The crystal structure of wild-type HER2-TAK285 (PDB ID: 3RCD; resolution: 3.21Å) was obtained

from the RCSB Protein Data Bank and prepared using “Prepare Protein” algorithm of Discovery Studio 2017 (DS2017). Using the “Define Site” tool of DS2017, the active site of HER2 structure was defined as a collection of several critical residues, such as Leu755, Thr798, and Cys805 etc., enclosed within a sphere of 11 Å radius around the co-crystallized ligandTAK285. Then, the co-crystalized ligand TAK285 and crystal water molecules were removed and only the wild-type HER2 structure was retained as the target template. Based on the structure of wild-type HER2, the structures of HER2 L755S, T798I and T798M mutants were constructed using the “Build Mutant” option implemented in DS2017. Subsequently, the qualities of these HER2 mutant structures were evaluated by the on-line UCLA SAVES server. All the HER2 mutant structures showed only 0.4% of residues in the disallowed region (Figure 2), which indicates that these constructed mutant structures have reasonable and stable conformations for the subsequent pharmacophore modeling. For active ligands, 40 known irreversible HER2 inhibitors were collected from previous literatures [16 , 17]. Their 3D structures were built using the sketch function of DS2017 and then optimized using “Prepare Ligand” algorithm. These 40 irreversible HER2 inhibitors, as shown in Figure 3, were named as compounds 1-40 according to their IC₅₀ (from the lowest to the highest).

Due to the lack of receptor-ligand binding information, each of the known 40 irreversible HER2 inhibitors was individually docked into the active sites of HER2 L755S, T798I and T798M mutant structures and a maximum of 10 binding poses were generated for each receptor-ligand complex using CDOCKER. For each receptor-ligand complex, the generated docking pose with the best CDOCKER interaction energy was selected for pharmacophore modeling. Thus, a total of 120 docking poses were selected for the generation of structure-based pharmacophore models using the “Receptor-Ligand Pharmacophore Generation” protocol of DS2017 [18 , 19]. For each receptor-ligand binding complex, a maximum of 10 pharmacophore models with 4-6 pharmacophore features were automatically generated with default parameters. Finally, a total of 36, 36, and 38 pharmacophore models based on HER2 L755S, T798I, and T798M mutant structures were constructed, respectively.

To select the best structure-based pharmacophore models for HER2 L755S, T798I, and T798M mutant, a test 3D database comprising both 40 known irreversible HER2 inhibitors and 254 decoys (from DecoyFinder-2.0) was constructed using “Build 3D Database” (maximum 255 conformations for each compound) of DS2017. Then, the constructed pharmacophore models were used for screening the test 3D database using “Screen Library” program of DS2017 with default parameters. Subsequently, Güner-Henry (GH) scoring and receiver operating characteristic (ROC) analysis methods were adopted to evaluate the screening results of each model. Finally, three models (named as models I, II and III), which were constructed based on the HER2 L755S-compound 39, HER2 T798I-compound 10 and HER2 T798M-compound 6 complexes, respectively, displayed the best GH scores and “excellent” ROC curves and were selected for further virtual screening.

The flowchart of the overall virtual screening procedure in this study is summarized in Figure 4. First, the NCI database were downloaded and filtered by both Lipinski’s rule-of-five [20] and Veber’s rule [21] to enhance the drug-like properties of the screening results [22]. Then, models I, II, and III were used as 3D search queries to identify novel irreversible HER2 inhibitors using “Screen Library” protocol of DS2017 with default parameters. To segregate irreversible HER2 inhibitors from the reversible ones, only compounds with Michael acceptors were selected by visual inspection. Finally, a total of 143, 72 and 191

compounds were selected from models I, II, and III, respectively, as potential irreversible HER2 inhibitors for the further molecular docking studies (Figure 4).

The selected 143, 72 and 191 potential irreversible HER2 inhibitors were separately docked into the active sites of HER2 L755S, T798I and T798M mutant structures, respectively, using the CDOCKER program of DS2017 with CHARMm force field. Subsequently, the CDOCKER interaction energies (kcal・mol⁻¹), which are also known as CDOCKER scores, were obtained from these docking studies. These docking results were further compared to those of the 40 known irreversible HER2 inhibitors, including HER2 L755S-compound 22, HER2 T798I-compound 14 and HER2 T798M-compound 19. These complexes displayed the lowest CDOCKER interaction energies among all of the 40 known irreversible HER2 inhibitors. Finally, the ones which exhibited lower CDOCKER interaction energies than those known irreversible HER2 inhibitors were selected as hit compounds.

