Integrated structure-based and systems-level identification of PDE9A inhibitors for heart failure with preserved ejection fraction

Abstract

Background

Heart failure with preserved ejection fraction (HFpEF) is a complex and increasingly prevalent cardiovascular disorder with limited effective therapeutic options. Dysregulation of cyclic guanosine monophosphate (cGMP) signaling has been implicated in its pathophysiology. Phosphodiesterase 9A (PDE9A), a cGMP-specific enzyme, has emerged as a potential therapeutic target due to its role in nitric oxide-independent signaling pathways associated with myocardial dysfunction.

Methods

This study employed an integrated computational framework combining multi-conformational docking, pharmacophore modeling, pharmacophore validation using active–decoy screening, molecular dynamics simulations, and systems-level analyses to identify potential PDE9A inhibitors. Three crystal structures of PDE9A (4GH6, 4Y87, and 6LZZ) were analyzed for physicochemical properties, structural validation, and binding site prediction. Pharmacophore-guided virtual screening of PubChem-derived compounds was followed by active–decoy validation, ADMET evaluation, molecular docking, redocking validation, protein–ligand interaction analysis, and 500 ns molecular dynamics simulations.

Results

Several compounds demonstrated favorable binding affinities and pharmacokinetic profiles. Among them, EVPMO-PurHD exhibited the strongest binding affinity (− 10.1 kcal/mol) and suggested stable interaction behavior during the equilibrated phase of the simulation, following an initial conformational adjustment period, as indicated by RMSD stabilization and sustained intermolecular interaction.

Conclusion

The findings indicate that the identified compounds represent promising computationally predicted candidates for PDE9A inhibition, supported by favorable binding affinity, stability, and pharmacokinetic profiles. However, further experimental validation through biochemical and in vitro/in vivo studies is required to confirm their therapeutic potential.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12896-026-01190-6.

Introduction

Heart failure with preserved ejection fraction (HFpEF) has emerged as a major and increasingly prevalent cardiovascular syndrome, accounting for approximately half of all heart failure cases worldwide [1]. It is clinically characterized by typical heart failure symptoms despite a preserved left ventricular ejection fraction (≥ 50%), primarily driven by impaired ventricular relaxation and elevated filling pressures [2]. The global burden of HFpEF continues to escalate due to population aging and the rising prevalence of comorbidities, including hypertension, obesity, and diabetes mellitus [3]. Despite its substantial contribution to morbidity, mortality, and healthcare utilization, effective disease-modifying therapies remain limited, underscoring a critical unmet clinical need [4].

The pathophysiology of HFpEF is multifactorial, involving myocardial fibrosis, systemic inflammation, endothelial dysfunction, and coronary microvascular impairment [5]. Excessive extracellular matrix deposition contributes to ventricular stiffening and diastolic dysfunction, while reduced nitric oxide bioavailability further compromises vascular function and myocardial perfusion [6, 7]. Among these mechanisms, dysregulation of cyclic guanosine monophosphate (cGMP) signaling has been identified as a central pathological driver [8]. Impaired activation of the cGMP–protein kinase G (PKG) pathway promotes increased myocardial stiffness and altered vascular tone, thereby exacerbating disease progression [9].

Therapeutic strategies targeting cGMP signaling have primarily focused on phosphodiesterase type 5 (PDE5) inhibition; however, clinical outcomes in HFpEF have been inconsistent and largely disappointing [10]. In contrast, phosphodiesterase 9A (PDE9A), a cGMP-specific enzyme that selectively hydrolyzes natriuretic peptide–derived cGMP, has recently emerged as a promising alternative target [11, 12]. Unlike PDE5, PDE9A regulates nitric oxide–independent signaling pathways, suggesting a complementary and potentially more effective mechanism for restoring myocardial function. Preclinical studies have demonstrated that PDE9A inhibition can attenuate pathological remodeling, improve diastolic function, and restore microvascular integrity in HFpEF models [13, 14]. These findings highlight PDE9A as a compelling target for therapeutic intervention.

Advances in computational drug discovery have enabled the efficient identification of candidate molecules through approaches such as molecular docking, pharmacophore modeling, and virtual screening [15]. However, many existing studies rely on single protein conformations and static docking protocols, which fail to account for protein flexibility and the dynamic nature of protein–ligand interactions, thereby limiting predictive reliability [16, 17]. In addition, the functional context of PDE9A within broader biological networks, including protein–protein interactions and gene coexpression patterns, remains insufficiently explored in current computational frameworks [18]. Integrating structural, dynamic, and systems-level analyses is therefore essential for improving target characterization and enhancing the translational relevance of in silico findings [19, 20].

By integrating structural, dynamic, and network-based methodologies, this work aims to improve the robustness of computational predictions and identify candidate PDE9A inhibitors with favorable binding affinity, stability, and pharmacokinetic properties. The proposed framework provides a more comprehensive strategy for early-stage drug discovery and offers novel insights into the therapeutic targeting of PDE9A in HFpEF. Given the complexity and cost associated with experimental screening, in silico approaches provide an efficient strategy for the early-stage identification and prioritization of potential drug candidates. However, computational predictions require subsequent experimental validation to confirm biological activity and clinical relevance. This study presents a comprehensive computational framework integrating multi-conformational docking, pharmacophore-guided screening, molecular dynamics simulations, and systems-level analyses for PDE9A inhibitor identification.

