Based on artificial intelligence-assisted generation and in-depth in-silico evaluation of potential inhibitors targeting Stearoyl-CoA desaturase 1 (SCD-1)

Abstract

Stearoyl-CoA desaturase 1 (SCD-1), a key rate-limiting enzyme in lipid metabolism, catalyzes the conversion of saturated fatty acids to monounsaturated fatty acids. It regulates membrane fluidity and modulates ferroptosis by influencing cellular antioxidant systems. Frequently overexpressed in malignancies, SCD-1 promotes tumor progression by driving metabolic reprogramming, activating proliferative pathways such as PI3K-AKT-mTOR, and suppressing ferroptosis, thereby enhancing tumor survival, invasion, and chemotherapy resistance. Clinical evidence links elevated SCD-1 expression with advanced TNM stage, lymph node metastasis, and poor prognosis in various cancers. Although SCD-1 inhibitors exhibit promising antitumor efficacy in preclinical studies, their clinical translation is hampered by off-target toxicity and drug resistance. Developing tumor-specific SCD-1 inhibitors is therefore crucial. This study employed an artificial intelligence (AI)-aided drug design strategy using the MolProphet platform to generate 399 novel compounds. Virtual screening via molecular docking, molecular dynamics simulations (200 ns), and binding free energy calculations (MM-PBSA) identified four candidates (Cpd1–Cpd4) with stable binding modes, low RMSD (0.2–0.5 nm), and high binding affinity (MM-PBSA < − 40 kcal/mol). These results demonstrate the strong potential of the selected compounds as SCD-1 inhibitors, offering promising leads for targeted anticancer drug development and a robust AI-driven design framework.

Introduction

Cancer progression is a highly dynamic process maintained by finely regulated metabolic pathways. Stearoyl-CoA desaturase 1 (SCD1), a key fatty acid desaturase that catalyzes the conversion of saturated fatty acids to monounsaturated fatty acids, serves as a central regulator of the fatty acid metabolic network^1–3. Aberrantly high SCD1 expression is closely associated with poor prognosis in various malignancies, including liver, breast, and ovarian cancers, conferring a survival advantage to cancer cells by modulating susceptibility to ferroptosis^1,4.

Under physiological conditions, SCD1 maintains metabolic homeostasis and tissue function by responding to endogenous stimuli mediated by its catalytic products, and participates in diverse processes including energy metabolism, embryonic development, autophagy, tumorigenesis, and inflammation⁵. In malignancies, aberrant transcriptional and epigenetic activation of SCD1 promotes intracellular lipid accumulation and drives malignant progression via signaling pathways such as AMPK/ACC, SIRT1/PGC1α, and NcDase/Wnt⁶.

Uncontrolled proliferation, a hallmark of cancer, requires substantial energy and continuous biosynthesis of macromolecules such as nucleic acids, lipids, and amino acids⁷. To support this demand, cancer cells undergo significant glycolipid metabolic reprogramming^8,9. SCD1, a rate-limiting enzyme in lipid metabolism, is frequently overexpressed in cancers, and its expression level correlates with invasive, metastatic, and drug-resistant phenotypes, establishing it as a critical regulatory molecule and a promising therapeutic target in oncogenesis¹⁰.

Ferroptosis is a regulated cell death driven by intracellular oxidative imbalance, constitutively regulated by glutathione peroxidase 4 (GPX4)¹¹. Recent studies indicate that SCD1 expression confers resistance to ferroptosis; in ovarian cancer, high SCD1 levels protect tumor cells from ferroptotic death, thereby promoting survival¹².

To date, a variety of SCD1 inhibitors have been developed and evaluated in pre-clinical and clinical stages, including small-molecule synthetic inhibitors (e.g., CAY10566) and natural product-derived inhibitors (e.g., Agrimonolide). CAY10566, a classic selective SCD1 inhibitor, exhibits potent inhibitory activity in pre-clinical models of breast and prostate cancer but lacks tumor tissue specificity; Natural inhibitors represented by Agrimonolide have low toxicity but limited potency and poor water solubility. Despite these advances, the clinical application of SCD1 inhibitors still faces several intractable challenges. (1) Off-target toxicity: Structural homology among desaturase family members (SCD1) leads to non-specific binding, disrupting normal fatty acid metabolism in normal tissues and causing adverse effects; (2) Compensatory resistance mechanisms: Tumor cells upregulate other desaturases or increase uptake of exogenous monounsaturated fatty acids (MUFAs) to bypass SCD1 inhibition, leading to drug resistance; (3) Mechanism-based toxicities: SCD1’s physiological expression in normal tissues (e.g., skin, eyes) results in on-target off-tumor toxicity, as suggested by adverse effects observed in animal models; (4) Pharmacokinetic defects: Most small-molecule inhibitors have poor oral bioavailability, short half-life, or limited tumor tissue penetration, reducing in vivo therapeutic efficacy. These limitations underscore the urgent need for novel SCD1 inhibitors with tumor specificity, high potency, low off-target effects, and favorable pharmacokinetic properties.

Despite its clear potential as an anti-cancer target, the clinical application of SCD1 inhibitors faces several challenges: (1) Structural homology among desaturase family members leads to off-target effects and disruption of normal fatty acid metabolism; (2) Compensatory resistance mechanisms, such as upregulation of other desaturases or increased uptake of exogenous monounsaturated fatty acids; (3) Mechanism-based toxicities due to SCD1’s physiological expression in normal tissues (e.g., skin, eyes), as suggested by adverse effects observed in animal models; (4) Potential for SCD1 target mutations or tumor cell adaptation through alternative signaling pathways during long-term treatment^13–16. These limitations underscore the urgent need for novel SCD1 inhibitors.

Artificial intelligence (AI) has emerged as a powerful tool for accelerating drug discovery, enabling efficient prediction and generation of potential drug molecules¹⁷. Structure-based and ligand-based drug design methods, along with pharmacophore analysis, significantly shorten development timelines, while machine learning techniques further validate the drug potential of AI-generated molecules^18–20. Integrating AI with computer-aided drug design (CADD) allows for the rational design of inhibitors targeting the unique structural features of SCD1’s active pocket, which can effectively improve target specificity and avoid cross-binding with other desaturase family members^21,22. This integrated strategy facilitates the development of safe and effective targeted therapies, potentially overcoming the clinical limitations associated with current inhibitors.

Therefore, this study aims to utilize the AI-assisted tool MolProphet to design and generate novel, safe SCD1-targeting inhibitor molecules²³. As a key enzyme regulating lipid metabolism in multiple malignancies, SCD1 inhibition represents a promising therapeutic strategy²⁴. For the selection of reference ligands in AI-driven molecular generation, we prioritized two scaffold classes—benzoylpiperidine and spiro[chroman-2,4’-piperidin]-4(3 H)-one—over other known SCD1 inhibitors (including clinical-stage compounds) based on three core considerations: (1) Structural compatibility with SCD1’s active pocket: These scaffolds contain a piperidine core that fits the hydrophobic cavity of SCD1 (PDB: 4ZYO) and functional groups (e.g., carbonyl, aromatic rings) capable of forming key hydrogen bonds and π-π interactions with residues such as TRP153 and ASN148, as validated in prior SCD1 inhibitor studies; (2) Balanced drug-likeness and structural flexibility: Compared with clinical-stage inhibitors (e.g., ARN-509, which has a rigid steroid skeleton) or natural products (e.g., Agrimonolide, with poor aqueous solubility), these two scaffolds have moderate molecular weight (300–400 Da), optimal Log P (2–3), and adjustable substitution sites (e.g., para-positions of the benzoyl group, 3-position of the chroman ring), providing sufficient flexibility for AI-driven structural optimization; (3) Low off-target risk potential: Unlike pan-desaturase inhibitors (e.g., CAY10566), benzoylpiperidine and spiro[chroman-2,4’-piperidin]-4(3 H)-one derivatives have been reported to exhibit higher SCD1 selectivity in preclinical studies, aligning with our goal of developing tumor-specific inhibitors. Based on the structural characteristics of existing SCD1 inhibitors and their clinical limitations, the generated molecules were designed to optimize three core properties: (1) Target specificity: matching the unique amino acid residues of SCD1’s active pocket to avoid off-target binding to other desaturases; (2) Pharmacokinetic properties: complying with Lipinski’s RO5 and optimizing ADMET properties to improve oral bioavailability and tumor penetration; (3) Anti-resistance potential: targeting the key functional regions of SCD1 to reduce the possibility of compensatory resistance. The inhibitory potential of the generated molecules will be systematically evaluated using CADD techniques, including molecular docking and molecular dynamics (MD) simulations, complemented by machine learning validation^25,26. Binding free energies will be calculated using MM-PBSA method to provide robust energetic evidence for high inhibitory activity^27–29. This research aims to provide novel SCD1 inhibitors with clinical translation potential and establish a feasible AI-assisted paradigm for targeted anti-cancer drug development³⁰. Figure 1 illustrates our workflow and the related results.

