Discovery of Novel AURKA Inhibitors for Triple-Negative Breast Cancer (TNBC) Therapy via a Hybrid Virtual Screening Pipeline, Biological Evaluation and Molecular Dynamics Simulation

Abstract

Aurora kinase A (AURKA) is a pivotal driver of malignant progression and poor prognosis in triple-negative breast cancer (TNBC). In this study, we developed a cascaded AI-driven virtual screening pipeline, integrating sequence-based affinity prediction (PSICHIC), equivariant deep learning docking (KarmaDock), and geometric rescoring (DeepDock) to identify novel AURKA inhibitor candidates. From an in-house 160,000-compound screening library assembled from commercially available collections, three leads (compounds 3, 5, and 8) were selected and subsequently validated via HTRF biochemical assays, exhibiting potent enzymatic inhibition with IC₅₀ values of 157 nM, 21.64 nM, and 46.03 nM, respectively. Cell-based assays demonstrated that compound 3 produced stronger short-term cell-growth inhibition in MDA-MB-231 (TNBC) cells compared to clinical benchmarks MLN8237 and CCT241736, whereas compounds 3 and 5 showed cell-growth inhibition in NIH/3T3 cells within the same concentration range as the reference inhibitors. Triplicate 500 ns molecular dynamics simulations supported stable binding modes of the identified leads in the AURKA binding pocket. Additional computational analyses further provided supportive information for subsequent lead optimization. This study provides a transparent and open-source workflow for AI-assisted identification of AURKA-active chemotypes.

1. Introduction

Breast cancer (BC) is the most prevalent malignant tumor worldwide and the main cause of cancer-related mortality in women. Specifically, according to the immunohistochemical (IHC) expression, BC can be classified into four subtypes including ER + PR + HER2+, ER + PR + HER2–, ER–PR–HER2+, and ER–PR–HER2–. The last subtype, ER–PR–HER2–, is referred to as triple-negative breast cancer (TNBC). Of note, about 15–20% of all BC cases is associated with TNBC, which has been recognized as an aggressive subtype with a high risk of distant metastasis. More significantly, among all types of BC, TNBC is related to the poorest prognosis. Thus, there is an urgent need for novel and effective treatment strategies for TNBC.

Aurora kinase A (AURKA) is an evolutionarily conserved kinase which demonstrated a crucial role in cell cycle regulation. It is integral to critical cellular processes and influence the G2-to-M phase transition. In the meantime, its dysregulation can impair mitotic checkpoint functions to induce genetic instability and tumorigenesis. It has been reported that the overexpression of AURKA contributed to drug resistance and the shortage of median survival time in about 73% of BC patients. Besides, kinase activity of AURKA is closely related to chemotherapy resistance in TNBC. , Collectively, the abnormal expression of AURKA is significant for the initiation and progression of BC through multiple mechanisms and it has emerged as a promising therapeutic target for TNBC treatment.

Recently, some AURKA-targeting inhibitors have been disclosed and exhibited promising results in both preclinical and clinical studies. MLN 8237, developed by Millennium Pharmaceuticals Ltd., is a selective AURKA inhibitor with an IC₅₀ value of 1.2 nM (Figure ). , PHA739358, a pan-Aurora kinase inhibitor, can effectively inhibit all three human Aurora kinases, AURKA (IC₅₀: 13 nM), B (IC₅₀: 79 nM), and C (IC₅₀: 61 nM). Additionally, AT9283, a multitarget kinase inhibitor, demonstrates potent inhibition of AURKA with an IC₅₀ value of 3 nM. These representative AURKA inhibitors have shown diverse therapeutic effects on tumors, though some have not demonstrated significantly greater efficacy compared to existing treatments in clinical trials. Nevertheless, some AURKA inhibitors have exhibited inevitable toxicity in clinical trials; for instance, myelosuppression and gastrointestinal adverse effects. Additionally, the activation of bypass signaling pathways also contributed to the development of drug resistance. These drawbacks resulted in the absence of any AURKA small-molecule inhibitors in market successfully for TNBC treatment. This highlights the urgent need for more effective compounds targeting AURKA.

1.

Representative AURKA inhibitor structures.

Computer-aided drug design (CADD) has accelerated modern drug discovery by expediting lead compound identification and optimization while reducing costs. , Virtual screening (VS) plays a crucial role in rapidly screening millions of compounds to identify promising candidates, thereby improving efficiency and success rates in the early stages of drug discovery. , Additionally, molecular dynamics (MD) simulations provide detailed insights into drug–target interactions and binding stability. By prioritizing molecules with optimal pharmacokinetics and target affinity, these techniques collaboratively address the limitations of traditional trial-and-error approaches. , As drug targets become increasingly complex, the multiscale approach of CADD–integrating VS, MD, and experimental validation–remains indispensable for developing therapies for challenging diseases such as cancer and neurodegenerative disorders.

In this work, we established a robust, cascaded AI-driven discovery framework that harmonizes sequence-based affinity prediction with geometric deep learning-based docking and rescoring. Our pipeline integrates PSICHIC, a sequence-based model for predicting and ranking binding affinity from protein sequences and ligand representations, with KarmaDock, an equivariant deep learning docking tool designed to capture complex ligand–protein interactions. To further refine the candidates, the geometric deep learning framework DeepDock was employed alongside structure-based similarity scoring for high-precision rescoring. Further bioactivity assays revealed that compounds 3, 5, and 8 exhibited potential AURKA biochemical inhibitory activity and growth-inhibitory effects in triple-negative breast cancer (TNBC) cell lines in vitro. In summary, our workflow provides an effective strategy for the discovery and development of structurally distinct AURKA inhibitor candidates.

2. Results and Discussion

2.1. Collection of AURKA Evaluation Data Set

To establish a high-quality set of active molecules for benchmarking docking software, we extracted 2107 AURKA-targeted bioactive compounds from the ChEMBL database. A systematic screening process retained only compounds with pIC₅₀ ≥ 6.0 (equivalent to IC₅₀ ≤ 1 μM), ensuring robust activity data. The structural composition and bioactivity profile of the data set are illustrated in Figure A, underscoring its chemical diversity.

2.

Distribution of the AURKA active data set. (A) The pIC₅₀ distribution of AURKA active compounds. (B) The t-SNE plot of the AURKA active data set.

Subsequently, t-SNE (t-distributed stochastic neighbor embedding) was employed for dimensionality reduction and structural clustering, partitioning the data set into 50 distinct groups based on molecular fingerprints (Figure B). From each cluster, a representative active molecule was chosen to maximize scaffold diversity. We utilized DeepCoy, a deep learning-based decoy generator, which produced 2400 decoy molecules. These decoys were carefully matched to their corresponding actives in physicochemical properties (e.g., molecular weight, logP) while preserving topological dissimilarity to preclude binding activity.