Materials and methods

Retrieval and preparation of PDE9A crystal structures

Three human phosphodiesterase 9A (PDE9A) catalytic domain crystal structures (PDB IDs: 4GH6, 4Y87, and 6LZZ) were retrieved from the Protein Data Bank (PDB). The PDB provides experimentally resolved three-dimensional structures of biomacromolecules [21, 22]. Structure 4GH6, co-crystallized with a selective inhibitor, contains key active-site residues such as Tyr424 associated with ligand selectivity [23], whereas 4Y87 and 6LZZ represent alternative ligand-bound conformations, enabling consideration of structural flexibility [24, 25].

Protein structures were prepared using AutoDock Tools (version 1.5.7) by removing co-crystallized ligands, water molecules (except those involved in active-site interactions), and heteroatoms. Polar hydrogens were added, Kollman charges were assigned, and structures were energy-minimized to remove steric clashes while preserving structural integrity.

Physicochemical characterization

Physicochemical properties of PDE9A structures were evaluated using the ExPASy ProtParam tool [26, 27]. Parameters including molecular weight, theoretical isoelectric point (pI), amino acid composition, instability index, aliphatic index, and grand average of hydropathicity (GRAVY) were calculated to assess protein stability, solubility, and suitability for molecular modeling.

Secondary structure prediction

Secondary structure elements were predicted using the PSIPRED server, which utilizes PSI-BLAST-derived position-specific scoring matrices combined with neural network-based classification [28, 29]. The distribution of α-helices, β-sheets, and coil regions was analyzed to identify structurally flexible segments potentially influencing ligand binding.

Structural validation

Structural quality was assessed using PROCHECK and ERRAT servers [30, 31]. Ramachandran plot analysis was performed to evaluate backbone dihedral angles (φ and ψ), while ERRAT analysis provided an overall quality factor based on non-bonded atomic interactions. Structures with high percentages of residues in favored regions and ERRAT scores above 95% were considered reliable for downstream analyses.

Binding site prediction

Potential ligand-binding sites were identified using the PrankWeb server, which employs the P2Rank machine-learning algorithm [32, 33]. Binding pockets were ranked based on score and probability, and the top-ranked sites were selected. Docking grid centers were defined based on predicted pocket coordinates to ensure accurate targeting of active sites.

Pharmacophore modeling and virtual screening

Pharmacophore modeling and virtual screening were performed using the Pharmit server, which integrates pharmacophore-based filtering with shape-based screening of large chemical libraries. Key interaction features, including hydrogen bond donors (HBD), hydrogen bond acceptors (HBA), hydrophobic regions, aromatic rings, and charged groups, were identified based on the geometry and interaction patterns of the PDE9A active site across multiple crystal structures.

These features were spatially arranged to define the pharmacophore model, and compounds were screened based on their ability to match these features within the Pharmit platform. Default geometric tolerance parameters were applied for feature matching, and only compounds satisfying the defined pharmacophore constraints were retained for further analysis.

Pharmacophore model validation

To evaluate the discriminative ability of the selected pharmacophore model, pharmacophore validation was performed using a dataset of experimentally reported PDE9A inhibitors and decoy compounds. Known PDE9A active compounds were retrieved from the ChEMBL database using the human PDE9A target entry CHEMBL3535. Compounds with reported IC50 or Ki values ≤ 1 µM were selected as actives. Property-matched decoy compounds were generated using the DUD-E decoy generation approach. The combined active–decoy dataset was screened against the selected pharmacophore model using the same screening conditions applied in the virtual screening workflow. The model performance was evaluated using sensitivity, specificity, ROC-AUC, and enrichment factors at 1% and 5% of the ranked database. Enrichment factors (EF1% and EF5%) were calculated as the ratio of active compounds retrieved within the top-ranked subset to the number expected from random selection.

Ligand retrieval and preparation

Ligand candidates identified through pharmacophore-based screening of a pre-filtered subset of the PubChem database were retrieved for further analysis. Three-dimensional structures were obtained using compound identifiers (CIDs), SMILES, or compound names. Ligands were converted to PDBQT format using Open Babel (version 3.1.1) and energy-minimized using the MMFF94 force field. Gasteiger charges were assigned, and rotatable bonds were defined prior to docking.

ADMET and drug-likeness evaluation

Pharmacokinetic and toxicity profiles were predicted using ADMETlab 3.0, which incorporates advanced predictive models, including deep learning approaches. Parameters assessed included absorption, distribution, metabolism, excretion, and toxicity properties.

Drug-likeness was evaluated using the SwissADME platform, including Lipinski’s rule of five and BOILED-Egg analysis for gastrointestinal absorption and blood–brain barrier permeability. Compounds satisfying drug-likeness criteria and exhibiting acceptable ADMET profiles were prioritized for further analysis.

In addition to standard ADMET parameters, detailed pharmacokinetic descriptors including lipophilicity (LogP), topological polar surface area (TPSA), cytochrome P450 (CYP) enzyme interactions, and toxicity-related endpoints (e.g., hepatotoxicity and mutagenicity) were evaluated to provide a comprehensive assessment of drug-likeness and safety. Drug-likeness was assessed based on established criteria, including Lipinski’s rule of five and commonly accepted thresholds for physicochemical and pharmacokinetic parameters.

Molecular docking

Molecular docking simulations were performed using PyRx (version 0.8) with the AutoDock Vina engine. Docking grids were centered on predicted binding pockets with dimensions of approximately 20 × 20 × 20 Å to fully encompass active sites.

Docking parameters included an exhaustiveness value of 8 and a maximum of nine binding poses per ligand. Binding affinities were calculated in kcal/mol, and the lowest-energy conformations with favorable interaction geometry were selected for further analysis.

To provide a reference for docking performance, previously reported PDE9A inhibitors including BAY-73-6691 and PF-04447943 were included as positive control compounds. The docking behavior of screened compounds was compared against these reference inhibitors to provide contextual evaluation of binding affinity and interaction patterns.

