Identification of novel PTP1B inhibitor for the treatment of LPS-induced myocardial apoptosis: machine learning based virtual screening and biological evaluation

Abstract

Sepsis-induced cardiovascular dysfunction (SICD) poses a serious threat to human life. Protein tyrosine phosphatase 1B (PTP1B) displays an essential role in SICD occurrence, so discovering novel inhibitors targeting PTP1B is an effective strategy for SICD treatment. In this research, we exploited a novel virtual screening pipeline consisting of both ligand-based and structure-based modules to find novel PTP1B inhibitors, and compound PI-2 with IC₅₀ = 4.1 ± 0.3 μM was successfully discovered. Enzymatic and cellular thermal shift assay showed PI-2 displayed a moderate PTP1B inhibitory activity and a good selectivity towards both PTP1B and TCPTP. Besides, PI-2 effectively protected Lipopolysaccharide (LPS) induced AC16 injury by reducing cell ROS levels and enhancing mitochondrial membrane potential. Overall, this research not only provides a novel virtual screening strategy for discovering novel PTP1B inhibitors, but also supplies a potential candidate for further optimisation for the treatment of SICD.

Introduction

Sepsis induced cardiovascular dysfunction (SICD) is one of the major complications in sepsis patients, posing a serious threat to human life¹. Epidemiological investigations have shown that SICD has become the main indicator of poor prognosis in sepsis patients, with a mortality rate as high as 70%-90%, which is 2–3 times higher than that of ordinary sepsis patients². However, the clinical management of SICD remains highly limited; exploring the molecular mechanisms and promising therapeutic targets, along with potential pharmacological interventions, holds significant values³.

SICD primarily manifests as impaired myocardial systolic and diastolic function, and a reduced left ventricular ejection fraction (LVEF%) is the most notable indicator for patients⁴. Consequently, restoration of myocardial contractility is critical for SICD prevention and treatment⁵. Mitochondria are the main source of cardiac energy, which is important for myocardial contraction⁶. Various studies have shown that mitochondrial dysfunction is a key factor for cardiac depression; therefore, improving myocardial mitochondrial function is the central focus of SICD treatment^7–9.

Protein tyrosine phosphatase 1B (PTP1B), an essential member of protein tyrosine phosphatase (PTP) superfamily^10,11, has been commonly recognised as a negative regulator of insulin and leptin receptor signaling^12–15. Recently, a growing number of evidences has proved that PTP1B is a promising target for mitochondrial functional adjustment^16–18. Inhibition of PTP1B promotes mitochondrial homeostasis in both diabetic cardiomyopathy and calcific aortic valve disease^19,20.

Despite, the regulatory function of PTP1B on mitochondria in SICD being unclarified, the importance of PTP1B in SICD treatment is well defined²¹. Studies have shown that deletion of PTP1B attenuates the activation level of p38 mitogen-activated protein kinase (p38 MAPK) and promotes phosphatidylinositol 3-kinase (PI3K)/protein kinase B (Akt) phosphorylation to alleviate SICD^22,23 illustrating that PTP1B is an essential target for SICD treatment.

In the past decade, numerous PTP1B inhibitors, containing carboxylate, sulphonate, and difluorophosphate groups, have been developed^24–28. These inhibitors display preferable PTP1B inhibitory activities; However, none of these compounds have been approved for market because of lacking PTP1B selectivity and poor druggability induced by the high electrification and homology of the binding pocket of PTP1B with other PTPs. Among them, compound Tegeprotafib and Osunprotafib are the only two PTP1B inhibitors which are in clinical phase I for cancer treatment, both of them are also potent T-cell protein tyrosine phosphatase (TCPTP) inhibitors²⁹, and compound Ertiprotafib and Trodusquemine are failed in clinical trials for diabetic treatment because of insufficient therapeutic effect or occurrence of side effects (Table 1) ³⁰. Despite, PTP1B has already identified as a potential target for SICD treatment, the studies of the anti-SICD activities and mechanisms of these compounds are limited, so developing novel PTP1B inhibitors and clarifying their activities in mitochondria for SICD is still valuable.