The final evaluation data set, encompassing imbalanced active–decoy pairs, underwent stringent quality control, including chemical validity and structural redundancy checks.

2.2. Docking Model Evaluation

2.2.1. Docking Program Validation with Multiple Proteins

To identify the best-performing docking software for the AURKA target, four representative docking tools were selected to evaluate the performance: AutoDock-GPU and Vina-GPU, which represent the GPU-accelerated evolution of classical scoring functions; and PLANET and KarmaDock, which represent the emerging class of deep learning-based docking paradigms, summarized in Figure . Receiver operating characteristic (ROC) curve analysis (Figure A) showed that the KarmaDock-6VPL combination achieved the highest ROC-AUC value (0.833), surpassing all other software–protein combinations. A t-test (Supplementary Table S1) gave a t-statistic of −11.57 and a p-value of 3.35 × 10̂^–30, indicating KarmaDock-6VPL’s significant advantage in discriminating active compounds from decoys. Box plots (Figure B) further revealed a clear separation between active molecules and decoys, underscoring KarmaDock’s robust capacity to identify bioactive compounds for AURKA.

3.

Evaluation results of docking programs and AURKA proteins. (A) The ROC curve for the molecular docking of multiple AURKA proteins with the selected docking programs using the AURKA validation data set; (B) the box plot of KarmaDock with 6VPL protein; (C) the redocking result of AURKA protein (PDB ID: 6VPL) using KarmaDock (Blue: original structure; Purple: redocking conformation).

2.2.2. Redocking of Co-Crystallized Ligand

To rigorously validate KarmaDock’s reproducibility and conformational accuracy, we performed redocking experiments using the 6VPL crystal structure and its cocrystallized ligand. Superimposing the native and postdocked ligand conformations (Figure C) showed high structural concordance, particularly in aligning key functional groups (e.g., hydrogen-bonding motifs and hydrophobic moieties) within the binding site. The root-mean-square deviation (rmsd) between the native and docked poses, calculated via OpenBabel (v3.1.1) using the Kabsch algorithm, was 1.89 Å–well below the 2.0 Å threshold commonly accepted for reliable pose prediction in structural biology.

These findings underscore the methodological robustness of KarmaDock in reproducing experimental binding modes. Collectively, the validation metrics–structural fidelity, low RMSD, and functional group preservation–establish KarmaDock as the core computational engine in our multitiered virtual screening workflow, facilitating the identification of high-confidence AURKA inhibitor candidates for subsequent molecular dynamics (MD) simulations and biochemical evaluation.

2.2.3. Early-Stage Enrichment Capability Assessment

To further scrutinize the early enrichment capabilities of these models, enrichment factor (EF) and partial AUC (pAUC) analyses were conducted at 1% and 5% thresholds (Figure ). At the 1% cutoff, the KarmaDock-6VPL and KarmaDock-6VPM combinations achieved the highest EF1% value of 23.491 (Figure A). For the present benchmark set, which comprises 50 actives and 2400 decoys (2450 compounds in total), the theoretical maximum EF1% is 49.00. Therefore, the observed EF1% value indicates strong early enrichment, although it does not reach the theoretical upper bound. In contrast, the classical scoring functions, AutoDock-GPU and Vina-GPU, showed substantially lower enrichment, whereas PLANET was less effective at prioritizing active compounds under this stringent condition. When the threshold was extended to 5% (Figure B), KarmaDock maintained leading performance across nearly all AURKA structures, further supporting its use as the primary docking engine for early stage hit prioritization.

4.

Analysis of the early enrichment capabilities of docking software. The heatmap of EF at (A) 1% and (B) 5%; the heatmap of pROC-AUC at (C) 1% and (D) 5% for selected docking programs.

The precision of these docking paradigms was further examined through pAUC analysis (Figure C,D), which focuses on early stage screening efficiency by restricting the false positive rate (FPR). At the most rigorous 1% FPR level, KarmaDock consistently outperformed the other three tools, maintaining pAUC values between 0.550 and 0.615 across diverse protein conformations. In comparison, the performance of AutoDock-GPU, Vina-GPU, and PLANET largely converged toward a value of 0.500, suggesting a limited capacity for high-confidence discrimination in the extreme early phase of screening.

Based on the above results, KarmaDock was integrated into our structure-based virtual screening (SBVS) pipeline as the primary docking engine.

2.3. Hybrid Virtual Screening Workflow

To prioritize high-quality candidates from a vast chemical space while balancing computational efficiency and scoring accuracy, a multistage hybrid virtual screening workflow was implemented (Figure ). An in-house virtual library was first compiled from the MCE drug-like and Maybridge screening collections, undergoing standardized preprocessing via RDKit, to yield a final set of 160,000 unique compounds in SMILES format. Because only a minimal molecular standardization workflow was applied at this stage, downstream hit interpretation should be made cautiously and should ideally be supported by additional orthogonal structure curation and follow-up validation.

5.

Hybrid virtual screening workflow of AURKA inhibitor discovery.

The initial screening phase utilized PSICHIC, a high-throughput deep learning model, to predict binding affinities based on protein sequences and ligand SMILES. Given the massive scale of the library, a top 15% threshold (affinity score >7.080) was applied to discard a large proportion of likely inactive molecules while retaining approximately 25,000 potential hits. Since sequence-based deep learning models are exceptionally fast but generally less precise than structure-based methods, this broader cutoff ensures that the workflow discards a vast majority of definitive nonbinders while maintaining a high recall rate. By retaining the top 15%, we preserved a diverse range of chemical scaffolds that might have been prematurely eliminated by a more aggressive sequence-based filter, while successfully reducing the library to a volume manageable for the subsequent docking phase.

In the second round, the 25,000 candidates were subjected to rigid docking using KarmaDock against the AURKA 6VPL structure. To ensure the selection of only the most geometrically and energetically compatible binders, a stringent top 5% cutoff was applied (docking score >81.88). As demonstrated in the enrichment analysis (Figure ), KarmaDock exhibits its most significant competitive advantage in the early enrichment phase, particularly within the top 1% to 5% of the ranked library. By applying this 5% cutoff, the workflow effectively capitalizes on KarmaDock’s high precision in identifying the most geometrically and energetically compatible binders for the AURKA pocket. This stage serves as the primary “bottleneck” of the funnel, shifting the focus from broad coverage to high-confidence hit identification, thereby isolating the most promising candidates from the thousands of molecules that passed the initial affinity screen.