Redocking and validation

Docking results were validated using the CB-Dock2 server, which integrates cavity detection with blind docking. Redocking was performed for selected ligands, and root mean square deviation (RMSD) values below 2.0 Å were considered indicative of reliable binding pose prediction. Consistency between docking and redocking results was used to confirm binding stability.

Protein–ligand interaction analysis

Protein–ligand interactions were analyzed using BIOVIA Discovery Studio Visualizer (version 2021). Hydrogen bonds, hydrophobic interactions, electrostatic interactions, and π–π stacking were identified and characterized. Interaction distances and key residues were recorded to evaluate binding stability and specificity.

Ligand optimization

Lead compounds were optimized using the DeepFrag platform, a deep learning-based fragment design tool. Fragment substitutions were generated based on binding site compatibility, and optimized ligands were re-docked to evaluate improvements in binding affinity and interaction profiles.

Protein–protein interaction analysis

Protein–protein interaction (PPI) networks were constructed using the STRING database [version 11.5], integrating experimental data, computational predictions, coexpression, and literature mining. High-confidence interactions (score > 0.7) were selected to identify functional partners and associated biological pathways.

Gene coexpression analysis

Gene coexpression analysis was performed using the STRING platform to evaluate transcriptional relationships between PDE9A and related genes across Homo sapiens and other organisms. This analysis provided insight into regulatory networks and functional associations relevant to nucleotide metabolism and cGMP signaling pathways.

Molecular dynamics simulation

Molecular dynamics (MD) simulations were conducted using the Desmond MD engine implemented in the Schrödinger software suite (Release 2025-1). Protein–ligand complexes were solvated in an orthorhombic simulation box using the TIP3P water model with a 10 Å buffer.

The system was neutralized with counter ions, and 0.15 M NaCl was added to mimic physiological conditions. Simulations were performed using the OPLS4 force field under NPT (constant number of particles, pressure, and temperature) ensemble conditions at 300 K and 1 atm pressure.

Energy minimization and equilibration were carried out using the default Desmond relaxation protocol, followed by a 500 ns production run with a 2 fs time step. Trajectory analyses included root mean square deviation (RMSD), root mean square fluctuation (RMSF), and radius of gyration (Rg, a measure of structural compactness).

Results

Overview of computational workflow

The overall computational workflow employed for the identification of PDE9A inhibitors is presented in Fig. 1. The integrated pipeline comprised protein preparation, binding-site prediction, pharmacophore modeling, pharmacophore validation, virtual screening, ADMET evaluation, molecular docking, redocking validation, protein–ligand interaction analysis, molecular dynamics simulation, and network-based functional analyses.

Fig. 1.

Integrated computational workflow for the identification of PDE9A inhibitors targeting HFpEF

Physicochemical characterization of PDE9A structures

The physicochemical properties of PDE9A crystal structures (4GH6, 4Y87, and 6LZZ) are summarized in Table 1. All structures exhibited acidic isoelectric points (pI range: 5.14–5.67), indicating a predominantly negatively charged environment under physiological conditions. Molecular weights ranged from 38.0 to 61.7 kDa, reflecting structural variability among conformations.

Table 1.

Physicochemical properties of PDE9A crystal structures retrieved from the Protein Data Bank

PDB ID	Amino Acids (n)	Theoretical pI	Molecular Weight (Da)	Molecular Formula	Instability Index	Aliphatic Index	GRAVY
4GH6	326	5.14	38,502	C₁₇₂₇H₂₆₅₈N₄₄₆O₄₉₂S₃₀	53.87	84.05	−0.249
4Y87	533	5.67	61,709	C₂₇₂₈H₄₃₁₀N₇₃₆O₈₁₁S₄₁	49.80	83.25	−0.348
6LZZ	322	5.14	38,044	C₁₇₀₄H₂₆₂₈N₄₄₂O₄₈₆S₃₀	54.03	85.09	−0.236

pI: Isoelectric point; GRAVY: Grand Average of Hydropathicity; Da: Dalton (molecular mass unit)

The instability index values > 40 indicate that the proteins may be unstable under in vitro conditions

Positive aliphatic index values suggest high thermostability of the protein structures

Negative GRAVY values indicate the hydrophilic nature of the proteins

Instability index values exceeding 40 suggest potential in vitro instability; however, elevated aliphatic index values (> 83) indicate substantial thermostability. Negative GRAVY values across all structures confirm their hydrophilic nature, which may facilitate ligand accessibility. Collectively, these properties support the suitability of the selected structures for structure-based drug discovery.

Secondary structure and structural validation

Secondary structure prediction revealed that α-helices constitute the dominant structural component (49.36–52.50%), with minimal β-sheet content and moderate coil regions (Table 2). The presence of coil regions suggests localized flexibility that may facilitate ligand accommodation.

Table 2.

Secondary structure composition of PDE9A crystal structures predicted using PSIPRED

PDB ID	α-Helix (%)	β-Sheet (%)	Coil (%)	Confidence Score (%)
4GH6	51.14	0.00	26.29	85.8
4Y87	49.36	2.03	30.76	88.0
6LZZ	52.50	3.10	28.40	87.8

α-Helix: Alpha-helical secondary structure

β-Sheet: Beta-strand secondary structure

Coil: Random coil or loop regions

Secondary structure elements were predicted using the PSIPRED algorithm based on amino acid sequences

The confidence score represents the average reliability of secondary structure prediction, with values above 80% indicating high prediction accuracy

The predominance of α-helical content suggests structural stability of the PDE9A catalytic domain, while coil regions may contribute to conformational flexibility and ligand binding

Structural validation using Ramachandran plot analysis demonstrated that more than 89% of residues were located in favored regions, while ERRAT quality factors exceeded 97% (Table 3), indicating high stereochemical reliability. These findings confirm the structural integrity of the selected models for downstream analyses.

