Use of a aurora kinase a inhibitor in the preparation of a medicament for treating triple negative breast cancer

Abstract

The application discloses application of aurora kinase A inhibitors in preparation of drugs for treating triple-negative breast cancer, and confirms that compounds 3, 5 and 8 have significant aurora kinase A inhibiting activity and can effectively inhibit the proliferation of triple-negative breast cancer MDA-MB-231 cells. 50 The IC of the three compounds on aurora kinase A is 157 nM, 21.64 nM and 46.03 nM respectively; and cell experiments prove that the three compounds can efficiently inhibit the growth of triple-negative breast cancer cells. The application expands the medical application range of the compounds, provides a new drug selection for clinical treatment of triple-negative breast cancer, and has important clinical value and industrialization prospect.

Description

Technical Field

[0001] This invention relates to the field of pharmaceutical technology, specifically to a new pharmaceutical use of a compound, and more particularly to the application of an aurora kinase A inhibitor in the preparation of a drug for treating triple-negative breast cancer. Background Technology

[0002] Breast cancer is the most common malignant tumor among women worldwide. Triple-negative breast cancer (TNBC) is highly malignant, has a high risk of distant metastasis, and a very poor prognosis, making clinical treatment options very limited. Because TNBC lacks target sites for estrogen receptors, progesterone receptors, and human epidermal growth factor receptor 2, endocrine therapy and targeted therapy have limited efficacy. Furthermore, this subtype of tumor is highly invasive, prone to metastasis, and has a poor overall prognosis, making it a challenging area for clinical cancer treatment.

[0003] Aurora kinase A (AURKA) is a key kinase regulating the cell cycle. It is a highly conserved kinase that plays a central role in cell cycle regulation, primarily involved in the transition from G2 to M phase. Abnormal expression of this protein disrupts spindle checkpoint function, leading to genetic instability and ultimately inducing tumorigenesis. Studies have confirmed that approximately 73% of breast cancer patients overexpress AURKA, a phenomenon that contributes to tumor drug resistance and shortens median survival. Furthermore, AURKA kinase activity is closely related to chemotherapy resistance in triple-negative breast cancer. In summary, AURKA participates in the occurrence and progression of breast cancer through multiple mechanisms, making it a highly promising drug target for the treatment of triple-negative breast cancer.

[0004] Several inhibitors targeting aurora kinase A have entered preclinical and clinical research stages and have shown promising application prospects. Millennium Pharmaceuticals' MLN8237 is a highly selective aurora kinase A inhibitor with an IC50... 50 The concentration is 1.2 nanomolar; PHA739358 belongs to the pan-photokinase inhibitor class, which can simultaneously inhibit three photokinases in the human body, and its IC50 for photokinases A, B, and C is 1.2 nanomolar. 50 The concentrations were 13 nanomolar, 79 nanomolar, and 61 nanomolar, respectively; the multi-target kinase inhibitor AT9283 also showed potent inhibitory activity against aurora kinase A, with an IC50 concentration of 13 nanomolar, 79 nanomolar, and 61 nanomolar, respectively. 50 The concentrations reach 3 nanomoles. While the aforementioned representative inhibitors have shown varying degrees of efficacy in anti-tumor studies, some have not demonstrated significantly superior efficacy compared to existing therapies in clinical trials, and also exhibit significant toxic side effects, such as bone marrow suppression and gastrointestinal adverse reactions. Furthermore, activation of alternative signaling pathways can lead to drug resistance. Due to these limitations, no small-molecule inhibitors targeting Aurora kinase A have been approved for the treatment of triple-negative breast cancer, making the development of novel Aurora kinase A inhibitors with superior activity an urgent priority. Summary of the Invention

[0005] In view of the shortcomings of the prior art, the purpose of this invention is to provide a compound with aurora kinase A inhibitory activity.

[0006] To achieve the above objectives, the present invention provides a compound with aurora kinase A inhibitory activity having the following structure: .

[0007] The compound with aurora kinase A inhibitory activity provided by this invention was screened from an existing database using a multi-level virtual screening method for kinase inhibitors incorporating deep learning, comprising the following steps: Step S1 Compound library preprocessing: Construct an initial compound library and use the RDKit tool for normalization preprocessing to remove failed-analyzing molecules and salt ions to obtain a normalized compound screening library; Step S2 First round of sequence affinity screening: Using the PSICHIC model based on kinase protein sequences, the binding affinity of all compounds in the standardized compound screening library is predicted, and an affinity score threshold is set to obtain the first-level candidate compound set; Step S3: Second round of deep learning docking screening: Using the crystal structure of kinase protein as the target, the KarmaDock isovariant deep learning docking tool is used to perform rigid molecular docking on the first-level candidate compound set, and a docking score threshold is set to obtain the second-level candidate compound set; Step S4 Final Screening and Validation: The docking conformations of the second-level candidate compound set are re-scored using the DeepDock geometric deep learning model; the similarity between the compound and the known kinase inhibitor active molecule library is calculated based on the Morgan / ECFP4 fingerprint; and the potential kinase inhibitor candidate molecules are finally obtained by manual screening based on the re-scoring results, structural similarity, pharmacophore matching degree, synthetic feasibility and skeleton novelty.