Table 1.

Reported PTP1B inhibitors that in clinical trial.

Compounds	Inhibitory activity (IC50)		Phase	Indication
	PTP1B	TCPTP
Tegeprotafib	1-10 nM	4.4 nM	I	cancer
Osunprotafib	2.5 nM	1.8 nM	I	cancer
Ertiprotafib	1.6 μM		Failed in Phase II	diabetes
Trodusquemine	1 μM	224 μM	Failed in Phase II	diabetes

To discover novel PTP1B inhibitors and evaluate the anti-SICD activities of the novel compound, a virtual screening pipeline consisting of pharmacophore filtering and molecular docking was established. To improve the screening efficiency, Naïve Bayesian Classification (NBC) was used to achieve the requirements of high enrichment of virtual screening. Then, a series of biological assays were conducted to elucidate the PTP1B inhibitory activity and selectivity on the enzymatic and cellular levels. Eventually, cell apoptosis assay, inflammatory cytokine assay, ROS detection, and mitochondrial membrane potential test were used to evaluate the anti-LPS induced myocardial injury potential and mechanisms of the compound with preferable PTP1B inhibitory activity.

Results and discussion

PTP1B crystal structure selection

Around 367 PTP1B crystal structures (homo sapiens, resolution < 2.5) were listed in RCSB Protein Data Bank (https://www.rcsb.org/). Only 76 PTP1B structures were presented with ligands at the active site. After structure alignment, 73 PTP1B crystals were kept. According to the RMSD and distance matrix, these 73 structures were further clustered by K-means algorithm. As shown in Figure 1(A), when K = 10, the silhouette score was maximum, so these crystals were divided into 10 clusters and proteins whose RMSD value was closest to the centroid within the same category was selected. Finally, 1KAK³¹, 2NTA³², 1NZ7³³, 2F70³⁴, 2CNE³⁵, 1G1F³⁶, 1C86³⁷, 1Q6M³⁸, 1NWL³⁹ and 1NL9⁴⁰ were chosen for further study.

Figure 1.

PTP1B crystal cluster and inhibitor database analysis. A. The clustering results of collected PTP1B crystals (The PTP1B was divided into 10 clusters and the representative protein was shown in green); B. Chemical space analysis of the training set (yellow) and the test set (red) by principle component analysis.

Database diversity analysis

Eventually, a database containing around 26 thousand compounds was generated. The database was further divided into training set and test set in the ratio of 4:1. Generally, the predictive ability of machine learning is affected by the quality of the database, so the principle descriptors including, ALogP, molecular weight, number of hydrogen donors, number of hydrogen acceptors, number of rotatable bonds, number of rings, number of aromatic rings, molecular fractional polar surface area were calculated and analysed. As shown in Figure 1(B), the physical and chemical properties of molecules in both the training set and the test set were distributed uniformly, indicating that the constructed database displayed a good chemical diversity.

Pharmacophore generation and evaluation

The pharmacophores were generated according to the crystal complex of 1C86, 1G1F, 1KAK, 1NL9, 1NWL, 1NZ7, 1Q6M, 2CNE, 2F70, and 2NTA. The pharmacophores with the highest training set AUC value for each complex were listed in Table 2. We could see that the predictive ability of these pharmacophores was significantly different. Four pharmacophores produced by 1C86, 1NL9, 1NWL, and 1Q6M displayed a good “discriminatory power” with AUC_Training value > 0.7. Among them, 1C86 exhibited the best predictive ability with AUC_Traing value = 0.856, which was much better than 1G1F and 2CNE. To further validate the “discriminatory power” of these pharmacophores, the AUC value of the test set was also calculated. Consistent with the training set, pharmacophores generated by 1C86, 1NL9, 1NWL, and 1Q6M also presented a good “discriminatory power” for the test set.

