Discovery of natural RORγt inhibitor using machine learning, virtual screening, and in vivo validation.

Discovery of natural RORγt inhibitor using machine learning, virtual screening, and in vivo validation - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice J Adv Res . 2025 Sep 5;84:1059–1071. doi: 10.1016/j.jare.2025.09.004 Discovery of natural RORγt inhibitor using machine learning, virtual screening, and in vivo validation Hojin Yoo a, 1 , Sang-Jun Han b, 1 , Jeong-Eun Lee c, 1 , Chaeyeon Cho d , Donggyun Hong a , Birang Jeong d , Sijin Kim b , Go-Yeon Jung b , Minjeong Ma b , Soeun Jung b , Beomjun Park a , Namgil Lee a, e , Hee-Seop Yoo f , Kwang-Jin Cho g , Min-Duk Seo f, ⁎ , Yeonseok Chung c, ⁎ , Byung-Seok Kim b, ⁎ , Heejung Yang a, d, ⁎⁎ Author information Article notes Copyright and License information a Bionsight, Inc., Chuncheon 24341, South Korea b Division of Life Sciences, College of Life Science and Bioengineering, Incheon National University, Incheon 22012, South Korea c Laboratory of Immune Regulation, Institute of Pharmaceutical Sciences, College of Pharmacy, Seoul National University, Seoul 08826, South Korea d Department of Pharmacy, Kangwon National University, Chuncheon 24341, South Korea e Department of Information Statistics, Kangwon National University, Chuncheon 24341, South Korea f Department of Molecular Science and Technology & College of Pharmacy, Ajou University, Suwon 16499, South Korea g Department of Biochemistry and Molecular Biology, Boonshoft School of Medicine, Wright State University, Dayton, OH 45435, USA ⁎ Corresponding authors. mdseo@ajou.ac.kr yeonseok@snu.ac.kr byungseokkim@inu.ac.kr ⁎⁎ Corresponding author at: Department of Pharmacy, Kangwon National University, Chuncheon 24341, South Korea. heejyang@kangwon.ac.kr 1 These authors contributed equally to this work. Received 2025 Jun 13; Revised 2025 Aug 20; Accepted 2025 Sep 2; Collection date 2026 Jun. © 2025 The Author(s) This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). PMC Copyright notice PMCID: PMC13227160  PMID: 40915561 Graphical abstract Open in a new tab Keywords: RORγt, Berberine, Coptisine, Machine learning, Ensemble model, Natural products Highlights • Integrates machine learning, chemotaxonomy, and experiments for discovering natural RORγt inhibitors. • Identifies novel RORγt inhibitors with therapeutic potential for autoimmune diseases like psoriasis. • Reveals mechanism of protoberberine alkaloids as RORγt inhibitors from natural products. • Shows preclinical efficacy of coptisine in a psoriasis mouse model, supporting translational promise. Abstract Introduction Retinoic acid receptor-related orphan receptor gamma t (RORγt) is a crucial transcription factor regulating Th17 cells, which secrete the cytokine IL-17. RORγt inhibitors are regarded as a therapeutic modality in a wide range of autoimmunity including psoriasis. Objectives The objective of the study is to investigate novel RORγt inhibitors from natural products (NPs), combining machine learning (ML)-based virtual screening, chemotaxonomic analysis, molecular docking, and molecular dynamics simulations, and biological validation. Methods This study employed an integrated approach combining ML-based ligand-based screening, docking study, molecular simulation, and chemotaxonomic analysis to identify RORγt inhibitors from NPs. Results ML ensemble models predicted potential RORγt inhibitors from an NP library; subsequent chemotaxonomic classification of top-ranked hits prioritized protoberberine alkaloids. Six protoberberine alkaloids, which are predicted to bind RORγt via docking studies, were selected for experimental validation. Among them, berberine ( Ber ) and coptisine ( Cop ) potently inhibited Th17 differentiation in vitro . Surface plasmon resonance analysis demonstrated that both Ber and Cop directly bind to RORγt, with Cop exhibiting a stronger affinity for RORγt than Ber . Moreover, Cop demonstrated therapeutic efficacy in a preclinical mouse model of psoriasis. These results validate an integrated workflow, combining ML, chemotaxonomy, and experimental testing in vitro and in vivo , for the efficient discovery of novel RORγt inhibitors. Introduction Psoriasis is a chronic and immune-mediated skin disease that primarily affects the skin and joints, with an estimated global prevalence of up to 4.4 % [ 1 ]. The predominant symptoms include itching, burning, and soreness on the skin, and the disease presents with various phenotypes, such