[0008] Step S5 Activity Verification: Homogeneous time-resolved fluorescence biochemical experiments, tumor cell proliferation experiments, and molecular dynamics simulations were sequentially performed on the candidate molecules obtained in Step S4 to verify the inhibitory activity of the compounds on kinases and their binding stability.

[0009] This invention successfully constructed an AI-based cascaded virtual screening system, integrating sequence-based affinity prediction (PSICHIC), equivariant deep learning docking (KarmaDock), and geometric rescoring (DeepDock) technologies to efficiently mine 10 candidate molecules from a massive compound library. Homogeneous time-resolved fluorescence biochemical experiments confirmed that three of the compounds (compounds 3, 5, and 8) possessed good aurora kinase A inhibitory activity; molecular dynamics simulations and docking analysis elucidated the binding mechanism between the molecules and the target; and cell experiments demonstrated that the candidate molecules could inhibit the short-term proliferation of tumor cells.

[0010] Compared with the prior art, the present invention has the following advantages: This invention is the first to demonstrate that known compounds 3, 5, and 8 possess potent aurora kinase A inhibitory activity, with an enzyme level IC50. 50 The activity can be as low as 21.64 nM, which is superior to some existing candidate inhibitors. According to ADMET prediction, compound 3 has a high intestinal absorption probability of 0.995, extremely low blood-brain barrier penetration rate, and low risk of central nervous system toxicity. Compound 5 has low risk of cardiotoxicity and has excellent overall drug development potential. Therefore, it is expected that this class of compounds can be used to treat triple-negative breast cancer, and can significantly inhibit the proliferation of typical triple-negative breast cancer cells MDA-MB-231, make up for the shortcomings of existing therapeutic drugs, and meet different clinical dosing needs. Attached Figure Description

[0011] Figure 1 Distribution of active compounds of Aurora Kinase A (A) Negative logarithmic half-maximal inhibitory concentration (pIC) of active compounds of Aurora Kinase A 50 (A) Distribution; (B) Scatter plot of t-distribution neighborhood embedding (t-SNE) of the Aurora Kinase A active compound dataset.

[0012] Figure 2 Evaluation results of docking procedures and Aurora Kinase A protein. (A) Receiver operating characteristic curves of molecular docking of multiple Aurora Kinase A proteins with the selected docking procedures using the Aurora Kinase A validation dataset; (B) Box plot of KarmaDock procedure with 6VPL protein; (C) Results of re-docking of Aurora Kinase A protein (protein database number: 6VPL) using KarmaDock (blue: original conformation; purple: re-docking conformation).

[0013] Figure 3Early enrichment capacity analysis of the docking software. Heatmaps of enrichment factors for the selected docking program at screening ratios of 1% (A) and 5% (B); heatmaps of the area under the local receiver operating characteristic curve at screening ratios of 1% (C) and 5% (D).

[0014] Figure 4 Hybrid virtual screening process for Aurora Kinase A inhibitors.

[0015] Figure 5 Bioactivity evaluation results of 10 screened compounds. (A) Bar graph of inhibition rate of 10 screened compounds (positive control: asteroidin); (B) Dose-response curves of compounds 3, 5, and 8 (positive control: CCT241736); Cell viability assay results of compounds 3, 5, and 8 on (C) JIMT-1 cell line, (D) MDA-MB-231 cell line, and (E) MDA-MB-453 cell line (positive control: CCT241736); Short-term cell viability assay results of compounds 3 and 5 on (F) NIH / 3T3 cell line and (G) MDA-MB-231 cell line (positive controls: MLN8237, CCT241736).

[0016] Figure 6 Molecular dynamics trajectory analysis of the complex of compound 3 and aurora kinase A protein. (A) Chemical structure of compound 3; (B) Root mean square (RMS) wave trajectory of protein residues after compound 3 binds to aurora kinase A; (C) Root mean square deviation trajectory analysis of aurora kinase A protein during molecular dynamics simulation; (D) Root mean square deviation trajectory analysis of compound 3 during molecular dynamics simulation.

[0017] Figure 7 Total energy analysis of the molecular dynamics of the compound 3-Aurora kinase A protein complex. Total energy decomposition results of the complex under three parallel experiments: (A) Group 1, (B) Group 2, (C) Group 3; Binding free energy decomposition results of the complex under three parallel experiments: (D) Group 1, (E) Group 2, (F) Group 3.

[0018] Figure 8 Analysis of the two-dimensional free energy surface, radius of gyration, and solvent-accessible surface area of ​​compound 3-Aurora kinase A protein complex. Two-dimensional free energy surface diagrams of compound 3 in three parallel experiments: (A) Group 1, (B) Group 2, (C) Group 3; (D) radius of gyration and (E) solvent-accessible surface area analysis results of the complex.

[0019] Figure 9Analysis of the Gibbs free energy contribution and contact occupancy of each amino acid residue after compound 3 binds to aurora kinase A protein. Energy contribution analysis of amino acid residues of aurora kinase A in three parallel experiments: (A) Group 1, (C) Group 2, (E) Group 3; Contact occupancy of compound 3 binding to aurora kinase A protein in three parallel experiments: (B) Group 1, (D) Group 2, (F) Group 3.

