Identification of potential inhibitors of 3‑mercaptopyruvate sulfurtransferase with a deep-learning based screening of natural products

Abstract

3-Mercaptopyruvate sulfurtransferase (3-MST), a key enzyme in sulfur metabolism, has recently gained attention as a potential anticancer target. However, reported 3-MST inhibitors remain limited, motivating the exploration of new scaffolds such as natural products. In this study, a library of 3744 natural products was virtually screened against human 3-MST using DiffDock (diffusion-model-based docking) followed by AutoDock Vina docking. Top-ranking candidates were further analyzed via molecular dynamics simulations and Molecular Mechanics Poisson–Boltzmann Surface Area binding free energy calculations. Methylophiopogonanone A (4), Daphnoretin (5), and L-asarinin (9) exhibited stable binding with favorable energetics, displaying binding free energies comparable to the reference ligand 7NC301. Binding mode analyses revealed that Methylophiopogonanone A primarily engaged in hydrophobic interactions, whereas Daphnoretin and L-asarinin formed extensive polar contacts, accompanied by higher desolvation penalties. In vitro cytotoxicity assays showed that Methylophiopogonanone A and L-asarinin reduced HCT116 cell viability by 40.3% and 26.3% at 25 µM, which is consistent with their inhibitory of 3-MST with IC₅₀ = 5.83 ± 0.69 µM and IC₅₀ = 19.11 ± 3.37 µM, respectively. These results suggested that the natural products identified in this study represent promising scaffolds for further optimization as potential 3-MST inhibitors. This work provides an AI-guided natural-product screening workflow for 3-MST and delivers prioritized inhibitor scaffolds for subsequent optimization and experimental validation.

Graphical abstract

Supplementary Information

The online version contains supplementary material available at 10.1007/s10822-026-00795-5.

Introduction

Hydrogen sulfide (H₂S) is increasingly recognized as a critical endogenous gasotransmitter, alongside nitric oxide (NO) and carbon monoxide (CO), regulating diverse physiological processes including vasodilation, neuromodulation, inflammation, and cellular metabolism [1–3]. In cancer biology, H₂S has emerged as a pro-oncogenic mediator, supporting tumor cell proliferation, migration, bioenergetics, and angiogenesis [4, 5]. Elevated H₂S levels have been documented in various tumor types, driven by overexpression of the enzymes responsible for its biosynthesis: cystathionine β-synthase (CBS), cystathionine γ-lyase (CSE), and 3-mercaptopyruvate sulfurtransferase (3-MST) [6–8]. 3-MST, which is localized in both the cytosol and mitochondria, has gained attention as a distinct and underexplored target for anticancer therapy, particularly in colorectal malignancies [6, 9].

3-MST catalyzes the conversion of 3-mercaptopyruvate (3-MP), derived from cysteine via aminotransferase activity, into pyruvate and a protein-bound persulfide intermediate [10–12]. This sulfane sulfur-containing species can subsequently release H₂S through physiological reductants such as thioredoxin or dihydrolipoic acid [9, 13, 14]. While CBS and CSE have been extensively studied in cancer, the role of 3-MST in tumor biology is less defined, despite evidence of its upregulation in colon, renal, and brain cancers [13, 15]. In colon cancer specifically, 3-MST expression correlates with increased H₂S production and tumor cell viability, positioning this enzyme as a potential target for small-molecule inhibition [16–18].

Recent advances have led to the identification of novel 3-MST inhibitors through structure-based drug design campaigns [19–21]. Structural and molecular dynamics studies have highlighted the significance of key interactions between inhibitor scaffolds and the active site cysteine (Cys248) of 3-MST, underscoring opportunities for structure-based drug design [22]. However, relatively few efforts have explored natural products as potential 3-MST inhibitors, despite their structural diversity and favorable pharmacological profiles.

In this context, this study integrates DiffDock, a deep-learning-based tool that enhances the screening process by improving ranking reliability, hit diversity, and hit quality [23]. We did not aim to benchmark DiffDock against traditional docking methods. Instead, DiffDock was used as a complementary pose-generation and triage step to improve conformational coverage and workflow efficiency, and the shortlisted candidates were subsequently evaluated by Vina refinement and MD/MM-PBSA to assess pose persistence and interaction stability. In contrast to traditional synthetic compound screening, natural products represent ideal candidates for 3-MST inhibitor screening due to their structural diversity and promising pharmacological properties [24]. Many of these compounds have already demonstrated excellent bioactivity and biocompatibility, making them a valuable source for developing new anticancer drugs. However, there has been limited reporting on inhibitors of 3-MST, which clearly does not meet these requirements. Using docking software alone for the screening of a large compound library would also result in relatively low efficiency. With the widespread adoption of AI and related technologies, the classical screening protocol can be partially or entirely replaced by AI tools.

