Identification of dual inhibitors of ROCK1 and TGFBR1 for overcoming metastasis: an integrated pharmacophore and docking driven CADD strategy

Abstract

Rho-associated coiled-coil kinase 1 (ROCK1) and transforming growth factor-beta receptor 1 (TGFBR1) plays crucial roles in epithelial-mesenchymal transition (EMT), cytoskeleton remodeling, cell adhesion, and tumor metastasis. Targeting both ROCK1 and TGFBR1 simultaneously offers a strategic approach to disrupt these pathways, potentially leading to more effective treatments for cancer metastasis. To identify dual inhibitors, a multifaceted computational approach integrating pharmacophore modeling, molecular docking, and molecular dynamics (MD) simulations was employed. Pharmacophore models were derived from the PDB structures of ROCK1 and TGFBR1 and validated based on their ability to discriminate between known active and inactive inhibitors from the ChEMBL database with performance evaluated using Receiver Operating Characteristic Area Under the Curve (ROC AUC) calculations. These validated pharmacophore models were used to virtually screen a diverse set of 1,48,648 compounds, identifying promising candidates for dual inhibition. Top-ranking compounds from the pharmacophore screening were subjected to molecular docking studies to evaluate their binding interactions with ROCK1 and TGFBR1. The most promising candidates, Hit-1, Hit-2, and Hit-3, were further analyzed through 200-ns molecular dynamics simulations to assess their stability and interaction dynamics with the hinge residues of both targets. Among the screened compounds, Hit-2, (4‐(1,3‐dimethyl‐1H‐pyrazol‐4‐yl)‐N‐[2‐(pyridin‐3‐yloxy)propyl]pyrimidin‐2‐amine) exhibited the most favourable binding affinity (− 155.801 kJ/mol) for ROCK1 and (− 213.264 kJ/mol) for TGFBR1, structural stability, and interaction profile, suggesting its potential as a dual inhibitor of ROCK1 and TGFBR1.

Supplementary Information

The online version contains supplementary material available at 10.1007/s40203-025-00414-5.

Introduction

Metastasis is a complex and multistep process that involves the dissemination of cancer cells from the primary tumor to distant organs, leading to the formation of secondary tumors (Fares et al. 2020). In the United States, approximately 90% of cancer deaths are attributable to metastasis (Dillekås et al. 2019). The survival rates for metastatic cancer are generally lower compared to localized cancer (De Silva and Alcorn 2022). Despite advancements in targeted therapies, immunotherapies, and personalized medicine, treating metastatic cancer remains a significant challenge in oncology.

Metastasis requires cancer cells to acquire specific traits that enable their migration and invasion through surrounding tissues. One critical process that facilitates this transformation is epithelial–mesenchymal transition (EMT). EMT enables cells to detach from the primary tumor, invade surrounding tissues, and ultimately enter the bloodstream or lymphatic vessels, facilitating their dissemination to distant sites (Banyard and Bielenberg 2015). EMT is involved in extracellular matrix reorganization. Activation of matrix metalloproteinases (MMPs), which degrade the extracellular matrix and facilitate tissue remodelling (Radisky and Radisky 2010).

The functional convergence of TGFBR1 and ROCK1 in driving cancer metastasis through EMT is relevant from their biological pathways. TGFBR1 (Transforming Growth Factor Beta Receptor 1), a transmembrane receptor activated by TGF-β ligands, initiates intracellular signaling through two primary pathways: the SMAD and non-SMAD pathways. In the SMAD pathway, TGFBR1 phosphorylates SMAD2 and SMAD3 proteins, which then form a complex with SMAD4 and translocate to the nucleus to regulate the transcription of genes involved in processes such as cell adhesion, migration, and proliferation (Kubiczkova et al. 2012; Rojas et al. 2010). The non-SMAD pathway involves a cascade of proteins, including ROCK1, and is primarily involved in regulating the expression of actin-myosin filament proteins, fibronectin, and kindling (Zhang 2010). These proteins play a crucial role in cell migration, adhesion, and cytoskeletal reorganization. The combined activation of the SMAD and non-SMAD pathways leads to a cellular phenotypic switch, characterized by the activation of EMT (Hao et al. 2019; Samarakoon and Higgins 2008). Thus, the inhibition of TGFBR1 and ROCK1, or Rho-associated coiled-coil kinase 1 has emerged as a promising therapeutic approach to counteract these pro-metastatic effects. Inhibitors of TGFBR1 have shown promise in preclinical studies, demonstrating their ability to suppress the pro-metastatic effects (Huang et al. 2022). Targeting ROCK1 and its associated pathways holds promise as a therapeutic strategy to impede the metastatic progression of cancer, offering potential avenues for the development of novel anti-metastatic interventions (Barcelo et al. 2023; Kim et al. 2021; Zhang et al. 2024).

While both ROCK1 and TGFBR1 have been studied individually in the context of metastasis, their dual inhibition remains an underexplored strategy in cancer therapeutics. As discussed above, ROCK1 and TGFBR1 signalling have emerged as crucial contributors for EMT in cancer cells, leading to metastasis process. Both the signalling pathways convergence towards cell cytoskeleton dynamics and EMT phenotypic switching in cancer cells, thus setting the rationale for dual inhibition of ROCK1 and TGFBR1 targets (Pascual et al. 2021). Dual inhibition of the targets could effectively impede EMT and hence metastasis, offering a promising strategy in the pursuit of effective cancer therapeutics. Figure 1 depicts the involvement of ROCK1 and TGFBR1 signaling pathway in metastasis and cell proliferation process.

Fig. 1.