[0020] Figure 10 . Fine-grained binding pattern analysis of compound 3 with aurora kinase A protein. Three-dimensional binding patterns of the complex in three parallel experiments: (A) Group 1, (D) Group 2, (G) Group 3; Two-dimensional interaction maps of the complex in three parallel experiments: (B) Group 1, (E) Group 2, (H) Group 3; Statistical analysis of the number of hydrogen bonds in the simulated trajectory in three parallel experiments: (C) Group 1, (F) Group 2, (I) Group 3.

[0021] Figure 11 The binding modes of compound 3 with aurora kinase A (protein database ID: 6VPL), aurora kinase B (protein database ID: 4AF3), and aurora kinase C (protein database ID: 6GR8) proteins are shown in the diagrams. (A) Three-dimensional binding diagram, (B) Two-dimensional interaction diagram: Compound 3 with aurora kinase A protein; (C) Three-dimensional binding diagram, (D) Two-dimensional interaction diagram: Compound 3 with aurora kinase B protein; (E) Three-dimensional binding diagram, (F) Two-dimensional interaction diagram: Compound 3 with aurora kinase C protein.

[0022] Figure 12 Structure-activity relationship analysis of compounds (A)3, (B)5, and (C)8 within the Aurora kinase A binding pocket. Detailed Implementation

[0023] The design and application of the present invention will be further described below with reference to embodiments, but this does not limit the present invention.

[0024] Example 1. Construction of Aurora Kinase A Evaluation Dataset The dataset was constructed and the software was evaluated using the integration tool Evaluation Master. Inhibitors of Aurora Kinase A (UniProt ID: O14965) were retrieved from the 2024 version of the ChEMBL database, and pICs were screened. 50 ≥6 (IC) 50 Compounds with a concentration of <1 micromolar) ensure reliable activity data.

[0025] The t-distribution neighborhood embedding algorithm was used to cluster active molecules into 50 groups based on the molecular core framework, and one representative molecule was selected from each group as the active set. The DeepCoy algorithm embedded in Evaluation Master was then used to generate 2450 high-quality decoy molecules based on the 50 active molecules. The decoy molecule generation parameters were: dataset source: ZINC library; number of generated molecules: 50; number of effective molecules: 50; batch size: 1; and the loaded model: DeepCoy_DUDE_model_e09.pickle.

[0026] Finally, RDKit is used to preprocess the dataset, performing deduplication, desalting, and removal of erroneous structures to ensure the quality and consistency of the dataset.

[0027] See results Figure 1 .

[0028] As can be seen, to construct a high-quality dataset of active molecules for benchmark testing of interface software, this invention extracted 2107 active compounds targeting aurora kinase A from the ChEMBL database. The negative logarithmic half-maximal inhibitory concentration (pIC) was retained during screening. 50 ≥6.0 (corresponding IC) 50 Compounds with a molecular weight ≤1 μmol are used to ensure reliable activity data. The structural composition and activity distribution of the dataset are shown below. Figure 1 A. The compounds exhibit good chemical diversity.

[0029] Subsequently, the t-distributed neighborhood embedding algorithm was used for dimensionality reduction and structural clustering, and the dataset was divided into 50 different groups based on molecular fingerprints. Figure 1 (B) Select one representative active molecule from each cluster to maximize molecular skeleton diversity. Using the deep learning-based decoy molecule generation tool DeepCoy, 2450 decoy molecules were generated. These decoy molecules have highly similar physicochemical properties (molecular weight, lipid-water partition coefficient, etc.) to the active molecules, but their topological structures are significantly different and they do not have target binding ability.

[0030] The final result is a non-equilibrium dataset containing active molecules and decoy molecules, and rigorous quality control is performed, including verification of the validity of chemical structures and removal of redundant structures.

[0031] Example 2: Evaluation of the docking model 1. Validation of multi-protein system docking software Experiments were conducted using the built-in automated evaluation module of EvaluationMaster. Four mainstream docking software programs—AutoDock-GPU, Vina-GPU, PLANET, and KarmaDock—were selected, along with 14 Aurora Kinase A protein crystal structures (obtained from the RCSB protein database) with resolutions better than 1.80 Å. Protein pretreatment was performed using PDBFixer: filling in missing residues and atoms, removing water of crystallization and heteroatoms, and protonating at physiological pH (7.4). The Propka algorithm was then used to minimize energy, eliminate steric hindrance, and preserve the native conformation of the protein backbone. PyMOL was used to assemble all protein structures based on the binding pocket, unifying the spatial coordinates of the active site.

[0032] Using the previously constructed Aurora Kinase A evaluation dataset, software performance was evaluated using the area under the receiver operating characteristic (ROC) curve and box plots, and statistical differences were analyzed using t-tests. Since the raw scores of KarmaDock and PLANET were positive, their scores were negative to match the scoring rules of AutoDock-GPU and Vina-GPU.