Table 2.

Pharmacophores generated by selected PTP1B crystal structures.

PDB Entry	Feature_Set	Selectivity_Score	AUC_Training	AUC_Test
1C86	AANNR	11.689	0.856	0.854
1G1F	ADDNNN	16.495	0.556	0.661
1KAK	AAHHN	9.3394	0.623	0.568
1NL9	AADDHN	13.026	0.724	0.741
1NWL	AAAADH	10.379	0.755	0.789
1NZ7	ADHNNP	15.583	0.549	0.721
1Q6M	AAHHNN	12.590	0.781	0.798
2CNE	AADDNN	14.349	0.591	0.703
2F70	AAADDN	13.301	0.675	0.622
2NTA	AAADHR	10.173	0.660	0.597

A: hydrogen bond acceptor; N: Negative ionisable group; R: Aromatic ring; D: Hydrogen bond donor; H: Hydrophobic group; P: Positive ionisable group; AUC_Training: the AUC value of the training set; AUC_Test: the AUC value of the test set.

As shown in Table 2 and Figure 2, the pharmacophore generated by 1C86 consisted of 5 features, including two hydrogen bond acceptors, two negative ionisable groups, and one aromatic ring. The negative ionisable group was also presented in pharmacophores produced by 1Q6M and 1NL9. Despite the negative ionisable feature was beneficial for improving PTP1B inhibitory activity, it was not a preferable fragment for ameliorating membrane permeability, which was one of the main reasons why most reported PTP1B inhibitors displayed poor activity in vivo, so the pharmacophore produced by 1NWL, which consisted of four hydrogen bond acceptors, one hydrogen bond donor and one hydrophobic group was chosen for further study.

Figure 2.

Pharmacophores produced by selected PTP1B crystal (Green: hydrogen bond acceptor; light blue: Hydrophobic group; dark blue: Hydrophobic group; yellow: Hydrogen bond donor; orange: Aromatic ring; gray: steric hindrance).

Validation of docking reliability

The docking reliability was reflected by the ability of docking programs to predict the correct conformation in the molecular docking. The RMSD between the docking pose and the crystal conformation of each ligand was calculated and analysed. As shown in Figure 3(A), Glide_SP and Glide_XP performed superior to other programs in terms of the success rates for the top-scored poses (poses with the highest docking score) and best poses (poses with the lowest RMSD with native conformation). We could see that in Figure 3(B), the top-scored poses were the best poses for 1C86, 1KAK, 1NL9, 1Q6M, and 2NTA. However, the RMSD for 2NTA was > 2 Å, which was considered a failed to predict. Eventually, 1C86, 1KAK, 1NL9, 1Q6M (Figure 3C) were chosen for further study.

Figure 3.

The docking reliability of various docking programs. A. The RMSD cumulative distribution of top-scored poses (Top Conformation) and best poses (Best Conformation) with various docking programs; B. RMSD of top-scored poses (left) and best poses (right) of all docking programs (From red to blue the RMSD increased); C. The superpose of the best poses (green, docked by Glide_SP and Glide_XP) and natural pose (Pink) of each ligand.

Docking accuracy evaluation and multi-conformational NBC model construction

Docking accuracy represented the ability of programs to discriminate active inhibitors from decoys. According to the results of docking reliability, the docking accuracy of 1C86, 1KAK, 1NL9, and 1Q6M by Glide_SP and Glide_XP were further evaluated by the Mann-Whitney U test. As shown in Figure 4(A) and (B), both Glide_SP and Glide_XP could effectively distinguish active inhibitors and decoys for all crystal structures with a very low p values.

Figure 4.