The third stage focused on pose refinement and consensus scoring using DeepDock for rescoring the docking poses generated in the previous round, applied to eliminate molecules with high docking scores but poor geometric reliability or unfavorable local interactions, allowed for a balanced distribution of high-scoring scaffolds while providing sufficient diversity for the final stage. Finally, through rigorous visual inspection by medicinal chemists, considering factors such as pharmacophore alignment, synthetic accessibility, and scaffold novelty, ten representative compounds were selected for biological activity validation. The chemical structures and multistage scoring data for these 10 leads are detailed in Figure and Table , respectively. The complete screening results for each stage can be found in our GitHub repository.

6.

Structures of the screened 10 molecules. Structures for compounds (A) 1; (B) 2; (C) 3; (D) 4; (E) 5; (F) 6; (G) 7; (H) 8; (I) 9 and (J) 10.

1. Screening Score Result of the Screened 10 Compounds.

entry	PSICHIC affinity	karma dock score	deep dock score	similarity score	@25 μM inhibition rate (%)
compound 1	7.549	103.91	–198.30	0.398	28.54 ± 1.44
compound 2	7.174	106.82	–248.71	0.580	52.08 ± 1.98
compound 3	7.388	106.18	–154.29	0.411	99.22 ± 2.80
compound 4	7.930	107.02	–217.52	0.427	42.22 ± 2.10
compound 5	7.081	100.69	–143.00	0.451	92.00 ± 2.33
compound 6	7.857	100.22	–249.92	0.472	23.19 ± 8.70
compound 7	7.777	93.21	–216.39	0.692	53.57 ± 1.43
compound 8	7.355	107.30	–144.50	0.417	103.35 ± 0.79
compound 9	7.798	106.01	–203.73	0.575	41.57 ± 0.54
compound 10	7.673	81.89	–61.44	0.361	90.93 ± 0.42
staurosporine	-	-	-	-	100.06 ± 0.56

^aThis entry in the table represents the positive control compound used in the enzyme inhibition assay at 25 μM.

^bSimilarity Score denotes the maximum Tanimoto similarity (Morgan/ECFP4 fingerprints) between each selected lead and the curated ChEMBL-derived AURKA-active reference data set (AURA_IC₅₀.csv) used in this study. Lower values indicate greater structural distinctiveness relative to previously reported AURKA-active compounds.

2.4. Biological Evaluation

The biological activity of selected compounds was assessed using the HTRF assay at a 25 μM concentration to measure AURKA inhibition. Staurosporine, a broad-spectrum ATP-competitive kinase inhibitor, served as the positive control. As shown in Figure A, compound 8 achieved the highest inhibition rate at 103.35%, followed by compounds 3, and 5 at 99.22% and 92.00%, respectively. Staurosporine showed an inhibition rate of 100.06%. IC₅₀ values were determined to further evaluate potency, with CCT241736 (an orally active dual FLT3/AURKA inhibitor) as the positive control.

7.

Biological evaluation results of the screened 10 compounds. (A) The Inhibition Rate histogram of the screened 10 compounds (positive reference: Staurosporine); (B) the dose–response curves for compounds 3, 5 and 8 (positive reference: CCT241736); the cell viability assay results for compounds 3, 5 and 8 for (C) JIMT-1 cell line, (D) MDA-MB-231 cell line, and (E) MDA-MB-453 cell line (positive reference: CCT241736). The short-term cell viability assay results for compounds 3 and 5 for (F) NIH/3T3 cell line, and (G) MDA-MB-231 cell line (positive reference: MLN8237 and CCT241736).

The dose–response curves for compounds 3, 5, 8, and CCT241736 (Figure B) revealed that compound 5 exhibited strong inhibitory activity against AURKA (IC₅₀ = 21.64 nM), followed by compound 8 (IC₅₀ = 46.03 nM) and 3 (IC₅₀ = 157.00 nM). CCT241736 had an IC₅₀ of 2.912 nM. Additionally, cell viability assays in breast cancer cell lines, including JIMT-1 (ER–PR–HER2+), MDA-MB-231 (ER–PR–HER2−), and MDA-MB-453 (ER–PR–HER2−), assessed the phenotypic activity of compounds 3, 5, 8, and CCT241736. In the JIMT-1 cell line, all tested compounds reduced viability, although compound 3 was less potent than CCT241736 (Figure C). In MDA-MB-231 cells (Figure D), compound 3 exerted significant antiproliferative activity (IC₅₀ = 0.8207 μM), while compounds 5 and 8 displayed moderate activity (IC₅₀ = 60.84 μM and 84.36 μM, respectively). In MDA-MB-453 cells, compound 3 remained active, and compounds 5 and 8 also showed measurable cell-growth inhibition (Figure E). These cellular observations are reported here as phenotypic activity rather than as direct evidence of cellular AURKA target engagement.

The short-term 48 h cell-growth inhibition of the newly identified leads were further examined using the noncancerous mouse embryonic fibroblast line NIH/3T3 (Figure F) and the TNBC cell line MDA-MB-231 (Figure G). As illustrated in Figure F, both compound 3 and compound 5 exhibited dose-dependent cell-growth inhibition in NIH/3T3 cells with IC₅₀ values of 3.71 μM and 4.67 μM, respectively. These values fall within the same concentration range as the reference Aurora kinase inhibitors MLN8237 (IC₅₀ = 2.01 μM) and CCT241736 (IC₅₀ = 4.67 μM). In the corresponding 48 h MDA-MB-231 assay (Figure G), compound 3 displayed the strongest short-term growth inhibition (IC₅₀ = 5.14 μM), whereas compound 5 was less active (IC₅₀ = 49.78 μM). Under these specific 48 h conditions, MLN8237 and CCT241736 showed weaker apparent growth inhibition than compound 3. Because no direct cellular target-engagement or isoform-selectivity measurements were performed, these data are interpreted conservatively as short-term phenotypic growth-inhibition results rather than as mechanistic proof of AURKA-dependent cellular activity.

Taken together, compounds 3, 5, and 8 showed measurable AURKA biochemical inhibition and variable cell-growth effects across the tested breast cancer models. Among them, compound 3 provided the most consistent short-term phenotypic activity in the TNBC-oriented assays used here, whereas the broader cellular mechanism and selectivity profile remain to be established experimentally.

Interestingly, although compound 5 exhibited stronger biochemical inhibition against AURKA compared to compound 3, it was compound 3 that demonstrated more consistent short-term phenotypic activity in MDA-MB-231 cells. This phenomenon indicates a discrepancy between biochemical potency and cellular efficacy, which may arise from variations in cellular permeability, intracellular exposure, compound efflux, or potential off-target effects that favor compound 3 in a complex cellular environment. Since direct cellular target-engagement or intracellular pathway-inhibition assays were not performed in the current study, these factors remain to be fully elucidated. Future investigations will incorporate these cellular mechanistic studies to better understand the intracellular behavior of these chemotypes.

