An original set of quinoxaline- and indole-based S1PR2 activators was obtained from the ChEMBL database (Mendez et al., 2018[22]) and was curated (Fourches et al., 2010[9]) by the following steps (i) removal of inorganic compounds (i.e., salts and mixtures), (ii) structural validation and cleaning, (iii) normalization of specific chemotypes, (iv) deletion of duplicates, and (v) final checking to obtain the final curated data set containing 11 active compounds along with their bioactivity values, represented by the half maximal effective concentration (EC₅₀) (Neubig et al., 2003[24]). EC₅₀ is a concentration that increases the S1P activity by 50 %. The EC₅₀ values were converted into pEC₅₀ values by taking the negative logarithm based 10 to normalize the data points. According to the compound's core structure, the dataset was separately preprocessed into two final datasets (i.e., scaffold A = 6 quinoxaline-based compounds and scaffold B = 5 indole-based compounds), Figure 2(Fig. 2).

Chemical structures of compounds in SMILES format were converted into MOL format using MarvinSketch (ChemAxon Ltd., 2018[5]). All structures were geometrically optimized to obtain the lowest energy conformers by density functional theory (DFT) computation with Becke's three-parameter Lee-Yang-Parr hybrid functional (B3LYP) and 6-31g(d) basis set using Gaussian 09 software (Frisch et al., 2009[11]). The optimized structures were subsequently used as input files for descriptor calculation.

Molecular descriptors are numerical values used to represent characteristics of the compounds in terms of connectivity, constitution, and physicochemical properties (Nantasenamat et al., 2009[23]). A set of 13 quantum chemical descriptors was calculated using an in-house developed script (i.e., mean absolute atomic charge (Q_m), molecular dipole moment (µ), electronegativity (χ), total energy (E_T), electron affinity (EA), ionization potential (IP), electron ionization (EI), highest occupied molecular orbital energy (HOMO), lowest unoccupied molecular orbital energy (LUMO), energy difference of HOMO  and LUMO (HOMO-LUMO_gap), absolute hardness (η), softness (S), and electrophilicity (ω). Additionally, the optimized structures were used as input files for calculation of molecular descriptor using Dragon 5.5 software (Mauri et al., 2006[20]) to obtain a set of 3,224 molecular descriptors, comprising 48 constitutional descriptors, 119 topological descriptors, 47 walk and path counts, 33 connectivity indices, 47 information indices, 96 2D autocorrelations, 107 edge adjacency indices, 64 Burden eigenvalues, 21 topological charge indices, 44 eigenvalue-based indices, 41 Randic molecular profiles, 74 geometrical descriptors, 150 RDF descriptors, 160 3D-MoRSE descriptors, 99 WHIM descriptors, 197 GETAWAY descriptors, 154 functional group counts, 120 atom-centered fragments, 14 charge descriptors, 29 molecular properties, 780 2D binary fingerprints, and 780 2D frequency fingerprints.

Descriptor selection is a process to select a set of informative descriptors from a large set of calculated descriptors to be used as final predictors. Correlation-based feature selection was initially performed to filter a set of descriptor variables that are highly correlated with bioactivity of the compounds. Pearson's correlation coefficient (r) values for each pair of descriptor value and bioactivity value (pEC₅₀) were calculated. Descriptors with |r| ≥ 0.8 were considered as highly correlated predictors and were selected for further selection process using stepwise MLR in PASW Statistics 18 software (SPSS Inc, 2009[40]) and M5 method in WEKA software (Frank et al., 2016[10]) to obtain a final set of informative descriptors. Finally, the values of selected descriptors along with the pEC₅₀ values were used to prepare final datasets for model construction.

Two final datasets were used as input files for construction of QSAR models by WEKA software 3.8.4 (Frank et al., 2016[10]) using MLR algorithm. The MLR model represents a linear relationship between multiple descriptors (X variables) and bioactivity (Y variable) as shown in equation (1).

Y = B₀ + ΣB_nX_n (1)

where Y is the pEC₅₀ of compounds, B₀ is the Y-intercept and B_n are the regression coefficients of descriptors (X_n).

The dataset was divided into training set and testing set by leave-one-out cross validation (LOO-CV). Since the datasets contain small samples with less than 50 compounds, the LOO-CV sampling method is reliable for model validation (Roy et al., 2015[36]). The training set is used to build the model whereas the testing set was used to validate the predictive performance of the built model. For each round, one sample was excluded from the whole dataset (N) to be used as a testing set (in which its activity was predicted using the trained model) whereas the remaining samples (N-1) were used as a training set to train and construct the model. The same process was repeated until every compound in the dataset was selected to be used as a testing set (Sammut and Webb, 2011[37]).

The constructed models were assessed for their predictive performance using two statistical parameters i.e., correlation coefficient (R) and root mean square error (RMSE). The first parameter reflects predictive correlation of the model whereas the second one represents predictive error (Rácz et al., 2015[34]).

To increase structural diversity, a set of 752 structurally modified compounds was rationally designed based on key descriptors presented in the constructed QSAR models. All newly designed compounds were undergone the same processes as their parent compounds (e.g., drawing, geometrical optimization, and descriptor calculation) to obtain their informative descriptor values, and their S1P activities (pEC₅₀ values) were subsequently predicted using the constructed models.

The newly designed compounds were predicted for their drug-likeness using in silico web-based tool SwissADME (Daina et al., 2017[6]). Chemical structures in SMILES format were uploaded for prediction. The drug-likeness of the compounds was assessed based on Lipinski's rule of five (LRo5).

Two final datasets of 6 quinoxaline-based (scaffold A) and 5 indole-based (scaffold B) S1PR2 activators were prepared for construction of two QSAR models, Figure 2(Fig. 2). Compounds were preprocessed to obtain their descriptor values. Bioactivity values were normalized into pEC₅₀ values. Descriptor selection was performed to select a final set of 6 informative descriptors for construction of scaffold A model (4 descriptors: R2v+, MATS5m, nR05, and Km) and scaffold B model (2 descriptors: H3e and E3v), Table 1(Tab. 1). Dataset and predicted activities of scaffolds A and scaffold B models are provided in Tables 2(Tab. 2) and 3(Tab. 3).

Two QSAR models were successfully constructed using MLR (i.e., scaffold A and scaffold B models). Two built models provided acceptable predictive performance as shown by their validated statistical values for both training and testing sets (R_tr = 0.9667-0.9997, RMSE_tr = 0.0046-0.0375, Q_cv = 0.7902-0.9989, and RMSE_cv = 0.0093-0.1057, Table 4(Tab. 4)). Comparative plots of experimental pEC₅₀ versus predicted pEC₅₀ values are provided in Figure 3(Fig. 3).

From scaffold A, four descriptors play roles as predictors for the activity of quinoxaline-based activators. According to regression coefficient, the most influential descriptor was noted to be a van der Waals volume descriptor (R2v+) followed mass descriptors (MATS5m and Km), and number of five-membered ring (nR05), respectively (Table 4(Tab. 4)). The high positive values of R2v+ and MATS5m together with low positive or high negative values of the nR05 and Km are required to provide potent activity (high pEC₅₀ value). For scaffold B, the van der Waals volume descriptor (E3v) also played major influence on S1PR activating effect, followed by the electronegativity (H3e) descriptor (Table 4(Tab. 4)). Herein, both predictors possessed positive regression coefficient (R) values suggesting their positive effects on the pEC₅₀ values.

Key descriptors of the constructed models were further used to guide the design of new derivatives. Modification strategies are conceptually depicted in supplementary information section 1.2 and 2.2 (Tables S2 and S5). Finally, two sets of structurally modified compounds were designed (i.e., scaffold A = 635 compounds, scaffold B = 117 compounds). Chemical structures of the newly designed compounds and their predicted pEC₅₀ values are provided in supplementary data and supplementary information, Figures S1-S11. In overview, most of the modified compounds exhibit improved predicted activity when compared to their parents.

Understanding structure-activity relationship (SAR) is essential for successful drug development (Macalino et al., 2015[18]; Roy et al., 2015[35]). Herein, SAR analysis was performed to reveal key features influencing S1PR activating effect of both compound series. For each series, a modification template is illustrated to facilitate effective discussion (as shown in the main panels of Figures 4(Fig. 4) and 6). Insight analysis of each sub-modified series was provided to suggest a potential strategy for rational design and development of new derivatives with improved properties. Additionally, the most potent compounds of each subseries are summarized in Figures 4(Fig. 4), 5(Fig. 5), and 6(Fig. 6).

From the original quinoxaline-based activators, compound 2A (pEC₅₀ = 5.77) is the most potent compound followed by 4A, 3A, 1A, 5A, and 6A (pEC₅₀ = 5.66, 5.62, 5.58, 5.41, and 5.34, respectively, Table 2(Tab. 2). An increased length of the alkyl chain amide linker at A site on ring D (Figure 4a(Fig. 4)) could improve activity via increasing MATS5m values (4A and 3A, Table 2(Tab. 2)). The presence of long-length alkyl chain at A site is required to give high R2V+ and MATS5m values, thereby, high pEC₅₀ values of the compounds (as observed for compounds 2A and 4A). The aromatic ring (i.e., benzene, and thiophene) nearly attached at B site is also essential for potent activity. Two compounds without (6A) or with aromatic ring attached at this position in further distance (5A) provided lower nR05 value (nR05 =1) leading to their lesser potency when compared to other compounds (1A-4A with nR05 =2), Figure 4a(Fig. 4) and Table 2(Tab. 2). Summarized SAR analysis of the original compounds 1A-6A is provided in supplementary information, Table S1.

An additional set of 635 modified quinoxaline-based activators were designed. The modified compounds were grouped by their main core structures into three main templates (Figure 4b-d(Fig. 4)) and their modification strategies are provided in supplementary information Table S2. Summarized SAR analysis of the modified compounds 1A-6A is provided in supplementary information, Table S3. In overview, the modified series 6A provided the most potent predicted activities (predicted pEC₅₀ = 5.262-5.772, supplementary data, Scaffold A). Notably, the replacement of nitrogen-containing ring D from the five-membered (pyrrole) to six-membered ring (piperidine or pyridine) gave various effects depending on the parent compounds. Improved activities were observed for the six-membered ring modified series 5A-6A, whereas decreased activities were found for those of the series 1A-4A. This phenomenon could be due to the decreased nR05 values (supplementary data, Scaffold A).

For the best modified series 6A, modifications were performed on A, B, and C sites to obtain 287 modified compounds. Most of them provided better activities (predicted pEC₅₀ = 5.262-5.772, supplementary data, Scaffold A) than the parent 6A (experimental pEC₅₀ = 5.340, Table 2(Tab. 2)). Compounds 6A-116 and 6A-115 were the two most potent compounds (predicted pEC₅₀: 6A-116 = 5.772 and 6A-115 = 5.769). The improved activity (higher predicted pEC₅₀) was suggested to be via decreasing the nR05 and Km values while increasing the R2v+ value (supplementary information, Table S3). A set of promising compounds in modified 6A series is summarized in Figure 5(Fig. 5). Some essential key modifications to achieve preferable activity of series 6A were revealed such as 1) a replacement of ring D with six-membered ring, 2) modification on A site by an insertion of a moiety containing terminal benzene ring, 3) a modification on B site by replacing the branched alkyl chain with -COH and -CN groups as well as 4) a modification on C site by substitutions of -COCH₃ or -COCF₃ gave compounds with better activities than those substituted with benzoyl group. The better activities were provided when the longer chain was substituted at A position (6A-116 > 6A-44 and 6A-115 > 6A-43, Figure 5(Fig. 5)). Substitution at B position with -COH group provided better activity than with the CN group (6A-116 > 6A-115 and 6A-44 > 6A-43, Figure 5(Fig. 5)). The replacement of COH group on B position of compound 6A-44 (predicted pEC₅₀ = 5.767) by the COCH₃ group gave compound 6A-45 (predicted pEC₅₀ = 5.755) with the lesser potency. Similarly, the bulky ring substitutions on A and B positions impaired activity of compound 6A-140 (predicted pEC₅₀ = 5.741) when compared to 6A-116 (predicted pEC₅₀ = 5.772). Additionally, the substitution at C position with acetyl group (COCH₃) gave the more potent compounds rather than the COCF₃ group (6A-43 > 6A-47 and 6A-44 > 6A-48). The lesser potency of these COCF₃ derivatives was suggested to be via a shift of MATS5m value to positive values and the increased Km value (supplementary data, Scaffold A).