The docking accuracy of Glide_SP and Glide_XP with various crystal structures and multiple conformation models. A. Distributions of the docking scores for active inhibitors (orange) and decoys (blue) with different crystal structures docked by Glide_SP; B. Distributions of the docking scores for active inhibitors (orange) and decoys (blue) with different crystal structures docked by Glide_XP; C. ROC curves of NBC based on a single conformation docked by Glide_SP; D. ROC curves of NBC based on a single conformation docked by Glide_XP; E. ROC curves of NBC based on multiple conformations docked by Glide_SP (Model1: 1NL9 and 1Q6M, Model2: 1NL9 and 1KAK, Model3: 1NL9 and 1C86, Model4: 1C86 and 1KAK, Model5: 1C86 and 1Q6M; Model6: 1Q6M and 1KAK, Model7: 1NL9, 1Q6M, and 1KAKA, Model8: 1NL9, 1Q6M, and 1C86, Model9: 1NL9, 1KAK, and 1C86, Model10:1Q6M, 1KAK, and 1C86, Model11: 1NL9, 1Q6M, 1KAK, and 1C86).

The NBC models, in which the docking score was set as an independent property, were further constructed to intuitively estimate the discrimination capability of each crystal structure. As described in Figure 4(C), the area under curve (AUC) values of 1NL9 and 1Q6M docked by Glide_SP were 0.811 and 0.732, respectively, which was superior or comparable to Glide_XP (Figure 4D). For both docking program, the AUC values were higher than 0.7, which illustrated that these models displayed a satisfactory degree of distinction. Given there was no significant difference between Glide_SP and Glide_XP and the docking efficiency of Glide_SP was superior to Glide_XP, so Glide_SP was used for further research.

To improve the docking accuracy, the NBC models based on multiple crystal structures under Glide_SP were developed. Figure 4(E) showed that the AUC values increased after the combination of multiple structures, especially for the models containing 1NL9; the AUC values for these models remained around 0.85, which was preferable for single structure models.

Virtual screening

Eventually, a virtual screening pipeline consisted of pharmacophore screening based on 1NWL and molecular docking based on multi-conformational NBC model was conducted. After pharmacophore screening, around 1% compounds were kept for further screening, and top 100 compounds were remained for manual selection after molecular docking. To guarantee the diversity of the selected compounds, the top 100 compounds were clustered by the fingerprint FCFP_6 into 10 cluster. Finally, 7 compounds (Figure 5A) in different clusters were selected and purchased. These compounds were found with structural diversity, and most of them possessed a carboxylic group, which was preferable for interacting with the polar PTP1B active binding pocket.

Figure 5.

The structures, activities and binding modes of selected compounds. A. The structures of selected compounds through virtual screening; B. PTP1B inhibitory rate of selected compounds at the concentration of 20 μM; C. The IC₅₀ value of PI-2 against PTP1B; D. The binding modes of PI-2 with PTP1B through molecular docking.

The result of inhibitory rate measurement showed that compounds PI-1, PI-2 and PI-7 displayed a moderate to potent inhibitory activity against PTP1B at 20 μM (Figure 5B). Among them, compound PI-2 was the most potent, the inhibitory rate was presented as 95.6% which was comparable with positive control NaVO₃ (Figure 5B). Enzymatic assay quantitatively clarified PI-2 was a moderate PTP1B inhibitor with IC₅₀ = 4.1 ± 0.3 μM (Figure 5C).

PTP1B binding mode and selectivity of PI-2

The binding mode of PI-2 with PTP1B was further studied by molecular docking. As shown in Figure 5(D), the carboxylic group of PI-2 effectively bound with the key residues in the active pocket of PTP1B through hydrogen bonds. Further, the phenyl group of PI-2 could interact with Phe 182 by hydrophobic interaction, and the benzimidazole group formed a hydrogen bond with Asn48/Asp48 or interacted with Tyr 46, Arg 47 through hydrophobic interaction to stabilise the binding conformation in different PTP1B crystal structures.

The inhibitory activity of PI-2 against TCPTP, the most homologous PTP with PTP1B, was measured to evaluate the PTP1B inhibitory selectivity of PI-2. As shown in Figure 6(A), the IC₅₀ of PI-2 against TCPTP was 5.9 ± 2.4 μM, which was slightly higher than PTP1B with a selective index = 1.4 (selective index = IC_{50 (TCPTP)}/IC_{50 (PTP1B)}). As illustrated in the docking results, the secondary noncatalytic arylphosphate binding site consisted by Asn 48/Asp 48, Tyr 46, and Arg 47, which is not conserved for all PTPs, displayed an essential role for PI-2 binding. To further evaluate the PTP1B selectivity of PI-2, the inhibitory activity of PI-2 against acidic phosphatase (ACP1), which also contains this binding pocket, was measured. Compared with PTP1B, the inhibitory potency of PI-2 against ACP1 was much lower, with IC₅₀ > 20 μM (Figure 6B), which means PI-2 exhibited a certain level of PTP1B selectivity at the enzymatic level.

Figure 6.

The PTP1B inhibitory selectivity of PI-2 at the enzymatic and the cellular level. A. The IC₅₀ value of PI-2 against TCPTP; B. The IC₅₀ value of PI-2 against ACP1; C-E. The effect of PI-2 on the stability of PTP1B and TCPTP in AC16 cells.

To further clarify the PTP1B inhibitory activity and selectivity of PI-2 in vitro, the cell thermal shift assay (CETSA) was conducted. As illustrated in Figure 6(C)–(E), with the increase of temperature, the proteolysis level of PTP1B and TCPTP in the PI-2 treated group was much lower than that of the untreated group, demonstrating that PI-2 could effectively enhance the PTP1B and TCPTP stability in vitro. Further, for the PI-2 treated group, the PTP1B stability was higher than TCPTP, especially for high-temperature treatment, indicating PI-2 also displayed a moderate PTP1B inhibitory selectivity at the cellular level.

The effect of PI-2 on LPS-induced AC16 apoptosis In vitro

To investigate the effectiveness of PI-2 on the LPS-induced myocardial injury, the expressions of inflammation factors IL-1β, TNF-α, and IL-6 of AC16 cells were tested. We found that after LPS administration, the levels of IL-1β, TNF-α, and IL-6 were increased significantly (Figure 7A). After 48 h of PI-2 treatment at a concentration of 2.5 μM, the level of TNF-α decreased significantly, and the contents of IL-1β and IL-6 were also recovered (Figure 7A), which preliminarily proved that PI-2 could effectively alleviate LPS-induced myocardial injury. Then, we observed the influence of PI-2 on LPS-induced apoptosis of AC16 cells by flow cytometric analysis. As described in Figure 7(B), the percentage of early and late apoptotic cells was elevated by LPS administration, and PI-2 treatment could ameliorate this situation at a low concentration, which was comparable with 10 μM BMOV, the general PTP inhibitor.

Figure 7.

The effect of PI-2 on LPS-induced AC16 injury. A. The effect of PI-2 on LPS-induced AC16 inflammation; B. The effect of PI-2 on LPS-induced AC16 apoptosis; C. The effect of PI-2 on AC16 ROS level.

The ROS level displayed a vital role in cell survival, so we further investigated the ROS scavenging activity of PI-2 in the AC16 cell. As expected, the level of ROS was significantly decreased after PI-2 treatment in a concentration-dependent manner (Figure 7C), suggesting PI-2 suppressed LPS-induced myocardial apoptosis through ROS scavenging.