In order to investigate the binding conformations in the active sites of the HER2 mutants, the selected hit compounds were subjected to further visual inspection based on the main criteria that the distance between Michael acceptors of each selected compound and the sulfhydryl group of HER2 Cys805 must be within the ideal range (5.5 Å) for Michael addition [23]. Finally, compounds with better binding affinity and favorable binding poses for Michael addition were retrieved as potential irreversible HER2 inhibitors.

MD simulations were performed to confirm the binding stabilities of the selected potential irreversible HER2 inhibitors using the Gromacs 2016.3 software package [24 , 25]. With Amber Tools 2017, the AMBER99SB-ILDN force field [26] was assigned for proteins and the general AMBER force field (GAFF) was assigned for ligands [27 , 28]. In a 10.0 Å dodecahedron periodic box, the system was solvated with TIP3P water model and neutralized by replacing solvent molecules with counter ions. Subsequently, the entire system was stabilized through energy minimization and the equilibration of each system was conducted at constant volume (NVT) for 200 ps and constant pressure (NPT) for 200 ps. Finally, the equilibrated systems were subjected to 50 ns production runs at constant temperature (300 K) and pressure (1 atm).

ADMET properties of the selected potential irreversible HER2 inhibitors were obtained by uploading the simplified molecular input line entry specification (SMILES) data of each compound to the online webserver admetSAR [29]. In this webserver, mathematical models (human intestinal absorption, CYP450 inhibition, hERG inhibition, AMES toxicity and Carcinogenicity) were used to quantitatively predict the properties of a set of rules that specify ADMET characteristics of the chemical structures for the selected potential irreversible HER2 inhibitors. Compounds with satisfactory ADMET properties were finally selected as safe and novel irreversible HER2 inhibitors for breast cancer treatment.

Structure-based pharmacophore models are generated from the structures of the protein-ligand complexes through investigating the important interactions and excluded volumes in their binding sites [30]. Comparing to ligand-based approach, structure-based approaches have advantages of more precise prediction, good filter to remove decoys and finding structurally novel ligands against target receptors [31]. To date, the X-ray crystallographic structures of HER2 L755S, T798I and T798M mutants have not yet been reported. In order to carry out structured-based pharmacophore modeling, the homology models of HER2 L755S, T798I and T798M mutants were generated from the wild-type HER2 structure (PDB ID: 3RCD) using “Build Mutant” protocol and validated by Ramachandran plot.

To obtain receptor-ligand docking poses, 40 known irreversible HER2 inhibitors were individually docked into the constructed HER2 L755S, T798I and T798M mutant structures using CDOCKER protocol. Then, the generated 120 docking complexes were applied to structure-based pharmacophore modeling, and a total of 36, 36, and 38 models based on HER2 L755S, T798I, and T798M mutant structures were constructed, respectively. Then, the performance of these models was further validated by GH scoring method and ROC analysis [32 - 35].

For HER2 L755S mutant, the optimal structure-based pharmacophore models (model I) was constructed based on the HER2 L755S-compound 39 complex (Figure 5(a)). Compound 39 displays hydrogen bonding interactions with Lys753, Thr862 and Asp863 in the active site of HER2 L755S mutant, which have been identified as critical residues for HER2 inhibitory. Based on the HER2 L755S-compound 39 complex, model I contains three hydrophobic (HY) features, one hydrogen bond acceptor (HA) anchoring the ligand with Lys753, and two hydrogen bond donors (HD) anchoring the ligand with Thr862 and Asp863 (Figure 5(d)). As shown in Table 1, model I can successfully recognize 39 out of 40 known irreversible HER2 inhibitors (sensitivity = 0.975) and correctly exclude 241 out of 254 decoys (specificity = 0.949). In addition, model I possess the best GH scores of 0.77 (Table 1) and excellent ROC curve with AUC values of 0.98, demonstrating that the quality of model I is acceptable for further virtual screening [36].

For HER2 T798I mutant, the optimal structure-based pharmacophore model (model II) was constructed from HER2 T798I-compound 10 complex (Figure 5(b)). Compound 10 forms hydrogen bonds with two important residues, Cys805 and Asp863. Based on the HER2 T798I-compound 10 complex, model II consists of three HY features, one HA with Cys805 and two HDs with Asp863 (Figure 5(e)). As shown in Table 1, model II can recognize 39 out of 40 known irreversible HER2 inhibitors (sensitivity = 0.975) and exclude 246 out of 254 decoys (specificity = 0.969). Besides, model II possess the best GH scores of 0.84 (Table 1) and excellent ROC curve with AUC values of 0.980, indicating that the quality of model II is acceptable for the subsequent virtual screening.