2.5. Molecular Dynamics

Based on the results of biological assays, compounds 3, 5 and 8 were selected for 500 ns molecular dynamics (MD) simulations with the AURKA protein to further investigate the stability of their binding to AURKA and the interactions with key amino acid residues, providing insights for subsequent structural optimization. The simulations were performed using GROMACS 2021.7.

Taking compound 3 as a representative model (Figure A), the triplicate MD simulations consistently showed that the residue fluctuations across the three replicas remained remarkably low, particularly within the active site (residues 140–220), where RMSF values were generally below 1.5 Å (Figure B). This indicates that the binding of compound 3 effectively rigidifies the catalytic pocket. Analysis of the protein and ligand RMSD (Figure C,D) showed that all three systems achieved structural convergence within the first 250 ns, with the ligand maintaining a stable pose characterized by RMSD values fluctuating between 0.2 and 0.4 nm, confirming a robust equilibrium suitable for detailed interaction analysis.

8.

Trajectory analysis of MD results of compound 3 with the AURKA protein. (A) The structures of compound 3; (B) the RMSF trajectory of AURKA protein in the complexes with compound 3; (C) the RMSD trajectory analysis of AURKA protein in MD; (D) the RMSD trajectory analysis of compound 3 in MD.

The strength of the binding was further estimated using the MM/GBSA method. The calculated binding free energy for compound 3 was favorable across the three replicas, ranging from −37.01 to −40.71 kcal/mol (Figure D–F). Energetic component decomposition (Figure A–C) suggested that van der Waals interactions (VDWAALS) and electrostatic contributions (EEL) were the primary favorable terms within this computational model outweighing the desolvation penalties. These data support a stable binding hypothesis that is consistent with the observed biochemical activity, but they should not be interpreted as a quantitative substitute for experimental thermodynamic measurements.

9.

Total energy analysis of the MD results for compound 3 in complex with the AURKA protein. The total energy decomposition for the complex of AURKA protein with compound 3 at replication (A) 1, (B) 2 and (C) 3; the binding free energy decomposition of the complex of AURKA protein with compound 3 at replication (D) 1, (E) 2 and (F) 3.

The global stability of the AURKA-compound 3 complex was further validated through free energy surface (FES) analysis and structural descriptor tracking. The 2D FES plots (Figure A–C) illustrate that the complex resides in a well-defined, deep-blue low-energy well, centered at an rmsd of approximately 0.32 nm and a radius of gyration (R _g) of 2.02 nm. This sharp energy minimum across all replicas underscores the thermodynamic stability of the predicted binding mode. Furthermore, the stability of the system was corroborated by the consistent R _g values (Figure D) and solvent accessible surface area (SASA, Figure E), which remained stable throughout the 500 ns trajectory, suggesting that the hydrophobic core of the protein remained shielded upon ligand binding.

10.

2D FES, Rg and SASA analysis of compound 3 in complex with the AURKA protein. The 2D FES plots for compound 3 at replication (A) 1, (B) 2 and (C) 3; the (D) R _g and (E) SASA analysis for the complex of compound 3 with AURKA protein.

To pinpoint the specific residues responsible for this affinity, per-residue energy decomposition and contact occupancy analyses were performed. As depicted in Figure A, C, and E, residues Ala213, Thr217, Arg220, and Leu263 emerged as the most significant contributors to the binding energy, with Ala213 and Thr217 often contributing more than −2.0 kcal/mol individually. Contact occupancy plots (Figure B, D, and F) further confirmed that the interaction with Ala213 (the hinge region) and Arg220 maintained near 100% occupancy throughout the simulations.

11.

Contribution to Gibbs energy and occupancy analysis for different residues of the AURKA protein in complex with compound 3. The contribution analysis of amino acid residues for AURKA protein at replication (A) 1, (C) 2 and (E) 3; the occupancy percentage of binding between compound 3 and AURKA protein at replication (B) 1, (D) 2 and (F) 3.

Detailed 3D binding poses (Figure A, D and G) and 2D interaction maps (Figure B, E and H) revealed a consistent hydrogen-bonding network involving the hinge residue Ala213 and the gatekeeper-related residue Thr217. The hydrogen bond count (Figure C, F, and I) consistently fluctuated between 3 and 5 throughout the trajectory, providing a structural explanation for the high stability and potency of compound 3. These MD-derived insights suggest that the synergistic effect of hinge-binding and the stabilization of the Thr217-Arg220 region are critical factors for the high inhibitory activity of compound 3 against AURKA.

12.

Detailed binding mode analysis of compound 3 in complex with AURKA protein. The 3D binding mode of AURKA protein with compound 3 at replication (A) 1, (D) 2 and (G) 3; the 2D binding plots of AURKA protein with compound 3 at replication (B) 1, (E) 2 and (H) 3; the analysis of hydrogen bond counts for compound 3 Trajectories at replication (C) 1 and (F) 2 and (I) 3.

Similar structural stability and binding patterns were observed for compounds 5 (Figures S1–S5) and 8 (Figures S6–S10), the detailed trajectory analyses of which are provided in the Supporting Information.

2.6. ADMET Prediction, Aurora-Family Docking Comparison and Pharmacophore Analysis

As summarized by the calculated descriptors from ADMETlab 3.0, all three compounds fall within the optimal range for oral drug-likeness in Table . Compound 3 possesses the highest molecular weight at 470.12 Da, yet remains below the 500 Da threshold defined by Lipinski’s Rule of Five. Its lipophilicity (logP = 2.67) is moderate, suggesting a well-balanced profile between solubility and membrane permeability. This is further evidenced by its exceptional HIA probability of 0.995, indicating nearly complete absorption in the human intestine. Furthermore, its extremely low BBB penetration probability (0.0005) suggests a minimal risk of central nervous system (CNS) side effects. However, a predicted hERG inhibition probability of 0.681 indicates a potential risk for cardiotoxicity that may warrant further structural refinement in future optimization cycles.

2. ADMET Prediction Results of Compounds 3, 5, and 8 .

entry	MW	logP	HIA	BBB	hERG
compound 3	470.12	2.67	0.995	0.0005	0.681
compound 5	386.12	3.51	0.751	0.0006	0.278
compound 8	381.17	4.83	0.370	0.0006	0.480

Compounds 5 and 8 show related but distinct predicted property profiles. Compound 5 combines a lower molecular weight (386.12 Da) with a predicted HIA of 0.751 and a lower predicted hERG score (0.278) than compounds 3 and 8, whereas compound 8 is the most lipophilic (logP = 4.83) and has the lowest predicted HIA (0.370). Because these readouts are model-based descriptors rather than experimental safety data, we use them here only for comparative prioritization within the current lead set.