as psoriasis vulgaris, guttate psoriasis, inverse psoriasis, pustular psoriasis, palmoplantar psoriasis, and erythrodermic psoriasis [ 2 ]. Severe or exacerbated psoriasis can lead to rheumatologic complications, such as psoriatic arthritis, as well as cardiovascular and psychiatric comorbidities [ 3 ]. Retinoid-related orphan receptor gamma-t (RORγt), a member of the nuclear receptor superfamily, plays a crucial role in regulating the development and function of T-helper 17 (Th17) cells, which are implicated in the pathogenesis of various autoimmune disorders [ [4] , [5] , [6] ]. Inhibitors of RORγt that modulate Th17 cell function have emerged as promising therapeutic candidates for treating Th17-related autoimmune diseases such as multiple sclerosis, rheumatoid arthritis, and psoriasis [ 5 , 7 ]. Ursolic acid and digoxin are two well-studied natural products that act as RORγt inverse agonists and selectively inhibit RORγt, thereby suppressing IL-17 production [ 8 ]. However, digoxin exhibits potent carcinogenic, extracardiac, and coronary vascular effects, while ursolic acid acts on multiple targets, potentially leading to off-target side effects. Consequently, efforts are ongoing to identify additional RORγt inhibitors, including both natural and synthetic compounds [ [9] , [10] , [11] , [12] , [13] , [14] ]. Natural products (NPs) offer unparalleled chemical diversity shaped by evolution, making them a valuable source for drug discovery [ 15 , 16 ]. NPs and their derivatives have historically made significant contributions to the development of newly approved therapeutics [ 17 ]. However, identifying novel bioactive compounds within the chemical space of NPs remains a major challenge in drug discovery, often requiring laborious and resource-intensive experimental screening efforts [ 18 ]. Recent advancements in machine learning (ML), encompassing a range of techniques from traditional algorithms to deep learning (DL), have revolutionized the drug discovery process [ 19 , 20 ]. Artificial intelligence (AI)-powered methods streamline various stages of drug discovery, including hit identification and lead optimization, by efficiently analyzing large datasets and predicting molecular properties, thereby reducing timelines and costs [ [21] , [22] , [23] ]. Specifically, ML models trained on chemical structures and associated bioactivity data can rapidly screen extensive virtual libraries to prioritize compounds for experimental validation [ 22 ]. Among these, ligand-based virtual screening (LBVS) approaches are widely used, particularly when ligand structure and properties are known but the target structure is not available [ 23 ]. Herein, we present an integrated in silico strategy to identify novel RORγt inhibitors from NPs, combining ML-based virtual screening, chemotaxonomic analysis, molecular docking, and molecular dynamics simulations. ML models were developed to predict RORγt inhibitory activity and applied to a large NP database. The top-scoring hits were subjected to hierarchical chemical classification and literature-based investigation, which led to the prioritization of six protoberberine alkaloids. Using in vitro and ex vivo assays to assess IL-17 production by T cells, we identified berberine ( Ber ) and coptisine ( Cop ) as potent and selective inhibitors of Th17 differentiation and function. Importantly, in vivo studies demonstrated that Cop significantly ameliorates Th17/γδT17-driven skin inflammation in a mouse model of psoriasis, correlating with RORγt inhibition. Materials and methods Data preparation Bioactivity data for model training were sourced from the PubChem database [ 24 ]. Specifically, data originated from assays associated with Assay ID (AID) 2604 ( https://pubchem.ncbi.nlm.nih.gov/bioassay/2604 ) that include the primary cell-based RORγt reporter assay (AID 2551) and a counter-screen assay utilizing the VP16 transactivation domain (AID 2546). These assays aimed to identify inhibitors of RORγt-mediated transcription while enabling the filtering of compounds potentially causing non-specific reporter activation or inhibition. Compound activity was defined based on the 'Activity Score' provided by PubChem, ranging from 0 to 100, reflecting the percentage inhibition of activation across a concentration range of 0.05 nM to 150.4 µM. For the primary RORγt screen (AID 2551), compounds with scores between 40 and 100 were designated as 'active' inhibitors (label 1), while all others were labeled 