[0033] The core interface parameters for the four software programs are as follows: 1. AutoDock-GPU: 20 runs, center coordinates (-7.041102564, 27.68179487, 80.28092308), box size 30, other parameters default; 2. Vina-GPU: 8000 threads, center coordinates and box size are the same as above, other parameters are default; 3. PLANET: Center coordinates are the same as above, other parameters are default; 4. KarmaDock: Center coordinates are the same as above, random seed 2023, score threshold 5000, other parameters are default.

[0034] The enrichment factor, Boltzmann enhanced receiver operating characteristic, and area under the precision-recall curve were calculated for different screening ratios to comprehensively evaluate the software's enrichment ability.

[0035] See the summary of results below. Figure 2 .

[0036] As can be seen, to screen the optimal docking tool for the Aurora Kinase A target, this invention selected four mainstream software programs for evaluation: two classic scoring function GPU-accelerated tools, AutoDock-GPU and Vina-GPU, and two emerging deep learning docking models, PLANET and KarmaDock. Receiver operating characteristic (ROC) curve analysis showed ( Figure 2 A) The area under the curve (AUC) for the KarmaDock-based combination of the 6VPL protein crystal structure reached 0.833, outperforming all other software-protein combinations. The t-test results showed a t-value of -11.57 and a p-value of 3.35 × 10⁻⁶. -30 This demonstrates that the combination's ability to distinguish between active molecules and decoy molecules is statistically significant. (Box plot) Figure 2 B) This also visually demonstrates the clear distinction between the scoring ranges of active molecules and decoy molecules, confirming KarmaDock's excellent ability to recognize active compounds of aurora kinase A.

[0037] 2 Eutectic ligand weight docking Protein crystal 6VPL and its co-crystal ligand were selected, and re-docking was performed using KarmaDock with a random seed of 2023 and a fraction threshold of 0. After docking, the original ligand conformation was superimposed and compared with the docked conformation. The root mean square deviation between the two was calculated using OpenBabel to evaluate the conformation prediction accuracy.

[0038] visible, To further verify the reproducibility and conformation prediction accuracy of KarmaDock, this invention utilizes the 6VPL crystal structure and its eutectic ligands to conduct re-docking experiments. The docking conformation is then superimposed on the native conformation. Figure 2 C), the overall structures of the two are highly compatible, and the key functional groups (hydrogen-bonding groups and hydrophobic groups) within the pocket are well aligned. The root mean square deviation of the conformation was calculated using OpenBabel software combined with the KarmaDock algorithm, and the result was 1.89 Å, which is lower than the 2.0 Å reliability threshold recognized in the field of structural biology, proving that KarmaDock can accurately reproduce the natural binding mode of the ligand.

[0039] Based on multiple indicators such as structural fidelity, low root mean square bias, and functional group matching, KarmaDock was selected as the core computational tool for the multi-level virtual screening process, used for subsequent molecular dynamics simulations and screening of candidate molecules for biochemical activity verification.

[0040] To evaluate the early enrichment performance of each model, enrichment factors and partial areas under the curve (AUCs) were calculated at screening ratios of 1% and 5% (Figure 3). At a screening threshold of 1%, the combination of KarmaDock-6VPL and KarmaDock-6VPM achieved the highest enrichment factor of 23.491. This benchmark dataset contains 50 active molecules and 2400 decoy molecules (a total of 2450 compounds). The theoretical maximum enrichment factor for this system at a screening ratio of 1% is 49.00. The experimental results demonstrate that KarmaDock possesses excellent early enrichment capabilities, but it did not reach the theoretical upper limit. In comparison, the enrichment effects of the two classic tools, AutoDock-GPU and Vina-GPU, were significantly lower, and PLANET's ability to prioritize active molecules under stringent screening conditions was also weak. When the screening ratio was increased to 5%, KarmaDock maintained optimal performance across various Aurora Kinase A protein conformations.

[0041] Partial area under the curve (AUC) analysis focused on the low false positive rate range to evaluate the efficiency of early screening stages. Under the most stringent 1% false positive rate condition, KarmaDock's partial AUC remained between 0.550 and 0.615 across different protein conformations; while the results from AutoDock-GPU, Vina-GPU, and PLANET were generally close to 0.500, indicating that these three tools struggled to achieve high-confidence molecular differentiation in the early stages of screening.

[0042] Based on the above results, this invention incorporates KarmaDock into a structure-based virtual screening process as the core docking tool.

[0043] Example 3: Hybrid Virtual Screening Process A self-built screening library was constructed by integrating the MCE drug molecule library and the Maybridge compound library. Preprocessing was performed using RDKit scripts: parsing simplified molecular linear input specifications, removing molecules with more than 80 atoms, and desalting. The final output was a standard simplified molecular linear input specification. This stage did not perform tautomer normalization, physiological pH protonation correction, or filtering of interfering compounds; only basic desalting was performed. A total of 160,000 unique molecules were obtained and stored in simplified molecular linear input specifications, Mol2, and PDBQT formats.

[0044] Based on the results of previous software evaluation, a three-level virtual screening process was established: the first stage used PSICHIC to complete high-throughput affinity ranking; the second stage used KarmaDock to conduct structure-based virtual screening; and the third stage used DeepDock rescoring, combined with similarity retrieval and manual screening by medicinal chemistry experts to obtain candidate compounds.