Fig. 1.

Workflow for the identification of SCD1 inhibitors using artificial intelligence and virtual screening. The schematic outlines the integrated process of molecular generation, machine learning, absorption, distribution, metabolism, excretion, and toxicity (ADMET) profiling, and molecular dynamics (MD) simulations.

Results and discussion

Generation of AI-based novel inhibitors

MolProphet is not limited to the chemical space of the input reference ligands but supports extensive scaffold hopping through two core mechanisms: (1) Multi-scale structural generation: The platform first extracts pharmacophore features (e.g., hydrogen bond donors/acceptors, hydrophobic groups, aromatic rings) from reference ligands that are critical for SCD1 binding, then decouples these features from the original scaffolds and recombines them with diverse heterocyclic cores (e.g., pyridine, indole, pyrazole) or aliphatic linkers (e.g., ethylene glycol, piperazine) to generate novel skeletons; (2) Reinforcement learning-guided exploration: The model is trained on a large-scale database of drug-like molecules (including 1.2 million compounds from ChEMBL, PubChem, and FDA-approved drugs), enabling it to explore chemical spaces beyond the input scaffolds while maintaining drug-likeness. In our study, although benzoylpiperidine and spiro[chroman-2,4’-piperidin]-4(3 H)-one were used as references, the AI-generated 399 compounds included 12 distinct scaffold types (e.g., indolylpiperidine, chromenopyridine, benzodiazepine), confirming the platform’s ability for scaffold hopping.

In our preliminary research, our team utilized the MolProphet AI platform with the known SCD1 inhibitors benzoylpiperidine and spiro[chroman-2,4’-piperidin]-4(3 H)-one as reference ligands^31–34, generating 399 potential novel SCD1 inhibitors based on their structural and physicochemical characteristics. Subsequent activity assessments confirmed that these novel inhibitors effectively suppress SCD1 biological function, thereby validating the efficacy of the MolProphet platform in novel inhibitor development.

As shown in Figs. 2 and 3, the distribution of key predicted physicochemical properties and activity parameters for the 399 generated compounds is presented, including the negative logarithm of the half-maximal inhibitory/effective concentration (pIC₅₀/pEC₅₀), number of hydrogen bond acceptors, number of hydrogen bond donors, partition coefficient (Log P), total polar surface area, and number of rotatable bonds. These parameters serve as core indicators for evaluating compound drug-likeness and target binding activity, and their interaction profiles are considered crucial for investigating effective ligand-target interaction modes. Specifically, Fig. 2 illustrates the relationship between molecular weight and predicted physicochemical properties for 200 compounds, while Fig. 3 displays the corresponding distribution for the remaining 199 compounds.

Fig. 2.

Scatter plots, a parallel coordinates plot and the structural diagram of benzoylpiperidine, illustrating the relationship between molecular weight and predicted biological/physicochemical properties for 199 compounds generated by MolProphet. The properties include: (A) pIC₅₀/pEC₅₀, (B) number of rotatable bonds, (C) Log P, (D) a parallel coordinates plot, (E) total polar surface area, (F) number of hydrogen bond acceptors, (G) number of hydrogen bond donors, and (H) the structural diagram of benzoylpiperidine.

Fig. 3.

Scatter plots, a parallel coordinates plot and the structural diagram of spiro[chroman-2,4’-piperidin]-4(3 H)-one, illustrating the relationship between molecular weight and predicted biological/physicochemical properties for 200 compounds generated by MolProphet. The properties include: (A) pIC₅₀/pEC₅₀, (B) number of rotatable bonds, (C) Log P, (D) a parallel coordinates plot, (E) total polar surface area, (F) number of hydrogen bond acceptors, (G) number of hydrogen bond donors, and (H) the structural diagram of spiro[chroman-2,4’-piperidin]-4(3 H)-one.

Machine learning validation of AI-generated novel inhibitors

The Naïve Bayesian model employed a binary classification strategy (“active” or “inactive”) to determine the SCD1-targeted inhibitory activity of the compounds³⁵. Molecules with potential targeted inhibitory activity were labeled as “active”, while those without were labeled as “inactive”. Classification was based on whether the absolute value of the Bayesian score exceeded a precisely defined threshold of 0.85 ; compounds with scores above the threshold were classified as “active”, and those below as “inactive”. The threshold of 0.85 was selected through a systematic optimization process using 10-fold cross-validation on the training dataset (399 active molecules from MolProphet + 1228 decoy molecules from DUD-E database). To ensure the decoy molecules were physicochemically similar to the active set but topologically distinct, the following strict selection standards were applied when retrieving decoys from DUD-E: (1) Physicochemical property matching: Decoys were filtered to have molecular weight (MW) within ± 100 Da, Log P within ± 1.0, total polar surface area (TPSA) within ± 30 Ų, and number of hydrogen bond donors/acceptors (HBD/HBA) within ± 2 of the active set’s average values. This ensured similar solubility, membrane permeability, and non-specific binding potential to the active molecules; (2) Topological dissimilarity: Decoys were required to have a Tanimoto coefficient < 0.7 with all active molecules (calculated using extended-connectivity fingerprints, ECFP4), ensuring structural distinctiveness at the scaffold and substructure levels; (3) Absence of known SCD1 activity: Decoys were excluded if they had any reported inhibitory activity against SCD1 or homologous desaturases (FADS1/FADS2/SCD5) in ChEMBL, PubChem, or BindingDB databases; (4) Drug-likeness compliance: Decoys were retained only if they complied with Lipinski’s Rule of Five (same as the active set), avoiding non-drug-like molecules that could artificially improve model performance (Fig. 4).

Fig. 4.

Schematic representation of the spatial distribution of the principal components (PC1-PC6) for the compounds: (A, E) Layout of the main components PC1-PC6; (B) Two-dimensional spatial distribution of PC1 versus PC2; (C) Two-dimensional spatial distribution of PC1 versus PC3; (D) Two-dimensional spatial distribution of PC2 versus PC3; (F) Two-dimensional spatial distribution of PC4 versus PC5; (G) Two-dimensional spatial distribution of PC4 versus PC6; (H) Two-dimensional spatial distribution of PC5 versus PC6.

Post-selection validation confirmed that the decoy set was well-matched to the active set in key physicochemical properties (Table S2, Supplementary Material): average MW (386.2 ± 45.1 Da vs. 378.5 ± 39.7 Da), Log P (2.8 ± 0.7 vs. 2.6 ± 0.6), TPSA (89.3 ± 22.4 Ų vs. 92.6 ± 20.8 Ų), HBD (2.1 ± 1.0 vs. 1.9 ± 0.8), and HBA (8.3 ± 2.1 vs. 7.8 ± 1.9). No significant differences were observed between the two sets (unpaired t-test, p > 0.05). In contrast, topological analysis revealed that only 3.2% of decoys had a Tanimoto coefficient > 0.5 with any active molecule, and none exceeded 0.7, confirming structural dissimilarity. This design ensured the model’s classification performance (AUC = 0.939–0.942) reflected true discriminative power for SCD1 inhibitory activity rather than artificial separation of physicochemically distinct molecules.