'inactive' (label 0). An analogous labeling procedure was applied to the counter-screen dataset (AID 2546) to identify compounds active in the non-specific assay. The resulting RORγt dataset comprised 270,773 unique compounds, containing 10,704 active inhibitors and 268,069 inactive compounds. The counter-screen dataset contained 300,701 compounds, with 32,933 compounds labeled as active and 267,768 as inactive. To ensure robust model evaluation based on chemical similarity, both datasets were partitioned into training (∼80 %) and test (∼20 %) sets using a scaffold splitting, where we have used the RDKit library to generate the scaffold for each molecule [ 25 ]. Approximately 20 % of the training set was further reserved as a validation set during hyperparameter optimization. Model performance was evaluated by averaging results from three independent scaffold-based splits. Model training and hyperparameter tuning Three distinct molecular representations were generated for model inputs. Two of these—AttentiveFP and Canonical featurizers—were used for GNN models. In these representations, molecules were encoded as undirected graphs where atoms serve as nodes and chemical bonds as edges; associated atom features (including element type, degree, formal charge, hybridization state, aromaticity, and number of attached hydrogens) and bond features (such as bond type, conjugation status, and ring membership) were transformed into numerical vectors suitable for graph-based learning. The third representation was based on molecular fingerprints. Specifically, molecular structures were converted into fixed-length binary vectors using the Extended-Connectivity Fingerprint (ECFP) algorithm from RDKit library. GNN models were built using the Deep Graph Library leveraging the utilities within the DGL-LifeSci package [ 26 , 27 ]. RF classifiers with ECFP fingerprint representations were implemented using the scikit-learn library [ 28 ]. The predictive performance of the optimized RF models and the GNN ensemble models was evaluated on the independent test set. Counter-screen models were similarly evaluated, and the corresponding GNN ensemble was used for filtering. For robust prediction, we selected the top three performing hyperparameter configurations for each model based on validation results. To further evaluate the generalization ability of our ensemble model, we conducted external validation experiments using three binary classification datasets (BBBP, HIV, and BACE) from the MoleculeNet benchmark [ 29 ]. Each dataset was split into training, validation, and test sets at an 8:1:1 ratio using scaffold-based partitioning. The model was trained with the best hyperparameter setting identified in our primary experiments, and predictive performance was assessed by calculating the area under the ROC curve (AUC) on the validation and test sets. For each dataset, three independent training runs were performed under the same hyperparameter configuration, and the resulting models were combined into an ensemble. The results demonstrate consistently high predictive performance across datasets, supporting the model’s generalization capability ( Supplementary Table S1 ). Preparation of six protoberberine-type compounds via virtual screening 276,518 compounds from NPs were downloaded from an open source database, LOTUS ( https://lotus.naturalproducts.net/ ) [ 30 ]. Of these, 196,450 compounds with the molecular weight of more than 200 were prepared using the OPLS4 force field, the Epik mode for ionization, and up to 32 combinations per a ligand generated using the LigPrep module in Maestro. Each compound in the prepared library was scored using the GNN ensemble models for both RORγt inhibition and non-specific activity (VP16 counter-screen). Candidate compounds were prioritized based on a high predicted probability of RORγt inhibition after filtering a low predicted probability of activity in the counter-screen (i.e., average ensemble probability <0.2). The top 1,000 compounds meeting these criteria were selected for further analysis. The selected 1,000 high-priority hits were subjected to automated chemical classification using the ClassyFire web server [ 31 ] to obtain a hierarchical chemical ontology (superclass, class, subclass) ( Supplementary Files 1 ). To quantitatively evaluate drug-likeness of 1,000 high-priority hits, we employed ‘Lipinski’s rule of five’ (no more than one violation among: molecular mass ≤500 Da, no more than 5 hydrogen bond