The complete ADMET prediction data set, encompassing 125 physicochemical and toxicological descriptors for all three compounds, has been compiled into a comprehensive table and is available for download at our GitHub.

Docking-based comparison across the Aurora kinase family was performed for compounds 3, 5, and 8 as a supplementary structural analysis to explore potential differences in their modeled binding patterns toward AURKA, AURKB, and AURKC. The results in Table suggest that these three compounds may display distinct docking-score profiles across the Aurora isoforms, with compound 8 showing a relatively more even distribution than compounds 3 and 5. However, because these values arise from a model-specific docking calculation, they should be interpreted only as relative within-model comparisons rather than as experimentally validated isoform preferences. In the current study, AURKA remains the primary target and the only Aurora isoform for which biochemical inhibitory activity was experimentally confirmed.

3. Karma Dock Score of Compounds 3, 5, and 8 with Different AURA Family Proteins.

entry	AURKA score	AURKB score	AURKC score
compound 3	108.94	128.51	93.83
compound 5	101.22	127.51	118.52
compound 8	109.05	116.24	111.66

The structural analysis of the corresponding binding modes (Figures , S14,S15) suggests possible explanations for the relative score differences, such as additional hydrogen-bonding opportunities in the AURKB model. These observations are useful for future design, but they do not establish biochemical selectivity in the absence of matched Aurora isoform assays.

13.

Binding mode of compound 3 with AURKA (PDB ID: 6VPL), AURKB (PDB ID: 4AF3) and AURKC (PDB ID: 6GR8) protein. The (A) 3D and (B) 2D binding mode of compound 3 with AURKA protein; the (C) 3D and (D) 2D binding mode of compound 3 with AURKB protein; the (E) 3D and (F) 2D binding mode of compound 3 with AURKC protein.

Kinome-wide activity was further estimated using a deep-learning-based prediction model and visualized through phylogenetic tree maps. Compound 3 showed the broadest predicted off-target profile (Figure S16), compound 5 the most restricted predicted profile (Figure S17), and compound 8 an intermediate pattern (Figure S18). These outputs are best viewed as computational prioritization tools for future counter screening rather than as confirmed kinome selectivity data.

The structural-activity relationship (SAR) of compounds 3, 5, and 8 within the AURKA binding pocket (PDB: 6VPL) highlights a highly conserved anchoring mechanism complemented by divergent peripheral interactions (Figure ). A fundamental requirement for potency across all three inhibitors is the formation of a robust bidentate hydrogen-bonding network with the hinge residue ALA213. The nitrogen and oxygen atoms of the core scaffolds consistently act as hydrogen bond donors and acceptors, maintaining distances between 3.0 Å and 3.3 Å. This hinge-binding motif serves as the essential anchor that orients the molecules within the ATP-binding site, ensuring that the primary pharmacophore is correctly positioned for further stabilizing interactions.

14.

Structural-activity relationship analysis of AURKA binding pocket with compounds (A) 3, (B) 5 and (C) 8.

Beyond the hinge region, the occupancy of the hydrophobic subpockets provides a clear differentiation in how these compounds utilize the available space within the AURKA domain. Compound 8 exhibits the most comprehensive hydrophobic complementarity, effectively shielding residues such as LEU139, LEU210, and LEU263. Its fused ring system appears optimized to fit deeply into these nonpolar regions, suggesting that increasing the lipophilic surface area of the core scaffold is a viable strategy for enhancing van der Waals stabilization. While compounds 3 and 5 also engage these hydrophobic residues, their occupancy is less exhaustive, particularly around the deeper LEU210 and LEU263 clusters, which may explain their slightly different energetic profiles compared to the more structurally optimized compound 8.

A tentative SAR observation is that auxiliary polar interactions outside the hinge region may contribute to activity differences within this small set. Compound 3 distinguishes itself by forming a hydrogen bond with GLU260 (3.11 Å) and by extending toward the ASN261/ASP274 region, whereas compound 5 lacks the same polar reach. Given the limited number of experimentally active compounds discussed here, we interpret this pattern as a preliminary structure-based hypothesis rather than as a definitive SAR model.

3. Conclusion

Given the pivotal role of AURKA in the malignant progression and poor prognosis of breast cancer, the discovery of potent inhibitors remains therapeutically relevant. In this study, we successfully developed and validated an integrated AI-driven virtual screening pipeline, incorporating sequence-based affinity prediction (PSICHIC), equivariant deep learning docking (KarmaDock), and geometric rescoring (DeepDock). This cascaded workflow efficiently navigated a large chemical space, leading to the identification of three structurally distinct leads (compounds 3, 5, and 8) with measurable AURKA enzymatic inhibitory activities, as confirmed by HTRF biochemical assays. The accompanying MD simulations and docking analyses provide a coherent structural hypothesis for AURKA engagement, while the cell-based assays indicate short-term growth-inhibitory activity under the tested conditions. The current data provide only preliminary insight into Aurora-family selectivity, kinome-wide selectivity, and cellular target engagement, all of which will require dedicated experimental validation in future studies. Overall, the present work contributes biochemically active starting points and an openly documented workflow for AI-assisted kinase inhibitor discovery. The revised manuscript therefore emphasizes the screening framework and the experimentally confirmed AURKA biochemical activity, while treating broader mechanistic and selectivity interpretations as future directions.

4. Materials and Methods

4.1. Data Set Curation and Preparation for AURKA Target

The construction of the evaluation molecular data set and the assessment of candidate docking software were conducted using the integrated tool EvaluationMaster. To generate a high-quality molecular data set for docking software evaluation, we retrieved reported AURKA inhibitors (Uniprot ID: O14965) from the ChEMBL database , (2024). Compounds with a pIC₅₀ value of at least 6 (equivalent to IC₅₀ < 1 μM) were selected to ensure the validity and reliability of the active molecular data.

Subsequently, the active molecules were clustered into 50 major groups using the T-SNE algorithm based on their core scaffolds. A representative molecule from each cluster was selected as part of the active component of the evaluation data set. Next, the DeepCoy algorithm integrated into EvaluationMaster was employed to generate 2450 high-quality decoy molecules based on the structures and relevant features of the 50 representative active compounds. The parameters used for decoy genaration as follow: --data set; zinc; --generate_number: 50; --valid_number: 50; --batch_size: 1; --restore: models/DeepCoy_DUDE_model_e09.pickle.

Finally, RDKit , was used to preprocess the data set, including deduplication, desalting, and removing erroneous structures, ensuring the quality and consistency of the data set.