Table 3.

Structural validation of PDE9A crystal structures based on Ramachandran plot and ERRAT analysis

PDB ID	3D Structure	Ramachandran Favored Residues (%)	ERRAT Quality Factor (%)
4GH6		89.3	97.95
4Y87		91.5	98.57
6LZZ		92.0	98.70

ERRAT: Structure validation tool based on non-bonded atomic interactions

Ramachandran plot analysis evaluates the stereochemical quality of protein backbone dihedral angles (φ and ψ). Values above 90% in favored regions indicate high structural reliability

ERRAT quality factor assesses the overall model quality based on non-bonded interactions; values above 95% indicate high-resolution and well-refined protein structures

The high validation scores across all PDE9A structures confirm their suitability for downstream molecular docking and simulation studies

Binding pocket identification

Binding site prediction using PrankWeb identified high-confidence ligand-binding pockets across all structures (Table 4). The top-ranked pockets exhibited high scores (> 52) and probabilities (> 0.98), indicating strong likelihood of biological relevance.

Table 4.

Predicted ligand-binding pockets of PDE9A crystal structures identified using PrankWeb

PDB ID	Pocket Rank	Score	Probability	Residues (n)	Pocket Center Coordinates (x, y, z)
4GH6	1	55.34	0.983	32	(80.79, 50.63, 39.39)
	2	54.82	0.983	30	(130.88, 44.66, 80.17)
4Y87	1	52.29	0.980	32	(81.54, 51.22, 39.55)
	2	48.13	0.974	26	(97.67, 26.21, 13.62)
	3	1.05	0.008	5	(87.39, 38.79, 29.85)
	4	0.92	0.005	5	(98.90, 44.62, 22.25)
6LZZ	1	57.74	0.985	26	(− 26.96, − 44.22, − 54.68)
	2	53.77	0.981	30	(− 2.06, − 28.24, − 28.71)
	3	2.16	0.050	6	(− 17.36, − 35.35, − 44.73)
	4	1.97	0.040	8	(− 23.41, − 41.16, − 43.82)
	5	1.91	0.037	6	(− 14.07, − 33.52, − 37.70)
	6	1.28	0.014	5	(4.57, − 36.30, − 18.58)
	7	1.25	0.013	5	(− 6.86, − 45.50, − 38.01)
	8	1.19	0.011	5	(− 21.96, − 51.68, − 65.34)
	9	1.11	0.009	7	(− 16.60, − 50.16, − 30.86)
	10	0.98	0.006	5	(− 8.23, − 44.95, − 45.96)

PDB: Protein Data Bank

Binding pockets were predicted using the PrankWeb server based on structural and physicochemical features of PDE9A

The score reflects the likelihood of a pocket being a ligand-binding site, while probability indicates prediction confidence

Higher scores and probabilities correspond to biologically relevant active sites suitable for molecular docking

Only top-ranked pockets (Rank 1–2) were considered for subsequent docking studies

These pockets comprised approximately 26–32 residues and were spatially well-defined. Only the top-ranked sites were selected for docking grid generation to ensure accurate targeting of catalytically relevant regions (Fig. 2).

Fig. 2.

Predicted binding pockets of PDE9A crystal structures identified using PrankWeb

Pharmacophore modeling and virtual screening

Pharmacophore models derived from PDE9A structures (Fig. 3) revealed essential interaction features, including hydrogen bond donors and acceptors, hydrophobic regions, and aromatic centers. A total of 30 compounds matching the pharmacophore criteria were retrieved and subjected to subsequent analysis (Supplementary Table S1). Rather than screening the entire PubChem database, pharmacophore-based filtering was applied to identify compounds that matched the defined spatial and chemical features, thereby generating a focused library of candidate molecules for further analysis. The selected pharmacophore features reflect key interaction requirements within the PDE9A active site and were used to prioritize compounds with complementary spatial and chemical properties.

Fig. 3.

Pharmacophore features of PDE9A crystal structures identified using Pharmit

The pharmacophore-based filtering effectively enriched ligands with complementary interaction patterns, thereby reducing the chemical search space and improving the efficiency of downstream docking analyses.

Pharmacophore validation

The reliability of the selected pharmacophore model was evaluated using an active–decoy validation dataset consisting of 32 known PDE9A inhibitors and 960 property-matched decoy compounds. A summary of the pharmacophore validation performance metrics is provided (Supplementary Table S2). Pharmacophore screening successfully identified 27 active compounds while excluding the majority of decoy molecules, resulting in a sensitivity of 84.38% and a specificity of 91.25%.

Receiver operating characteristic (ROC) analysis yielded an area under the curve (ROC-AUC) value of 0.81, indicating good discriminatory performance of the pharmacophore model in distinguishing active PDE9A inhibitors from inactive compounds (Supplementary Figure S6). Enrichment factor analysis further demonstrated preferential retrieval of active compounds among the highest-ranked screening hits, with EF1% and EF5% values of 12.50 and 6.80, respectively. These findings support the suitability of the pharmacophore model for virtual screening and prioritization of potential PDE9A inhibitors for subsequent docking and ADMET analyses.

ADMET screening and molecular docking

Selected ligands were evaluated for pharmacokinetic properties and binding affinities (Supplementary Table S3). Several compounds demonstrated favorable ADMET profiles and compliance with Lipinski’s rule of five, indicating acceptable drug-like characteristics. According to Lipinski’s criteria (molecular weight ≤ 500 Da, LogP ≤ 5, hydrogen bond donors ≤ 5, and hydrogen bond acceptors ≤ 10), the selected lead compounds satisfied key thresholds associated with oral bioavailability.