Intersection of TGFBR1and ROCK1 signaling pathway. TGFBR1 is a transmembrane receptor that binds to TGF-β ligands, initiating a cascade of intracellular signaling events. The pathway can be broadly divided into the Smad pathway and the non-Smad pathway. In the Smad pathway, TGFBR1 activates SMAD2 and SMAD3 proteins through phosphorylation. These phosphorylated SMAD proteins form a complex with SMAD4 and translocate to the nucleus, where they bind to specific DNA sequences, regulating the expression of target genes involved in various biological processes. In the non-Smad pathway, TGFBR1 activates the small GTPase RhoA, which further activates ROCK1, a kinase involved in cytoskeletal reorganization. ROCK1 phosphorylates myosin light chain (MLC), leading to actin stress fiber formation and changes in cell shape and motility. The activation of both the SMAD and non-SMAD pathways leads to a variety of cellular responses, including changes in cell shape and motility, loss of cell junctions and EMT, and cell migration, growth, and proliferation, leading to metastatic cancer

However, current drug discovery efforts have largely focused on individual inhibitors targeting either ROCK1 or TGFBR1, with limited emphasis on identifying compounds capable of dual inhibition. This lack of dual-targeted therapeutic strategies represents a significant gap in anti-metastatic drug development, particularly given the converging roles of ROCK1 and TGFBR1 in EMT and cytoskeletal reorganization. To address this gap, we adopted an integrated in silico drug discovery strategy that combines pharmacophore-based and structure-based computer-aided drug design (CADD). The pharmacophore modeling enables the identification of key molecular features required for dual inhibition, while structure-based virtual screening ensures precise docking and binding affinity estimation against both ROCK1 and TGFBR1. This complementary approach increases the likelihood of identifying potent dual inhibitors with optimized interactions at both target sites.

In this study, pharmacophore models were built for ROCK1 and TGFBR1 inhibition, based on the PDB structures in complex with their inhibitors. These models were evaluated by their ability to discriminate the known active and inactive inhibitors from ChEMBL library. The best pharmacophore models were used for screening 148,684 compounds from various libraries. The common compounds retrieved by both ROCK1 and TGFBR1 pharmacophore models were then subjected to docking in both the targets followed by molecular dynamic simulations and binding energy calculations to identify compounds with dual inhibition potential (Fig. 2).

Fig. 2.

Integrated computer-aided drug design strategy employed in the study to identify dual inhibitors of ROCK1 and TGFBR1

Materials and methods

Retrieval of target proteins

The three-dimensional structures of ROCK1 and TGFBR1 were retrieved from the RCSB Protein Data Bank (PDB, https://www.rcsb.org) through keyword search for each protein. The retrieved PDB IDs with the co-crystal ligand information are shown in Supplementary Table S1.

Retrieval of ligands

The ChEMBL (https://www.ebi.ac.uk/chembl/) database is a repository of bioactivity data, offering an extensive collection of small molecule and their biological targets. It contains collective data from scientific literature and patents. Known ROCK1 and TGFBR1inhibitor along with their bioactivity data were collected from ChEMBL. They were classified into three groups: active (IC₅₀ < 1 µM), intermediate (IC₅₀ between 1 and 10 µM), and inactive (IC₅₀ > 10 µM) molecules. Compounds with IC₅₀ > 10 μM were considered inactive for modeling purposes, following conventional thresholds used in virtual screening. However, we acknowledge that such compounds may still possess weak or context-specific activity, and that this cutoff is an operational simplification. Table 1 provides the summary of known ROCK1 and TGFBR1 inhibitors retrieved from ChEMBL database. Supplementary excel tables S1 provide detailed list of the inhibitors along with their activity.

Table 1.

ROCK1 and TGFBR1 inhibitors from ChEMBL database used in this study

Target	Activea	Intermediate activeb	Inactivec	Total compounds
ROCK1	755 (66%)	221 (19%)	170 (15%)	1146
TGFBR1	437 (65.3%)	161 (24%)	71 (10.6%)	669

^aIC₅₀ < 1 μM, ^bIC₅₀ between 1 to 10 μM, ^cIC₅₀ > 10 μM

Generation and validation of pharmacophore models

The 3D structures of ROCK1 and TGFBR1 in complex with their co-crystal ligands (Supplementary Table S1) were utilized to generate both interaction-based and ligand-based pharmacophore (IBP and LBP) models using the Biovia Discovery Studio v2020 software. For IBP generation, the Interaction-Based Pharmacophore module was employed to create models based on complementary interactions between the protein and co-crystal ligand. Ten models models were generated from each co-crystal structure. These were then clustered using the compare pharmacophore module. To ensure chemical and spatial diversity, representative pharmacophore models were selected from distinct clusters generated using feature-based RMSD comparisons in the ‘Compare Pharmacophore Hypotheses’ module of Biovia Discovery Studio (https://www.3ds.com/products/biovia/discovery-studio). In parallel, the Ligand-Based Pharmacophore (LBP) module was used to generate models based on the co-crystal ligands of ROCK1 and TGFBR1, also derived from PDB structures (Supplementary Table S1). The LBP models were generated with a maximum count of ten and a feature distance of 0.5 Å. This module identifies essential ligand features such as hydrogen bond donors and acceptors, hydrophobic regions, and aromatic rings, allowing for alignment of common chemical motifs. The generated pharmacophores were further validated.

Validation of pharmacophore models

To validate the generated IBP and LBP models, the ability of each model to pick the known active and inactive ROCK1 and TGFBR1 inhibitors from ChEMBL library (Table 1 and supplementary excel tables S2) during screening was investigated. The 3D library of the ChEMBL database compounds was built using Build 3D Database module of Biovia Discovery Studio v2020 (Giordano et al. 2022). The pharmacophores were used to screen the 3D libraries of ROCK1 and TGFBR1 compounds. From the screening results, compounds are classified into four categories: True Positives (TP), False Positives (FP), True Negatives (TN), and False Negatives (FN) (Ntie-Kang et al. 2016; Saravanan et al. 2024). Based on the above criteria, to assess the performance of the pharmacophore models, several key metrics are calculated: True Positive Rate (TPR), False Positive Rate (FPR), Precision and Specificity (Karki et al. 2023; Mafethe et al. 2023; Moussa et al. 2021; Pandit et al. 2006). To visualize the trade-off between sensitivity and specificity, Receiver Operating Characteristic (ROC) curves are plotted. These curves illustrate the relationship between TPR and FPR at various threshold settings, with the Area Under the Curve (AUC) quantifying the overall performance of the model; an AUC of 1 indicates perfect model performance, while an AUC of 0.5 suggests performance no better than random chance (Tai et al. 2012; Temml et al. 2017).

All these metrics were computed to allow a comprehensive comparison, the final selection of pharmacophore models primarily emphasized using three criteria: ROC-AUC, to assess the overall discriminative ability of each model; specificity, to ensure accurate rejection of inactive compounds and thus minimize false positives; and precision, to prioritize models that efficiently identify true active compounds. Pharmacophore models performing better at least three of these key parameters were subsequently retained for pharmacophore based virtual screening.