4.2. Docking Protocols Validation

4.2.1. Validation of Docking Programs with Multiple Proteins

To identify the optimal docking software for the AURKA target which exhibits the highest capability in distinguishing active molecules from inactive ones, and to determine the most structurally informative AURKA protein structure, we employed the automated Evaluation Workflow module integrated within EvaluationMaster. This evaluation was conducted on four widely used docking programs: Autodock-GPU, , Vina-GPU, , PLANET, and KarmaDock, as well as 14 protein structures with resolutions better than 1.80 Å retrieved from the RCSB Protein Data Bank (PDB). All protein Structures were prepared using PDBFixer to ensure structural integrity, including repairing missing residues and atoms, removing crystallized water molecules and heteroatoms (desalting), and adding missing heavy atoms. Then the structures were protonated at a physiological pH of 7.4 using the Propka algorithm, followed by a brief energy minimization to resolve steric clashes while maintaining the experimental backbone conformation. Subsequently, all protein structures were aligned in PyMOL based on their respective binding pockets. This structural alignment aimed to standardize the spatial coordinates of the active sites, thereby simplifying the configuration of search space parameters within the EvaluationMaster assessment pipeline.

The performance of these docking software and protein structures was assessed using the AURKA evaluation molecular data set, comprehensively constructed in Section . The effectiveness of each docking method was evaluated using Receiver Operating Characteristic-Area Under Curve (ROC-AUC) and box plots, followed by t tests to further analyze statistical significance. To facilitate ROC-AUC and t-test calculations, the scores of KarmaDock and PLANET were converted to their negative values to align with the scoring conventions of AutoDock-GPU and Vina-GPU. In reality, the original scores of KarmaDock and PLANET are positive.

The Docking parameters of these docking programs as follows.

• (1)AutoDock-GPU: --nrun: 20; --center_x/y/z: −7.041102564, 27.68179487, 80.28092308; --size_x/y/z: 30; other parameters remain at their default settings;
• (2)Vina-GPU: --thread: 8000, --center_x/y/z: −7.041102564, 27.68179487, 80.28092308; --size_x/y/z: 30; other parameters remain at their default settings;
• (3)PLANET: --center_x/y/z: −7.041102564, 27.68179487, 80.28092308; other parameters remain at their default settings
• (4)KarmaDock: --center_x/y/z: −7.041102564, 27.68179487, 80.28092308;, --randon_seed: 2023; --score_threshold: 5000; other parameters remain at their default settings.

To further evaluate the enrichment performance of various software tools on the AURKA target, we calculated the Enrichment Factor (EF), Boltzmann-Enhanced Discrimination of Receiver Operating Characteristic (BEDROC), and Area Under the Precision-Recall Curve (AUPRC).

The enrichment factor at a given fraction x of the ranked list was calculated as

$$\mathrm{E}{\mathrm{F}}_x=\frac{Hit{s}_x/N_x}{A/N}$$

where Hits _x is the number of active compounds recovered within the top fraction N _x of the ranked library, A is the total number of active compounds in the data set, and N is the total number of compounds. In the present benchmark data set, A = 50 and N = 2450 (50 actives and 2400 decoys). Accordingly, the theoretical maximum EF_1% is 49.00, which corresponds to the ideal case in which all compounds ranked within the top 1% are active.

4.2.2. Redocking of the Co-Crystallized Ligand

To further evaluate the accuracy and reliability of KarmaDock docking, we performed a redocking of the 6VPL protein and its cocrystallized ligand using KarmaDock. The docking parameters were set with a random seed of 2023 and a threshold of 0.

Following the redocking process, the original conformation of the cocrystallized ligand was visually compared with the redocked conformation by superimposing both structures. Additionally, OpenBabel was used to calculate the root mean square deviation (rmsd) between the ligand conformations before and after docking.

4.3. Hybrid Virtual Screening Workflow Design

We constructed an in-house database comprising compounds from the MCE drug-like molecule library and the Maybridge database. The in-house screening library was preprocessed using an RDKit-based script. In this step, all input structures were first parsed from SMILES, and entries that failed SMILES parsing were excluded. Molecules containing more than 80 atoms were removed. For the retained structures, salts were stripped using RDKit SaltRemover, and the resulting structures were exported as RDKit canonical SMILES for subsequent screening. No dedicated tautomer canonicalization, protonation-state normalization at physiological pH, or explicit rule-based filtering for PAINS, aggregators, reactive groups, assay-interfering compounds, organometallics, or mixtures was performed in this preprocessing step beyond salt removal. The final database contains 160,000 unique molecules, which was stored in SMILES, Mol2 and PDBQT format using RDkit, OpenBabel 3.0 and AutoDock Tool 1.5.7. Because only minimal molecular standardization was applied at this stage, downstream hit interpretation should be made cautiously and supported by additional curation and follow-up validation.

Based on the evaluation results, we established a comprehensive virtual screening workflow. Initial high-throughput filtering was performed using PSICHIC, a sequence-based affinity-ranking step, followed by KarmaDock in the second stage for structure-based virtual screening. DeepDock was then used in the third stage for rescoring. Finally, a similarity search was conducted to compare selected molecules with existing AURKA inhibitors, followed by a manual inspection by medicinal chemists.

4.3.1. PSICHIC

PSICHIC is a graph neural network (GNN) framework based on physicochemical constraints, designed to directly infer interaction fingerprints from protein sequences and ligand SMILES representations. By leveraging a geometric–adaptive graph attention layer, the model dynamically aggregates physicochemical complementarity (e.g., electrostatic matching, van der Waals contacts) and spatial geometric relationships between neighboring atomic nodes. Additionally, a differentiable energy constraint module (e.g., binding free energy decomposition terms) is incorporated to ensure that the learned interaction fingerprints adhere to data-driven patterns and fundamental thermodynamic principles. This method eliminates dependence on experimental or predicted structural data, making it well-suited for cross-target affinity prediction and virtual screening applications.

PSICHIC inference was conducted using the pretrained model trained_weights/PDBv2020_PSICHIC with a batch size of 16 on cuda:0. The run was performed with save_cluster = False and save_interpret = True. In this study, PSICHIC scores were used only for within-model ranking and compound prioritization prior to downstream docking and rescoring.

4.3.2. KarmaDock

KarmaDock is a geometric deep learning-based molecular docking framework that integrates implicit energy field modeling with a dynamic graph attention mechanism to predict protein–ligand binding modes and affinities accurately.