In addition, lipophilicity (LogP) values were within the optimal range for membrane permeability, while topological polar surface area (TPSA) values were consistent with favorable gastrointestinal absorption (typically < 140 Ų). Predicted cytochrome P450 (CYP) interaction profiles indicated minimal inhibition of major isoforms, suggesting a lower risk of metabolic drug–drug interactions. Toxicity predictions further indicated low risks of hepatotoxicity and mutagenicity for the prioritized candidates.

Collectively, these results support the pharmacokinetic suitability and drug-likeness of the identified compounds for further investigation.

Docking analysis revealed that multiple ligands exhibited strong binding affinities ( ≤ − 9.0 kcal/mol) across PDE9A structures (Supplementary Table S4). Among these, EVPMO-PurHD, BFMP, Irsenontrine, and CHOX-CHP-TPA consistently displayed high binding affinity across multiple conformations.

Redocking validation using CB-Dock2 confirmed the reliability of docking predictions, with consistent binding energy profiles observed for selected ligands (Supplementary Table S5).

Lead compound prioritization

An integrated evaluation of docking scores, ADMET properties, and redocking results identified several high-priority candidates. Although several screened compounds exhibited more favorable predicted docking scores than reference PDE9A inhibitors, docking energies alone do not necessarily correlate with biological potency and require experimental validation (Table 5). EVPMO-PurHD is an existing compound retrieved from the PubChem database (CID: 144509063) during pharmacophore-guided virtual screening and is not a newly designed or synthesized molecule. The compound was prioritized based on its favorable docking affinity, pharmacokinetic profile, and molecular dynamics performance.

Table 5.

Prioritized PDE9A inhibitor candidates and reference inhibitors identified through integrated pharmacophore screening, docking, ADMET, and redocking analyses

Ligand	CID	Docking (kcal/mol)*	ADMET/Lipinski	Redock (kcal/mol)	Rank
EVPMO-PurHD	144,509,063	−10.1 (4Y87)	Acceptable / Yes	−9.1	Lead
BFMP	91,031,059	−10.1 (4Y87)	Acceptable / Yes	−9.0	Lead
Irsenontrine	136,253,996	−9.8 (4Y87)	Acceptable / Yes	−10.2	Lead
CHOX-CHP-TPA	166,210,258	−9.3 (6LZZ)	Acceptable / Yes	−10.3	Lead
FPCHM-MBA	70,829,620	−9.7 (4Y87)	Acceptable / Yes	—	Promising
CFPEA-PTP	133,305,091	−9.4 (4GH6)	Partial / Yes	—	Moderate
FPEA-MTDP	52,209,272	−9.1 (4Y87)	Partial / Yes	—	Moderate
PF-04449613	137,796,288	−9.4 (4Y87)	Acceptable / Yes	—	Moderate
DMP-TPT-PEt	137,538,996	−9.0 (4GH6)	Acceptable / Yes	—	Moderate
BAY-73-6691	73,234,858	−7.5 (4GH6)	Partial / Yes	—	Low
PF-04447943	135,564,558	−8.8 (4GH6)	Rejected / Yes	—	Low
NBTP	79,882,083	−7.8 (4GH6)	Rejected / No	—	Rejected

BAY-73-6691 and PF-04447943 are previously reported PDE9A inhibitors included as reference compounds (positive controls) for comparative evaluation

Additional candidates, including BFMP and Irsenontrine, exhibited comparable binding affinities and favorable pharmacokinetic profiles, while CHOX-CHP-TPA showed strong binding to the 6LZZ structure. Fragment-based optimization using DeepFrag further enhanced the interaction potential of selected ligands.

Protein–ligand interaction analysis

Detailed interaction analysis of the EVPMO-PurHD–PDE9A complex (4Y87) revealed multiple stabilizing interactions within the active site (Table 6; Fig. 4). Key hydrogen bonds were observed with ASN301, ASP293, and ASP402, with distances below 3.5 Å, indicating strong and stable polar interactions.

Table 6.

Protein–ligand interaction profile of the EVPMO-PurHD–PDE9A (4Y87) complex obtained after redocking analysis

No.	Interacting Residues (Protein–Ligand)	Interaction Category	Interaction Type	Distance (Å)
1	ASN301–O (ligand)	Hydrogen bond	Conventional H-bond	3.00
2	ASP293–H (ligand)	Hydrogen bond	Conventional H-bond	2.82
3	ASP402–H (ligand)	Hydrogen bond	Conventional H-bond	2.55
4	ASP293–C (ligand)	Hydrogen bond	C–H bond	3.33
5	TYR424–C (ligand)	Hydrogen bond	C–H bond	3.68
6	ASP293	Electrostatic	π–anion	3.87
7	ASP402	Electrostatic	π–anion	3.51
8	PHE456	Hydrophobic	π–π stacked	5.30
9	PHE251	Hydrophobic	π–π T-shaped	5.07
10	THR363, ASP364	Hydrophobic	Amide–π stacked	4.56
11	MET365	Hydrophobic	π–alkyl	5.47
12	ILE403	Hydrophobic	π–alkyl	5.38
13	LEU420	Hydrophobic	π–alkyl	4.99

Å: Angstrom (distance unit); H-bond: Hydrogen bond

π–π: Aromatic stacking interaction

Interactions were analyzed using BIOVIA Discovery Studio based on the redocked complex of EVPMO-PurHD with PDE9A (PDB ID: 4Y87)

Hydrogen bond distances below 3.5 Å indicate strong and stable polar interactions

Electrostatic π–anion interactions contribute to binding specificity and stabilization within the active site

Hydrophobic interactions, including π–π stacking and π–alkyl contacts, enhance ligand affinity by stabilizing the ligand within the hydrophobic pocket

The presence of multiple interaction types suggests a stable and energetically favorable binding mode

Fig. 4.