For HER2 T798M mutant, the optimal structure-based pharmacophore model was built from the HER2 T798M-compound 6 complex (Figure 5(c)). Unlike HER2 T798I mutant, the sulfur group of Met798 and the 3-chloro-4-fluoroaniline substituent group of compound 6 bound with each other through S/π interaction, which has been identified as prevalent and important stabilizing interaction [37]. Such interaction can twist the binding conformation of compound 6 in the HER2 T798M mutant. Meanwhile, the nitrogen atom of pyrimidine ring forms a hydrogen bond with Cys805 and the carbonyl substructure forms hydrogen bond with Lys753. Based on such binding conformation, the generated model III consists of two HY features, two HAs with Lys753 and Cys805, and two HDs with Asn850 and Asp863 (Figure 5(f)). As shown in Table 1, model III can successfully recognize 39 out of 40 known irreversible HER2 inhibitors (sensitivity = 0.975) and exclude 242 out of 254 decoys (specificity = 0.953). Furthermore, model III possess the best GH scores of 0.79 (Table 1) and excellent ROC curve with AUC values of 0.983, suggesting that the quality of model III is good enough for further virtual screening. The above validation results all confirm that models I-III exhibit high predictive efficiency and can be served as reliable 3D queries for the further virtual screening procedures [38].

D is number of compounds in a database, A is the number of active compounds in the database, Ht is the number of hits retrieved, Ha is the number of actives in hit list, %Y is the fraction of hit relative to the size of database (hit rate or selectivity), %A is the ratio of actives retrieved in hit list, EF is the enrichment of active bin by model relative to random screening, GH is the Güner-Henry score. ^a [ ( H a 4 H t × A ) ( 3 A + H t ) × ( 1 − ( H t − H a D − A ) ) ] ; GH score > 0.7 indicates a statistically good model.

In order to investigate the influence caused by point mutation, compounds 6, 10, 39 were docked into the active sites of WT HER2 structure. Then the docking conformations were compared with those of HER2 mutants. As shown in Figure 6(a), the residues of WT HER2 and HER2 L755S mutant were almost overlapped; therefore, compound 39 displayed similar docking conformations in both binding sites of WT HER2 and HER2 L755S mutant. The results suggest that compound 39 is a very good irreversible HER2 inhibitor for both WT HER2 and HER2 L755S mutant. As shown in Figure 6(b), Met798 residue is bulkier than Thr798, which caused slight steric collision for compound 10, leading to moving upward to reduce such steric collision. As for the docking poses of compound 6, T798M mutation also caused critical steric collision to the ligand, resulting in breaking the hydrogen bond between compound 6 and Met801 residue, which further caused the different binding conformations for WT HER2 and HER2 T798M mutant (Figure 6(c)).

Virtual screening is a versatile technique to identify novel and potent lead compound for particular drug target [39]. In order to identify novel irreversible HER2 inhibitors, NCI database (265,242 compounds) was utilized for our virtual screening and the flowchart is shown in Figure 4. For primary screening, all compounds were filtered with Lipinski’s rule-of-five and Veber’s rule to obtain the drug-like compounds. Then the optimal pharmacophore models I-III were used as 3D queries to screen the newly drug-like compounds library with the best search mode to discover the novel scaffold of HER2 inhibitors. Subsequently, visual inspection was carried out to retrieved potential irreversible HER2 inhibitors with Michael acceptors, which are essential to form covalent interactions with HER2 mutants. Finally, these potential irreversible HER2 inhibitors were subjected to further docking studies to investigate their binding conformations.