In this study, we employed an integrative AI-based in silico and in vitro approach to identify and characterize natural products as inhibitors of 3-MST (Fig. 1). A library of natural compounds was screened against the 3-MST active site via an AI-based screening tool (DiffDock) [25] running on deep-learning and then a docking-based (AutoDock Vina) screening, followed by molecular dynamics simulations and free energy calculations [26]. Hit compounds were prioritized based on their predicted binding stability and interaction energetics, which were further validated by in vitro anticancer evaluation using HCT116 colon cancer cells. Our findings provide new insights into natural-product-based 3-MST inhibition and lay a foundation for further development of selective modulators targeting H₂S signaling in colorectal cancer.

Fig. 1.

Overview of the workflow of this study

Materials and methods

Protein and ligand preparation

To support the virtual screening and biological evaluation described below, a standardized computational-experimental workflow was applied. The crystal structure of human 3-MST (PDB ID: 5WQK) was prepared in UCSF Chimera by adding hydrogens, completing missing loops, and assigning Gasteiger charges. Specifically, missing heavy atoms and structural gaps/loops were repaired using PDBFixer (default settings). The structure was then processed using AmberTools (pdb4amber), and hydrogen atoms were added with protonation states assigned using the built-in reduce program under default settings (physiological conditions). Crystallographic waters and other non-protein entries were treated using default cleanup settings; no cofactors were present in the selected structure, and any alternate conformations were resolved by retaining the major-occupancy state. A natural-product library (TargetMol) containing 3744 compounds was processed using the Prepare Ligands module in Discovery Studio 2016 to generate optimized 3D conformations. Ligands were prepared using the default Prepare Ligands settings, including pH-based ionization (Change Ionization: True; Ionization Method: pH Based; Remove Charges: Protonate Acids, Deprotonate Bases), tautomer generation (Generate Tautomers: True; Enumerate What: AllTautomers; Maximum Number: 10), stereoisomer handling (Generate Isomers: True; EnumerateWhat: UnknownStereoatoms, UnknownStereobonds; AcceptInputStereoAs: True), 3D coordinate generation (Generate Coordinates: 3D; Rearrange Hydrogens: True), and duplicate removal (Duplicate Structures: Remove).

Virtual screening with DiffDock and AutoDock Vina

Virtual screening was performed using DiffDock-L v1.1 (GitHub commit d3791a8). For each ligand, 10 poses were generated; no confidence threshold was applied and the top-ranked pose (Rank 1) was selected for downstream analysis. The screening was executed on a local workstation with an NVIDIA GeForce RTX 4070 Ti (12 GB VRAM), and the random seed was set to 42 for reproducibility. DiffDock was used as an AI-assisted pose-generation and triage module to generate diverse plausible binding conformations and provide a model confidence score for pose selection, thereby reducing manual pose inspection prior to Vina refinement. The binding pocket was defined based on the co-crystallized inhibitor-binding site in PDB ID: 5WQK. The docking grid box was generated automatically to encompass the co-crystallized ligand and the surrounding binding cavity using the software’s default settings. Docking was performed using AutoDock Vina with default parameters.

Molecular dynamics simulations and free energy calculations

Top-ranked complexes were initially refined through 10-ns molecular dynamics simulations in AMBER, followed by MM-PBSA calculations using snapshots from the equilibrated trajectories. For the prioritized complexes, MM-PBSA (Poisson-Boltzmann model) was performed based on the 100 ns production trajectories; snapshots were uniformly extracted from the last 20 ns (e.g., every 100 ps, 200 frames). The dielectric constants were set to ε_in = 1.0 and ε_out = 80.0, and the salt concentration was set to 0.15 M to match the MD conditions. The complexes were solvated in explicit TIP3P water in a truncated octahedral periodic box with a 10 Å padding. Counterions (Na⁺/Cl⁻) were added to neutralize the system (and 0.15 M NaCl when specified). A 2 fs time step was used with SHAKE applied to constrain bonds involving hydrogen atoms. After energy minimization and standard NVT/NPT equilibration, a 10-ns production run was performed at 300 K and 1 bar using default AMBER thermostat/barostat settings. Based on the promising results from the 10-ns simulations, the top three compounds were selected for further analysis. To improve the equilibration of the protein–ligand complexes and obtain more reliable binding free energy estimates, 100-ns MD simulations were performed on these selected compounds. This extended simulation period allowed for a more thorough exploration of the binding dynamics, ensuring a more accurate and stable assessment of the binding affinity. MM-PBSA calculations were subsequently conducted on the snapshots from these longer trajectories to derive more precise binding free energy estimates.

Cytotoxicity evaluation in HCT116 colorectal cancer cells

For preliminary biological evaluation, the most promising natural products were assessed for their cytotoxic effects on HCT116 colorectal cancer cells using the CCK-8 assay. HCT116 cells were maintained in RPMI 1640 medium supplemented with 10% fetal bovine serum (FBS) and 1% penicillin–streptomycin at 37 °C in a 5% CO₂ atmosphere. Upon reaching approximately 80% confluence, cells were seeded into 96-well plates at a density of 2,000 cells per well and incubated for 24 h to allow cell attachment. The culture medium was then replaced with fresh medium containing the test compounds at concentrations ranging from 1 to 5,000 µM. After 72 h of exposure, cell viability was quantified by adding 10 µL of CCK-8 reagent (Biosharp) to each well and incubating for 1 h at 37 °C. Absorbance was measured at 450 nm using a microplate reader (BioTek) to assess cell viability. Detailed experimental conditions and parameters are provided in the Supporting Information.

Results

Analysis of the 3-mercaptopyruvate sulfurtransferase (3-MST)

3-Mercaptopyruvate sulfurtransferase (3-MST) is a key enzyme in sulfur metabolism. Its crystal structure, resolved at 2.15 Å, reveals a two-domain architecture characteristic of the rhodanese-like fold, with N-terminal (residues 1–138) and C-terminal (residues 165–285) domains connected by a linker (residues 139–164) that wraps around and stabilizes both domains. The active site is located in the interdomain cleft, where 3-mercaptopyruvate (3-MP) binds [20] (Fig. 2A). In complex with compound 7NC301, Cys248 is observed in a persulfurated state, engaging in an unusual interaction with the 4-pyrimidone-like aromatic ring of the ligand, characterized by a perpendicular orientation and a sulfur-π distance exceeding 3.45 Å. Additional stabilization arises from direct hydrogen bonding with R188 and S250, water-mediated interactions involving E195 and R197, and van der Waals contacts with multiple active-site residues (Fig. 2B).

Fig. 2.

Structure of human MST. A Crystal structure of MST displayed in a ribbon drawing (N-terminal domain shown in cyan; C-terminal domain shown in green; linker region between the two domains shown in magenta) (PDB: 4JGT). B Crystal structure of 3-MST in complex with inhibitor 7NC301 (green stick), and the key interactions were shown with hydrogen bonds in green, π-cation in orange and π-alkyl in pale magenta (PDB: 5WQK)

Virtual screening

A natural product library of 3,744 compounds was screened against the 3-MST protein using DiffDock, a GPU-accelerated deep learning-based docking tool. DiffDock frames docking as a diffusion-based generative modeling problem over ligand translations, rotations and torsions, sampling multiple candidate poses that are subsequently ranked by a learned confidence model [25, 27]. The confidence score has been reported to correlate with pose quality and can be used for selective filtering in downstream applications. The top 200 molecules ranked by DiffDock scores were further analyzed using AutoDock Vina. The crystal structure of 3-MST complexed with the known ligand 7NC301 (PDB: 5WQK) served as the receptor model. To validate the docking protocol, the co-crystallized ligand was redocked into the binding site. The resulting pose closely aligned with the experimental conformation, yielding a root-mean-square deviation (RMSD) of 0.973 Å and an docking score of − 9.3 kcal/mol. An RMSD below 2.0 Å supports the reliability of the docking protocol, and the redocked ligand reproduced the principal interaction pattern observed in the crystal structure, including contacts in the Cys248/Arg188/Ser250 region and the key hydrophobic packing within the validated pocket.

The top 10 hits, ranked by binding energy, are summarized in Table 1, with their chemical structures shown in Fig. 3. These compounds exhibited docking scores ranging from − 9.9 to − 8.7 kcal/mol, indicating favorable docking scores for ranking and prioritization of candidates. Among the top hits, (-)-Zuonin A (1) demonstrated the highest predicted affinity (-9.9 kcal/mol), exceeding that of the reference ligand 7NC301. Other notable compounds include Licarin B (− 9.2 kcal/mol), Chrysophanol 8-O-glucoside and Methylophiopogonanone A (both -8.9 kcal/mol), as well as Daphnoretin and Plantagoside (-8.8 kcal/mol). The comparable or superior docking scores of these candidates suggest their potential as scaffolds for further development of 3-MST inhibitors.

Table 1.

The scores predicted by DiffDock and AutoDock Vina docking scores and MM-PBSA calculation of top 10 hits

No	Compound	ID	DiffDock score	Binding energy (kcal/mol)
				AutoDock vina	MM-PBSA
1	(-)-Zuonin A	ZINC000257520722	9.996295	− 9.9	− 18.76 ± 3.81
2	Licarin B	ZINC000001587485	9.990738	− 9.2	− 19.56 ± 3.26
3	Chrysophanol 8-O-glucoside	ZINC00004098657	9.985181	− 8.9	− 8.47 ± 3.69
4	Methylophiopogonanone A	ZINC000013481899	9.975920	− 8.9	− 21.08 ± 3.10
5	Daphnoretin	ZINC000000689683	9.966658	− 8.8	− 20.24 ± 3.87
6	Plantagoside	ZINC000257543460	9.957396	− 8.8	− 0.27 ± 5.05
7	Mebendazole	ZINC00000121541	9.948134	− 8.7	− 15.32 ± 3.91
8	Anisindione	ZINC0010015486	9.938873	− 8.7	− 13.63 ± 3.3
9	L-asarinin	ZINC00001668768	9.925906	− 8.7	− 20.41 ± 2.73
10	6'-O-β-D-Glucosylgentiopicroside	ZINC00085799038	9.912940	− 8.6	− 1.03 ± 0.37
11	7NC301	\		− 9.3	− 24.95 ± 2.75

Fig. 3.

Chemical Structures of the top 10 hit Compounds

MD simulation and MM-PBSA analysis

To assess the dynamic stability and binding behavior of the top 10 hit compounds within the 3-MST active site, 10 ns molecular dynamics (MD) simulations were performed for each protein–ligand complex. While molecular docking provides useful insights into potential binding conformations, its limitations—particularly the rigid treatment of the receptor—can result in suboptimal binding poses. In contrast, MD simulations allow both ligand and protein flexibility under near-physiological conditions, offering a more accurate depiction of complex behavior over time.

MD trajectory analysis focused on root-mean-square deviation (RMSD) values to evaluate structural stability and convergence. The backbone RMSD of the protein in all complexes stabilized within 1.0–2.5 Å, indicating minimal global conformational drift during the simulation. Ligand RMSDs remained within the 0.5–2.0 Å range for most compounds, except for 6'-O-β-D-Glucosylgentiopicroside (10), which exhibited slightly higher fluctuations (average RMSD ~ 3.0 Å), suggesting less stable binding or greater conformational flexibility. Figure 4 illustrates the RMSD profiles of selected complexes. Notably, (-)-Zuonin A (1), Licarin B (2), Methylophiopogonanone A (4), Anisindione (8), and L-asarinin (9) displayed minimal ligand RMSD values and rapid equilibration, supporting the reliability of their initial docking poses and conformational stability throughout the simulation. All systems achieved equilibrium within the first 3 ns and remained stable for the duration of the simulations, with backbone fluctuations consistently below 2.0 Å.

Fig. 4.

Root-mean-square deviation (RMSD) of protein backbone (A) and ligand (B) with respect to the initial structure from 10 ns MD simulation

These findings suggest that the selected natural compounds are not only capable of occupying the substrate-binding site of 3-MST but also form stable complexes under dynamic conditions, supporting their stable binding behavior in the MD simulations for further development.

MM-PBSA binding free energy.

The molecular mechanics Poisson-Boltzmann surface area (MM-PBSA) technique was used to calculate the binding free energy of protein–ligand complexes from the trajectory of the last 4 ns of MD simulation (Table 1). Subsequently, the energy contribution of individual residues was evaluated. The binding energies were calculated and compared with native ligand 7NC301 as a positive control. 7NC301 yielded the most favorable MM-PBSA estimate and was used as the reference (− 24.95 ± 2.75 kcal/mol). The binding free energy of the top 10 compounds ranged from − 0.27 ± 5.05 to − 21.08 ± 3.10 kcal/mol, which were less favorable than the co-crystallized/native ligand; however, they still represented top-ranked, pose-plausible candidates within the validated pocket and were therefore prioritized for further MD and experimental evaluation as potential 3-MST inhibitors.

Several glycosylated compounds, including Chrysophanol 8-O-glucoside (3), Plantagoside (6), and 6′-O-β-D-Glucosylgentiopicroside (10), exhibited relatively weak MM-PBSA binding free energies compared with the prioritized hits (Table 1). This observation may be attributed to the increased polarity and conformational flexibility introduced by the sugar moieties, which can lead to higher polar solvation/desolvation penalties and consequently reduce the net binding free energy in MM-PBSA estimation. Consistent with this interpretation, 6′-O-β-D-Glucosylgentiopicroside (10) also showed larger ligand RMSD fluctuations during MD simulations, suggesting less stable binding behavior. Therefore, these glycosylated candidates were deprioritized for subsequent extended MD validation and experimental prioritization.

Methylophiopogonanone A (4) exhibited the most favorable MM-PBSA estimate among the screened hits of − 21.08 ± 3.10 kcal/mol, followed by Daphnoretin (5) and L-asarinin (9), with binding energy below − 20 kcal/mol. The interactions of these molecules with a binding pocket were primarily due to hydrophobic interactions. Moreover, (-)-Zuonin A (1) and Licarin B (2) also provided excellent binding affinity with binding energies of − 18.76 ± 3.81 and − 19.56 ± 3.26 kcal/mol, respectively. Our studies revealed that molecular dynamics and MM-PBSA calculation results supported the feasibility of these compounds to prioritize compounds 4, 5 and 9 for in vitro experiments.

Analysis of molecular dynamics and binding free energy of candidate compounds

Based on the initial 10-ns screening simulations, the top three compounds, Methylophiopogonanone A (4), Daphnoretin (5), and L-asarinin (9), were selected for extended evaluation together with the reference ligand 7NC301. To address concerns regarding sampling length and to better assess pose persistence, 100-ns MD simulations were performed for these complexes. As shown in Fig. 5A, the protein Cα RMSD profiles remain within a stable range after equilibration, indicating that the overall protein structures are well maintained across the trajectories. Residue-level flexibility was further examined by RMSF analysis (Fig. 5B). All complexes display broadly similar fluctuation patterns to the 7NC301-bound form, with larger fluctuations mainly located in solvent-exposed loop regions, while residues around the active site remain comparatively stable. Notably, the Daphnoretin (5) complex shows moderately increased fluctuations in several regions, suggesting enhanced local dynamics relative to compounds 4 and 9; however, overall complex integrity is preserved throughout the 100-ns simulations. In addition, we assessed global compactness by radius of gyration (Rg), which showed no apparent drift after equilibration, indicating that the overall fold remained stable during the simulations. We further analyzed protein–ligand hydrogen bonds (default criteria), and the key polar contacts between the ligands and the binding pocket were maintained during the equilibrated trajectories, supporting stable binding behavior. Collectively, the extended MD results support the stability of the binding poses for the prioritized hits and provide a more robust basis for subsequent trend-level energetic assessment.

Fig. 5.

Stability assessment from 100-ns MD simulations of 3-MST complexes with the top three hits and the reference ligand 7NC301. A Protein Cα RMSD as a function of simulation time. B Residue-wise RMSF calculated over the 100-ns trajectories

Binding energetics were estimated using the MM-PBSA method, and the results are summarized in Table 2. Overall, Methylophiopogonanone A (4) exhibits the most favorable net binding energy among the prioritized hits, followed by Daphnoretin (5) and L-asarinin (9). For compound 4, the favorable profile is mainly driven by van der Waals interactions together with a relatively smaller polar solvation penalty. In contrast, Daphnoretin (5) shows stronger electrostatic contributions, but these are substantially offset by a larger polar desolvation cost. L-asarinin (9) displays weaker electrostatic contributions and greater variability across snapshots, which is consistent with its less favorable net energy and larger uncertainty. We note that several MM-PBSA components, particularly the electrostatic term, show relatively large standard deviations, reflecting the sensitivity of Coulombic interactions to instantaneous geometry and solvent exposure; importantly, electrostatic contributions are often partly compensated by the polar solvation term. Moreover, entropic contributions were not explicitly included in the present MM-PBSA analysis. Therefore, the MM-PBSA results are interpreted primarily for qualitative ranking and trend comparison rather than as absolute binding affinities.

Table 2.

Energy Contribution of the Various Components to the Total Binding Free Energies of the Simulated Systems

Compound	(kcal/mol)	(kcal/mol)	(kcal/mol)	(kcal/mol)	(kcal/mol)
Methylophiopogonanone A (4)	− 29.71 ± 2.65	− 21.50 ± 6.81	+ 34.64 ± 4.46	− 4.28 ± 0.25	− 20.85 ± 3.26
Daphnoretin (5)	− 21.05 ± 2.13	− 21.05 ± 2.13	− 21.05 ± 2.13	− 21.05 ± 2.13	− 17.44 ± 2.91
L-asarinin (9)	− 26.65 ± 5.88	− 26.65 ± 5.88	− 26.65 ± 5.88	− 26.65 ± 5.88	− 15.55 ± 6.09
7NC301	− 32.39 ± 3.31	− 32.39 ± 3.31	− 32.39 ± 3.31	− 32.39 ± 3.31	− 22.41 ± 3.35

The protein–ligand interactions of the three hit compounds were depicted in Fig. 6. Methylophiopogonanone A (4) primarily engaged in hydrophobic and π-π stacking interactions, notably with Leu38, Val251, Trp36, and His74. Additional contacts with Tyr108 and Arg112 likely contributed through dispersion and electrostatic effects. The absence of classical polar interactions was consistent with its favorable van der Waals energy and moderate polar solvation penalty, indicating a hydrophobic-driven binding mode. In contrast, Daphnoretin (5) formed an extensive polar interaction network, including hydrogen bonds with Arg184 and Lys40, π-π stacking with Pro196, and ionic interactions with Arg188 and Glu199, correlating with substantial electrostatic contributions. However, a high polar solvation penalty offset its binding affinity. L-asarinin (9) exhibited both aromatic (Trp36, Tyr108) and π-cation (Arg197) interactions, yielding balanced van der Waals and electrostatic contributions, though attenuated by solvation effects. Collectively, Methylophiopogonanone A’s hydrophobic binding conferred thermodynamic favorability, whereas Daphnoretin and L-asarinin incurred greater desolvation penalties despite stable polar contacts.

Fig. 6.

Calculated binding modes of Methylophiopogonanone A (4), Daphnoretin (5) and L-asarinin (9) in complex with 3-MST

To link these interaction patterns to known catalytic features of 3-MST, we overlaid the MD-representative poses of compounds 4, 5, and 9 with the co-crystallized inhibitor 7NC301 in the interdomain catalytic cleft (Figure S1). All three hits occupy the active-site pocket in close proximity to the catalytic residue Cys248 (persulfidated in the 5WQK structure), consistent with a competitive active-site binding mode. The overlays indicate that the hits share the same cleft region as 7NC301 while adopting distinct anchoring patterns at the cleft entrance, including contacts in the Arg188/Ser250 region. Notably, none of the natural-product scaffolds contains an obvious electrophilic warhead for covalent capture of Cys248, supporting a predominantly non-covalent inhibition mechanism.

Evaluation of cytotoxicity against colon cancer cells

The cytotoxic potential of three hit compounds was assessed against the human colon cancer cell line HCT116 at a concentration of 25 µM using the CCK-8 assay (Fig. 7A). Among the three candidates, Daphnoretin (5) displayed the strongest inhibitory effect on HCT116 cell growth (61.67%), consistent with previously reported anticancer activity of this scaffold in colorectal cancer-related models. Methylophiopogonanone A (4) also showed appreciable cytotoxicity, reducing cell viability by 40.34%, whereas L-asarinin (9) induced a comparatively modest decrease (26.29%). These results suggest that the three natural products exert measurable antiproliferative effects in colon cancer cells, with compound 5 exhibiting the most pronounced cellular phenotype under the screening condition.

Fig. 7.

Cellular cytotoxicity profiling, biochemical 3-MST inhibition, and validation of 3-MST overexpression. A Single-dose screening of growth inhibition in HCT116 cells treated with compounds 4, 5, and 9 (25 µM) and 5-FU (10 µM) using the CCK-8 assay. B Concentration–response curves for biochemical inhibition of 3-MST by compounds 4, 5, and 9. C Dose–response growth inhibition curves of compounds 4, 5, and 9 in HCT116 cells (CCK-8). D–F Relative 3-MST protein expression in HCT116 cells under control, compound treatment, 3-MST overexpression (3-MST OE), and OE plus compound treatment conditions

To determine whether the observed cellular effects are associated with direct inhibition of the proposed target, we next evaluated their inhibitory activity against 3-MST in a biochemical assay (Fig. 7B). All three compounds inhibited 3-MST with micromolar potency. Notably, compound 4 exhibited stronger biochemical inhibition than compound 9 (IC₅₀ = 5.83 ± 0.69 µM vs 19.11 ± 3.37 µM), supporting compound 4 as the more effective enzymatic inhibitor within this pair. Daphnoretin (5) showed an intermediate level of 3-MST inhibition (IC₅₀ = 9.59 ± 1.22 µM), which aligns with the trend suggested by our MM-PBSA calculations. Importantly, however, the cellular antiproliferative activity did not perfectly track the biochemical rank order: despite slightly weaker 3-MST inhibition than compound 4, compound 5 produced stronger growth inhibition in HCT116 cells. This divergence suggests that the cellular phenotype may not be solely determined by 3-MST enzymatic inhibition and could reflect additional mechanisms. In line with this interpretation, Daphnoretin has been reported to exhibit agonistic activity toward protein kinase C (PKC), which may contribute to enhanced suppression of cancer cell proliferation through multi-pathway modulation. Taken together, these results indicate that the three hits inhibit 3-MST enzymatic activity in vitro, and their cellular phenotypes are consistent with 3-MST pathway perturbation, while compound 5 may achieve superior cellular efficacy through concurrent engagement of additional growth-regulatory signaling.