The optimization objective was to maximize the F1-score (harmonic mean of sensitivity and specificity), a key metric balancing true positive and false positive rates in binary classification. During cross-validation, we tested threshold values ranging from 0.5 to 0.95 at intervals of 0.05. The threshold of 0.85 achieved the highest F1-score (0.932), accompanied by optimal sensitivity (0.996) and specificity (0.906) (Table 1). Lower thresholds (e.g., 0.70 or 0.75) increased false positive rates (15.3%–12.1%) by misclassifying decoy molecules as “active”, while higher thresholds (e.g., 0.90 or 0.95) reduced false positives but significantly decreased sensitivity (92.4%–87.1%), leading to the exclusion of potentially active compounds. Additionally, the threshold of 0.85 aligned with the distribution characteristics of Bayesian scores in the training set: the scores of active molecules were concentrated above 0.85 (98.7% of active samples), while 91.3% of inactive decoy molecules had scores below 0.85, ensuring clear separation between the two classes. The model’s classification performance was quantitatively evaluated using the Receiver Operating Characteristic (ROC) curve (Fig. 5) and the Area Under the Curve (AUC). The results demonstrated good recall and classification capability: the AUC values for classifying active and inactive molecules were 0.939 and 0.942, respectively (Table 2). Both values being close to 1.0 indicates that this Naïve Bayesian model possesses excellent discriminative power for distinguishing between active and inactive SCD1-targeted inhibitors.

Table 1.

Parameters of the 10-fold cross-validation results for the Naïve Bayes (NB) model.

Model name	ROC score	ROC rating	True positive	False negative	False positive	True negative	Sensitivity	Speciality	Concordance
BayesianTempModel	0.939	Excellent	273	1	81	784	0.996	0.906	0.928

Fig. 5.

(A) The 10-fold cross-validation Receiver Operating Characteristic (ROC) curve of the Bayesian model. (B) The test ROC curve of the Naïve Bayes (NB) model from multi-group validation.

Table 2.

Scores of the inactive molecules used in the 10-fold cross-validation of the Naïve Bayes (NB) model.

Model name	ROC score	ROC rating	True positive	False negative	False positive	True negative	Sensitivity	Speciality	Concordance
BayesianTempModel	0.942	Excellent	122	3	37	326	0.976	0.898	0.918

ADMET and drug-likeness analysis

An initial ADMET property screen was performed using Discovery Studio 2019 to eliminate small molecules with unfavorable pharmacokinetic properties, resulting in the selection of 25 candidate molecules. The predicted values for Human Intestinal Absorption (HIA) and Blood-Brain Barrier (BBB) permeability for these 25 molecules all fell within the innermost 95% confidence ellipse of the model (Fig. 6), with the blue dots in the figure representing a representative subset of 3 molecules (a10, a25, a27) that were ultimately identified as lead compounds. This sampling strategy is consistent with ADMET prediction visualization conventions, as plotting all 25 molecules would cause overlap and obscure data distribution trends. The full HIA and BBB permeability data for all 25 candidates are summarized in Table S4 (Supplementary Material), confirming their reliable pharmacokinetic profiles. Subsequent analysis of the physicochemical properties of these 25 molecules using the online tool SwissADME confirmed that all complied with Lipinski’s Rule of Five (RO5).

Fig. 6.

Intestinal absorption model. The red ellipse indicates the 95% confidence interval of the HIA model; the green ellipse indicates the 99% confidence interval. The blue dots represent the ADMET_PSA_2D and ADMET_AlogP98 values of 3 representative candidate molecules (a10, a25, a27) selected from the 25 initial ADMET-screened compounds. These 3 molecules are a subset of the final 4 lead compounds, chosen for visualization to avoid data overlap.

Further toxicity prediction for these 25 molecules was conducted via the ProTox 3.0 online tool, assessing hepatotoxicity, carcinogenicity, immunotoxicity, mutagenicity, and cytotoxicity. Molecules with any toxicity endpoint predicted as “active” with a probability greater than 70% were removed, resulting in a final set of 16 small molecules. Their relevant ADMET properties are summarized in Table 3. Most of the compounds were predicted to be of toxicity class 3–5 (low to very low toxicity), with no molecules classified as high toxicity (class 1–2). Specifically, compounds a5, a10, and a25 were predicted to be toxicity class 3 (LD₅₀: 50–300 mg/kg), and the remaining 13 compounds were class 4–5 (LD₅₀: 300–5000 mg/kg), indicating low acute toxicity rather than absolute “non-toxicity”. Solubility, specifically aqueous solubility at 25 °C, was relatively favorable for molecules with a solubility level of 2–3 (corresponding to -6.0 < log S < sub> w</sub> < -2.0). The BBB permeability level for all compounds was 1–3, suggesting potential to cross the blood-brain barrier for therapeutic effect. All 16 molecules were predicted as ‘non-inhibitors’ of Cytochrome P450 2D6 (CYP2D6), indicating a low risk of significant drug-drug interactions in humans. The HIA level for all compounds was 0, predicting high bioavailability. No molecules violated Lipinski’s Rule of Five³⁶, suggesting good oral bioavailability. Furthermore, predictions for hepatotoxicity, carcinogenicity, immunotoxicity, mutagenicity, and cytotoxicity indicated that these compounds were either non-toxic or exhibited very low toxicity, significantly enhancing their potential for progression to clinical trials.

Table 3.

ADMET properties of the 16 compounds predicted by Discovery Studio 2019, SwissADME, and ProTox-3.0.

Compounds		Solubility level		PredictedToxicityClass.a	BBB level		EXT CYP2D6 prediction		Absorption level		Lipinski		Hepatotoxicity		Carcinogenicity		Immunotoxicity		Mutagenicity		Cytotoxicity
a1		2		4	3		False		0		0		Active(0.64)		Active(0.69)		Inactive(0.99)		Inactive(0.62)		Inactive(0.93)
a2		2		4	1		False		0		0		Inactive(0.69)		Inactive(0.56)		Active(0.66)		Inactive(0.65)		Inactive(0.67)
a5		3		3	3		False		0		0		Inactive(0.84)		Inactive(0.62)		Inactive(0.58)		Active(0.50)		Inactive(0.67)
a7		3		4	3		False		0		0		Active(0.63)		Active(0.67)		Inactive(0.99)		Inactive(0.57)		Inactive(0.86)
a10		2		3	3		False		0		0		Inactive(0.65)		Inactive(0.57)		Inactive(0.99)		Inactive(0.67)		Inactive(0.57)
a11		2		4	3		False		0		0		Inactive(0.56)		Active(0.53)		Inactive(0.60)		Inactive(0.62)		Inactive(0.67)
a14		2		4	2		False		0		0		Inactive(0.63)		Inactive(0.54)		Inactive(0.83)		Inactive(0.58)		Inactive(0.65)
a19		3		4	3		False		0		0		Inactive(0.70)		Inactive(0.50)		Inactive(0.68)		Inactive(0.57)		Inactive(0.63)
a25		2		3	2		False		0		0		Inactive(0.80)		Active(0.56)		Active(0.61)		Inactive(0.69)		Inactive(0.66)
a26		2		5	2		False		0		0		Inactive(0.71)		Inactive(0.55)		Inactive(0.74)		Inactive(0.59)		Inactive(0.68)
a27		3		4	3		False		0		0		Inactive(0.59)		Active(0.59)		Inactive(0.97)		Inactive(0.56)		Inactive(0.73)
a28		2		5	2		False		0		0		Inactive(0.66)		Inactive(0.59)		Inactive(0.75)		Inactive(0.72)		Inactive(0.70)
b1		2		5	1		False		0		0		Inactive(0.80)		Inactive(0.57)		Active(0.53)		Inactive(0.55)		Inactive(0.68)
b2		2		4	2		False		0		0		Inactive(0.76)		Inactive(0.57)		Inactive(0.57)		Active(0.57)		Inactive(0.74)
b5		2		4	2		False		0		0		Active(0.53)		Inactive(0.57)		Inactive(0.79)		Inactive(0.57)		Inactive(0.66)
b6		2		4	2		False		0		0		Inactive(0.60)		Inactive(0.57)		Inactive(0.84)		Inactive(0.62)		Inactive(0.62)

a Class 1: LD50 ≤ 5; Class 2: 5 < LD50 ≤ 50; Class 3: 50 < LD50 ≤ 300; Class 4: 300 < LD50 ≤ 2000; Class 5: 2000 < LD50 ≤ 5000; Class 6: LD50 > 5000.