donors, no more than 10 hydrogen bond acceptors, logP ≤5) via the QikProp module within in Maestro Schrödinger suite version 2024–2 (Schrödinger, LLC, New York, NY) [ 32 ]. Additionally, drug-likeness was also quantified using the Quantitative Estimate of Drug-likeness (QED) as implemented in RDKit library [ 25 ]. Compounds exhibiting Lipinski’s rule of five scores of 0 or 1, together with those with a QED score of 1, were preferentially selected as filtering criteria ( Supplementary Files 2 ). After filtering hits based on drug-likeness results and literature survey, we prioritized the ‘Alkaloid and derivatives’ superclass and used the NPASS database ( https://bidd.group/NPASS/ ) to manually review target information of compounds within each subclass that had more than 30 activity records. Among these, we ultimately selected the ‘Protoberberine alkaloids and derivatives’ subclass. Six commercially available protoberberine alkaloids (phellodendrin, palmatine, coptisine, columbamine, jatrorrhizine, berberine) were selected for structural diversity within the subclass and purchased from Sigma Aldrich (St. Louis, MO, USA; Purity ≥95 %). Mice C57BL/6 mice were purchased from Orient Bio (Seongnam, South Korea) or KOATECH (Pyeongtaek, South Korea). Il17a EGFP reporter mice were purchased from Jackson Laboratories (USA). Il17a Cre x Rosa26 YFP mice were previously described [ 33 ]. All the animal procedures were conducted in accordance with the animal use guidelines approved by the Institutional Animal Care and Use Committee of Seoul National University (SNU-191210-3-11), Lee Gil Ya Cancer and Diabetes Institute (LCDI-2022-0088), and Incheon National University (INU-ANIM-2024-0026-01). Th17 and Th1 differentiation in vitro Splenic naive CD4 + T cells were isolated from Il17a EGFP mice or C57BL/6 mice using naive CD4 + T cell isolation kit (Miltenyi Biotec). Purified naive CD4 + T cells were cultured in a flat-bottomed 96 well plate pre-coated with anti-CD3ε (1.0 µg/ml, 145-2C11, BioXCell) and anti-CD28 (1.0 µg/ml, 37.75, BioXCell) antibodies in the presence of recombinant human TGF-β (0.5 ng/ml, Peprotech), recombinant mouse IL-6 (50 ng/ml, Peprotech). Titrated doses of candidate compounds or vehicle (DMSO) were added to the culture from day 0. On day 4, cells were harvested and were subjected to flow cytometric analysis to detect EGFP expression levels. IL-17 levels in the culture supernatant were measured by mouse IL-17A ELISA kit (Thermo Fisher Scientific). To detect IL-17 expression by flow cytometry, cells were further stimulated with PMA (50 ng/ml, Sigma) and ionomycin (1 μg/ml, Sigma) for 3–4 h in the presence of Brefeldin A (Thermo Fisher Scientific) and monensin (Thermo Fisher Scientific). After fixation with IC Fixation buffer (Thermo Fisher Scientific), cells were permeabilized and intracellularly stained with APC anti-mouse IL-17 (TC11-18H10.1, Biolegend) and Alexa 488 anti-mouse IFN-γ (XMG1.2, Biolegend). For Th1 differentiation, naive CD4 + T cells were stimulated with plate-bound anti-CD3ε (1.0 µg/ml) and anti-CD28 (1.0 µg/ml) antibodies in the presence of recombinant murine IL-12 (10 ng/ml, Peprotech) and recombinant murine IL-2 (10 ng/ml, Peprotech). On day 4, IFN-γ levels in the culture supernatant were measured by mouse IFN-γ ELISA kit (Thermo Fisher Scientific). The cultured cells were subjected to intracellular cytokine staining with Alexa 488 anti-mouse IFN-γ (XMG1.2, Biolegend) after restimulation with PMA and ionomycin in the presence of Brefeldin A and monensin. MOG-specific Th17 restimulation assay Il17a Cre x Rosa26 YFP mice were immunized with myelin oligodendrocyte glycoprotein peptide (MOG 35-55, RS synthesis, USA) emulsified in CFA as previously described [ 33 ]. Eight days after immunization, single cells prepared from draining lymph nodes were restimulated with MOG 35-55 peptide (25 µg/ml) and recombinant mouse IL-23 (20 ng/ml, Thermo Fisher Scientific) in the presence of titrated doses of Cop, Ber, or vehicle (DMSO). At day 5 after culture, IL-17 levels in the culture supernatant were measured by ELISA. Cells were further restimulated with PMA and ionomycin in the presence of Brefeldin A and monensin for flow cytometric analysis of cytokine production. Cell surface staining was performed with Pacific Blue anti-mouse CD4 (GK1.5, Biolegend), PerCP-Cy5.5 anti-mouse CD44 (IM7, Biolegend), and eFluor506-fixable viability dye (Thermo Fisher Scientific). After cell surface staining, cells were permeabilized