Three-dimensional representation of EVPMO-PurHD interactions within the active site of PDE9A (PDB ID: 4Y87)

Electrostatic π–anion interactions involving ASP293 and ASP402 contributed to binding specificity, while hydrophobic interactions—including π–π stacking with PHE456 and PHE251 and π–alkyl interactions with MET365, ILE403, and LEU420—further stabilized the complex. Interaction distances ranging from 2.5 to 5.5 Å support a stable and energetically favorable binding configuration.

Pharmacokinetic profiling (BOILED-Egg model)

Pharmacokinetic evaluation of EVPMO-PurHD using the BOILED-Egg model (Fig. 5) indicated its localization within the white region, suggesting favorable gastrointestinal absorption. The absence of the compound in the yolk region indicates limited blood–brain barrier permeability.

Fig. 5.

Predicted pharmacokinetic properties of EVPMO-PurHD using the BOILED-Egg model

These results suggest that EVPMO-PurHD possesses desirable oral bioavailability while minimizing potential central nervous system exposure.

Molecular dynamics simulation analysis

Molecular dynamics simulation of the EVPMO-PurHD–PDE9A complex (4Y87), the complex exhibited an initial equilibration phase followed by stable behavior during the latter portion of the simulation (Fig. 6). The RMSD profile exhibited an initial phase of relatively large fluctuations, particularly for the ligand, with values reaching up to ~ 6 Å during the first ~ 200 ns. This behavior suggests a conformational adjustment period in which the ligand undergoes reorientation within the binding pocket. Following this phase, the RMSD values gradually decreased and stabilized, indicating equilibration and the formation of a more stable protein–ligand complex.

Fig. 6.

Molecular dynamics simulation analysis of the PDE9A–EVPMO-PurHD complex (4Y87). (A) Root mean square fluctuation (RMSF) of protein residues indicating structural flexibility. (B) RMSF of ligand atoms reflecting ligand stability within the binding pocket. (C) Protein–ligand RMSD profile showing stability of the complex over the simulation time. (D) Hydrogen bond analysis illustrating the persistence of intermolecular interactions throughout the simulation. (E) Radius of gyration (Rg) plot representing compactness and structural stability of the protein–ligand complex over time

After ~ 200 ns, the ligand RMSD remained comparatively stable with reduced fluctuations, supporting a consistent binding mode during the later stages of the simulation. Protein RMSD values remained within ~ 1–3 Å throughout the simulation, indicating overall structural stability of the system.

RMSF analysis indicated moderate flexibility in specific regions, while ligand fluctuations remained minimal, suggesting stable positioning within the binding pocket. Hydrogen bonding interactions were observed throughout the simulation, particularly during the equilibrated phase, indicating their contribution to stabilizing the ligand within the binding pocket. However, a detailed quantitative hydrogen bond occupancy analysis was not performed in the present study.

Additionally, the radius of gyration remained relatively constant, indicating maintained structural compactness. Extended analyses, including secondary structure evolution and contact profiling, are provided in Supplementary Figure S5.

Protein–protein interaction and gene coexpression analysis

Protein–protein interaction network analysis revealed strong functional associations between PDE9A and proteins involved in nucleotide metabolism and cGMP signaling pathways (Table 7; Fig. 7). Key interacting partners included APRT, GMPR, GUK1, and GMPS, highlighting the role of PDE9A in nucleotide metabolism. Importantly, additional interacting proteins such as GUCY1A2, GUCY1B1, and NPR1 are directly involved in cyclic GMP (cGMP) synthesis and natriuretic peptide signaling pathways, which are critical regulators of myocardial relaxation and vascular function. Dysregulation of these pathways is strongly associated with the development and progression of HFpEF.

Table 7.

Protein–protein interaction network of PDE9A and its associated functional partners identified using the STRING database

No.	Protein Symbol	Protein Name / Function	Interaction Score
1	APRT	Adenine phosphoribosyltransferase; involved in AMP formation	0.913
2	GMPR	GMP reductase 1; catalyzes NADPH-dependent deamination of GMP to IMP	0.907
3	ITPA	Inosine triphosphate pyrophosphatase; hydrolyzes non-canonical nucleotides	0.905
4	HPRT1	Hypoxanthine-guanine phosphoribosyltransferase; converts guanine to GMP	0.904
5	GUK1	Guanylate kinase; catalyzes phosphorylation of GMP to GDP	0.903
6	GMPS	GMP synthase; involved in de novo guanine nucleotide synthesis	0.903
7	ALDH7A1	Alpha-aminoadipic semialdehyde dehydrogenase; involved in metabolic detoxification	0.881
8	GUCY1A2	Guanylate cyclase soluble subunit α2; catalyzes cyclic GMP synthesis	0.828
9	NPR1	Natriuretic peptide receptor 1; regulates cardiovascular signaling	0.818
10	GUCY1B1	Guanylate cyclase soluble subunit β1; mediates NO-dependent cGMP signaling	0.816

AMP: Adenosine monophosphate; GMP: Guanosine monophosphate; GDP: Guanosine diphosphate; NO: Nitric oxide; STRING: Search Tool for the Retrieval of Interacting Genes/Proteins

Interaction scores represent the confidence level of predicted protein–protein associations derived from combined evidence sources, including experimental data, co-expression, and curated databases

Scores closer to 1.0 indicate stronger confidence in functional association

The identified interaction partners are primarily involved in nucleotide metabolism and cyclic GMP (cGMP) signaling pathways, highlighting the functional role of PDE9A in cardiovascular and metabolic regulation

Fig. 7.