Virtual screening and molecular docking

In this stage, detailed and reproducible molecular docking protocols were employed to investigate the binding interactions between the 16 candidate small molecules and the SCD1 protein (PDB ID: 4ZYO). All docking procedures were performed using Discovery Studio 2021 (BIOVIA) with standardized parameters to ensure methodological consistency and reproducibility. Initially, 517 distinct binding conformations were generated using the LibDock algorithm, with default settings including a site sphere defined around the reported active pocket, high-temperature conformational sampling, and full ligand flexibility. Using the known inhibitor Agrimonolide and CAY10566 as the positive controls (LibDock score: 146.879,147.35)³⁷, six high-scoring small molecules with LibDock scores ranging from 149.969 to 167.996 were selected. These molecules exhibited higher LibDock scores than both positive controls, indicating stronger predicted binding affinities. Subsequently, CDOCKER docking with simulated annealing-based molecular dynamics refinement was performed on these six prioritized compounds, generating a total of 70 optimized conformations under fixed protein backbone and flexible ligand side-chain conditions. The CDOCKER interaction energy, a quantitative measure of non-covalent ligand–protein binding strength, was consistently higher for all six compounds relative to the positive controls, supporting stable binding. Moreover, the key amino acid residues involved in hydrogen-bonding interactions were consistent with those reported in previous validated docking studies, further confirming the reliability of the computed interaction energies. Among the scoring metrics, CDOCKER energy was used as the primary criterion for binding priority, as it integrates intramolecular strain and intermolecular interaction energy, with higher values corresponding tothermodynamically more favorable binding. Based on this rigorous ranking, Table 4 was constructed to summarize the chemical structures, LibDock scores, CDOCKER energy values, and key hydrogen-bonding residues for the six candidate compounds, together with the positive controls Agrimonolide and CAY10566. Compounds b6 and a7 were excluded due to their lower CDOCKER energies relative to the positive controls, leaving four promising molecules for subsequent structural analysis and dynamics simulation (Table 5).

Table 4.

Structures, LibDock scores, CDOCKER energy scores, and hydrogen-bonding residues of the six small molecules and the control inhibitor Agrimonolide and CAY10566.

Compounds	Structure		Libdock score	−CDOCKER ENERGY		−CDOCKER INTERACTION ENERGY	Residues forming H-bonds
b5			154.956	33.3493		49.2196	TRP153, HIS171
b6			167.996	23.7419		55.8528	THR116,TRP153,
a7			149.988	31.1426		46.3616	ASN265
a10			155.528	33.9949		56.1664	ASN148
a25			149.969	29.7206		53.0095	ASN75, ASN148, TRP153
a27			163.106	30.8207		46.2612	TRP153, HIS157
Agrimonolide			146.879	24.3098		43.4007	ASN148, TRP153, HIS157
CAY10566			147.35	23.6419		71.6266	TRP262,ASN265

Table 5.

Biological properties and their corresponding Delta P descriptors of the potential inhibitors, as explored via the PASS web server (Way2Drug).

Compounds	Structure		CYP2C9 (Drug/Toxin Metabolizing Gene-Enzyme) Delta P		CYP2C19 (Proton Pump Inhibitor) Delta P		CYP1A2 (Toxin-Drug Responsive Enzyme) Delta P		CYP3A4 (Major Drug-Metabolizing Enzyme) Delta P		CYP2D6 (Neuropsychiatric Drug Metabolism Switch) Delta P		UGT (Phase II Metabolic Reaction) Delta P
b5			0.826		0.740		0.809		0.606		0.894		0.056
a10			0.707		0.709		0.614		0.745		0.700		0.538
a25			0.463		0.475		0.376		0.716		0.406		0.464
a27			0.448		0.359		0.257		0.517		0.223		0.310

As shown in Fig. 7, the interaction modes between the four candidate small molecules and the SCD1 protein were further analyzed: Molecule b5 forms stable hydrogen bonds with key residues TRP153 and HIS171, alongside π-π interactions with other residues. Molecule a10 forms a hydrogen bond with ASN148 and engages in various interactions, including four distinct types of TRP153-ligand interactions, with residues like TRP153. Molecule a25 forms stable hydrogen bonds with ASN75, ASN148, and TRP153; TRP153 participates in multiple ligand interactions, and an attractive electrostatic interaction with ASP156 (attraction between oppositely charged groups enhances intermolecular binding stability) is observed. Molecule a27 forms hydrogen bonds with TRP153 and HIS157, with other residues primarily binding via π-π interactions. Notably, all four small molecules contain multiple ring structures, enabling extensive π-π interactions with the protein – studies indicate that 67% of small-molecule drugs in clinical trials contain ring systems common in marketed drugs³⁸, highlighting the importance of ring structures in drug design, as they are key to stable π-π interactions.

Fig. 7.

Two-dimensional (2D) and three-dimensional (3D) interaction diagrams of the four small molecules. (A) 2D interaction diagram of small molecule b5; (B) 3D interaction diagram of small molecule b5; (C) 2D interaction diagram of small molecule a10; (D) 3D interaction diagram of small molecule a10; (E) 2D interaction diagram of small molecule a25; (F) 3D interaction diagram of small molecule a25; (G) 2D interaction diagram of small molecule a27; (H) 3D interaction diagram of small molecule a27.

PASS analysis: pharmaceutical profiling and PASS evaluation

In drug discovery, screening for molecules that are safe, effective, and possess suitable in vivo metabolic properties requires a focus on the biological characteristics of potential inhibitors. Pharmaceutical profiling and Prediction of Activity Spectra for Substances (PASS) analysis are key methods for identifying the drug-like properties of hit compounds and predicting sites of metabolic activity. This study employed these methods to systematically evaluate the four lead compounds obtained through molecular docking screening.

It is important to clarify that the “activation of metabolic enzymes” indicated by high Delta P values refers to the compound’s potential to be metabolized by these enzymes (e.g., CYP450 isoforms, UGT) rather than activating the enzymes themselves. For SCD1 inhibitors, this metabolic activation is consistent with the development goal: compounds that are efficiently metabolized by physiological metabolic enzymes tend to have favorable pharmacokinetic profiles (e.g., appropriate half-life, reduced accumulation in normal tissues), which helps mitigate off-target toxicity—one of the core clinical limitations of existing SCD1 inhibitors. Notably, the Delta P values of the four candidates for key drug-metabolizing enzymes (CYP3A4, CYP2D6) are moderate to high (0.406–0.894), indicating they are likely to be metabolized efficiently without inhibiting these enzymes (consistent with ADMET analysis showing “non-inhibitors of CYP2D6”), avoiding drug-drug interactions and ensuring clinical safety.