followed by intracellular staining with PE anti-mouse IL-17 (TC11-18H10.1, Biolegend). Imiquimod-induced psoriasis-like skin inflammation C57BL/6 mice were treated daily with Aldara cream (5 % imiquimod, w/w, BK Pharm, South Korea) in the left and right ears for six consecutive days. Cop (50 mg/kg) was administered intraperitoneally (i.p.) daily from day 0 (prevention model) or day 2 (therapeutic model) after disease induction. The same volume of vehicle (DMSO) was injected i.p. as a control. Ear thickness of both ears was measured daily. On day 6 after disease induction, mice were sacrificed for analysis. Single cell suspensions prepared from draining cervical lymph nodes were incubated with PMA and ionomycin for 3 h in the presence of Brefeldin A and monensin for flow cytometric analysis or 24 h without Brefeldin A and monensin for detection of IL-17 levels in the culture supernatant by ELISA. For surface staining of ear cells, cells were stained with PE anti-mouse TCRγδ (GL3, Biolegend), BUV395 anti-mouse CD3ε (145-2C11, BD Bioscience), PE-Cy7 anti-mouse CD4 (GK1.5, Biolegend), APC-Cy7 anti-mouse CD45 (30-F11, Biolegend) for 30 min at 4 °C. After surface staining, cells were fixed with IC fixation buffer (Thermo Fisher Scientific) for 20–30 min at 4 °C, and then intracellular staining was performed with FITC anti-mouse-IL-17A (TC11-18H10.1, Biolegend), APC anti-mouse-IFN-γ (XMG1.2, Biolegend). Surface staining of restimulated lymph node cells was performed using PE anti-mouse TCRγδ (GL3, Biolegend), APC anti-mouse/human CD44 (IM7, Biolegend), and Pacific Blue anti-mouse-CD4 (GK1.5, Biolegend). Intracellular staining was performed using FITC anti-mouse-IL-17A (TC11-18H10.1, Biolegend) and PE-Cy7 anti-mouse IFN-γ (XMG1.2, Biolegend). Histological and immunohistochemical analysis For histological analysis, tissues were fixed in 10 % neutral-buffered formalin, processed using standard paraffin embedding, and stained with hematoxylin and eosin (H&E) on an automated slide stainer (MICROM HMS 70) at the Core-facility for Cell to In-vivo imaging, Gachon University. The stained slides were digitized with a whole-slide scanner (PANNORAMIC SCAN II, 3DHISTECH). Images were randomly selected and analyzed at 30× magnification using the CaseViewer software (3DHISTECH). For immunohistochemistry, paraffin sections were incubated with a mouse anti-Ki-67 primary antibody (B56, BD Biosciences), followed by a biotinylated secondary antibody. The sections were then treated with the VECTASTAIN Elite ABC-HRP kit (PK-6200, Vector Laboratories), and immunoreactivity was visualized using 3,3′-diaminobenzidine (DAB) as the chromogen. Finally, the sections were counterstained with hematoxylin to visualize cell nuclei. Protein purification Complementary DNA encoding the RORγt ligand binding domain (RORγt-LBD) (residues 262–507) was cloned into the pET-28a vector and transformed into E. coli strains BL21 (DE3) cells. The cells were grown in LB medium at 37 °C. Overexpression of RORγt-LBD was induced with 0.5 mM isopropyl-β-D-1-thiogalactopyranoside (IPTG) at 15 °C for 16 h when the optical cell density reached an OD600 of 0.8. The cells were collected and resuspended in a buffer containing 50 mM Tris-HCl, 500 mM NaCl, 5 % Glycerol, 0.4 mM Tris(2-carboxyethyl)phosphine hydrochloride (TCEP), 0.1 % Nonidet P-40 (NP40), pH 7.3 with Protease inhibitor cocktail tablet (Roche) to be disrupted by sonication. Cell debris was removed by centrifugation and the supernatant was loaded onto a Ni-affinity column. Poly-histidine-tagged RORγt-LBD was eluted with raising concentration of imidazole up to 500 mM. The poly-histidine tags were cleaved by thrombin protease at 4 °C for 16 h. After that, the RORγt-LBD was diluted to 50 mM NaCl and purified by ion exchange chromatography with gradient increasing concentration of NaCl up to 1 M. The eluate was concentrated and further applied to size exclusion chromatography using Superdex 75 increase 10/300 GL with PBS buffer (pH 7.4). The peak fractions of the purified proteins were collected and concentrated to 1 mg/ml for SPR experiments. Surface plasmon resonance SPR was employed to measure binding affinities between RORγt-LBD protein and Cop and Ber using Biacore T200 (Cytiva). Cop and Ber were prepared to concentration of 50 mM in 100 % DMSO for complete dissolution. The ligand protein was diluted with 10 mM sodium acetate and immobilized on CM5 chip (Cytiva) at a flow rate of 30 μl/min. Each of Cop and Ber was associated