[0045] Phase 1 PSICHIC PSICHIC is a graph neural network model based on physicochemical constraints that can directly use protein sequences and ligands to simplify the linear input of molecules for canonical prediction of interaction characteristics. This model combines a geometrically adaptive graph attention layer to integrate interatomic physicochemical complementarity and spatial geometric relationships, and introduces a differentiable energy constraint module to ensure that the prediction results conform to thermodynamic laws. It does not rely on the three-dimensional structure of proteins and is suitable for cross-target affinity prediction and high-throughput virtual screening.

[0046] This invention calls the pre-trained model PDBv2020_PSICHIC, with a computing device of cuda:0, a batch size of 16, and disables cluster saving while enabling parsing result saving; the model score is only used for molecular ranking in the screening stage.

[0047] Phase Two KarmaDock KarmaDock is a geometric deep learning docking framework that integrates an implicit energy field model and a dynamic graph attention mechanism to accurately predict ligand-protein binding conformations and affinities. Its core architecture is a multi-scale isovariant graph neural network that directly resolves three-dimensional atomic coordinates, identifying local environments such as hydrogen bonds and hydrophobic pockets, as well as long-range interactions. Simultaneously, it introduces physical constraints such as torsional potential energy and van der Waals repulsion to ensure conformational rationality. This tool can directly generate binding conformations based on simplified linear molecular input specifications, without requiring pre-defined ligand three-dimensional structures, thus avoiding biases introduced by initial conformations.

[0048] Phase 3 DeepDock DeepDock is a geometric deep learning docking model that uses a dual-flow graph convolutional network to analyze the protein surface mesh (local chemical microenvironment) and ligand molecular map (atomic bonding features) respectively, constructing an atomic-level statistical potential model to predict binding conformation and affinity. The DeepDock score is defined as the potential energy corresponding to the negative log-likelihood of the atomic distance; the lower the score, the more stable the binding conformation.

[0049] Phase 4 Similarity Scoring A Morgan / Extended Connectivity Fingerprinting algorithm combined with the Tanimoto similarity algorithm was used to calculate the structural similarity between candidate molecules and a reported library of active molecules of Aurora Kinase A. Molecules were converted into fixed-length binary fingerprints, and the Tanimoto coefficient (intersection / union ratio) was calculated to characterize structural similarity. The similarity score for each candidate molecule in this invention is its maximum Tanimoto coefficient with molecules in the reference library. The reference dataset AURA_IC50.csv has been uploaded to GitHub.

[0050] The process is as follows: Figure 4 As shown.

[0051] visible, This invention constructs a multi-stage hybrid virtual screening system to screen high-quality candidate molecules from a massive number of compounds while balancing computational efficiency and scoring accuracy. First, it integrates the MCE drug molecule library and the Mébridge screening library, using RDKit for standardization preprocessing, ultimately obtaining a self-built molecular library containing 160,000 unique compounds (in simplified linear input format). Since this stage only performs basic molecular standardization, subsequent activity interpretation requires caution and must be combined with structure correction and activity verification.

[0052] Phase 1: The high-throughput deep learning model PSICHIC was used to predict binding affinity based on protein sequence and ligand structure. Due to the large size of the compound library, a screening threshold was set to retain the top 15% of molecules (affinity score > 7.080), ultimately yielding approximately 25,000 potential candidates. While sequence-based deep learning models are fast, their accuracy is lower than structure-based methods. Therefore, a lenient screening standard was adopted to eliminate the vast majority of inactive molecules while ensuring high recall, retaining diverse chemical frameworks, and preventing the premature elimination of active molecules.

[0053] The second stage involved targeting 25,000 candidate molecules with the 6VPL crystal structure of aurora kinase A. Rigid docking was performed using KarmaDock, with a strict threshold set to retain only the top 5% (docking score > 81.88). Preliminary enrichment analysis confirmed that KarmaDock showed the greatest advantage in the top 1%–5% range. This step, focusing on binding molecules with optimal geometry and energy matching, is the core filtering step in the entire screening process.

[0054] Phase 3: DeepDock was used to re-score and optimize the docking conformations, eliminating molecules with high docking scores but unreasonable spatial conformations or poor local interactions, while ensuring molecular skeleton diversity and reserving sufficient selection space for the final screening. Finally, medicinal chemistry experts conducted manual screening based on indicators such as pharmacophore matching, synthetic feasibility, and skeleton novelty, identifying 10 compounds for in vitro activity verification. The structures of the 10 candidate molecules and the scoring data at each stage are shown in Table 1.

[0055] Table 1. Scoring results of 10 screened compounds at each stage Note: Astrocytocin is a positive control for this enzyme activity experiment; the similarity score is the maximum Tanimoto similarity (Morgan fingerprint / extended connectivity fingerprint 4) between the target molecule and the reported Aurora kinase A active molecule library. The lower the score, the stronger the structural novelty.

