time consuming. Therefore, DFT/BPV86 using 6-31G basis set is the best method which offers the best prediction for the independent variables to be more accurate. In addition, BPV86 is included as hybrid GGA (generalized gradient approximation) functional (Sajjad et al., 2017) which integrate Becke and Perdew’s 1986 functional density with electron correlation replaced with Vosko.

3.2 Geometrical optimization and descriptors calculation

Generating QSAR equations always charge high attentions in the entire process. The Hartree-Fock complicity in Hybrid density functional (H-GGA) methods upgrades a conventional GGA used in this research (BPV86) into the next level sophistification for the derivatives. H-GGA let a percentage of Hartree-Fock exchange to be fitted semiempirically with experimental ionization potentials, total atomic energies, proton affinities and other data representating a group of small molecules (Becke, 1993).

The data of several distance modifications between two atoms are performed in Table S1 (see the table in Supplementary material). The distance between C₇ and O without optimization is quite narrow (1.338 Å). Following the optimization, the distance becomes wider (1.412 Å) placing the atoms into more relax position and reduce the high level energy.

3.3 Model development and validation

The QSAR equations robustness is predicted by calculating the capability of the training set equation to predict log IC₅₀ of the test set. Backward elimination has become the most promising method to progressively eliminate the descriptors and also gives better performance to correlation coefficient compared with forward selection and stepwise regression. In the end of the process, only a few descriptors are left in order to reduce redundant, the risk of overfitting from noisy and irrelevant descriptors.

3.4 Model development and validation of anti-colon cancer (anti-HCT)

The best five models resulted from various training and test set packs of anti-colon cancer activity are listed in Table S2 (see the table in Supplementary material). The lowest r_train value is 0.872 and the lowest r²_train is 0.760. It explains the goodness of statistical significance that is>87.2% and the r²_train reveals the closeness of the data to the fitted regression. This study provided the lowest F_cal/F_tab values of all models is 3.588 explaining a good statistically significant. Every model is clustered as having good accuracy in predicting Log IC₅₀ of anti-HCT.

The external validation was done. In Table 1, the data which satisfied the standard were merely model 4. Model 4 presented r²_test 0.8307, with RMSE 0.177, and PRESS 0.157. The PRESS and RMSE value of this model was occupied as the smallest value considering that model 4 gave the least error. In addition, the accuracy of the model 4 was described by correlation coefficient (r = 0.882) that meant statistical significance>88% (Fig. S3, see the figure in Supplementary material). Therefore, model 4 was considered as a validated, robust, and predictive model. (Model 4, Table S2)

The descriptors that affect the molecule's activities are C15 and C17 of anilino group and Log P (Eq. 1, Fig. 3). This discovery is suitable with molecular docking of Cao et al. which shows that anilino forms hydrophobic interactions with amino acids residues Ala316 (Cao et al., 2016). The atomic charge shift significantly changes the interaction ability of this ring with its receptor. Another critical factor was the solubility. Unsatisfactory log P of a drug may cause it never reach the receptor. The calculated log IC₅₀ values of model 4 of all compounds were compared with the experimental activities giving a slope of 0.7559 that shows that this model is able to favourably predict 75.59% of the data (Fig. 1).

Close

Fig. 1

The correlation of experimental and predicted Log IC₅₀ of anti-HCT training and test set.

Export to PPT

3.5 Model development and validation of Anti-Hepatoma activity (anti-HEP)

Table S3 (see the table in Supplementary material) provides the best models of various training and test sets for anti-hepatoma activity resulted by SPSS software. The lowest r values is 0.845 indicating a good statistical significance which is >84.5%. Every model in Table S3 shows a very good determination coefficient that the lowest is 0.747. The other case, the whole ratio values of F_cal/F_tab show numbers between 3.588 and 7.400 which meet the standard value. In addition, the highest SEE value among all candidate is 0.209 meaning the errors rate of all equation are quite small.

The external validation was done which model 1, 2, and 3, presented r²_test > 0.6 (Table 2). Yet, the only accepted RMSE value was belonged to model 2 that was 0.298. This model also had the smallest PRESS values (0.445). The both values meant that its error was the mildest among the rest equations. Despite, the r²_test (0.8223) occupied as the highest r²_test among others. In addition, model 2 has statistical significance >86% (Fig. S5, see the figure in Supplementary material). Therefore, model 2 was considered as a robust, validated and predictive model.

The descriptors that affect the molecules activities are C2, C17, C18 and Log P. C17 and C18 build the anilino ring (Eq. 2, Fig. 3). According to Cao et al., this anilino ring is very important for derivative’s anticancer activity due to its hydrophobic interaction with amino acid Ala316. (Model 2, Table S3)

Log P descriptor could be further modified to provide better anticancer activity. The picture above was the comparison between calculated log IC₅₀ from entire derivatives and observed values providing r² = 0.6922. This determination coefficient indicates that the model was able to favourably predict 69.22% of the data (Fig. 2).