Protein–protein interaction network of PDE9A generated using the STRING database

Gene coexpression analysis further demonstrated coordinated expression patterns between PDE9A and associated genes across Homo sapiens and other organisms (Fig. 8). These findings indicate that PDE9A is functionally integrated within regulatory networks governing nucleotide metabolism and cGMP signaling pathways. Notably, key interacting proteins such as GUCY1A2, GUCY1B1, and NPR1 are directly involved in cyclic GMP synthesis and cardiovascular signaling, which are critical processes in the pathophysiology of HFpEF. This functional association highlights the potential impact of PDE9A modulation on myocardial relaxation, vascular tone, and cardiac remodeling.

Fig. 8.

Gene coexpression analysis of PDE9A across Homo sapiens and other organisms

Discussion

Heart failure with preserved ejection fraction (HFpEF) remains a significant clinical challenge due to the absence of effective therapies capable of substantially improving patient outcomes. Accumulating evidence highlights phosphodiesterase 9A (PDE9A) as a promising therapeutic target, owing to its role in regulating cyclic guanosine monophosphate (cGMP) signaling derived from natriuretic peptide pathways [34]. Dysregulation of this signaling axis contributes to myocardial fibrosis, hypertrophy, and impaired diastolic function, which are central features of HFpEF pathophysiology [35]. Moreover, elevated PDE9A expression in HFpEF myocardium has been associated with inflammatory remodeling and disrupted cardiac signaling, further supporting its pathological relevance [36].

In contrast to phosphodiesterase type 5 (PDE5) inhibitors, which primarily modulate nitric oxide–dependent cGMP signaling and have shown limited clinical efficacy in HFpEF, PDE9A selectively hydrolyzes nitric oxide–independent cGMP pools. This mechanistic distinction suggests that PDE9A inhibition may provide a complementary and potentially more effective therapeutic strategy [19, 37]. Preclinical studies have demonstrated that selective inhibition of PDE9A attenuates myocardial remodeling and improves diastolic function, reinforcing its potential as a viable drug target [38].

The present study employed an integrated computational framework to systematically identify and evaluate potential PDE9A inhibitors. A key methodological strength lies in the incorporation of multiple crystal structures (4GH6, 4Y87, and 6LZZ), enabling the consideration of conformational variability within the catalytic domain. This multi-structure approach addresses a critical limitation of conventional docking studies that rely on a single static conformation and may therefore overlook alternative binding modes [39]. The identification of high-confidence binding pockets across all structures further supports the robustness of the docking strategy.

Compared with existing studies that typically employ single-structure docking or short-duration simulations, the present work adopts a multi-level strategy that integrates structural, dynamic, and systems-level analyses [18, 22]. The use of multiple crystal structures enables consideration of conformational variability, while extended molecular dynamics simulations provide deeper insight into interaction stability. Furthermore, the incorporation of protein–protein interaction and gene coexpression analyses offers additional biological context that is often lacking in conventional computational studies. Although the present molecular dynamics analysis provides insight into structural stability and interaction persistence, more advanced analyses such as binding free energy calculations (MM-PBSA/MM-GBSA) and principal component analysis (PCA) were not conducted. These approaches would allow quantitative estimation of binding energetics and characterization of dominant conformational motions, respectively. Incorporating such analyses in future studies would provide a more comprehensive understanding of protein–ligand interactions and further strengthen the predictive reliability of the results.

Pharmacophore-guided virtual screening combined with ADMET evaluation enabled efficient prioritization of compounds with favorable physicochemical and pharmacokinetic properties. The reliability of the pharmacophore model was further supported through validation using an active–decoy dataset comprising known PDE9A inhibitors and property-matched decoy compounds. The model achieved an ROC-AUC value of 0.81, together with enrichment factors of 12.50 and 6.80 at the top 1% and 5% of screened compounds, respectively. These results indicate good discriminatory performance and support the suitability of the pharmacophore model for prioritizing potential PDE9A inhibitors during virtual screening. Among the identified candidates, EVPMO-PurHD, BFMP, Irsenontrine, and CHOX-CHP-TPA demonstrated consistently strong binding affinities across multiple protein conformations. Notably, EVPMO-PurHD exhibited the highest docking affinity (− 10.1 kcal/mol), along with favorable ADMET characteristics, indicating its potential as a lead compound. The inclusion of detailed pharmacokinetic descriptors such as LogP, TPSA, CYP interaction profiles, and toxicity predictions further strengthens the reliability of compound prioritization and enhances the translational relevance of the findings.

At the molecular level, interaction analysis revealed that EVPMO-PurHD forms a stable and well-defined binding network within the PDE9A active site. The presence of multiple hydrogen bonds with key residues (ASN301, ASP293, and ASP402), along with electrostatic and hydrophobic interactions, suggests strong binding specificity and stabilization. These interaction patterns are consistent with previously reported binding modes of PDE9A inhibitors, supporting the structural validity of the identified ligand [40].

Importantly, molecular dynamics simulations provided dynamic validation of docking results, demonstrating that the EVPMO-PurHD–PDE9A complex remains stable over a 500 ns simulation period. The observed RMSD convergence, low ligand fluctuation, and sustained hydrogen bonding interactions indicate persistent binding within the active site. These findings highlight the importance of incorporating long-timescale simulations to capture the dynamic behavior of protein–ligand systems and improve the reliability of computational predictions [41].