[0056] The compound structure is as follows: Example 4: Evaluation of bioactivity 1. Homogeneous time-resolved fluorescence aurora kinase in vitro enzyme activity assay Homogeneous time-resolved fluorescence spectrometry was used to determine the inhibitory activity of compounds against aurora kinase A, a method that is resistant to fluorescence interference and has high sensitivity. The experimental system contained aurora kinase A, S2 substrate, europium-labeled antiphosphorylated substrate antibody, and XL665-labeled streptavidin; the buffer composition was 50 mmol / L 4-hydroxyethylpiperazine ethanesulfonic acid (pH 7.5), 5 mmol / L magnesium chloride, and 1 mmol / L dithiothreitol. All test compounds and reference inhibitors were prepared as 10 mmol / L stock solutions using dimethyl sulfoxide.

[0057] The experiment was conducted in 384-well white polystyrene plates with a total reaction volume of 10 μL. The compound was added using an acoustic pipetting workstation, with a final dimethyl sulfoxide (DMSO) volume fraction of 0.5%. Aurora kinase A (final concentration 2.00 nmol) and the compound were added first, and the mixture was pre-incubated at 25°C for 10 minutes. Then, a substrate / ATP mixture (S2 substrate final concentration 1 μmol, ATP final concentration 4 μmol) was added, and the reaction was carried out at 25°C for 40 minutes. The reaction was terminated by adding a detection reagent, and after incubation for another 60 minutes, fluorescence signals at 620 nm (donor) and 665 nm (acceptor) were detected.

[0058] The positive control was a system with complete enzyme activity, and the negative control was a system without enzyme.

[0059] A four-parameter nonlinear regression model was used to fit the dose-response curve, and the IC was calculated. 50 The fitting software used was GraphPadPrism 8.0; all experiments were set with duplicate wells, and the results were labeled with 95% confidence intervals and goodness of fit.

[0060] 2. Cell Culture JIMT-1, NIH / 3T3, MDA-MB-231, and MDA-MB-453 cells were all cultured in high-glucose Dalberg modified Eagle medium supplemented with 10% fetal bovine serum, 100 units / mL penicillin, and 100 μg / mL streptomycin. Culture conditions: 37℃, 5% carbon dioxide, and saturated humidity.

[0061] 3. Cell proliferation experiment 72-hour cell proliferation assay: Tumor cells were seeded at 3500 cells per well in 96-well plates and cultured for 72 hours after adding the test compound. The initial concentration of the compound was 20 μmol, serially diluted 3-fold, for a total of 7 concentration points. Cell viability was assessed using a cell counting kit-8, and the IC50 inhibition rate was calculated using ActivityBase software. 50 Each experiment was conducted in triplicate.

[0062] 48-hour short-term proliferation assay (NIH / 3T3, MDA-MB-231): Cells were seeded at 3000 cells per well, and drug was added 24 hours after adhesion, with a final volume of 200 μL. The compound was serially diluted 5-fold to a total of 8 concentrations (maximum 50 μmol), with a final dimethyl sulfoxide (DMSO) volume fraction of 0.1%. After 48 hours of drug treatment, the cells were added to a cell counting kit-8, and absorbance was measured at 450 nm after 30 minutes of incubation. MLN8237 and CCT241736 were used as reference drugs in the experiment; the results were interpreted only as short-term cell phenotypic changes.

[0063] The results are shown in Table 1 and appendix. Figure 5 .

[0064] As can be seen, this invention employs homogeneous time-resolved fluorescence experiments to determine the inhibitory activity of candidate compounds against aurora kinase A at a concentration of 25 μmol, selecting astrosarcine, a broad-spectrum ATP competitive kinase inhibitor, as a positive control. As shown in Table 1, compound 8 exhibited the highest inhibition rate (103.35%), followed by compounds 3 and 5, with inhibition rates of 99.22% and 92.00%, respectively; the positive control astrosarcine showed an inhibition rate of 100.06%.

[0065] Further determination of the IC50 of the compound 50 The orally active dual-target inhibitor CCT241736 was selected as the reference drug (see attached). Figure 5 The dose-response curve results showed that compound 5 had the strongest inhibitory activity, with an IC50 value of [missing value]. 50 The concentration was 21.64 nanomolars; compound 8 was second, with an IC50 value of 21.64 nanomolars. 50 The concentration of compound 3 was 46.03 nanomolar; the concentration of compound 4 was 157 nanomolar; the IC50 of the reference drug CCT241736 was 46.03 nanomolar. 50 It is 2.912 nanomoles.

[0066] Cell viability assays were performed on three types of breast cancer cells (JIMT-1, MDA-MB-231, and MDA-MB-453). In JIMT-1 cells, all tested compounds reduced cell viability, with compound 3 showing weaker activity than CCT241736. In triple-negative breast cancer MDA-MB-231 cells, compound 3 exhibited the best anti-proliferative activity, with an IC50 value of [missing value]. 50 The concentration was 0.8207 μmol; compounds 5 and 8 exhibited moderate activity, with an IC50 value of 0.8207 μmol. 50 The concentrations were 60.84 μmol and 84.36 μmol, respectively. In MDA-MB-453 cells, compound 3 maintained good activity, while compounds 5 and 8 also showed some inhibitory effects on cell proliferation. These cell experiment results are for phenotypic purposes only and do not constitute direct evidence that the compounds target aurora kinase A.