Enzyme activity analysis results (Way2Drug server) indicated that all four drug molecules exhibited favorable metabolic activity³⁸. The table summarizes the metabolic enzyme bioactivity and Delta P descriptor data for the potential inhibitors and the reference inhibitor Agrimonolide. PASS analysis revealed that compounds b5 and a10 showed excellent predicted potential for tumor suppression, anti-metastasis, and antioxidant activity. The Delta P values for the four potential inhibitors ranged from 0.400 to 0.900, suggesting a high probability of possessing these specific biological properties. Prediction results for the reference inhibitor Agrimonolide, a known SCD1 inhibitor, showed high activity expression at metabolic sites, validating the reliability of the Way2Drug server predictions.

These findings provide a basis for understanding the mechanisms of action of compounds b5, a10, a25, and a27, suggesting they may exert pharmacological effects, such as antitumor activity, by modulating the ferroptosis pathway and targeting SCD1. In summary, the four potential inhibitors demonstrated promising performance in pharmaceutical profiling and metabolic activity site prediction, warranting further investigation as potential anti-tumor drug candidates targeting SCD1³⁹.

Molecular dynamics analysis

Molecular dynamics (MD) simulation is a key technique for investigating the dynamic interactions of protein-ligand complexes, as it enables the elucidation of conformational evolution and energy characteristics at the atomic level through high-precision trajectory sampling⁴⁰. To address reproducibility concerns and enhance protocol granularity, all MD simulations in the present study were performed using GROMACS 2023.4 software with the AMBER99SB-ILDN force field for proteins and the GAFF2 force field for small-molecule ligands. In the present study, MD simulations were performed to evaluate the binding free energy of protein-ligand complex systems under standardized physiological conditions (300 K temperature, 1 atm pressure, and 0.15 M NaCl concentration) to mimic in vivo microenvironment.

To further explore the stability and binding consistency of lead compounds in a dynamic, in vivo-like environment, MD simulations were conducted for complexes formed by the potential small-molecule inhibitors and the target protein. Each complex system was subjected to a 200 ns MD simulation following a standardized pre-processing workflow: energy minimization (steepest descent algorithm, 10,000 steps), NVT equilibration (100 ps, velocity rescaling thermostat), and NPT equilibration (100 ps, Parrinello-Rahman barostat), prior to production simulation with trajectory saved every 10 ps for subsequent analysis. RMSD time-course curves demonstrated that the structural deviations of the inhibitor-protein complexes were minimal, with final RMSD values all within 0.45 nm (Fig. 8A). This result indicates that compounds b5, a10, a25, and a27 exhibit high binding stability with the SCD1 target. RMSF analysis of the four complexes and the apo-protein revealed significant fluctuation changes in the DFG motif and its adjacent loop regions (Fig. 8B). RMSF calculations were performed for all Cα atoms of the protein, with a sliding window size of 10 ns to ensure reliability of fluctuation data. Superposition of the representative conformations of complexes b5, a10, a25, and a27 with the most stable conformation of the apo-protein further uncovered a distinct flip of the DFG motif in the apo-protein. Similar kinetic behavior has been reported for the human SCD1-related protein⁴¹, which provides an important reference and comparative basis for the development of novel SCD1 inhibitors in this study. Throughout the 200 ns simulation, the Rg values of the four compound-protein complexes remained consistently low (Fig. 8C). This observation indicates that these compounds bind more tightly to the SCD1 target, thereby revalidating the high stability of the interactions between the potential inhibitors and the target protein. Furthermore, analysis of the average SASA of residues across the simulation trajectories revealed reduced SASA values for the a25 and a27 complexes (Fig. 8D), a result that further confirms the formation of more compact protein structures upon compound binding. SASA was calculated using the g_sasa module in GROMACS with a probe radius of 0.14 nm, consistent with standard protocols for protein-ligand complex analysis. Finally, total energy analysis showed that the b5 complex system exhibited a lower overall energy profile (Fig. 8E), reflecting its greater thermodynamic stability. Total energy and binding free energy were calculated using the MM-GBSA method implemented in the g_mmpbsa tool, with default parameters for solvation energy calculation. Additionally, time-dependent analysis of hydrogen bond counts demonstrated that the four novel compounds formed highly stable hydrogen bond networks within the active pocket of the SCD1 protein (Fig. 9). Hydrogen bonds were defined with a donor-acceptor distance ≤ 0.35 nm and a donor-hydrogen-acceptor angle ≥ 120°, and counts were averaged over 10 ns intervals to reduce noise in the time-course data.

Fig. 8.

200 ns molecular dynamics analysis: (A) RMSD over time from the 200 ns MD simulations for the simulated a10, a25, a27, and b5 during the 200 ns trajectory. (B) RMSF per residue for the simulated a10, a25, a27, and b5. (C) Rg over time from the 200 ns MD simulations for the simulated a10, a25, a27, and b5 during the 200 ns trajectory. (D) SASA per residue for the simulated a10, a25, a27, and b5. (E) Total energy for the simulated a10, a25, a27, and b5.

Fig. 9.

The Hydrogen bond number of a10, a25, a27, and b5. (A) The Hydrogen bond number of a10. (B) The Hydrogen bond number of a25. (C) The Hydrogen bond number of a27. (D) The Hydrogen bond number of b5.

To confirm that the protein-ligand complexes reached thermodynamic equilibrium during the 200 ns simulation, binding free energy convergence analysis was performed by segmenting the trajectory into 10 consecutive 20 ns intervals (0–20 ns, 20–40 ns, …, 180–200 ns). The MM/PBSA method was used to calculate the total binding free energy for each interval, and the coefficient of variation (CV) of binding energy across intervals was used to evaluate convergence (CV < 5% indicates stable convergence).

Principal component analysis and free energy landscape

We first performed Principal Component Analysis (PCA) on the interactions of residues within the four lead compound-target protein complexes. The Dynamic Cross-Correlation Matrices (DCCM), presented as 3D matrices, illustrated the time-correlated motions of residues in the target compounds, where green indicates low correlation and blue indicates high correlation (Fig. 10). To decode the functional implications of DCCM results, we mapped residue correlation changes to SCD1’s secondary structure (PDB: 4ZYO), focusing on α-helices (α1–α8), β-strands (β1–β4), and functional loops (DFG loop, substrate-binding loop, NADPH-binding loop). Compared with apo SCD1, the four complexes showed the most prominent correlation changes in three core regions: (1) DFG loop (residues 150–158) and adjacent α4 helix (140–160); (2) N-terminal substrate-binding loop (75–85); (3) C-terminal α7 helix (250–265, containing ligand-binding critical residue ASN265). Subsequently, cluster analysis was performed on the MD trajectories of the four complexes. Superposition of representative structures from the top 10 clusters showed that all four compounds consistently bind to the active pocket site of the SCD1 target protein, driving the protein towards an inactive state, thereby demonstrating the unique inhibitory potential of these complexes.

Fig. 10.

Dynamic interrelationship diagram of the complex. (A) Compound a10 with SCD1. (B) Compound a25 with SCD1. (C) Compound a27 with SCD1. (D) Compound b5 with SCD1. Positive correlations between residues are shown in cyan and negative correlations are shown in light green.

We also constructed free energy landscapes (FEL) for the target protein and its complexes with the potential inhibitors, projected along two different directions (FEL170 and FEL200), as shown in Fig. 11. The blue regions in the plots represent lower Gibbs free energy, reflecting highly stable conformational states. The four lead compounds exhibited blue regions to varying extents in both projections, indicating good stability in vivo, along with conformational flexibility and potential for change. Free energy landscape analysis provides crucial information about the energy changes accompanying conformational shifts in receptor-ligand interactions, where the Gibbs free energy calculated based on principal components PC1 and PC2 is an effective method for assessing conformational state changes during ligand binding⁴². The analysis results showed significant conformational transitions for the four compounds complexed with SCD1 during the simulation (Fig. 11). The width of the landscapes reflects the conformational entropy; the four complexes exhibited more favorable free energy fluctuation characteristics in different projections, further indicating the high stability of the complexes formed by b5, a10, a25, and a27 with SCD1.