To further quantify cellular potency beyond the single-dose screen, we performed dose–response CCK-8 assays (Fig. 7C). The resulting concentration–response curves confirmed dose-dependent growth inhibition for all three compounds and enabled estimation of cellular IC₅₀ values. Based on cellular IC₅₀, compound 5 showed the highest potency, followed by compound 4 and then compound 9 in HCT116 cells. Thus, the dose–response profiling corroborated the single-dose screening outcome and provided a more robust quantitative basis for comparing cellular activities among the hits.

In addition, to strengthen cellular target-related validation, we performed a 3-MST overexpression experiment and quantified 3-MST protein levels (Fig. 7D–F). 3-MST overexpression significantly increased 3-MST protein abundance, confirming successful upregulation of the target at the protein level. Notably, treatment with compounds 4, 5, or 9 did not markedly change 3-MST protein expression either under basal conditions or in the overexpression background. These results indicate that the observed cellular effects are unlikely to arise from reduced 3-MST protein abundance and are more consistent with functional inhibition of 3-MST activity.

Overall, these results indicate that the three retained natural products exhibit measurable cytotoxicity against HCT116 cells and directly inhibit 3-MST in biochemical assays, supporting their further evaluation as 3-MST-related chemical probes and/or lead candidates for colorectal cancer-oriented optimization. Among them, Methylophiopogonanone A (4) represents the most potent biochemical 3-MST inhibitor in this set, whereas Daphnoretin (5) demonstrates superior cellular antiproliferative activity, potentially benefiting from additional target engagement.

Discussion

In this study, we identified three natural-product scaffolds as 3-MST inhibitory candidates through an integrated DiffDock-Vina-MD workflow and validated their inhibitory activity in biochemical assays with complementary cellular profiling. We discuss our findings in the context of prior 3-MST inhibitor studies, highlight points of agreement and difference, and outline key limitations. Comparison with previous 3-MST inhibitor studies. Previous studies have reported selective 3-MST inhibitors identified through high-throughput screening (HTS) and mechanistic characterization, highlighting that 3-MST inhibition can involve interactions with the active-site cysteine/persulfidated states and that specific chemotypes can achieve selectivity over other H₂S/sulfane sulfur-producing enzymes [28]. In addition, pharmacological studies using reported 3-MST inhibitors (e.g., I3MT-3, DPHE) illustrate both the utility of chemical inhibition and the importance of carefully interpreting cellular phenotypes, as off-target effects or alternative targets may contribute in some contexts [29]. In this context, our work differs by leveraging a diffusion-model docking framework (DiffDock) as a pose-generation/triage step followed by docking refinement and MD-based stability/energetics evaluation to prioritize natural-product scaffolds for further study.

Natural products are a valuable source for drug discovery; therefore, we briefly summarize the botanical origins and enriched medicinal parts of the prioritized hits. The three top candidates, 3-MST hits Methylophiopogonanone A (4), Daphnoretin (5) and L-asarinin (9), are components of traditional medicinal plants and are rich in specific medicinal parts. Methylophiopogonanone A (4) is reported from the dried tuberous roots of Ophiopogon japonicus. Recent pharmacological studies have reported its potential activities in immune regulation, antioxidation, and antitumor research [30–32]. Daphnoretin (5) is reported from the flower buds and root barks of Daphne genkwa. The isolation of toxic components (e.g., genkwanin) from Daphne genkwa’s active constituents has long been a key focus of research. As a non-toxic active component, Daphnoretin has been identified in recent years to possess significant antitumor and anti-inflammatory activities [33–35]. L-asarinin (9) is reported from Asarum spp. medicinal materials (whole herbs). Recent studies have demonstrated that its components (including L-asarinin) exhibit definite activities in inhibiting tumor cell proliferation and anti-platelet aggregation [36–39].

Building on the initial in silico prioritization, we further evaluated the top complexes using 100-ns MD simulations. Across these trajectories, the retained ligands remained stably accommodated in the 3-MST active site, supporting the robustness of the predicted binding poses. The dominant interaction patterns were consistent with a hydrophobic anchoring mode complemented by a limited number of polar contacts, which is favorable for maintaining binding while avoiding excessive desolvation penalties. In line with this interaction balance, MM-PBSA analyses suggested that compound 4 benefits from strong van der Waals contributions and a relatively smaller polar solvation cost, consistent with its stable binding behavior during MD.