[0067] This invention simultaneously detected the short-term growth inhibitory effects of compounds on normal mouse embryonic fibroblast NIH / 3T3 cells and triple-negative breast cancer MDA-MB-231 cells after 48 hours of treatment. In NIH / 3T3 cells, compounds 3 and 5 inhibited cell growth in a concentration-dependent manner, with an IC50 value of [missing value]. 50 The concentrations were 3.71 μmol and 4.67 μmol, respectively, within the same concentration range as the reference drugs MLN8237 (2.01 μmol) and CCT241736 (4.67 μmol). In the 48-hour experiment on MDA-MB-231 cells, compound 3 showed the strongest short-term growth inhibition effect (IC50). 50 =5.14 μmol), compound 5 has weaker activity (IC50). 50 =49.78 μmol); Under these conditions, the inhibitory effects of MLN8237 and CCT241736 were weaker than those of compound 3. Since this experiment did not conduct cell-level target binding and kinase subtype selectivity verification, this result is only interpreted as short-term cell phenotypic inhibition and does not confirm that its mechanism of action depends on Aurora kinase A.

[0068] In summary, compounds 3, 5, and 8 all possess aurora kinase A inhibitory activity and exhibited differentiated cell growth inhibition effects in different breast cancer cell models. Among them, compound 3 showed the best overall phenotype in triple-negative breast cancer-related cell experiments, but its overall mechanism of action and target selectivity still require further experimental verification.

[0069] Example 5: Molecular Dynamics Simulation The ligand-protein conformation with the best KarmaDock score was selected as the initial structure, and classical molecular dynamics simulations were performed at 500 nanoseconds using GROMACS 2021.7. The protein was subjected to the AMBER14SB force field, and the ligand to the universal AMBER2 force field. The stability of the system was evaluated by calculating parameters such as root mean square deviation, root mean square fluctuation, number of hydrogen bonds, and radius of gyration.

[0070] The results are attached. Figure 7-11 .

[0071] As can be seen, based on the in vitro activity results, compounds 3, 5, and 8 were selected to conduct 500 nanosecond molecular dynamics simulations with Aurora kinase A protein to explore the ligand-target binding stability and key amino acid interactions, providing a basis for subsequent structure optimization. The simulation software used was GROMACS 2021.7.

[0072] Analysis using compound 3 as a representative example: Three sets of parallel simulation results show that the overall residue fluctuation of the protein is small, and the root mean square fluctuation of residues in the active pocket region (residues 140-220) is generally below 1.5 Å, indicating that compound 3 can significantly stabilize the catalytic pocket structure after binding. Root mean square deviation analysis of the protein and ligand shows that all simulation systems reach conformational convergence within 250 nanoseconds; the root mean square deviation of the ligand conformation is stable in the range of 0.2-0.4 nanometers, proving that the overall binding conformation is stable and can be used for subsequent interaction analysis.

[0073] The binding free energy was calculated using molecular mechanics / generalized Born surface area method. In three parallel experiments, the binding free energy of compound 3 ranged from -37.01 to -40.71 kcal / mol. The negative binding energy indicates that the binding process can occur spontaneously. Energy component breakdown results show that van der Waals interactions and electrostatic interactions are the main beneficial binding forces, sufficient to offset the adverse effects of solvation energy. This result supports the stability of the binding mode, but it cannot be completely equated with thermodynamic experimental data.

[0074] Free energy surface analysis, radius of gyration, and solvent-accessible surface area results show that the complex system is concentrated in a single low-energy steady-state range. The radius of gyration and solvent-accessible surface area remain stable throughout the 500 nanosecond simulation, proving that the hydrophobic core structure of the protein does not change significantly after ligand binding, and the overall thermodynamic properties of the complex are stable.

[0075] Residue energy decomposition and contact occupancy analysis showed that alanine 213, threonine 217, arginine 220, and leucine 263 were key residues involved in binding. Among them, the single residue contribution energy of alanine 213 and threonine 217 exceeded -2.0 kcal / mol. The contact occupancy of alanine 213 (hinge region) and arginine 220 with the ligand was close to 100%, which are the core sites for maintaining binding.

[0076] Three-dimensional binding conformation and two-dimensional interaction maps show that compound 3 can form a continuous hydrogen bond network with alanine 213 in the hinge region and threonine 217, the gatekeeper residue. The number of hydrogen bonds remained stable at 3-5 throughout the simulation, which is the structural basis for the stable binding and excellent inhibitory activity of compound 3. The study also found that the synergistic stabilizing effect of the hinge region binding + threonine 217 - arginine 220 region is the key to compound 3's inhibition of aurora kinase A.

[0077] Compounds 5 and 8 also exhibited similar binding stability and interaction modes.