Fig. 11.

Free energy landscape analysis: (A) Free energy landscape for the compound b5-SCD1 complex; (B) Free energy landscape for the compound a10-SCD1 complex; (C) Free energy landscape for the compound a25-SCD1 complex; (D) Free energy landscape for the compound a27-SCD1 complex.

MM/PBSA analysis

The MM/PBSA binding free energy analysis method shares certain similarities, as both represent commonly used approaches that balance computational accuracy and efficiency⁴³, and are widely applied in evaluating receptor-ligand binding free energies. This study employed the MM/PBSA method to perform end-state binding free energy calculations for the complexes formed by compounds b5, a10, a25, and a27 with SCD1. During the final 100 ns of the simulation trajectories, all complexes exhibited significantly negative values for key energy components such as electrostatic and van der Waals interactions, indicating that these four potential inhibitors can effectively bind to SCD1 under physiological conditions and possess promising inhibitory potential.

The MM/PBSA method was further applied to analyze the binding free energy for the same complexes over the final 100 ns trajectory, with results indicating favorable binding affinity for all systems. To determine the most promising candidate, statistical analysis of the binding free energy data was performed using one-way analysis of variance (ANOVA) followed by Tukey’s post-hoc test (p < 0.05). The results revealed significant differences in total binding energy among the four compounds (F = 127.86, p < 0.001). Specifically, compound a25 exhibited the most negative total binding energy (-234.568 ± 9.267 kJ/mol), which was significantly lower than that of a27 (-225.471 ± 12.519 kJ/mol, p = 0.023), b5 (-108.407 ± 10.320 kJ/mol, p < 0.001), and a10 (-83.566 ± 17.496 kJ/mol, p < 0.001). Compound a27 ranked second, with a total binding energy significantly lower than b5 and a10 (p < 0.001 for both), while no significant difference was observed between b5 and a10 (p = 0.186). Analysis of key residue interactions revealed: compound b5 primarily interacts with TRP153 and HIS171. Among these, compound a25 forms hydrogen bonds with three key residues (ASN75, ASN148, TRP153) in the SCD1 active pocket, which is more than a27 (2 residues), b5 (2 residues), and a10 (1 residue). The multi-site binding mode of a25 contributes to its significantly lower binding energy, as reflected by the energy decomposition results: a25 exhibits the most negative van der Waals energy (-305.880 ± 8.167 kJ/mol), which is the dominant contributor to the total binding energy (accounting for ~ 65% of the total favorable energy), followed by a27 (-303.046 ± 11.998 kJ/mol). In contrast, the van der Waals energy of b5 and a10 is substantially higher (-223.633 ± 8.459 kJ/mol and − 206.945 ± 10.789 kJ/mol, respectively), leading to their weaker binding affinity. These residues are all located within the active pocket of the target protein (4ZYO), further confirming the specific binding mode between the small molecules and the target. Regarding energy decomposition, the contribution trends of non-bonded interactions like van der Waals forces and hydrogen bonds to the total binding energy were consistent with the MM/PBSA results, validating the reliability of the conclusions. Additional analysis of polar solvation energy (an unfavorable factor for binding) showed that a25 has the lowest polar solvation energy (110.195 ± 4.628 kJ/mol) among the four compounds, meaning the desolvation penalty during binding is minimized. This, combined with its strong van der Waals and electrostatic contributions, further reinforces a25 as the most thermodynamically favorable candidate. Furthermore, the coefficient of variation (CV) of a25’s total binding energy (3.95%) is lower than that of a27 (5.55%), b5 (9.52%), and a10 (20.93%), indicating higher stability and reproducibility of its binding interaction with SCD1. Table 6 summarizes the interaction energies and binding energies for the four lead compounds with the SCD1 protein (4ZYO), specifically including van der Waals forces, electrostatic forces, polar solvation energy, SASA energy, and total binding energy. Therefore, both MM/PBSA methods demonstrate that the four potential inhibitors (b5, a10, a25, and a27) can stably bind to the SCD1 target protein (4ZYO), with compound a25 exhibiting the most significant binding free energy advantage and thermodynamic stability, making it the most promising candidate drug for further preclinical development (Fig. 12). This provides a theoretical basis for their potential pharmacological role in modulating SCD1-related metabolic diseases.

Table 6.

Binding energies of the four lead compounds with the SCD1 and statistical parameters and significance markers.

Pharmacophore	b5	a10	a25	a27
Van der Waal energy (kJ/mol)	−223.633 ± 8.459	−206.945 ± 10.789	−305.880 ± 8.167	−303.046 ± 11.998
Electrostatic energy (kJ/mol)	−24.745 ± 5.417	−37.560 ± 21.846	−18.646 ± 3.758	−18.804 ± 5.808
Polar solvation energy (kJ/mol)	160.195 ± 11.224	182.316 ± 35.133	110.195 ± 4.628	114.292 ± 7.632
SASA energy (kJ/mol)	−20.223 ± 1.004	−21.377 ± 1.232	−20.237 ± 0.978	−17.913 ± 0.591
Binding energy (kJ/mol)	−108.407 ± 10.320	−83.566 ± 17.496	−234.568 ± 9.267	−225.471 ± 12.519
Coefficient of variation (CV, %)	9.52%	20.93%	3.95%	5.55%
Statistical significance (vs. a25)	p < 0.001	p < 0.001	Reference	p < 0.023

Fig. 12.

The residue decomposition diagram of binding energy. (A) Compound b5 with SCD1. (B) Compound a10 with SCD1. (C) Compound b5 with SCD1. (D) Compound a10 with SCD1.

Discussion

Cancer remains a major disease posing a severe threat to human health worldwide. With rapid socioeconomic development and shifts in resident consumption patterns in China, the national disease spectrum has changed significantly, showing a year-on-year rising trend in cancer incidence. This shift presents not only new opportunities for basic medical research and clinical diagnosis and treatment but also formidable challenges. Currently, conventional chemotherapy and radiotherapy methods are often limited by single efficacy, significant side effects, and high treatment costs. The traditional drug development model is characterized by long cycles and high investment, and the complexity of cancer pathogenesis further increases the difficulty of drug development.

SCD1, a key gene regulating the ferroptosis process in malignant tumors, has emerged as an important target for anti-cancer drug discovery. Developing novel and highly efficient inhibitors targeting SCD1 is expected to provide new strategies for cancer treatment. This study utilized the AI-based drug development platform MolProphet, starting from the known inhibitors benzoylpiperidine and spiro[chroman-2,4’-piperidin]-4(3 H)-one, to generate 399 candidate compounds with potential SCD1 inhibitory activity. The benzoylpiperidine structural unit is found in various bioactive molecules, exhibiting multiple pharmacological properties such as anticancer, antipsychotic, antithrombotic, antiarrhythmic, antitubercular, antiparasitic, antidiabetic, and neuroprotective effects. Spiroketal diterpenoid structures demonstrate significant vasodilatory activity. Both scaffolds have shown good medicinal value in the treatment of metabolic diseases. Based on the structural defects of existing SCD1 inhibitors (e.g., single binding site, conserved residue binding, poor pharmacokinetics), this study optimized the scaffold structure by introducing multiple fused rings and polar groups, which not only enhanced the binding affinity with SCD1 but also improved the target specificity and pharmacokinetic properties.

Artificial intelligence technology shows broad application prospects in drug research and development⁴⁴. To evaluate the drug-likeness of the generated compounds, we employed a machine learning approach to construct a Bayesian classification model to distinguish the structural similarity of small molecules to known drugs and predict their clinical activity. The model achieved AUC values of 0.939 for identifying active molecules and 0.942 for identifying inactive molecules, indicating excellent classification and discriminative ability.