The mechanism of PI-2 on LPS-induced AC16 apoptosis

Mitochondria are the main organelle to produce ROS and display a vital role in cell survival⁴¹. As JNK regulates the activity of mitochondrial apoptosis-related proteins directly⁴². Previous studies have demonstrated that inhibition of PTP1B attenuated activation of JNK⁴³, so we speculated that PI-2 regulated LPS-induced AC16 apoptosis by maintaining mitochondrial potential through PTP1B/JNK pathway. To evaluate the function of PI-2 on AC16 mitochondria, first, we investigated the effect of PI-2 on the phosphorylation level of JNK. As shown in Figure 8(A), there is no obvious change for the expression of JNK after LPS and PI-2 treatment. However, the phosphorylation level of JNK was elevated by LPS and decreased after PI-2 treatment. These results further illustrated PI-2 was a preferable PTP1B inhibitor which could attenuate the activation of JNK. Then, the ΔΨm of mitochondrial membrane was tested by JC-1 staining (Figure 8B). In comparison with the Blank, LPS decreased the ΔΨm significantly. After the treatment of PI-2, the membrane potential of AC16 recovered, and a similar effect was seen with the treatment of BMOV. These results demonstrated that PI-2 was a potent PTP1B inhibitor that alleviated LPS-induced apoptosis by protecting mitochondrial membrane potential.

Figure 8.

The effect of PI-2 on mitochondrial membrane potential through PTP1B/JNK pathway. A. The effect of PI-2 on the phosphorylation level of JNK; B. The effect of PI-2 on mitochondrial membrane potential.

Besides, the study clarifies the anti-SICD activity of the novel inhibitor through PTP1B/JNK pathway. All the studies are carried out by LPS induced AC16 injury model, which could simulate the inflammatory factor releasing, oxidative stress processes induced by endotoxin release due to Gram-negative bacterial infection. However, the occurrence of SICD is more complicated, the model lacks real intercellular interactions and microenvironments in vivo, so further in vivo studies should be carried out and structure optimisation studies based on AI assisted or traditional drug design should be continued to find more potent PTP1B inhibitors.

Conclusions

PTP1B is a potential target for sepsis-induced cardiovascular dysfunction. Many efforts have been made to discover potent and selective PTP1B inhibitors; However, the effect of these inhibitors on SICM is unclear.

In this study, we established a novel virtual screening pipeline with a multi-conformational NBC model. Compound PI-2, with IC₅₀ = 4.2 μM was successfully screened by using this pipeline. Enzymatic and cellular assay illustrated that PI-2 displayed a moderate selectivity against both PTP1B and TCPTP, which was more than 4 times potent than ACP1. Biological assay demonstrated PI-2 was a potential compound for LPS-induced cardiovascular dysfunction treatment by improving mitochondrial membrane potential through inhibition of JNK phosphorylation. In summary, this study provides a potential PTP1B inhibitor and offers a new screening pipeline for discovering novel PTP1B inhibitors.

These findings serve as a pivotal stepping stone for the development of novel anti-SICD drugs. The anti-LPS induced cardiovascular dysfunction activity and substantial PTP1B inhibitory activity makes PI-2 as a potential candidate for disease involving PTP1B/JNK pathway. Future studies should delve into structure optimisation for bioactivity enhancement and nanomedicine and combination therapies should be considered to promote the clinical use of PI-2.”

Materials and methods

PTP1B crystal structure collection

After obtaining the distance matrix, the K-means algorithm (an unsupervised learning method) was applied to cluster the data and select representative PDB structures. To identify the optimal number of clusters (K), values ranging from 2 to 14 were evaluated. Clustering quality was assessed using the Silhouette Score, a metric that measures how well each data point fits within its assigned cluster compared to others. The score ranges from −1 to 1, with higher values indicating more well-defined and distinct clusters.

PTP1B inhibitor dataset generation

PTP1B inhibitors were collected from ChEMBL (https://www.ebi.ac.uk/chembl/) database. Given that the IC₅₀ was affected by assay condition, compounds with K_i value less than 10 μM were retained. The dataset was further refined by omitting allosteric inhibitors, and 501 inhibitors were kept. The decoys were generated by DUD-E (https://dude.docking.org/) with a ratio of 50 per 1 active compound. Finally, the database was divided into the training set and the test set with the ratio of 4:1, and the principal components of each set were analysed by Discovery Studio2019 (https://www.3ds.com/products-services/biovia/products/molecular-modeling-simulation/biovia-discovery-studio/).