with running buffer containing PBS, 0.005 % Tween20, and 5 % DMSO. The binding of Cop and Ber to RORγt-LBD was assessed across a series of concentrations: 6.25, 12.5, 25, 50, 100, 200, 400, and 800 μM and 3.125, 6.25, 12.5, 25, and 50 μM, respectively. The binding curves for SPR measurement were collected and analyzed by using Biacore T200 Control and Biacore T200 evaluation software. Docking study We employed a structure-based virtual screening through molecular docking approach using the crystal structure of RORγt protein with ligand-binding domain in complex with an inverse agonist (PDB code: 4QM0) in the RCSB PDB homepage ( https://www.rcsb.org/ , December 2021). The protein structure was prepared with the default parameters including hydrogens addition, water deletion, protonation state adjustment and restrained minimization using the Protein Preparation Wizard workflow in Maestro Schrödinger suite version 2024–2 (Schrödinger, LLC, New York, NY) [ 32 ]. After the Receptor Grid Generation step for limiting the ligand-binding site using the original ligand for 4QM0, Glide SP (standard precision) workflow was used for the ligand docking study. Ber and Cop , which showed inhibitory effects on IL-17 production by T cells in in vitro and ex vivo assays, were docked on the crystal structure of RORγt protein using Glide SP. Molecular dynamics study Molecular dynamics (MD) study was performed to investigate the stability of Cop in the binding pocket of RORγt protein (PDB code: 4QM0) for 200 ns using the Desmond software implemented in Maestro [ 32 ]. The protein–ligand complex was prepared with the TIP4P solvent model, the boundary condition of 10 Å distances using Orthorhombic Box Shape function and neutralized using the OPLS4 force field. After solvation, the complex was subjected to equilibrium runs using the Nose-Hoover chain thermostat and Martyna-Tobias-Klein baristat methods with NPT ensemble class [ 34 , 35 ]. The equilibration of the protein–ligand complex was analyzed using the root mean square deviation (RMSD). The interactions between the protein and Cop were monitored using Protein-Ligand contacts and Ligand-Protein contacts plots throughout the simulation time. Results Development of Machine learning models for RORγt inhibitor discovery To identify novel RORγt inhibitors, we first established robust ML models to predict compound activity based on chemical structure ( Fig. 1 A). We explored different modeling strategies, including traditional algorithms using molecular fingerprints and deep learning approaches utilizing molecular graphs. Specifically, random forest (RF) was trained using extended-connectivity fingerprints (ECFP) [ 36 ], while graph neural networks (GNNs), such as graph convolutional network (GCN) [ 37 ], graph attention networks (GAT) [ 38 ], Attentive FP [ 39 ], MPNN [ 40 ], NF [ 41 ], and Weave [ 42 ], were trained directly on molecular graph representations ( Supplementary Table S2 ). Both RF and GNN models were trained to distinguish RORγt inhibitors using the screening dataset of RORγt transcriptional activity (PubChem ID: AID2551) and were also trained on a counter-screen dataset (AID2546) to identify and filter out compounds exhibiting non-specific activity ( Fig. 1 B). Model performance was evaluated three times using distinct scaffold-based splitting, which is described in detail in the Materials and methods section. Comparing the performance on a test set, the GNN-based models, especially AttentiveFP, and Weave, demonstrated superior predictive capability based on AUC-ROC and F1 score for this specific task ( Fig. 1 C and D). Due to their comparatively lower accuracy, RF and MPNN models were excluded from further prediction tasks. To evaluate the robustness and reproducibility of our GCN model predictions, we employed the Integrated Gradients (IG) method for an atom-level importance analysis [ 43 , 44 ]. This interpretability assessment was performed on pairs of structurally similar active compounds identified by Tanimoto similarity to test whether the model learned consistent chemical patterns. The analysis revealed that the GCN model consistently attributed high importance to analogous atoms and substructures across structurally similar molecules ( Supplementary Fig. S1 ). This remarkable consistency in the importance patterns indicates that the model's predictions are based on stable and chemically meaningful features rather than spurious correlations. Fig. 1. Open in a new tab The machine