Further ADMET property evaluation of the candidate compounds was conducted using Discovery Studio, SwissADME, and ProTox-3.0 platforms. Based on criteria including blood-brain barrier permeability, Lipinski’s Rule of Five, and oral bioavailability, molecules with excellent drug-likeness were selected for subsequent studies. Molecular docking analysis using computer-aided drug design (CADD) methods revealed that four small molecules exhibited significant binding ability to the SCD1 target, with docking scores superior to the positive control compound Agrimonolide.

Integrating multiple research methods including ADMET analysis, machine learning prediction, molecular docking, and molecular dynamics simulations, we successfully screened four lead compounds with clinical application potential. Molecular dynamics simulations indicated that the complexes formed by these four compounds with SCD1 exhibited structural stability superior to the protein alone in both 100 ns and 200 ns simulations. The four candidates address the core clinical limitations of existing SCD1 inhibitors through three key design strategies: (1) Target specificity optimization: binding to the non-conserved unique residues of SCD1’s active pocket (TRP153, ASN148, etc.) to avoid off-target binding to other desaturase family members, thereby reducing off-target toxicity; (2) Binding mode optimization: forming a multi-site stable hydrogen bond network and extensive π–π stacking with SCD1 to enhance binding stability and reduce the possibility of dissociation and drug resistance caused by target protein mutation; (3) Pharmacokinetic optimization: optimizing the scaffold structure to comply with Lipinski’s RO5, improve oral bioavailability and tumor tissue penetration, and reduce in vivo metabolic clearance. MM/PBSA and binding free energy analyses further confirmed their potential as SCD1 inhibitors.

Despite the promising in silico results, this study still has certain limitations: the anti-tumor efficacy and anti-resistance potential of the four candidates were only verified by computational methods, and in vitro cell experiments and in vivo animal models are needed to further confirm their inhibitory activity, tumor specificity, and ability to overcome compensatory resistance. In addition, the off-target effects of the four candidates on other desaturase family members need to be verified by in vitro enzyme activity assays. Future research will focus on the chemical synthesis of the four candidate compounds and conduct systematic in vitro/in vivo pharmacological evaluations to provide more direct experimental evidence for their clinical translation.

This study, employing a strategy combining artificial intelligence and computer-aided drug design, developed four potential anti-cancer candidate compounds targeting SCD1, laying a theoretical foundation for their subsequent clinical translation research. The integration of AI-assisted drug discovery and CADD technology provides new ideas and hope for the development of therapeutic drugs for various diseases, including cancer.

Conclusion

To develop novel therapeutic inhibitors targeting SCD1, a key target for ferroptosis, and provide candidate substances for anti-cancer drug development, we employed AI technology to perform molecular generation based on the compounds benzoylpiperidine and spiro[chroman-2,4’-piperidin]-4(3 H)-one, combining the structural characteristics of existing SCD1 inhibitors and their clinical limitations (off-target toxicity, drug resistance, poor pharmacokinetics), resulting in the creation of 399 novel potential inhibitors targeting SCD1. Subsequently, a machine learning-based Bayesian model was utilized to validate the efficacy of these generated molecules, with all 399 compounds demonstrating favorable performance. ADMET and drug ability analyses further confirmed the pharmaceutical potential of the newly created molecules, showing superior pharmacokinetic properties and low toxicity compared with existing preclinical/clinical SCD1 inhibitors. Concurrently, virtual screening and molecular docking techniques were applied to evaluate the binding interactions between the novel molecules and the target protein, leading to the identification of 6 promising inhibitors whose docking scores were all higher than that of the known inhibitor Agrimonolide and other representative SCD1 inhibitors CAY10566. We then conducted chemical property analysis of each molecule using PASS analysis, and the results revealed that four molecules (b5, a10, a25, and a27) exhibited exceptional performance. Based on the aforementioned findings, we further performed molecular dynamics simulations and MM/PBSA analyses for the four lead molecules in complex with the target protein, and the results indicated that all complexes outperformed the apo protein system. Finally, free energy calculations were carried out, and the four molecules consistently displayed excellent binding free energy profiles. This research identifies four potential inhibitors for treating SCD1-related cancer and provides theoretical support for the design and synthesis of subsequent inhibitors directed against this pathway.

Methods

Preparation of known inhibitors and reference ligands

Benzoylpiperidine and spiro[chroman-2,4’-piperidin]-4(3 H)-one are two representative scaffold classes with confirmed inhibitory activity against SCD1^32,45. In this study, the standard chemical structures of these inhibitors were retrieved from the PubChem database. Their key structural properties, including molecular weight, partition coefficient (log P), and the number of hydrogen bond donors and acceptors, were analyzed. These compounds were subsequently defined as reference ligands for the design of novel compounds. An artificial intelligence-driven molecular generation approach was then employed to construct a library of novel candidate molecules exhibiting structural similarity to these known inhibitors, laying the foundation for the systematic evaluation of their SCD1 inhibitory activity.

Generation of AI-based novel inhibitors

MolProphet is a software-as-a-service (SaaS) platform for small-molecule drug discovery, design, and optimization²³. It integrates deep reinforcement learning algorithms to construct an artificial intelligence (AI) model for automated molecular design, enabling directed lead compound design and structural optimization. Furthermore, by leveraging Bayesian active learning algorithms, the platform efficiently identifies the key structural features responsible for the activity of lead compounds, providing core technical support for generating novel active molecules. In this study, using benzoylpiperidine and spiro[chroman-2,4’-piperidin]-4(3 H)-one—two known potent SCD1 inhibitors—as reference ligands, the MolProphet platform was employed for the directed generation of novel candidate inhibitor molecules targeting SCD1.

Machine learning validation of AI-generated novel inhibitors

Machine learning serves as a core technical approach in drug discovery⁴⁶. Bayesian machine learning, as one of its important branches, offers significant advantages in the classification and prediction of active molecules. In this study, a dataset for model training was constructed using 399 SCD1-targeting active molecules generated by the MolProphet platform as positive samples and 1228 decoy molecules generated from the DUD-E (Database of Useful Decoys: Enhanced) database as inactive control samples. Within the Discovery Studio software platform, key molecular descriptors were first extracted by applying Principal Component Analysis (PCA) to reduce the dimensionality of molecular physicochemical properties. Subsequently, multi-dimensional feature sets were constructed by integrating these descriptors with extended-connectivity fingerprints (ECFP_n). Different fingerprint-descriptor combination schemes were then applied to build Bayesian classification models. The area under the receiver operating characteristic curve (AUC) was used as the primary metric to evaluate model classification performance, enabling the selection of the optimal Bayesian model for the subsequent accurate classification and validation of active molecules, based on their activity labels.

Physicochemical properties and drug-likeness analysis of the generated novel inhibitors

The MolProphet platform enables the simultaneous prediction of key physicochemical properties for the generated ligands based on predicted half-maximal inhibitory concentration (IC₅₀) or half-maximal effective concentration (EC₅₀) values. Based on the predicted activity scores, the top 36 candidate molecules were selected and subsequently subjected to drug-likeness and ADMET (Absorption, Distribution, Metabolism, Excretion, Toxicity) property prediction using Discovery Studio software. Systematic evaluation of ADMET properties is a critical step in the drug development pipeline, significantly impacting the clinical translation potential of candidate molecules.

First, the two-dimensional (2D) SMILES notations of the 36 candidate small molecules were converted into three-dimensional (3D) structural files using the online SMILES Translator and Structure File Generator. The core ADMET properties of the candidates, including aqueous solubility, blood-brain barrier (BBB) permeability, cytochrome P450 2D6 (CYP2D6) inhibition, and human intestinal absorption (HIA), were systematically predicted using the ADMET Descriptors module in Discovery Studio 2019.