Compared with previous computational studies, which often rely on single-structure docking and limited simulation analyses, the present work integrates multi-conformational modeling, pharmacophore-based screening, ADMET profiling, and extended molecular dynamics simulations. This comprehensive approach enhances predictive robustness and addresses key limitations related to protein flexibility and insufficient pharmacokinetic evaluation [11, 42]. Nevertheless, it is important to acknowledge that computational predictions inherently depend on model assumptions and may not fully capture biological complexity. The favorable ADMET profiles observed for the lead compounds, particularly in relation to Lipinski’s rule of five and key pharmacokinetic thresholds, further support their potential as orally bioavailable drug candidates.

In addition to structural and dynamic analyses, systems-level investigations provide important biological context for the role of PDE9A in HFpEF. The protein–protein interaction network revealed strong associations with key regulators of nucleotide metabolism and cyclic GMP signaling, including GUCY1A2, GUCY1B1, and NPR1. These proteins play central roles in the synthesis and regulation of cGMP and natriuretic peptide signaling, which are essential for maintaining myocardial relaxation, vascular tone, and cardiac homeostasis.

In HFpEF, impaired cGMP signaling contributes to increased myocardial stiffness, reduced diastolic function, and microvascular dysfunction. The observed interactions suggest that PDE9A operates within this regulatory network, and its inhibition may enhance cGMP availability, thereby improving cardiac relaxation and reducing pathological remodeling.

Furthermore, gene coexpression analysis demonstrated coordinated expression patterns between PDE9A and these functionally related genes, suggesting shared regulatory mechanisms within cardiovascular and metabolic pathways. This network-level evidence supports the hypothesis that inhibition of PDE9A may enhance cGMP signaling and thereby mitigate key pathological processes underlying HFpEF. Collectively, these findings strengthen the biological relevance of the identified targets and highlight the importance of integrating systems-level analyses in computational drug discovery [43].

Collectively, the results support the hypothesis that integrated computational strategies can effectively identify and prioritize candidate molecules targeting PDE9A. The convergence of docking, molecular dynamics, and network-based analyses strengthens the confidence in the identified lead compounds and provides a comprehensive framework for early-stage drug discovery in HFpEF.

Limitations

Despite the comprehensive computational framework employed in this study, several limitations should be acknowledged. First, the present work is based entirely on in silico approaches, and no experimental validation was performed. Therefore, the biological activity, selectivity, and safety of the identified compounds remain to be confirmed through biochemical assays, cell-based studies, and in vivo investigations.

Second, although long-timescale molecular dynamics simulations were conducted, these simulations may not fully capture the complexity of physiological environments or long-term conformational changes. Additionally, quantitative binding free energy calculations (e.g., MM/PBSA or MM/GBSA) and comparative simulations across multiple ligands were not performed, which may limit the accuracy of binding affinity estimation.

Third, ADMET predictions were derived from computational models and may not fully reflect in vivo pharmacokinetic and toxicity profiles. While multiple protein conformations were considered, the complete extent of conformational flexibility, including potential allosteric effects, may not be fully represented.

Future research should focus on experimental validation of the identified lead compounds, including enzyme inhibition assays, cellular functional studies, and animal models of HFpEF. In addition, advanced computational techniques, such as binding free energy calculations, enhanced sampling simulations, and structure–activity relationship (SAR) analysis, should be employed to further refine and optimize candidate molecules. Integrating experimental and computational approaches will be essential to translate these findings into clinically relevant therapeutic strategies. Furthermore, advanced molecular dynamics analyses, including MM-PBSA/MM-GBSA binding free energy calculations, principal component analysis (PCA), and detailed hydrogen bond occupancy analysis, were not performed, which may limit the depth of dynamic and energetic interpretation.

Conclusion

This study presents a comprehensive computational framework for the identification of potential phosphodiesterase 9A (PDE9A) inhibitors targeting heart failure with preserved ejection fraction (HFpEF). By integrating multi-structure protein analysis, pharmacophore-guided virtual screening, molecular docking, fragment-based optimization, molecular dynamics simulations, and ADMET evaluation, the approach enabled systematic prioritization of candidate compounds. Among the identified ligands, EVPMO-PurHD demonstrated the most favorable binding affinity, stable interaction profile, and consistent dynamic behavior, highlighting its potential as a lead candidate. The incorporation of structural, dynamic, and systems-level analyses enhances the robustness of predictions and provides additional insights into PDE9A-targeted drug discovery. Overall, these findings support PDE9A as a promising therapeutic target in HFpEF and demonstrate the utility of integrative in silico approaches in early-stage drug discovery. The identified compounds represent promising computational candidates; however, further experimental validation is essential to confirm their therapeutic efficacy and translational potential.

Electronic Supplementary Material

Below is the link to the electronic supplementary material.

Abbreviations

HFpEF

Heart Failure with Preserved Ejection Fraction

PDE9A

Phosphodiesterase 9A

cGMP

Cyclic Guanosine Monophosphate

ADMET

Absorption, Distribution, Metabolism, Excretion, and Toxicity

RMSD

Root Mean Square Deviation

PDB

Protein Data Bank

PKG

Protein Kinase G

Lipinski

Lipinski’s Rule of Five

BOILED-Egg

A model for drug permeability prediction

Author contributions

Xiaoyan Lei: Conceptualization, methodology, data analysis, writing—original draft preparation. Rujie Wang: Methodology, data analysis, validation, writing—review and editing. Yongmei Li: Supervision, project administration, writing—review and editing, funding acquisition. All authors have read and approved the final manuscript.

Funding

This research received no external funding.

Data availability

All data generated or analyzed during this study are included within the manuscript.

Declarations

Ethical approval

No human participants or animals were involved in this study.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.