learning model for the discovery of RORγt inhibitors from natural products space. (A) Development of machine learning models for predicting RORγt activity and counter-screening to eliminate non-specific candidates. Both models are built using ensemble approaches that combine graph neural networks (e.g., GCN, GAT, MPNN, Weave, NF) and molecular fingerprint-based random forest classifiers. (B) The pie charts for data distributions, active, inactive, and inconclusive sets for predicting RORγt activity (PubChem ID: AID 2551) and counter-screening dataset (AID 2546), respectively. (C) ROC-AUC plot evaluating model performance against tests by scaffold-splitting. (D) Precision-Recall (PR) curves and F1 scores on the scaffold-split test set. These metrics provide a robust evaluation of model performance, particularly for the imbalanced nature of the dataset. Additionally, to investigate the potential selectivity of Cop for RORγt over its related isoforms, RORα and RORβ, we performed an in silico structural similarity analysis against curated libraries of known active compounds from PubChem. Structural similarity was quantified using the Tanimoto coefficient based on Morgan fingerprints. While the absolute Tanimoto scores were modest, the top-ranked compounds in the RORγt active set shared significant and visually apparent structural scaffolds with Cop ( Supplementary Fig. S2 , top row). In contrast, the most similar compounds identified in the active sets for RORα and RORβ displayed both low similarity scores and a clear lack of structural analogy to Cop ( Supplementary Fig. S2 , middle and bottom rows). These results suggests that Cop is preferentially aligned with established RORγt ligands, providing indirect evidence for its selectivity. Consequently, an ensemble model averaging the predictions from the best-performing models was constructed for the subsequent virtual screening procedure. The corresponding ensemble model trained on the counter-screen data was used for filtering non-specific compounds. Virtual screening of natural products and selection of protoberberine alkaloids We utilized a combination of an ensemble model to selectively identify NPs targeting RORγt and employed an integrative strategy incorporating multiple in silico screening steps to discover potential RORγt inhibitors ( Fig. 2 A). We performed an ML-based virtual screening using our ensemble model against a library of 196,450 processed NPs derived from the LOTUS database [ 30 ]. The top 1,000 high-scoring candidate compounds underwent further in silico analysis via hierarchical chemical classification tool, ClassyFire [ 31 ], and manually inspected in the view of the chemotaxonomy by the comparison with literature ( Fig. 2 B). The structures of 1,000 candidates were classified into 11 superclasses, with the two largest groups being ‘Lipids and lipid-like molecules’ and ‘Phenylpropanoids and polyketides’. These groups were deprioritized due to their low drug-likeness scores according to the drug-likeness properties, such as Lipinski's Rule of Five and QED scores ( Supplementary Files 2 ), as well as lack of novelty. Moreover, lipid-like compounds have been reported to bind members of the ROR subfamily, such as RORβ, suggesting potential off-target effects [ 45 , 46 ]. Similarly, compounds in the ‘Phenylpropanoids and polyketides’ class, many of which are terpenoid or steroid derivatives such as hydroxycholesterol, ursolic acid, and digoxin, have been shown to act as agonists or inverse agonists of RORα and RORγt [ [47] , [48] , [49] ]. We decided to focus on the ‘Alkaloid and derivatives’ superclass, known for its diverse pharmacological roles, and manually investigated the activity records of compounds within its subclasses. Using the NPASS database ( https://bidd.group/NPASS/ ), which links chemotaxonomy information with activity records, we examined targets of compounds with more than 30 activity records. Among these, we focused on ‘Protoberberine alkaloids and derivatives,’ which have not been studied in relation to RORγt and represent the largest subset within the ‘Alkaloid and derivatives’ superclass ( Fig. 2 C). Protoberberine alkaloids are recognized as bioactive compounds with low toxicity and have been experimentally validated in various disease models [ [50] , [51] , [52] , [53] ]. Accordingly, we selected six commercially available protoberberine alkaloids for further evaluation of their biological activity related to RORγt ( Fig. 2 D). Fig. 2. