Design, synthesis and biological evaluation of novel 1H-1,2,4-triazole, benzothiazole and indazole-based derivatives as potent FGFR1 inhibitors viafragment-based virtual screening

Abstract Fibroblast growth-factor receptor (FGFR) is a potential target for cancer therapy. We designed three novel series of FGFR1 inhibitors bearing indazole, benzothiazole, and 1H-1,2,4-triazole scaffold via fragment-based virtual screening. All the newly synthesised compounds were evaluated in vitro for their inhibitory activities against FGFR1. Compound 9d bearing an indazole scaffold was first identified as a hit compound, with excellent kinase inhibitory activity (IC50 = 15.0 nM) and modest anti-proliferative activity (IC50 = 785.8 nM). Through two rounds of optimisation, the indazole derivative 9 u stood out as the most potent FGFR1 inhibitors with the best enzyme inhibitory activity (IC50 = 3.3 nM) and cellular activity (IC50 = 468.2 nM). Moreover, 9 u also exhibited good kinase selectivity. In addition, molecular docking study was performed to investigate the binding mode between target compounds and FGFR1.


Pharmacophore modeling
The selection of a suitable training set was the most important step to pharmacophore model, as it determined the quality of generated pharmacophore. A series of known FGFR1 inhibitors were obtained from the literatures [1][2][3][4][5][6][7] . Most of these inhibitors were launched or in clinical trial. Figure S1. The dataset to build Pharmacophore The three-dimensional pharmacophore Generation protocol in Discovery Studio 4.0 (DS) was used to build the pharmacophore models. The conformational set was generated for each molecule using the 'best-quality conformational analysis' method, based on the CHARMM force field. On the basis of the chemical features of compounds in the training set and the proposed mechanism of action, four pharmacophore features, hydrogen bond acceptor (HBA), hydrogen bond donor (HBD), hydrophobic (HY), and ring aromatic group (RA) were defined for the creation of pharmacophore models. All other parameters used were set at the default values. Pharmacophores were then computed and the top 10 scoring hypotheses were obtained. To check the pharmacophore's capability of distinguishing active inhibitors from inactive compounds and predicting their activity values, the developed pharmacophore models were validated by three methods: Cost analysis, Test set prediction and Fischer's randomization test. A best pharmacophore hypothesis should meet the following requirements: the total cost close to the fixed cost value and away from the null cost value. Meanwhile, the testing set was also prepared using the same protocol and utilized to verify the predictability of the best pharmacophore hypothesis. Fischer's randomization test was performed to obtain cross validation via randomizing the data.   (Fig S2a) and NVP-BGJ398 (Fig S2b) mapped the pharmacophore model. Pharmacophore features colored as follow: HBA (green), HBD (magenta), hydrophobic (cyan).

Fragment-Based Virtual Screening
The docking reliability was evaluated by calculating the root-mean-square deviation (heavy atoms) difference between the reference position of the ligand in the crystal structures and that predicted by the docking software. Finally, 4ZSA was selected as the template protein for follow-up virtual screening which had low re-docking value of RMSD.
The fragment library was derived from kinase hinge region directed library provided by enamine (https://enamine.net/index.php), which contained 11809 fragments. As the pharmacophore were constructed, we initially utilized this model for the first round of virtual screening. The well-validated pharmacophore model was used as a 3D structural query for retrieving potent compounds and estimating their inhibitory activity of FGFR1. The database screening was performed using the Ligand Pharmacophore Mapping protocol in DS 4.0. The fit values were calculated based on the chemical substructures map the location constraints of the pharmacophoric features and their distance deviation from the feature centers (Table S1). Finally, only the compounds both showed good fit values and the chemical features mapped the location distance constraints in pharmacophore model, could enter the next stage of virtual screening calculation ( Figure S3). Figure S3. The fragment-based virtual screening protocol.
The Surflex-dock module in Sybyl-x 2.0 was used for molecular docking. For checking the robustness of docking protocol, self-docking was performed in which the bound ligand was re-docked into the catalytic site of protein. In both cases, the ligandbinding site search region was defined to center on the ligand in the crystal structure.
The crystal structure of the FGFR1 (PDB ID: 4ZSA) was considered for this study and prepared using Protein Preparation Wizard module. Hydrogen atoms were added, water molecules in all the system were removed, followed by energy minimization and optimization by MMFF94 force field. LigPrep module was employed to prepare the compounds for molecular docking. To prepare ligand structures, hydrogens were added, 3D geometries, ionization and tautomeric states were generated. Finally, the ligand structures were minimized using MMFF94 with 5000 iterations and minimum RMS gradient 0.05. Other parameters were as default. The compounds were selected based on the docking score, seven fragments were select for the next filtration (Table S1)..
Finally, as for these seven fragments, we use SciFinder search tool to evaluate the difficulty of synthesis of target compounds, therefore only three target fragments for the next experimental research, including 1H-1,2,4-triazole, benzothiazole and indazole scaffold ( Figure S4).