Importantly, the computational findings were complemented by biochemical and cellular validation. In a biochemical 3-MST assay, all three prioritized compounds inhibited 3-MST with micromolar potency, with compound 4 showing the strongest enzymatic inhibition, followed by compound 5 and compound 9. In HCT116 colon cancer cells, single-dose screening at 25 µM revealed measurable growth inhibition for all three compounds, and subsequent CCK-8 dose–response profiling confirmed concentration-dependent antiproliferative activity. The cellular potency ranked compound 5 as the most active, followed by compound 4 and compound 9, thereby providing a quantitative cellular context beyond single-dose screening. Notably, biochemical potency did not fully translate into the same rank order in cells. While compound 4 exhibited the strongest 3-MST enzymatic inhibition, compound 5 showed superior antiproliferative activity in HCT116 cells. Such divergence is not uncommon and can arise from differences in cellular permeability, intracellular exposure, or parallel pathway modulation. In this context, Daphnoretin has been reported to modulate PKC signaling, which may augment its cellular antiproliferative effect beyond 3-MST inhibition alone. This multi-pathway contribution could rationalize why compound 5 displays stronger cellular efficacy despite slightly weaker enzymatic inhibition than compound 4. At the protein level, 3-MST overexpression increased 3-MST abundance, and compounds 4, 5, and 9 did not markedly affect 3-MST levels under basal or overexpression conditions, consistent with a functional inhibition mode rather than altered protein abundance.

Based on the binding-mode analysis together with the MD and MM-PBSA trends, we propose several binding-mode-guided design considerations rather than definitive SAR rules. First, the retained hits suggest that maintaining a hydrophobic anchoring motif is beneficial for engaging the predominantly hydrophobic environment of the 3-MST active site and for strengthening van der Waals interactions, consistent with the favorable nonpolar contributions observed for compound 4. Second, our MM-PBSA component analysis indicates that changes in polar solvation can substantially influence the overall binding energetics; therefore, the number and placement of polar functionalities should be balanced to preserve key polar contacts while avoiding excessive desolvation penalties. For example, for the Daphnoretin scaffold, selective masking of non-essential hydroxyl groups may be a practical polarity-tuning strategy and is worth testing in follow-up in silico analog docking and/or experimental studies.

This study screened anti-colon cancer natural products using 3-MST as the target, with the core basis being the specific high expression of 3-MST in colon cancer and its key regulatory role in H₂S synthesis. Existing studies have confirmed that in colon cancer cell lines (e.g., HCT116 and SW480) and colon cancer tissue samples, the mRNA and protein expression levels of 3-MST are 2–threefold higher than those in normal colonic epithelial cells-and its expression level is positively correlated with tumor stage (expression in patients with stage III-IV is significantly higher than that in stage I-II) [39]. This high expression directly leads to increased H₂S concentration in the tumor microenvironment (up to 1.5–2.0 fold that of normal tissues). As an oncogenic gas signaling molecule, H₂S can promote tumor cell proliferation and inhibit apoptosis by activating the PI3K/Akt/mTOR pathway; simultaneously, it can induce angiogenesis by upregulating VEGF expression, thereby providing nutritional support for tumor growth. In vitro assays in this study support the feasibility of targeting 3-MST in a colorectal cancer context and support the potential of natural products to exert anti-colon cancer effects, at least in part, through 3-MST inhibition.

Collectively, the MD-based stability assessment together with the biochemical and cellular assays provide a coherent activity profile, while the biochemical-cellular rank divergence highlights the role of cellular exposure and potential polypharmacology in interpreting natural product phenotypes.

In conclusion, this study identified Methylophiopogonanone A (4), Daphnoretin (5), and L-asarinin (9) as putative 3-MST inhibitors through an integrated in silico workflow including deep-learning-based screening, docking, 100-ns MD simulations, and MM-PBSA analyses. All three hits exhibited stable binding at the 3-MST active site and inhibited 3-MST enzymatic activity with micromolar potency, and their antiproliferative effects in HCT116 cells were supported by CCK-8 profiling. Among them, compound 4 emerges as the most promising starting point for developing 3-MST-focused inhibitors, as it combines the strongest biochemical inhibition with MD/MM-PBSA-supported binding stability and a favorable hydrophobic interaction profile with relatively limited polar solvation penalties, while retaining measurable cellular activity.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Changkang Wang proposed the concept, designed the methodology, and wrote the original draft. Xiao Chen completed the cell experiments and the WB experiments.Yu Yin was responsible for data curation and validation. Huimin Ding contributed to the investigation and visualization. Zhensuo Sha performed the formal analysis. Yifan Zhu carried out part of the computational work (software) and assisted with review and editing. Xin Xue supervised the study and acquired funding. Dongliang Zhang provided overall supervision and project administration. All authors have read and agreed to the published version of the manuscript.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.