Pharmacophore based virtual screening (PBVS)

Pharmacophore-based virtual screening (PBVS) involves aligning the pharmacophore features of the model with the compounds 3D features and prioritize the compounds that match the model features as hits (Giordano et al. 2022; Pal et al. 2019). A total of 148,684 compounds from Drug Bank (https://go.drugbank.com/), Kinase, Kinaset, KinaCore, and Chembridge (https://chembridge.com/targeted-and-specialty-libraries/kinase/) were used for pharmacophore screening. Each of these databases ligands were subjected to Build 3D Database module in Biovia Discovery Studio v2020 to generate 3D conformation library for pharmacophore screening. Search 3D Database module in Biovia Discovery Studio v2020 with the default parameters were used for the screening. Supplementary Table S2, provides comprehensive information about the library compounds used for virtual screening.

Molecular docking

Protein preparation

The 3D structure of four ROCK1 (PDB Ids: 6E9W, 3V8S, 4YVE and 5BML) and four TGFBR1 (PDB Ids: 5QIL, 1PY5, 1RW8 and 3KCF) proteins were used for the docking study. The target structure was prepared with the Protein Preparation Wizard in the Schrödinger suite 2020.3. Ions, cofactors and water molecules were deleted and hydrogen atoms were added (Madhavi Sastry et al. 2013). For PDB entries with unresolved side-chain atoms or loop regions, additional refinement was performed. The protein structural integrity was refined, and the missing side-chain atoms within the protein residues were predicted by what if server (https://swift.cmbi.umcn.nl/).

Ligand preparation

Ligand preparation was carried out using the LigPrep module from the Schrödinger Suite 2020-3. LigPrep was employed to generate low-energy, three-dimensional (3D) conformers of the input compounds (in 3D SDF format). The module optimized molecular geometry by adding hydrogen atoms, refining bond lengths and angles, and generating all possible tautomers, ionization states, and stereoisomers. Ionization states were assigned using the Epik tool at a physiological pH of 7.0 ± 2.0, ensuring accurate protonation (Johnston 2023).

For each ligand, LigPrep explored a maximum of 32 stereoisomers for molecules with undefined stereogenic centers. The tool also ensured proper assignment of chiral centers, ring conformations, and stereochemistries, selecting the lowest-energy conformers for each variation. All prepared ionization states, tautomers and stereoisomers were retained for use in subsequent molecular docking studies.

Receptor grid generation

Receptor grid for the docking was created at the active site of the protein structures. A cubic grid box was centered at the centroid of the active site guided by the already present co-crystal ligand. In glide, the grid was generated keeping the default parameters of van der Waals forces scaling factor 1.0 and charge cut-off 0.25. The inner grid that defines the space explored by the ligand centroid during docking was set to 10 × 10 × 10 Å³ by default. The outer grid, which encompasses the full ligand pose, was adjusted automatically to ensure all ligand atoms fit within the volume, typically around 30 Å, based on ligand size.

Glide XP (extra precision) docking

Extra precision (XP) was used as the scoring function in ligand docking and was carried out using Schrödinger Glide version XP5.0. The non-planar conformations were penalized and Van der Waals scaling factor and partial charge cut-off were set at 0.80 and 0.15, respectively, for the ligand atoms. Epik state penalties were incorporated to docking score. Final scoring was performed on energy-minimized conformations and displayed as glide score and docking score. The docked conformation with the lowest docking score was selected for each ligand (Friesner et al. 2006; Halgren et al. 2004).

Visualization and analysis of protein–ligand interactions

Protein–ligand 2D interaction map for the docked complex was generated using the 2D ligand–protein interaction option of Schrödinger suite. In addition, 3D structure of ROCK1 and TGFBR1 ligand complexes and their interactions were visualized with PyMOL v 3.1 tool.

Molecular dynamics simulations

The molecular dynamics simulations of the protein–ligand complexes were conducted over 200 ns using GROMACSv 2019 (Lindahl et al. 2001). The GROMOS force field was used to define the protein topologies, including atom types, partial charges, bond lengths, angles, and dihedrals. Ligand topologies were generated using the Automated Topology Builder (ATB) platform (https://atb.uq.edu.au/) (Malde et al. 2011). Each protein–ligand complex was centered in a cubic simulation box with a 1.0 nm buffer from the protein surface to the box edge to ensure sufficient solvation and support periodic boundary conditions (PBC). The system was solvated with the Simple Point Charge (SPC) water model and neutralized by adding appropriate Na⁺/Cl⁻ ions. Prior to production runs, energy minimization was carried out using 500 steps of the steepest descent algorithm, followed by a two-phase equilibration protocol. The first phase was conducted under NVT (constant Number of particles, Volume, and Temperature) conditions, and the second phase under NPT (constant Number of particles, Pressure, and Temperature) conditions. The production simulation was run under isothermal-isobaric (NPT) conditions at 300 K and 1 atm pressure, maintained using the Nosé–Hoover thermostat and Martyna–Tobias–Klein barostat with isotropic pressure coupling (Pronk et al. 2013). Simulations were performed using an integration time step of 2 femtoseconds (0.002 ps), amounting to 100 million steps (nsteps = 100,000,000) for the 200 ns trajectory. All simulation outputs were analyzed using built-in GROMACS analysis tools, and plots were generated in the Python environment using Matplotlib for visualization.

MM-PBSA binding energy calculation

The binding free energy (ΔG) of protein–ligand complexes was assessed using the g_mmpbsa tool, which applies the Molecular Mechanics Poisson–Boltzmann Surface Area (MM-PBSA) method (Genheden and Ryde 2015). This tool calculates the binding free energy by evaluating various components of the system’s energy profile. Specifically, it determines the potential energy contributions from solvation, including both electrostatic and van der Waals interactions, as well as the free energy associated with solvation, which encompasses both polar and non-polar solvation energies. These calculations were performed using the molecular dynamics simulation trajectories ranging from 100 to 200 ns with frames extracted every 100 ps intervals for every protein–ligand complexes and were analysed.