Molecular docking is a well-established method to predict the molecular-level interactions of small molecules in the receptor active sites. In order to investigate the binding interactions between receptors and ligands, these potential irreversible HER2 inhibitors were individually docked into the active sites of HER2 L755S, T798I and T798M mutant structures and CDOCKER interaction energies (kcal・mol⁻¹) obtained in the docking experiments were utilized to evaluate their binding interactions as presented in Table 2. For HER2 L755S mutant, compound 22, which showed the lowest CDOCKER interaction energies of −61.859 kcal/mol, was the optimal compound among 40 known irreversible HER2 inhibitors. Among the selected 143 compounds, NSC278329 (−62.748 kcal/mol), NSC381866 (−65.115 kcal/mol), NSC642003 (−63.009 kcal/mol) and NSC659397 (−62.345 kcal/mol) showed lower CDOCKER interaction energies than that of compound 22, thus were considered to exhibit stronger binding interactions. In addition to the docking scores, the essential covalent interactions of these docking complexes were also predicted by measuring the distances between the Michael acceptors of the inhibitors and the Cys805 residues of HER2 L755S mutant. According to previous literature, the distance for Michael addition should be within the ideal range of 5.5 Å. Among these binding complexes, the distances were 5.0 Å for NSC278329 and 4.5 Å for NSC642003 (Figure 7(a) and Figure 7(b)). Therefore, we anticipated that only NSC278329 and NSC642003 can execute Michael addition by forming covalent bonds between their Michael acceptors and the sulfhydryl group of Cys805 in HER2 L755S mutant. In addition, NSC278329 displayed strong hydrogen bonding interactions with Ala730, Lys753, Thr862, Asp863 and hydrophobic interactions with Val734, Leu796, Cys805 residues in the active site of HER2 L755S mutant (Figure 8(a)). NSC642003 formed hydrogen bonding interactions with Lys753, Asn850, Asp863 and hydrophobic interactions with Leu726, Ala751, Met774, Leu785, Leu796, Leu800, Cys805, Leu852, Phe864, Phe1004 in the HER2 L755S mutant (Figure 8(b)). All together, the above results suggest that both of NSC278329 and NSC642003 were in the vicinity of the active site containing the critical residues, such as Leu726, Gly729, Val734, Ala751, Lys753, Cys805, Arg849, Leu852, Thr862, Asp863. Finally, HER2 L755S-NSC278329, HER2 L755S-NSC642003 docking complexes were subjected for MD simulations to estimate their binding stabilities.

When docking to the HER2 T798I mutant structure, NSC702108 (−60.981 kcal/mol), NSC702137 (−60.748 kcal/mol) and NSC718305 (−68.817 kcal/mol) all had lower CDOCKER interaction energies than the optimal known irreversible HER2 inhibitor compound 14 (−59.644 kcal/mol) (Table 2). Based on their docking results, these compounds with lower CDOCKER interaction energies were considered to exhibit stronger binding interactions than all of the 40 known irreversible HER2 inhibitors, thus were retrieved for further visual inspection. The distance between Michael acceptor of NSC718305 and Cys805 of HER2 T798I mutant was about 4.3 Å, indicating that NSC718305 can execute Michael addition by forming covalent bond between their Michael acceptors and the sulfhydryl group of Cys805 in HER2 T798I mutant (Figure 7(c)). However, NSC702108 and NSC702137 did not have such favorable binding poses for Michael addition through CDOCKER results, indicating that they have fewer chances to form covalent bonds with HER2 T798I mutant. NSC718305 displayed hydrogen bonding interactions with Lys753, Ser783, Asp863, Phe864 and hydrophobic interactions with Val734, Ala751, Leu796, Leu852 in the active site of HER2 T798I mutant. It also formed pi-sulfur interaction with Cys805 residues in the active site of HER2 T798I mutant (Figure 8(c)). Finally, HER2 T798I-NSC718305 docking complex was subjected to further MD simulations to estimate its binding stability.

For the HER2 T798M mutant, NSC659505 (−60.220 kcal/mol), NSC692910 (−59.838 kcal/mol) and NSC718305 (−66.841 kcal/mol) all had lower CDOCKER interaction energies than compound 19 (−59.719 kcal/mol) (Table 2). From visual inspection, the distance between Michael acceptor of NSC718305 and Cys805 of HER2 T798M mutant was about 4.4 Å (Figure 7(d)), which strongly suggests that NSC718305 can execute Michael addition with HER2 T798M mutant. However, no such favorable binding poses for Michael addition were observed for both NSC659505 and NSC692910 through CDOCKER, which reveals that they have fewer chances to form covalent bonds with HER2 T798M mutants. NSC718305 also bound in the active site of HER2 T798M mutant by forming hydrogen interactions with Lys753, Asp863, Gly865 and hydrophobic interactions with Ala751, Leu796, Met798, Cys805, Arg849 (Figure 8(d)). Finally, HER2 T798M-NSC718305 docking complex was subjected to further MD simulations to estimate its binding stability.