Pharmacophore generation and assessment

Pharmacophores were generated according to the receptor-ligand complexes by Discovery Studio2019. At most 10 pharmacophores were generated for each complex. The number of features in each pharmacophore was set as 4 to 6 and validated by the training set. The mode of conformation generation for each ligand in the training set was “FAST” and fitted to the pharmacophore “Flexible”.

The first pharmacophore generated by each complex was further validated by the test set. After pharmacophore fitting, the FitValue of each ligand was recorded. For those ligands that failed to fit in the pharmacophore, the FitValue was set as −10. Collecting all the ligands and the property of “Real_activity” was added for all ligands. The value of “1” was defined to active PTP1B inhibitors, and “0” was set for decoys. 5 folds cross-validation was used to generate Bayesian Model according to the “FitValue” and “Real_activity” and the AUC was calculated for each pharmacophore.

General docking procedure

The docking prediction ability of four docking programs, including CDOCKER (Discovery Studio2019), Glide, Autodock (version 4.2.6, http://autodock.scripps.edu), and Autodock_Vina (version 1.1.2, https://vina.scripps.edu/) was assessed. The docking set was determined by the receptor-ligand complex. All the docking parameters for each program were set to default values unless otherwise stated.

CDOCKER: Proteins were prepared by the “Prepare Protein” module in Discovery Studio2019 with the default values, and ligands were prepared by the “Prepare Ligands” module with the pH range from 7.3 to 7.5. Molecular docking was conducted with the default parameters.

Glide: Proteins were prepared by the “Protein Preparation Wizard” module with the OPLS3 force field. Ligands were prepared by the “LigPrep” with pH = 7.4 ± 0.1. The “Glide Generation” module was used to generate grid files of each protein. Two docking modes, standard precision (SP) and extra precision (XP), were employed without any constraint.

Autodock: AutoDock_Tools_1.5.7 was used for protein and ligand preparation, including deleting water, adding hydrogen, assigning Gasteiger charges, and setting torsion count for ligands. The numbers of grid points in the xyz-direction were set according to the ligand in the crystal, and the grid spacing was 0.375 angstrom. The Lamarckian genetic algorithm (LGA) was used for conformation calculation.

Autodock_vina: The procedure was the same with Autodock, except the size of the search volume box was set as 20 Å × 20 Å × 20 Å.

Assessment of docking predictions

The docking reliability and accuracy of these docking programs for each crystal structure were assessed. For docking reliability, the native ligand in the complex was extracted and re-docked into the defined docking site. Then, the RMSD of the docked poses with the native conformation of each ligand was calculated by Discovery Studio 2019. The top-scored pose with the highest docking score and the best pose that had the lowest RMSD with the native conformation were evaluated.

For docking accuracy, all the ligands in both the training set and the test set were docked into each protein with the Glide programs. The Mann-Whitney U test was used to evaluate the difference in the distribution of docking scores between active inhibitors and decoys. The receiver operating characteristic (ROC) curve and area under the curve (AUC) of each protein were generated and calculated through the NBC model. For this model, the docking score of each ligand was set as an independent property, and the “Real_activity” was set as a dependent property.

Virtual screening

The final virtual screening pipeline consisted of Pharmacophore filtering produced by 1NWL, and Glide docking_SP with 1KAK, 1C86, 1NL9, and 1Q6M. After docking, the top 100 ligands were kept. Finally, the compounds were selected by comprehensively considering the FitValue and docking_score.

Assessment of PTP1B inhibitory activity

Inhibitory rate: Phosphatase inhibitory activities were performed in the buffer solution containing 50 mM Hepes and 1 mM DTT at pH of 7.0. All the compounds (10 μM as the final concentration) were incubated with the tested enzyme for 10 min at 37 °C (solution without test compounds was set as blank control). Para-nitrophenyl phosphate (pNPP) was added (the final concentration was 5 mM) and incubated. Finally, the reaction was stopped by adding 3 M NaOH solution. The absorbance was measured at 405 nm, and the inhibitory rate was calculated.