Its core architecture employs a multiscale equivariant graph neural network (GNN), which directly processes three-dimensional atomic coordinates while encoding local chemical environments (e.g., hydrogen bonds, hydrophobic pockets) and long–range interactions. Physical constraints such as torsional potential energy and van der Waals repulsion are incorporated to ensure conformational accuracy. Unlike traditional search-based methods, KarmaDock utilizes an S-E(3) equivariant neural network that takes SMILES strings as input to generate binding poses de novo. Therefore, the process is independent of any initial 3D ligand conformation, eliminating potential starting-geometry bias.

4.3.3. DeepDock

DeepDock is a geometric deep learning-based molecular docking framework that employs a dual–stream graph convolutional network (GCN) to process protein surface meshes (encoding local chemical microenvironments) and ligand molecular graphs (representing atomic bonding properties). DeepDock constructs an atomic-level statistical potential model by integrating these representations to predict binding conformations and affinities.

The DeepDock score is defined as a potential U­(x) derived from the negative log-likelihood of atom distances; thus, a lower U­(x) value indicates a more probable and stable binding conformation.

4.3.4. Similarity Searching Scoring

Tanimoto similarity search based on Morgan fingerprints is an efficient method for assessing compound structural similarity. This approach first converts the target inhibitor molecule and the compounds in a chemical library into fixed-length binary Morgan fingerprints, typically derived from ECFP4 or similar atomic environment encodings with predefined radii. Structural similarity is then quantified by calculating the Tanimoto coefficient (intersection-overunion ratio) between the fingerprints.

This method is particularly well-suited for rapidly identifying potential active compounds from ultralarge libraries (>10⁶ molecules), providing an efficient primary screening strategy for virtual screening.

In this study, tanimoto similarity was calculated using Morgan/ECFP4 fingerprints, with each selected lead compared against the curated ChEMBL-derived AURKA-active reference data set used in this study (AURA_IC₅₀.csv). For each lead, the reported Similarity Score corresponds to the maximum pairwise similarity observed against that reference set. The AURA_IC₅₀.csv data set is available in our GitHub repository.

4.4. Biological Assays

4.4.1. In Vitro AURKA Kinase Assay by HTRF

The inhibitory activity of the candidate compounds against the AURKA kinase was evaluated using a Homogeneous Time-Resolved Fluorescence (HTRF) assay. This ratiometric approach was selected to ensure high sensitivity and to minimize potential interference from compound autofluorescence.

The AURKA kinase, S2 substrate, and detection reagents (Eu-labeled antiphospho-substrate antibody and XL665-labeled streptavidin) were utilized according to the manufacturer’s instructions. The assay buffer consisted of 50 mM HEPES (pH 7.5), supplemented with 5 mM MgCl₂ and 1 mM DTT. All candidate compounds and reference inhibitors (MLN8237 and CCT241736) were dissolved in 100% DMSO as 10 mM stock solutions.

The enzymatic reactions were performed in 384-well white polystyrene plates with a final reaction volume of 10 μL. A 10-point serial dilution of each compound was prepared, and 50 nL of each solution was precisely transferred into the assay plate using an Echo 655 acoustic liquid handler (Labcyte, USA), resulting in a final DMSO concentration of 0.5% (v/v). A 2.5 μL volume of the AURKA kinase (final working concentration: 2.00 nM) was added to the wells and preincubated with the compounds for 10 min at 25 °C. The reaction was initiated by adding 2.5 μL of a substrate/ATP mixture to achieve final concentrations of 1 μM S2 substrate and 4 μM ATP (near the K _m for AURKA). The plates were then incubated for 40 min at 25 °C.

The kinase reaction was quenched by the addition of 5 μL of HTRF detection reagent. The mixture was incubated for an additional 60 min at 25 °C to allow for the formation of the FRET complex. The time-resolved fluorescence signals were measured at 620 nm (Cryptate donor) and 665 nm (XL665 acceptor) using a multimode microplate reader. The HTRF ratio (665 nm/620 nm × 10⁴) was employed to calculate the percentage of inhibition as follows

$$\mathrm{I}\mathrm{n}\mathrm{h}\mathrm{i}\mathrm{b}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}(\%)=100-\left(\frac{S_{\mathrm{c}\mathrm{o}\mathrm{m}\mathrm{p}\mathrm{o}\mathrm{u}\mathrm{n}\mathrm{d}}-S_{\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}}}{S_{\mathrm{n}\mathrm{e}\mathrm{g}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}}-S_{\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}}}\right)\times 100$$

where S _compound is the HTRF signal value of the wells treated with the tested compounds. S _positive and S _negative represent the mean signal values of the positive control wells and the negative control wells, respectively.

Dose–response curves were generated by fitting the inhibition values to a nonlinear regression model (four-parameter logistic equation with variable slope) using GraphPad Prism 8.0

$$Y=\mathrm{B}\mathrm{o}\mathrm{t}\mathrm{t}\mathrm{o}\mathrm{m}+\frac{\mathrm{T}\mathrm{o}\mathrm{p}-\mathrm{B}\mathrm{o}\mathrm{t}\mathrm{t}\mathrm{o}\mathrm{m}}{1+{10}^{({\log \mathrm{I}\mathrm{C}}_{50}-X)\times \mathrm{H}\mathrm{i}\mathrm{l}\mathrm{l}\mathrm{S}\mathrm{l}\mathrm{o}\mathrm{p}\mathrm{e}}}$$

where Y is the calculated percentage of inhibition, and X is the logarithmic value of the tested compound concentration. The parameters Top and Bottom define the maximum and minimum plateaus of the fitted dose–response curve, respectively. The logIC₅₀ represents the logarithmic concentration of the compound that yields 50% of the maximal inhibitory effect, and HillSlope denotes the Hill coefficient, which describes the steepness of the curve. All assays were performed in duplicate, and IC₅₀ values are reported with 95% confidence intervals (CI) and goodness-of-fit metrics (R ²).

4.4.2. Cell Culture Conditions

JIMT-1 was maintained Dulbecco’s Modified Eagle’s medium (DMEM; gibco) supplemented with 10% fetal bovine serum (FBS, cellbox), 100 U/mL penicillin (gibco), and 100 μg/mL streptomycin (gibco). JIMT-1 was cultured at 37 °C under an atmosphere of 5% CO₂ in a humidified cell incubator. Similarly, NIH/3T3, MDA-MB-231 and MDA-MB-453 cells were maintained under the same conditions.

4.4.3. Cell Proliferation Assays

For human cancer cell lines, cells were seeded in 96-well plates at the optimized density for cell growth in the recommended growth medium provided by the vendor. For JIMT-1, MDA-MB-231, and MDA-MB-453 cells, 3500 cells per well were seeded and treated for 72 h with DMSO or test compounds. Test compounds were applied at a starting concentration of 20 μM, followed by 3-fold serial dilution to generate seven concentration points (20, 6.66667, 2.22222, 0.74074, 0.24691, 0.0823, and 0.02743 μM). All compounds were subjected to three parallel experiments on each cell line. After 72 h, cell proliferation was assessed using CCK8 according to the manufacturer’s instructions. The growth inhibitory IC₅₀ values for the compounds were determined using ActivityBase (IDBS).