Chemical space and ADME analysis

The identified hits were analyzed based on fingerprints calculated using RDKit with MACCS keys. The chemical space was visualized by plotting t-SNE plot of these hits alongside known ROCK1-active and TGFBR1-active compounds based on PCA analysis of the calculated features to assess their structural similarity and potential clustering within the chemical space. Additionally, ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity) properties were evaluated to determine the drug-likeness and pharmacokinetic profiles of the identified hits using SwissADME (http://www.swissadme.ch/) server.

Results

Target proteins and ligands for ROCK1 and TGFBR1

For ROCK1, eleven PDB entries with a resolution ranging from 2.29 to 3.70 Å were retrieved. The co-crystal ligands associated with these structures include Fasudil (PDB ID: 2ESM), a known ROCK inhibitor approved for cerebral vasospasm following subarachnoid haemorrhage (Zhao et al. 2006). While, for TGFBR1, ten PDB structures with resolutions ranging from 1.45 to 3.70 Å were retrieved and the co-crystal ligands include potent TGFBR1 inhibitor SB431542 (PDB ID: 3TZM) and staurosporine (PDB ID: 5E8X) (Supplementary Table S1). From ChEMBL database, a total of 1146 ligands for ROCK1 inhibition with their IC₅₀ data were retrieved (Table 1).

Generation and validation of ROCK1 and TGFBR1 pharmacophore models

Two types of pharmacophore models were build using the PDB structures of ROCK1 and TGFBR1: interaction based pharmacophores (IBP) and ligand based pharmacophores (LBP). Each protein structure has generated a varying number of pharmacophores. One or more pharamcophore model for each PDB structure was selected based on the diversity of features after clustering them. The selected model has different combination of aromatic (arom), hydrophobic (H), donor (D), acceptor (A), ionisable groups (I) and ring (R) features.

In the IBP for ROCK1, a total of 90 pharmacophores are generated. The lowest number of pharmacophore models generated for a PDB ID is 2 (for PDB IDs 2ESM and 4YVC). The highest number of pharmacophore models is 10 (for PDB IDs 6E9W, 5UZJ, 3V8S, 3NDM, 4YVE, 5MBL, 4W7P, and 5WNG). The feature distribution pattern shows that, most common features are “AaromaromHH” appears frequently (in PDB IDs 7JOU, 5UZJ, 4YVE, 4W7P) and “AAAaromH” also has multiple occurrences (6E9W, 4YVE, 4YVE, 5WNG). However, certain features are less common or unique to specific PDBs (e.g., “ARaromR” in 4YVC and “AADaromH in 2ESM). In the LBP model for the ROCK1 inhibitors, a total of 55 pharmacophores are generated. The common features are “DaromaromHH” that appears multiple times across different PDBs (in IDs 6E9W, 7JOU, 4YVE, and 5BML). Other frequently seen features include variations of “AADaromaromH” and “DDaromaromHH”. Certain features are unique to specific PDBs, such as “ADDDaromarom” in 3V8S and “AADDaromaromaromH” in 4W7P.

In the interaction based pharmacophore for TGFBR1, a total of 64 pharmacophores are generated. The lowest number of pharmacophore models generated for a PDB ID is 4 (for PDB ID IPY5). The highest number of pharmacophore models is 10 (for PDB IDs 2X7O, 3KCF, and 3TZM). The feature distribution pattern shows that, most common features are “AaromaromHH” which appears multiple times (in PDB IDs IPY5 and 3KCF) and “AADaromH” and variations (e.g., “AADHH”, “AADaromHH”) are also recurrent (seen in PDB IDs 5QIL, 1RW8). While, some features are unique to specific PDBs, such as “AAADD” in 5QIM and “AAAADaromH” in 5E8Z. In the ligand based pharmacophore model for the TGFBR1 inhibitors, a total of 50 pharmacophores are generated. The most common feature maps are “DaromaromaromHH” (in PDB ID 5QIL and IPY5), “aromaromaromHHH” (in PDB ID 1RW8) and “DDaromaromHH” that appears several times, particularly in PDB IDs 5QIL and 2X7O. Additionally, unique entries like “ADD” and “AAD,” were only found in specific models (5QIM). The developed pharmacophore models are listed in Supplementary Tables S3–S6.

The IBP and LBP pharmacophore models for ROCK1 and TGFBR1 were validated prior to virtual screening by evaluating their ability to distinguish between known active and inactive compounds from the ChEMBL database. This discrimination performance was assessed for each model, and the results are summarized in the Supplementary excel tables S2. Receiver Operating Characteristic (ROC) curves were plotted from the true positive rate (TPR) and false positive rate (FPR) for each model based on their ability to discriminate between the active and inactive compounds from the ChEMBL database (Supplementary Fig. S1). Based on the AUC (Area Under the Curve) values, accuracy, sensitivity, precision, specificity and the number of active compounds retrieved, three ROCK1 models (6E9W.1, 3V8S.1, 4YVE.2) from IBP and three models from LBP (3V8S.4, 4YVC.4, 5MBL.1) were selected. Similarly, for the TGFBR1, three models (5QIL.1, IPY5.2, 3KCF.10) from IBP and three models from LBP (IPY5.5, 1R8W.2, 3KCF.1) were selected (Supplementary Tables S7 and S8).

The selected pharmacophore models of ROCK1 have 2 aromatic features, while TGFBR1 models mostly showcase 2–3, with some models possessing only aromatic characters. The ROCK models incorporated either or both donor and acceptor features, highlighting a robust hydrogen bonding capability for the inhibitors. A minimum of 0 to a maximum of 3 acceptor and donor features along with aromatic features were present in the ROCK1 models. Similarly, TGFBR1 models also carried hydrogen bonding acceptor and donor groups, ranging from 0 to 2. However, some of the TGFBR1 models were devoid of hydrogen bonding acceptor and donor groups (eg., 1R8W.2, 3KCF.1 and 3KCF.10). The ROCK1 models display donor and acceptor arrangements along with aromatic/hydrophobic interaction, whereas TGFBR1 models lean heavily on aromatic/hydrophobic interactions, suggesting differing interaction dynamics with their respective ligands. The selected 12 pharmacophore models of ROCK1 and TGFRB1 were mapped with their co-crystal ligands and shown in the Supplementary Fig. S2.