Activity of the original indole-based compounds (scaffold B) was ranked as 4B > 3B > 2B > 5B > 1B (pEC₅₀ = 5.66, 5.54, 5.38, 5.33, and 5.13, respectively, Table 3(Tab. 3)). The two most potent compounds (4B and 3B) are chloro-containing derivatives indicating that substitutions on ring A with chlorine (Cl) atom provides better improved activity than substitutions with alkyl (2B) or hydrogen (1B) groups. This could be via the increasing values of both H3e and E3v descriptors (Figure 2(Fig. 2) and Table 3(Tab. 3)). Moreover, an insertion of additional chloro group on the A ring of 3B gave the 4B compound with increased activity via increasing the value of van der Waals volume (4B > 3B: H3e value: 4B = 2.832, 3B = 2.74, Table 3(Tab. 3)). Summary SAR analysis of original compounds 1B-5B is provided in supplementary information, Table S4.

The QSAR scaffold B model was used to guide a design of 117 modified compounds (supplementary information, Table S4) and are subdivided into 5 subseries (1B-5B). The most potent compounds of each subseries are summarized in Figure 6b(Fig. 6). In overview, the modifications provided compounds with improved activities in every subseries. The substitution on Xa position of ring A with the ring-containing moiety (aliphatic ring: 1B-3, 2B-3, and 3B-4 or aromatic ring: 4B-48) provided the best activities when compared to other types of moieties (i.e., -OH, -OCH₃, -OCF₃, and alkyl chain). Notably, improvement was observed when the -CH₃ group of the parent 2B (experimental pEC₅₀ = 5.38) is replaced by a cyclohexane ring to give compound 2B-3 (predicted pEC₅₀ = 8.025). The same improvement was noticed when comparing parent mono-chloro containing compound 3B (experimental pEC₅₀ = 5.54) with the modified 3B-4 (predicted pEC₅₀ = 8.269). The improved activity of the 3B-4 is also influenced by a substituted fluorine (F) atom on the Yb moiety of ring B. The type of halogen atom substituted on this Yb position also affected the activity of the compounds as observed for modified subseries 3B. The best activity is achieved by substitution with fluorine (F: 3B-4: pEC₅₀ = 8.269) followed by chlorine (Cl: 3B-5: pEC₅₀ = 8.223) > bromine (Br: 3B-6: pEC₅₀ = 8.148) > iodine (I:3B-7: pEC₅₀ = 8.061), supplementary information, Table S5 and supplementary data, Scaffold B). For modified 4B series, the compounds with COCF₃ substitution on Xa position (4B-48 to 4B-55: predicted pEC₅₀ = 6.275-6.795, supplementary information, Figure S10) provided the best improvement among others. This could be due to their high E3v value (all compounds possess E3v value > 3, greater than those of other series).

For compound 5B possessing different core structure, the modifications effectively improved activity (as shown by high predicted pEC₅₀ = 5.794-8.784, supplementary information, Table S5 and supplementary data, Scaffold B). The best improvements were noted when the X moieties are substituted by alkyl chain (5B-9) or aromatic ring (5B-7 and 5B-11) moieties, while the Y moieties are replaced with long (5B-7) or branched alkyl (5B-9 and 5B-11) chain, Figure 6b(Fig. 6).

Regarding the Lipinski's rule, newly designed compounds were predicted for their drug-like parameters including molecular weight (MW), Ghose-Crippen-Viswanadhan octanol-water partition coefficient (AlogP), number of hydrogen bond donors and acceptors (nHBDon and nHBAcc) (Lipinski et al., 2001[17]). The parameter values were visualized as distribution plots (Figure 7(Fig. 7)). It was shown that all compounds are accepted by the rule of five, suggesting that they are drug-like molecules with the possibility for further development.

Supplementary information and supplementary data are available on EXCLI Journal's website.

There are no conflicts to declare.

Rattanawan Tangporncharoen is supported by the Royal Golden Jubilee (RGJ) Ph.D. Programme (grant no. PHD/0066/2561), through the National Research Council of Thailand (NRCT), Thailand Research Fund (TRF) and Thailand Science Research and Innovation (TSRI). Veda Prachayasittikul was supported by Office of the Permanent Secretary, Ministry of Higher Education, Science, Research and Innovation, Research Grant for New Scholar (grant no. RGNS 64-167), and by Mahidol University (Fundamental Funds: fiscal year 2023 by National Science Research and Innovation Fund (NSRF)). Chuleeporn Phanus-Umporn was supported by Office of the Permanent Secretary, Ministry of Higher Education, Science, Research and Innovation (grant no. RGNS 65-145).