For the exploratory 48 h assays in NIH/3T3 and MDA-MB-231 cells, cells were maintained in complete medium (DMEM supplemented with 10% fetal bovine serum and 1% penicillin/streptomycin). Cells were seeded into 96-well plates at 3000 cells per well in 100 μL complete medium and allowed to attach/grow for 24 h. Subsequently, 100 μL of compound-containing complete medium was added to each well, resulting in a final working volume of 200 μL per well before CCK-8 addition. Test compounds were prepared in 5-fold serial dilutions to yield eight final concentrations: 50,000, 10,000, 2,000, 400, 80, 16, 3.2, and 0.64 nM (starting final concentration: 50,000 nM). Here, “blank medium” refers to complete medium without test compound, used to keep well volume and solvent conditions consistent. At the highest compound concentration (50 μM), the final DMSO concentration was 0.1% (v/v), which is generally considered noncytotoxic under these conditions. After 48 h treatment, cell proliferation was measured using CCK-8 according to the manufacturer’s instructions: 20 μL of CCK-8 solution was added per well (final assay volume ∼220 μL per well), and absorbance at 450 nm was recorded after 30–45 min. The same reference compounds (MLN8237 and CCT241736) and the same CCK-8 analysis pipeline were used; results were conservatively interpreted as short-term phenotypic growth-inhibition outcomes.

4.5. Molecular Dynamics Simulation Protocol

The initial structure of the complexes of AURKA protein with compounds 3, 5 and 8 were obtained from the pose with the highest docking score generated by KarmaDock. A 500 ns classical molecular dynamics simulation was performed using the GROMACS 2021.7 software package to investigate the stability and interaction mechanisms of the complexes. The protein was parametrized with the AMBER14SB force field, , and the ligand was parametrized with the Generalized AMBER Force Field 2 (GAFF2). The rmsd, hydrogen bonding, and radius of gyration (R _g) were calculated to assess the system’s structural stability. The formula for calculating R _g was as follows

$$R\mathrm{g}=\sqrt{\frac{\sum _{i=1}^N{m}_i{(r_i-r_{\mathrm{c}\mathrm{o}\mathrm{m}})}^2}{\sum _{i=1}^N{m}_i}}$$

where N represents the number of atoms in the selected atomic group, m _i denotes the mass of the i-th atom, while r _i is its position vector in three-dimensional space. The term r _com refers to the center of mass (COM) of the atomic group, which is determined by taking the mass-weighted average of the position vectors of all constituent atoms. Mathematically, the center of mass is calculated as

$$r_{\mathrm{c}\mathrm{o}\mathrm{m}}=\frac{\sum _{i=1}^N{m}_i{r}_i}{\sum _{i=1}^N{m}_i}$$

The binding free energy of the complex was estimated using the molecular mechanics/Generalized–Boltzmann surface area (MM/GBSA) method using gmx_MMPBSA tool, using equilibrated snapshots from the final 10 ns of each trajectory (490–500 ns; 101 snapshots; 100 ps interval). A consistent continuum-solvent setting was applied across all replicas with dielectric constants of 1.0 and 78.5 and an ionic strength of 0.15 M. Per-residue decomposition was used only to support qualitative interaction analysis, and the absolute MM/PBSA values should be interpreted cautiously.

$${\mathrm{\Delta }G}_{\mathrm{b}\mathrm{i}\mathrm{n}\mathrm{d}\mathrm{i}\mathrm{n}\mathrm{g}}={\mathrm{\Delta }E}_{\mathrm{m}\mathrm{m}}+{\mathrm{\Delta }G}_{\mathrm{s}\mathrm{o}\mathrm{l}\mathrm{v}}-T\mathrm{\Delta }S$$

Detailed trajectory-processing settings and supporting analyses are provided in the Supporting Information.

Data will be made available at https://github.com/HMC-HuangLab/Discovery-of-AURKA-Inhibitors; EvaluationMaster could be found at https://github.com/shenzheyuan2020/EvaluationMaster; PLANET could be found at https://github.com/ComputArtCMCG/PLANET; KarmaDock could be found at https://github.com/schrojunzhang/KarmaDock; PSICHIC could be found at https://github.com/huankoh/PSICHIC; DeepDock could be found at https://github.com/OptiMaL-PSE-Lab/DeepDock. ADMETlab 3.0 be found at https://admetlab3.scbdd.com. KinomeMETA could be found at https://kinomemeta.alphama.com.cn.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsomega.6c00415.

• Additional experimental details for molecular dynamics simulation; ROC-AUC and P-value for docking program evaluation; certificate of analysis, LCMS report, Scaffold similarity and predicted kinase selection analysis for compounds 3, 5 and 8; MD trajectory and energy analysis for compounds 5 and 8 (PDF)

∥.

P. L. and Y. L. contributed equally to this work. Pei Liu: Investigation: Conducted the primary experiments and generated key data for the study. Writing: original draft: Contributed to writing the original draft of the manuscript, particularly the introduction and results sections. Yanqing Liu: Investigation: Conducted the primary experiments and generated key data for the study. Writing: original draft: Contributed to writing the original draft of the manuscript, particularly the experimental methods and results sections. Xin Zhang: Conceptualization: Designed the research plan and formulated the overall research strategy. Investigation: Contributed to the experimental work, assisting in data collection and experimental design. Runmei Chen: Investigation: Conducted the main experiments, including biological evaluation, and contributed to manuscript writing. Youlu Pan: Methodology: participated in experimental design and contributed to specific methodological development. Wenhai Huang and Xiangwei Xu: Writingreview and editing: Reviewed and edited the final manuscript, contributing to the refinement and polishing of the text. Supervision: Directed the overall research progress and ensured alignment with project objectives.

This work was supported by Zhejiang Medicine and Health Science and Technology Project (2023RC297) and National Health Commission Scientific Research Fund - Zhejiang Provincial health major science and technology plan project (WKJ-ZJ-2557) and the Basic Foundation of Hangzhou Medical College (KYYB202208, KYZD202204).

Statements: the MDA-MB-231 and MDA-MB-453 cell lines were all purchased from HealZen Therapeutics Co., Ltd. (Hangzhou, China), and JIMT-1 was purchased from Zhejiang Meisen Cell Technology Co., Ltd., while NIH/3T3 was purchased from Applied Biological Materials (abm) Inc. (Richmond, BC, Canada; Cat. No. T1065).

The authors declare no competing financial interest.