Pharmacophore-based virtual screening

The selected 12 pharmacophore models of ROCK1 and TGFBR1 were employed to conduct virtual screening against diverse chemical libraries, including Drug Bank (11,172 compounds), Kinase (36,324 compounds), Kinaset (12,249 compounds), KinaCore (10,614 compounds), and Chembridge (78,325 compounds). A total of 148,684 unique compounds were subjected for the pharmacophore-based virtual screening (PBVS). The number of hits identified by each model from each library is summarised in Table 2.

Table 2.

Summary of pharmacophore-based virtual screening results

Library name	No. compounds	No. compounds retrieved												Common compounds in ROCK1 and TGFBR1 models
		ROCK1 models						TGFBR1 models
		Ligand-based			Interaction-based			Ligand-based			Interaction-based
		3V8S,4	4YVC,4	5BML,1	6E9W,1	3V8S,1	4YVE,2	1RW8,2	1PY5,5	3KCF,1	5QIL,1	1PY5,2	3KCF,10
Drug Bank	11,172	433	313	314	172	200	345	940	1182	248	534	138	713	713
Kinase3D	36,324	756	3098	161	2137	785	1509	407	1167	1329	6303	983	11,393	1801
Kinaset3D	12,249	247	560	59	478	500	796	393	729	3027	1811	244	3106	176
KinaCore3D	10,614	507	143	338	147	174	9	1865	157	1795	2076	543	125	101
Chembridge 3D	78,325	13,850	406	224	569	1789	1076	1287	5631	503	2401	127	6698	1291
Total Compounds	148,684													4082

As the objective of the work is to identify compounds that exhibit dual interactions or have the potential to modulate both ROCK1 and TGFBR1 pathways, compounds retrieved by at least one of the ROCK1 and TGFBR1 models, i.e., common compounds, were considered for further investigation. Collectively, the models retrieved 4082 compounds from the various databases, which represent approximately 2.74% of the total compounds available across all libraries as dual ROCK1 and TGFBR1 hits. From the Drug Bank 172 (1.54%), Kinase3D 1801 (4.95%), Kinaset3D 176 (1.44%), KinaCore3D 125 (1.18%) and from Chembridge 3D 1291 (1.65%) compounds were identified as hits for both the targets.

Docking-based virtual screening (DBVS)

As a next step, the 4082 hits from PBVS was subjected to docking-based virtual screening (DBVS) using Glide XP module to identify the compounds that could interact with both ROCK1 and TGFR1 targets. For this purpose, 4 ROCK1 PDB structures (3V8S, 4YVE, 5BML and 6E9W) and 4 TGFBR1 PDB structures (5QIL, 1PY5, 3KCF and 1R8W) from the 12 selected pharmacophore models were used. The Glide XP docking protocol was first validated by docking the co-crystal ligands present in the ROCK1 and TGFBR1 PDB structures. Then, the docked pose and co-crystal conformation of the ligands were superimposed and the RMSD between them calculated for each ligand. The results showed that for both the ROCK1 and TGFBR1, co-crystal inhibitors redocked with an RMSD < 2 Å. The redocking experiment validates that the Glide XP docking protocol can recapture the co-crystal ligand conformation in all the eight instances. Moreover, the ROCK inhibitor present in 4YVE showed the lowest docking score of − 10.6 kcal/mol and was selected as the cut-off score for choosing ROCK1 hits from DBVS. For the TGFBR1, cut-off docking score of -8.1 kcal/mol was chosen for selecting the hits from DBVS and is based on the docking score of the ligand present in the 1RW8 TGFBR1 structure. Supplementary Figs. S3 and S4, shows the superimposed docked pose of the ligands onto the co-crystal conformation of the ligands present in ROCK1 and TGFBR1 PDB structures.

The common ROCK1 and TGFBR1 hits from PBVS, i.e., 4082 compounds, were subjected to DBVS using the Glide XP protocol. The resulting compounds with a docking score above the cut-off of − 10.6 (for ROCK) and − 8.1 (TGFBR1) kcal/mol, were retrieved for each PDB structures. The compounds commonly picked in at least two of the PDB structures of ROCK1 and TGFBR1 each were retrieved. A total of 18 compounds were identified, among them, 7 were from DrugBank, 3 each from Kinase 3D, Kinaset 3D and Kinacore 3D, and 2 from Chembridge library (Table 3 and Supplementary Tables S9 and S10).

Table 3.

Summary of docking-based virtual screening results

Library name	No. compounds	No. compounds retrieved								Common compounds retrieved in at-least two ROCK1 and two TGFBR1 docking
		ROCK1 − 10.6 kcal/mol (cut-off docking score)				TGFBR1 − 8.1 kcal/mol (cut-off docking score)
		6E9W	3V8S	4YVE	5BML	5QIL	1PY5	3KCF	1RW8
Drug Bank	713	13	15	17	22	27	10	18	22	7
Kinase3D	1801	16	12	13	10	22	13	10	7	3
Kinaset3D	176	12	18	17	13	7	5	11	17	3
KinaCore3D	101	9	6	12	8	10	6	15	12	3
Chembridge 3D	1291	15	2	18	10	11	3	7	14	2
Total compounds	4082									18

For each of the 18 compounds, the H-bond interacting residues were meticulously documented, allowing for a comprehensive understanding of their binding to ROCK1 and TGFBR1 (Supplementary Tables S9 and S10). Among these compounds, particular attention was given to those exhibiting H-bond interactions with the hinge regions of both ROCK1 (Met156) and TGFBR1 (His283). The hinge region interactions are critical for the kinase inhibitors. In the case of ROCK1 and TGFBR1, Met156 and His283, respectively are the hinge region residues and H-bond was observed with the hinge residues in the co-crystalized ligands of both ROCK1 (6E9W, 5BML, 3V8S, 4YVE) and TGFBR1 (3KCF, 5QIL, 1PY5, 5QIL) structures (Supplementary Figs. S5, S6).