[0078] Example 6: Prediction of absorption, distribution, metabolism, excretion and toxicity, docking comparison with the Aurora Kinase family and pharmacophore analysis This invention utilizes ADMETlab 3.0 to predict the physicochemical properties and druggability of three candidate compounds (Table 2). All three meet the optimal range of the five rules for oral drug formulations. Compound 3 has a molecular weight of 470.12 Daltons, below the critical value of 500 Daltons; its lipid-water partition coefficient is 2.67, indicating moderate lipid solubility and a good balance between solubility and membrane permeability; its intestinal absorption probability is as high as 0.995, indicating near-complete intestinal absorption; its blood-brain barrier penetration probability is only 0.0005, indicating extremely low risk of central nervous system toxicity. However, this compound has a potassium channel inhibition probability of 0.681, suggesting potential cardiotoxicity, which requires close attention in subsequent structural optimization.

[0079] Compound 5 has a molecular weight of 386.12 Daltons, an intestinal absorption probability of 0.751, and a human potassium channel inhibition probability of only 0.278, indicating a lower risk of cardiotoxicity than the other two compounds. Compound 8 exhibits the strongest lipid solubility (lipid-water partition coefficient of 4.83) and the lowest intestinal absorption probability (0.370). All data are model predictions and are used only for internal comparison and ranking of candidate molecules; they are not equivalent to experimental safety data.

[0080] Table 2. Predicted absorption, distribution, metabolism, excretion, and toxicity of compounds 3, 5, and 8 Docking comparisons were conducted on the Aurora Kinase family (Aurora Kinase A, B, and C) (Table 3) to explore the differences in binding to different isoforms of the three compounds. The results showed that the docking fraction distributions of the three compounds differed across isoforms, with compound 8 exhibiting the smallest difference in binding fractions among the three isoforms. These results are based solely on model calculations and are only used for relative comparisons within the system; they cannot be used as experimental evidence for isoform selectivity. This invention only experimentally confirms that the compounds possess inhibitory activity against Aurora Kinase A. Binding mode analysis revealed that the Aurora Kinase B pocket can form additional hydrogen bonds, which is the structural reason for the differences in binding fractions and can provide a reference for subsequent selective molecule design.

[0081] Table 3. KarmaDock docking scores of compounds 3, 5, and 8 with various proteins in the Aurora Kinase Family. Structure-activity relationship analysis was conducted based on the Aurora Kinase A binding pocket (PDB number: 6VPL), and the results are shown in the appendix. Figure 12 .

[0082] It is evident that the three compounds exhibit a conserved core-binding mode, while their peripheral interactions differ. The formation of a bidentate hydrogen bond network with alanine 213 in the hinge region is essential for their inhibitory activity. The nitrogen and oxygen atoms on the molecular backbone act as hydrogen bond donors and acceptors, respectively, with the hydrogen bond distances remaining stable at 3.0–3.3 Å. This hinge-binding mode precisely anchors the molecule within the ATP-binding pocket, ensuring the core pharmacophore functions effectively.

[0083] Outside the hinge region, the occupancy of the hydrophobic subpocket by the compounds determines the difference in activity: compound 8 has the best complementarity with hydrophobic residues such as leucine 139, leucine 210, and leucine 263, and its fused ring structure can penetrate deep into the hydrophobic region to maximize van der Waals interactions; compounds 3 and 5 have weaker binding effects on the above-mentioned hydrophobic residues, which is also the reason for the difference in energy characteristics among the three.

[0084] Preliminary structure-activity relationship (SOR) hypothesis suggests that polar auxiliary interactions outside the hinge region also affect molecular activity. Compound 3 can form hydrogen bonds with glutamic acid 260 and extend to the asparagine 261 / aspartic acid 274 region, while compound 5 does not exhibit such polar interactions. Due to the limited number of active compounds in this study, these conclusions are only preliminary assumptions based on structure and do not yet constitute a complete SOR model.

[0085] In summary, Aurora kinase A plays a crucial role in the malignant progression and poor prognosis of breast cancer, and the development of potent inhibitors targeting this target has significant clinical value. This invention successfully constructed and validated an AI-based cascaded virtual screening system, integrating sequence-based affinity prediction (PSICHIC), equivalent deep learning docking (KarmaDock), and geometric rescoring (DeepDock) technologies. This multi-stage process efficiently mines candidate molecules from a massive compound library, ultimately yielding three novel lead compounds. Homogeneous time-resolved fluorescence biochemical experiments confirmed that all three compounds possess Aurora kinase A inhibitory activity; molecular dynamics simulations and docking analysis elucidated the binding mechanism between the molecules and the target; and cell experiments demonstrated that the candidate molecules can inhibit short-term tumor cell proliferation. Overall, this invention not only yielded biochemically active Aurora kinase A inhibitor lead compounds but also established a complete, open-source, and reusable AI-assisted kinase inhibitor screening process, which can provide a reference for the development of similar drugs.

Claims

1. The use of a compound of Formula I in the preparation of an aurora kinase A inhibitor, wherein the structure of Formula I is shown below: (Equation I) 2. The application according to claim 1, characterized in that, The compound with the structure of Formula I is compound 3, compound 5, or compound 8.

3. The use of an aurora kinase A inhibitor in the preparation of a drug for treating triple-negative breast cancer, wherein the aurora kinase A inhibitor has the following structure: 。 4. The application according to claim 3, characterized in that, The aforementioned aurora kinase A inhibitor is compound 3, compound 5, or compound 8.

Images