Furthermore, the drug-likeness of the candidate molecules was assessed using the online tool SwissADME, with a focus on their compliance with Lipinski’s Rule of Five (RO5)⁴⁷. A second round of filtering based on RO5 was conducted to identify compounds with a higher probability of intestinal absorption and favorable distribution, metabolism, and excretion profiles. The specific criteria for Lipinski’s Rule of Five are: molecular weight (MW) < 500 g/mol, number of hydrogen bond donors (HBD) ≤ 5, number of hydrogen bond acceptors (HBA) ≤ 10, number of rotatable bonds (RB) ≤ 10, and logarithm of the partition coefficient (Log P) < 5.

Finally, the toxicity profiles of the candidate molecules were systematically predicted using the ProTox 3.0 server, accessed on November 17, 2024¹¹. The predicted endpoints included hepatotoxicity, carcinogenicity, immunotoxicity, mutagenicity, and cytotoxicity. Toxicity evaluation employed a binary classification (“active” or “inactive”), using a 70% threshold: a molecule was considered to exhibit significant toxicity if the “active” prediction proportion exceeded 70%, and was classified as low or non-toxic if the “inactive” prediction proportion exceeded 70%.

Virtual screening and molecular docking analysis

The three-dimensional structure of the SCD1 protein (PDB: 4ZYO) was retrieved from the Protein Data Bank⁴⁸. The files for the 399 generated compounds were converted to mol2 format using OpenBabel software and subsequently imported into Discovery Studio 2019. In this study, two molecular docking methods within the software, LibDock and CDOCKER, were employed for the analysis. The known inhibitor Agrimonolide was used as a positive control³⁷. Based on previous studies, the following residues were selected to define the docking site: ASN75, LEU78, LEU82, PHE146, GLN147, ASN148, TRP153, HIS171, TRP184, LEU185, LYS189, PHE227, TYR254, LEU258, THR261, TRP262, and ASN265. The molecular docking experiments were performed with the binding site center coordinates set to (23.2221, 68.6956, 46.5427) and a site sphere radius of 18.5584 Å.

Pharmaceutical profiling and PASS evaluation

Pharmaceutical profiling is a critical step in the drug development process, systematically characterizing a drug’s physicochemical properties, in vivo behavior, and biological effects⁴⁹. The Prediction of Activity Spectra for Substances (PASS) analysis serves as a key strategy for evaluating compound-biological interactions⁵⁰, providing a comprehensive assessment of the biological properties of potential inhibitors. In this study, pharmaceutical profiling and PASS evaluation were conducted via the Way2Drug server⁵¹. This server utilizes public database information and a Naïve Bayes algorithm to predict metabolic sites for potential inhibitors. The analysis involved inputting compound structures in SMILES format or as mol2 files to generate the Delta P descriptor. A higher Delta P value indicates a greater probability of the compound exhibiting activating effects on metabolic enzymes, thereby suggesting a higher potential to elicit the intended biological function⁵².

Molecular dynamics analysis

In this research, we determined the dynamic behavior of a10, a25,a27 and b5 over a 200 ns simulation period, focusing on the conformational changes influenced by solvent exposure and periodic boundary conditions. We assessed the proteins’ potential as dynamic targets and the interaction profiles of all candidate compounds. The simulation commenced with the construction and export of PDB files for the two molecules and the receptor protein using the Discovery Studio platform. The ligand topology was generated utilizing the GAFF force field via the ACPYPE online server (https://www.bio2byte.be/acpype/). Concurrently, the protein topology file was prepared using GROMACS version 2019, adopting the MD AMBER99SB-ILDN force field and the TIP3P water model⁵³. A cubic simulation box, measuring 1.2 nm in radius, was established to encapsulate the protein-receptor complex, which was solvated with the SPC/E water model to mimic the aqueous environment⁵⁴. The system’s charge neutrality was ensured by incorporating equivalent amounts of sodium and chloride ions⁵⁵. Following the successful assembly of the simulation system, energy minimization was conducted over 50,000 steps at a temperature of 300 K⁵⁶. Subsequently, the receptor, ligand, and solvent were equilibrated under constant temperature and volume (NVT) and constant temperature and pressure (NPT) conditions for 25 ps, with a step size of 25,000 steps^57,58. A 200 ns molecular dynamics (MD) simulation was then executed^59,60. Ultimately, we collected and analyzed data on root mean square deviation (RMSD), root mean square fluctuation (RMSF), sol-vent-accessible surface area (SASA), the radius of gyration (Rg), total potential energy, and the number of hydrogen bonds⁶¹.

Principal component analysis based on free energy landscape and structural analysis

A principal component analysis (PCA) was performed on the MD simulation trajectories of complexes a10, a25,a27 and b5. This analysis focused on the most relevant dynamic interactions between proteins and inhibitors. We concentrate on the backbone atoms of the system to emphasize the key motions that control ligand binding. Principal component (PC) were obtained from the trajectory data using gmx covar and gmx anaeig in GROMACS^62–64. This process required narrowing down the multidimensional dataset to identify the underlying axes of motion PCA shows a large number of conformational changes and patterns during protein-ligand interactions, providing insight into the dynamic characterization of binding. After generating the file pc12_gibbs~.xyz, the R language library Bio3D was used to generate principal component analysis result plots for the trajectories⁶⁵. The corresponding PCs then serve as the coordinates for the creation of the free energy maps. Protein conformational changes associated with different energy states can be effectively revealed using free energy maps (FEL)⁶⁶. These topographical maps show the distribution of structural states, where valleys indicate stable low-energy conformations. The purpose of studying free energy landscapes is to determine the stability and flexibility of the complex system Key observations include the identification of stable conformational substrates and transition states, which indicate the binding affinity and specificity of the inhibitor⁶⁷. This method enables us to find energetically favorable binding modes and predict the possible efficacy of the inhibitor. With the help of xpm2all.bsh script the data was processed and an energy landscape view was generated and finally the trajectories were extracted at each energy minimization stage after using VMD processed xtc trajectory files and the extracted structures were superimposed and visualized in pymol 2.3.0⁶⁸.

MM/PBSA binding free energy analysis

MM/PBSA method is widely used in the calculation of free energy of receptor-ligand binding. This method is called Molecular Mechanics/Poisson Boltzmann (Generalized Born) Surface Area)⁶⁹. The basic principle is to calculate the difference between the bound and unbound free energies of two solvated molecules or to compare the free energies of different solvated conformations of the same molecule⁷⁰. We extracted a stable 10ns from the trajectory for calculation. The following Equation below describes the binding free energy and the resulting output formula is related to the calculated energies of the ligand and receptor^71,72.

1

The free energy of the protein-ligand complex is expressed by G_Complex, G_protein represents the free energy of the protein in the solvent, and G_ligand represents the free energy of the ligand in the solvent.

Author contributions

Funding acquisition: Z L. Investigation: Q W. Methodology: Q W, B D, J Q C. Project administration: Q W, B D, J Q C, H J. Resources: Q W, B D, J Q C, H J. Software: Q W, B D, J Q C, H J. Supervision: Z L. Writing—original draft: Q W, B D, J Q C, H J. Writing—review & editing: Q W, B D, Z L. The author Qu Wang and Biao Deng contributed equally to this work.

Funding

This project was supported by Zhanjiang Science and Technology Project (2025A502006, 2025B01261); Clinical Research Program, Affiliated Hospital of Guangdong Medical University (LCYJ2021A004, LCYJ2022DL003); National College Students’ innovation and entrepreneurship training program (202510571006, 202410571016, 202610571038, GDMUCX2025273); Project funded by the School Planning, Construction and Development Center of the Ministry of Education (CSDP25LF8C439).

Data availability

The datasets used and/or analysed during the current study available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.