From the group of 18 compounds, six (DB02794, 12,328,542/Hit-1, 15,174,553, 24,428,386/Hit-2, 88,483,473/Hit-3 and 88,809,706) were identified that demonstrated hinge region H-bond interactions in both the targets. Further evaluation focused on the docking poses of these six compounds. The goal was to assess the consistency of their binding across the various PDB structures. This examination revealed that only three hits 1–3 displayed consistent docking poses and interactions in ROCK and TGFBR1 structures with the hinge residues (Fig. 3 and Supplementary Figs. S7–S9). Consistent docking poses are indicative of stable interactions and suggest a higher likelihood of biological activity. Table 4 lists the top three putative dual ROCK1 and TGFBR1 binders identified through virtual screening, based on favorable docking scores, and key binding interactions. While docking provided an initial assessment of ligand orientation and potential binding interactions, it is further complemented with MD simulations and MM-PBSA energy calculations to better approximate realistic protein–ligand behavior.

Fig. 3.

Docking pose of hits 1–3 in a ROCK1 and b TGFBR1. Top panel (Hit-1), middle panel (Hit-2) and bottom panel (Hit-3). H-bond with the hinge residues is represented as grey dotted lines. Docking was carried out using Glide XP and the image generated using PyMOL v 3.0

Table 4.

Three virtual screening hits identified as potential dual ROCK1 and TGFBR1 inhibitors

Compound ID	Average docking score in ROCK1 (kcal/mol)	Average docking score in TGFBR1 (kcal/mol)	Interacting residues in ROCK1 (H-bond)	Interacting residues in TGFBR1 (H-bond)
12,328,542 (Hit-1)	− 10.27	− 10.82	MET156	HIS283, ASP290, LYS337
24,428,386 (Hit-2)	− 10.06	− 9.69	MET156	HIS283
88,483,473 (Hit-3)	− 9.37	− 9.56	MET156, ASP160	HIS283
Fasudil (ROCK1 Std)	− 11.30	–	PHE87, MET156	–
J2V (TGFBR1 Std)	–	− 9.80	–	HIS283, LYS232

MD Simulation of virtual screening hits in ROCK1 and TGFBR1 proteins

Molecular dynamics simulation (MDS) was employed to investigate the ROCK1 and TGFBR1 proteins and to evaluate the stability of their complexes with the top three virtual screening hits (Hit-1, Hit-2, and Hit-3) over a 200-ns simulation period (Supplementary Mmaterial 3). Key parameters such as Root Mean Square Deviation (RMSD) of the backbone, Root Mean Square Fluctuations (RMSF) of the side chains, Radius of gyration (Rg) of the protein were analysed. Hydrogen bond interactions of the hits with the protein, specifically with the hinge region residues of the proteins were analysed.

Both the protein maintains a RMSD of less than 1 nm throughout the period of 200 ns with all the three hits (Fig. 4). The overall stability of the protein backbone RMSD profiles suggests that the proteins are well-folded and maintain their structural integrity. The RMSD graphs shows that, Hit-2 in ROCK1 and TGFBR1 make the most stable interaction through out the simulation (RMSD < 0.5 nm), compared to Hit-1 and Hit-2. However, in case of TGFBR1, stabilization was observed post 120 ns and may indicate structural rearrangements within loop regions that result in a more ordered active site conformation. However, the protein-only RMSD cannot provide information about ligand dynamics, which are addressed separately in the ligand RMSD plots.

Fig. 4.

Molecular dynamics simulation results of virtual screening Hits 1–3 for 200 ns. a ROCK1-Hit and b TGFBR1-Hit complex. Fasudil is the standard ROCK1 inhibitor and J2V is the standard TGFBR1 inhibitor used for comparison. The top panel represents the RMSD of Cα atoms of the protein backbone and the middle panel represent the ligand RMSD during the simulation time. Bottom panel shows the frequency of H-bond with the Met156 (ROCK1)/His283 (TGFBR1) with three hits identified over the simulation period. Simulation was carried out using GROMACS and the figures generated using Matplotlib

In ligand RMSD in ROCK1, Hit-1 and Hit-2 demonstrate better stability with RMSD values consistently around 0.05–0.1 nm, indicating minimal deviation from the original comformation. While Hit-3 shows higher fluctuations early on but stabilizes around 0.15 nm. Fasudil shows a slightly higher RMSD (~ 0.15 nm). In TGFBR1, Hit-1 shows the lowest RMSD (0.1 nm), indicating stable conformation. Hit-2 show fluctuations but remain relatively stable (~ 0.1–0.15 nm). Hit-3, however, shows significant initial fluctuations and higher RMSD values (~ 0.25–0.3 nm), suggesting weaker or less consistent conformation. J2V, in TGFBR1 shows fluctuations of 0.15 nm (Fig. 4). Further, the RMSF and gyration profile shows overall stability and compactness of the protein-Hit-2 complex as compared to the standard inhibitors in both ROCK1 and TGFBR1 (Supplementary Fig. S10).

Further, to analyse the stability of protein–ligand interaction, protein–ligand complex structures were extracted from the MD trajectory at the intervals of 100 and 200 ns and overlaid with that of the 0 ns (docking conformation) and RMSD was calculated and it were all observed to be below < 2 Å (Fig. 5). This shows stable ligand orientation within the active site of ROCK1 and TGFBR1 for all the three hits, similar to the standard inhibitors. Moreover, mapping the protein–ligand interactions showed that all the three hits maintained hinge region H-bond interaction with ROCK1 (Met156) and with TGFBR1 (His283) during the simulation period in a manner similar to the standard ROCK1 and TGFBR1 inhibitors (Table 5). In addition to the H-bond interactions, hydrophobic, pi-cation and other interactions were also observed between the protein and Hits.

Fig. 5.

Superimposition of protein–ligand complexes extracted from 200 ns MD simulation trajectories at the interval of 0, 100 and 200 ns for the hits 1–3. a ROCK1-Hit and b TGFBR1-Hit complexes. The zoomed in figures show the respective ligand orientations and the RMSD values corresponds to comparison between ligand poses at 0 and 200 ns

Table 5.