Molecular dynamics (MD) simulation is an accurate method for investigating the physical movements of proteins-ligand complexes, giving a view of the dynamical evolution of the binding processes. In this study, HER2 L755S-NSC278329, HER2 L755S-NSC642003, HER2 T798I-NSC718305 and HER2 T798M-NSC718305 docking complexes were subjected to 50 ns MD simulations in an explicit hydration environment to investigate their binding stabilities. The dynamic stabilities and MD simulation trajectories of these complexes were analyzed by their backbone root mean square deviations (RMSD) [40]. Small RMSD values along with small fluctuations during the entire MD courses represent a stable binding system. As shown in Figure 9, HER2 L755S-NSC278329, HER2 L755S-NSC642003, HER2 T798I-NSC718305 and T798M-NSC718305 protein-ligand systems all reached the converged stage in the last 10 ns, with average RMSD values of 2.143 Å, 2.15 Å, 2.054 Å and 2.338 Å, respectively. Based on their favorable binding poses and stable RMSD values, NSC278329, NSC642003, NSC718305 were considered to have stable binding modes in these HER2 mutants.

ADMET properties were calculated to investigate the pharmacokinetics of a prospective drug compound in the human body [41]. In this study, NSC278329, NSC642003 and NSC718305 were subjected to ADMET prediction using admet SAR and then compared to the FDA approved HER2 inhibitor, lapatinib. For absorption, all of these three compounds showed good human intestinal absorption (HIA), which suggests good oral delivery possibility. In the toxicity predictions, lapatinib, NSC278329 and NSC718305 are non-mutagenic (AMES toxicity test) and non-carcinogenic (carcinogenic profile) (Table 3) [42]. Since the US Food and Drug Administration (FDA) published the first in vitro/in vivo drug interaction guidance

documents in 1997 and 1999, respectively, off-target transporter interactions and drug-drug interactions have been considered as the important criteria for drug development [43]. In human heart, hERG potassium channels are essential for regulating electrical activity [44]. According to the literature, the inhibition of human Ether-à-go-go-Related Gene (hERG) potassium channel may result in Long QT syndrome (LQTS), which is known as the lethal side effects in clinical studies [45]. The results of hERG prediction I/II showed that NSC278329, NSC642003 and NSC718305 are non-hERG inhibitors, suggesting that they have lower risk of causing LQTS than that of lapatinib. CYP450 enzymes-based unanticipated DDIs and drug metabolism problems are also a common cause of adverse drug events (ADE) [43]. According to our prediction, all of these three compounds can penetrate the pathway, which is regulated by CYP3A4 enzyme, suggesting that they may compete with other drugs. Furthermore, both NSC278329 and NSC718305 can inhibit CYP3A4 enzyme, which reveals that they can affect the metabolism of other drugs using the same CYP3A4 pathway, which is correlated to the function of platelet in human. NSC278329 and NSC642003 were predicted to inhibit CYP2C19 enzyme, suggesting that they can affect the function of platelet and should be carefully used. Compared tolapatinib, both NSC278329 and NSC718305 showed lower CYP inhibitory promiscuity inhibition, indicating that they have lower risks of causing CYP450 enzymes-based unanticipated drug-drug interactions and metabolism problems [46]. Overall, both NCS278329 and NSC718305 showed better ADMET properties than lapatinib.

Cross docking analysis was performed to gain a deeper insight into the specificity of these selected irreversible HER2 inhibitors. In this study, NSC718305 was docked into HER2 L755S mutant while NSC278329 was docked into HER2 T798I and HER2 T798M mutants. Based on the results of cross docking, NSC718305 exhibited good CDOCKER scores of −71.664 kcal/mol by forming hydrogen bonding interactions with Lys753, Ser783, Cys805, Asp863 residues of HER2 L755S mutant (Figure 10(a)). On the other hand, NSC278329 exhibited strong binding with HER2 T798I mutant by forming hydrogen bonds with Lys753, Asn850, Thr862, Asp863, Gly865 residues (Figure 10(b)) with a good CDOCKER score of −70.482 kcal/mol. While docked into HER2 T798M mutant, NSC278329 formed hydrogen bonds with Lys753, Glu770, and Asp863 residues (Figure 10(c)) with a CDOCKER score of −66.760 kcal/mol. Therefore, both NSC278329 and NSC718305 were considered as potent irreversible inhibitors targeting HER2 L755S, T798I and T798M mutants.

Interestingly, NSC718305 was the only compound to be screened by model II and model III. These two pharmacophore models totally differ from each other not only by the number of their features but also by their geographies. As shown in Figure 11, the docking conformations of NSC718305 in HER2 T798M mutant displayed completely different orientations and binding angles that in HER2 T798I mutant. Such different binding conformations may provideNSC718305 the possibility for mapping both pharmacophore models II and model III simultaneously.