IC₅₀: The procedure was the same as the inhibitory rate assay except the concentrations were set as 10 μM, 5 μM, 2.5 μM, 1.25 μM, and 0.62 μM.

Cell culture

AC16 human cardiomyocytes (Procell, Cat. NO.: CL-0790, CVCL_4U18) were cultured in Dulbecco’s modified Eagle’s medium (DMEM) supplemented with 10% foetal bovine serum (Procell, Cat. NO.: 164220), 100 units/mL penicillin, and 100 μg/mL streptomycin at 37 °C with 5% CO₂ atmosphere incubator.

Cell apoptosis assay

Cell apoptosis was detected by the Annexin V-FITC/PI apoptosis detection kit. After 24 h of treatment by 50 μg/mL LPS and tested compounds, cells were collected and washed with PBS. Cells were re-suspended in 100 μL Binding Buffer. Then, 5 μL Annexin V-FITC and 10 μL PI staining solution were added and reacted for 15 min. Finally, the results were measured by BD FACSCelesta flow cytometry (New York, USA). The data was analysed by FlowJo_V10.

Reactive oxygen species (ROS) assay

The intracellular ROS was assessed using a ROS assay kit. After 24 h of treatment, cells were washed with PBS and stained with dichlorodihydrofluores-ceindiacetate (DCFH-DA) at 37 °C for 30 min in the dark. Then, samples were detected by FACSCelesta flow cytometry.

Mitochondrial membrane potential assay

The mitochondrial membrane potential was measured with JC-1 MMP assay kit. Cells were treated with the tested compounds for 24 h. and stained with JC-1 at 37 °C for 20 min. Then, the cells were washed with JC-1 staining buffer for twice. Finally, cell culture medium was added and evaluated through fluorescence microscope.

Western blot analysis

The primary antibody PTP1B (Rabbit, dilution: 1: 2000, RRID number: AB_10642566, Cat. NO.: 11334–1-AP) and TCPTP (Rabbit, dilution: 1: 500, RRID number: AB_2173235, Cat. NO.: 11214–1-AP) were purchased from Proteintech. JNK (Rabbit, dilution: 1:2000, Cat. NO.: YM8450) and p-JNK (Rabbit, dilution: 1:2000, Cat. NO.: YP0157) were purchased from Immunoway. The secondary antibody was purchased from Invitrogen (Goat, dilution: 1: 5000, RRID number: AB_1185567, Cat. NO.: Cat. NO.: 32460). The regular range protein markers were purchased from Yeasen (Cat. NO.: 20351ES72) for PTP1B and TCPTP and Thermo (Cat. NO.: 26616) for JNK and p-JNK.

After 24 h of treatment, proteins were extracted by RIPA cell lysis buffer (Beyotime, P0013B) with protease (MCE, HY-K0010) and phosphatase inhibitor cocktail (MCE, HY-K0010). After concentration determination by BCA assay, equal amounts of proteins were loaded for SDS-PAGE and transferred onto PVDF membrane. The membrane was further blocked with commercial blocking solution (Beyotime, China) and incubated with primary antibodies and secondary antibodies subsequently. Bands were visualised by ECL chemiluminescence kit (MCE, HY-K1005) and quantified by ImageJ software.

Statistical analysis

Statistical analysis was performed by GraphPad Prism 8 (GraphPad Software, Inc, La Jolla, CA, USA). All statistics are shown as mean ± SD with n = 3 unless otherwise specified. Analysis of variance (ANOVA) was used for comparisons of multiple groups. P < 0.05 was considered significant differences.

Funding Statement

The work was supported by the [Yangzhou City Social Development Science and Technology Research Project #1] under Grant [number YZ2023074]; [National Natural Science Foundation of China #2] under Grant [number 82304458]; and National Natural Science Foundation of China #3] under Grant [number 82500428].

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

Data will be made available on request.