Protein–ligand interacting residues and free energy of binding for the three hits calculated from the 200 ns MDS. 0, 100 and 200 ns frames were extracted from the trajectory and compared

Hit	Target protein	Interacting residues@			Total binding free energy (ΔG, kJ/mol)
		0 ns	100 ns	200 ns
Hit-1	ROCK1	MET156, ILE82, VAL90, LEU205, ASP202, PHE368	MET156, ILE82, PHE368	ILE 82, VAL 90, LEU205, ASP202, PHE368, MET156, ASP202, ASP216	− 160.69 ± 16.09
	TGFBR1	LYS232, SER287, ASP351, HIS283, LYS232, ILE211, VAL219, ALA230, LYS232, LEU260, LEU340	ILE211, VAL219, ALA230, LYS232, SER287, ASP351, HIS283	ILE211, LYS232, LEU340, TYR282, SER287, ASP351, HIS283, LYS232	− 186.79 ± 16.69
Hit-2	ROCK1	PHE87, MET156, ILE82, VAL137, LEU205, ALA215, ASP216, PHE368, LYS105	VAL137, ALA215, PHE87, MET156, LYS105	LEU205, PHE368, PHE 87, MET156, ASP216, LYS105	− 155.80 ± 22.55
	TGFBR1	HIS283, ILE211, LEU430	HIS283, LEU430	ARG294, ILE211, VAL219, TYR282, HIS283	− 213.26 ± 12.37
Hit-3	ROCK1	ARG84, MET156, PHE368, VAL90, ALA103, LEU205	ARG84, MET156, VAL90, ALA103, LEU205, ASP216	ARG84, LYS105, ILE82, VAL90, LEU205, ASP216, PHE368, MET156	− 191.24 ± 22.14
	TGFBR	HIS283, LYS232, ILE211, VAL219, ALA230, LEU340, ASP351	LYS 232, ARG 294, GLY 286, ASP 290, ASP 351, HIS 283	LYS 232, ARG294, GLY286, GLU228, LEU340, LYS232, SER280, HIS283	− 178.16 ± 13.50
Fasudil	ROCK1	ARG84, MET156, PHE368, VAL90, ALA103, LEU205, LYS105	MET156, LYS105, VAL137, ALA215, PHE87	MET156, PHE368, LEU205, LYS105	− 177.85 ± 24.87
J2V	TGFBR1	HIS283, LYS232, ARG294, GLY286, GLU228, LEU340, LYS232, SER280	ARG294, GLY286, GLU228, LEU340, LYS232, SER280	HIS283 ARG294, GLY286, GLU228, LEU340	− 181.03 ± 28.70

^@Character code: Bold, Hydrogen Bond; Italics, Hydrophobic; Underlined, Pi-cation and Bolditalics, pi-pi stacking interactions

Next, the total binding free energies of the protein–ligand complexes were calculated from the MDS trajectories. The ROCK1 standard ligand, Fasudil exhibited ΔG = − 178 kJ/mol, while the TGFBR1 standard ligand J2V exhibited ΔG = − 181 kJ/mol binding energy (Table 5, Supplementary Table and Fig. S11). Compared to the Fasudil, Hit-3 (ΔG = − 191 kJ/mol) had better binding to ROCK1. While, compared to J2V, Hit-1 (ΔG = − 187 kJ/mol) and Hit-2 (ΔG = − 213 kJ/mol) had better binding to TGFBR1. The major contributing factor for the binding affinity of the ligand is the Van der Waals interaction energy, followed by the SASA energy and then the electrostatic interaction energy between the protein and the ligands.

Chemical space and ADME analysis of the hits

The selected hit compounds (Hit-1, Hit-2, and Hit-3) exhibit favorable drug-like properties and occupy a relevant region of the chemical space associated with ROCK1 and TGFBR1 active molecules. The t-SNE plot illustrates that the hits are located at the intersection of the clusters formed by ROCK1 and TGFBR1 actives, indicating their structural and physicochemical similarity to both target classes. This positioning supports their potential as dual inhibitors capable of modulating both kinases. The radar plot comparison of key molecular descriptors further confirms their drug-likeness. All three hits demonstrate balanced properties, including acceptable molecular weight, limited Lipinski’s rule violations, favorable topological polar surface area (TPSA), and reasonable synthetic accessibility (Fig. 6). These characteristics are indicative of good oral bioavailability and ease of synthesis. Overall, the hits not only show promising spatial distribution within the dual-target chemical landscape but also possess physicochemical profiles that support their advancement for further biological evaluation.

Fig. 6.

Three hits identified from the virtual screening, showing the relative position within the chemical space of known ROCK1 and TBFBR1 inhibitors from ChEMBL database. The t-SNE (t-distributed stochastic neighbour embedding) plot was generated based on the PCA analysis of the MACCS fingerprints of the active ROCK1 and TGFBR1 inhibitors. The plot was made using Matplotlib and Seaborn Python package. The ADMET properties of the hits were predicted using Swiss-ADME server

Discussion

The primary goal of this study was to explore the feasibility of discovering dual inhibitors that can simultaneously target ROCK1 and TGFBR1, two proteins intricately involved in driving epithelial–mesenchymal transition and metastasis. By employing an integrated ligand and structure-based virtual screening approach, the study aimed to overcome limitations of single-target therapies and identify potential dual inhibitors of ROCK1 and TGFBR1.

To achieve this, a large number of pharmacophore models were generated based on both ligand and protein interaction data for ROCK1 and TGFBR1. These models incorporated multiple molecular features such as aromatic rings, hydrophobic groups, hydrogen bond donors and acceptors, ionizable groups, and ring structures, allowing for a comprehensive representation of essential interactions. In general, the selected ROCK1 models showed more number of hydrogen bond acceptor features and TGFBR1 showed more aromatic features, suggesting different pharmacophoric features essential for interaction with their targets. Model validation through ROC curve and parameters like precision and specificity analysis helped to prioritize pharmacophore models with the highest discriminatory power, ensuring that only the most predictive models were used for virtual screening. Out of 148,684 compounds screened, by the selected models, 4082 molecules were identified to match pharmacophoric requirements for both ROCK1 and TGFBR1. These compounds were subsequently docked into four PDB structures of each target to account for structural flexibility and enhance the reliability of binding predictions. This multi-conformational docking led to the identification of three promising hits (Hit-1, Hit-2, and Hit-3) that consistently formed key hydrogen bonds with the hinge region residues Met156 in ROCK1 and His283 in TGFBR1. These interactions are critical for kinase inhibition, as supported by literature and observed in co-crystal ligand structures. The hinge residue is a key amino acid for the kinase and interactions with this residue often dictate the potency of the inhibitors (Xing et al. 2015; Zhao and Bourne 2023). ROCK1 inhibitors are known to make H-bond with the hinge region Met156 and TGFBR1 inhibitors with His283 residue (Beljkas et al. 2024; Leonardo-Sousa et al. 2025). Hit-1’s isoquinoline ring made H-bond with the hinge region residue in both the targets. In the case of Hit-2, pyridine ring made H-bond with Met156 (ROCK1) and the pyrimidine ring made H-bond with His283 (TGFBR1). While, in Hit-3, pyridine ring made H-bond with Met156 (ROCK1) and the imidazole ring made H-bond with His283 (TGFBR1). This suggest that the three hits are chemically and target interaction wise distinct. During the 200 ns MD simulations of protein–ligand complexes, the hinge-region H-bonds were stable for all the three hits, similar to the known ROCK1 (Fasudil) and TGFBR1 (J2V) inhibitors.