Close

Fig. 2

The correlation of experimental and predicted Log IC₅₀ of anti-hepatoma training and test set.

Export to PPT

3.6 Drug design

Analyzing Eqs. 1 and 2, hydrophobic and electronic were the physicochemical properties that should be further discussed. The hydrophobic properties of a drug is vital to provide it’s easiness to cross cell membranes and in receptor interaction. Despite, a huge part of all living mass consists of water leading to every biochemical reactions are in aqueous phase. The new molecules should have certain moieties that alter the Log P becoming lower than the reference yet maximize its performance to cross lipid barriers. Therefore, Lipinski’s “rule of five” is also implemented to design new molecules.

A drug polarity is highly affected by the electronic. These characteristics give an effect on how easily a drug could pass cell membranes or how strongly it could bind to a receptor. According to Eqs. 1 and 2 (Fig. 3), the potential of the compounds is highly correlated with the partial atomic charge of the benzene rings. Hence, the moieties chosen to create new molecules with certain atomic charges. The overall result of new compound designs are presented in Fig. 4. The most promising structure for anti-coloncancer is structure 25 with a predicted IC₅₀ = 820 nM and for anti-hepatoma is structure 19 with predicted IC₅₀ = 440 nM.

Close

Fig. 3

Atomic charge points of each equation.

Export to PPT

Close

Fig. 4

New compound designs.

Export to PPT

The retrosynthesis of newly designed compounds 25 and 19 is depicted in Figs. 5 and 7 respectively. Referring to them, aniline from the lactone (coumarin) ring should be disconnected firstly. The disconnection gives aniline derivative and 4-chlorocoumarin (Godhani et al., 2015; Holm and Straub, 2011). The substitution is applied to be easily performed via chloride than hydroxyl due to hydroxyl’s poor leaving group ability (Luo et al., 2017).

Close

Fig. 5

Retrosynthesis of new designed compound 25.

Export to PPT

The pentagonal ring of 1,2,4-triazole can be made by nucleophilic substitution of thiocyanate into chloro-carbonyl moiety followed with cyclization and oxidation (Godhani et al., 2015). The carboxylation into coumarin ring cannot be easily made, therefore it needs an FGI from –CN. The last disconnection builds O₁-C₂ and C₃-C₄ in the cyclic shape of the coumarin ring. The lactone ring cyclization requires an aromatic o-hydroxy carbonyl compound and a two-carbon fragment (C₂ and C₄ builder) in Knoevenagel and Perkin reactions. Despite, the reaction of aldehyde and triazide chlorosilane (TACS) was suggested via the formation of siloxy azide, and subsequent formation of gem-diazoalkane to give later cyclization to tetrazole derivatives (El-Ahl et al., 1995). These probable synthesis routes of the proposed compound has been completely written as Figs. 6 and 8.

Close

Fig. 6

Reagents and conditions of new anti-coloncancer compound (25) synthesis:(a) NaH, DEC, 120 °C, 3 h; (b) HNO₃, H₂SO₄ (c) H₂, Pd, C, Ac₂O; (d) HCl, H₂O; (e) HCl, NaNO₂, Cu(I)CN; (f) NaOH; (g) SOCl₂; (h) NH₄SCN; (i) N₂H₄; (j) HNO₃, H₂O, 45 °C; (k) POCl₃, 60 °C, 2 h; (l) K₂CO₃, DMF, 75 °C, 2 h. (Luo et al., 2017; Godhani et al., 2015; Holm and Straub, 2011; Warren, 1982; Potts, 1961).

Export to PPT

Close

Fig. 7

Retrosynthesis of new designed compound 19.

Export to PPT

Close

Fig. 8

Reagents and conditions of new anti-hepatoma compound (19) synthesis: (a) NaH, DEC, 120 °C, 3 h; (b) ZnCl₂, CH₂O, HCl; (c) SiCl₄, NaN₃, CH₃CN, 25 °C, 12 h; (d) POCl₃, 60 °C, 2 h; (e) K₂CO₃, DMF, 75 °C, 2 h (Luo et al., 2017; El-Ahl et al., 1995; Warren, 1982).

Export to PPT

Cancer treatment like chemotherapy have major limitations. Chemotherapy strategy is powerful in killing cancer cells but also has curing problems caused by cytotoxic side effects on normal cells (Hendouei et al., 2019). This research has considered this risk by selecting the main molecule skeleton with low toxicity towards normal human cells (HUVEC). It is highly recommended to follow any cancer regimens regulation from the medication institution including any involvement of antipsychotic drug to relieve nausea and other side effect of the treatment. Furthermore, for those who are at high risk of chemotherapy it is recommended to consider colonoscopy for colon cancer screening and surveillance to detect the colonic lesions before receiving any treatments.