Among the identified hits, Hit-2 emerged as the most promising candidate. Apart from its ability to form stable hinge-region interactions in both targets across all docking and MD simulations, during 200 ns molecular dynamics (MD) simulations, these hydrogen bonds remained stable, comparable to interactions observed with reference inhibitors Fasudil (ROCK1) and J2V (TGFBR1). Importantly, while these reference inhibitors are potent against their respective targets, they did not exhibit dual-target binding in our study, highlighting Hit-2’s distinct dual-inhibition potential. Further, MM/GBSA-based binding free energy calculations supported Hit-2’s strong and stable binding affinities to both ROCK1 and TGFBR1. The molecular interactions observed suggest that Hit-2 is not only chemically distinct from other inhibitors but also positioned at the interface of active ROCK1 and TGFBR1 inhibitors in the tSNE chemical space analysis plot, suggesting it has chemical features for dual inhibition. Its favorable drug-likeness, synthetic accessibility, and predicted ADMET profile further strengthen its potential as a viable lead for drug development. Moreover, none of the ChEMBL ROCK1 and TGFBR1 bioactive compounds used in the study had dual inhibition potential based on the IC₅₀ data, suggesting that Hit-2 could be a potential dual inhibitor of ROCK1 and TGFBR1. Overall, the findings from this study highlight Hit-2’s potential as a dual ROCK1 and TGFBR1 inhibitor using in silico approaches, warranting further in vitro investigation into its application in treating metastatic cancer.

In conclusion, given the critical roles of ROCK1 and TGFBR1 signaling in cancer metastasis, the dual inhibition of these targets represents a promising strategy to enhance therapeutic efficacy while minimizing the side effects and pharmacokinetic challenges associated with monotherapy. By modulating both targets, dual inhibitors may provide a more comprehensive blockade of metastatic pathways, potentially leading to improved outcomes in treatment and metastasis control. An integrated traditional computer-aided drug design (CADD) methodologies such as pharmacophore modelling and molecular docking effectively helped to identify molecules that can modulate both the targets. Furthermore, dual inhibition can significantly reduce the risk of drug resistance, a common challenge in monotherapy where cancer cells adapt to the selective pressure exerted by a single agent. By targeting multiple targets in different pathways simultaneously, the likelihood of resistance is diminished, allowing for sustained treatment effectiveness over time (Ye et al. 2023; Zhao et al. 2017). Additionally, this approach can improve safety profiles; utilizing lower doses of each agent made possible through their combined action can help minimize toxicity, which is particularly beneficial for patients sensitive to the side effects of higher drug concentrations (Zhong et al. 2021). When two different drugs are administered, their absorption, distribution, metabolism, and excretion (ADME) profiles may vary considerably, leading to potential inconsistencies in therapeutic efficacy and side effects. These differences can complicate treatment regimens, making it difficult to achieve optimal drug concentrations within the target tissues (Tseng et al. 2024). In contrast, a dual inhibitor, being a single entity, provides a more predictable pharmacokinetic profile. This uniformity not only enhances the therapeutic profile of the treatment but also improves patient compliance, as patients are often more willing to adhere to simpler regimens that involve fewer medications.

While this study presents a promising computational identification of potential dual inhibitors targeting ROCK1 and TGFBR1, several limitations must be acknowledged. First, the findings are derived entirely from in silico approaches, including virtual screening, molecular docking, and molecular dynamics simulations. Although these techniques are powerful and widely used in early drug discovery, they cannot fully replicate the complexity of biological systems. Therefore, experimental validation particularly in vitro enzyme inhibition assays and cell-based functional studies is essential to confirm the inhibitory activity, selectivity, and therapeutic potential of the identified hits, especially Hit-2. Second, despite the large size of the compound library screened, it may not capture the entire chemical space, potentially overlooking novel chemotypes with dual-target activity. Future work could benefit from expanding the chemical diversity through the inclusion of de novo-designed compounds or larger, more diverse libraries. Third, the predicted ADMET properties and synthetic accessibility scores, while favorable, are only indicative and require experimental verification to ensure drug-likeness, safety, and feasibility for clinical development. Lastly, While docking provided an initial assessment of ligand orientation and potential binding interactions, we recognize its limitations and therefore complemented it with MD simulations and MM-PBSA energy calculations to better approximate realistic protein–ligand behaviour. Moreover, 200 ns molecular dynamics simulations provided valuable insights into the stability of the protein–ligand complexes, longer simulation times or multiple replicates may be needed to capture slower conformational transitions and ensure the convergence and reproducibility of binding free energy calculations. Addressing these limitations in future studies will be critical for advancing the computationally identified candidates, particularly Hit-2, toward experimental validation and potential development as dual-targeted therapeutic agents for metastatic cancer.

Despite these limitations, the study presents a compelling case for the rational design of dual-target inhibitors in the treatment of metastatic cancer. By identifying Hit-2, we provide a strong foundation for further experimental research and optimization. With continued investigation, dual inhibitors such as Hit-2 may offer a more effective and durable therapeutic strategy in oncology, especially in cases where complex signaling crosstalk contributes to disease progression and treatment failure.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

BK: study conception and design, data collection, analysis, result interpretation and draft manuscript preparation MSC: Study conception and design, result interpretation, editing and finalization of manuscript and project supervision. All authors reviewed the results and approved the final version of the manuscript.

Data availability

Data is provided within the manuscript or supplementary information files. All authors reviewed the results and approved the final version of the manuscript.

Declarations

Conflict of interest

The authors declare that there is no conflict of interest.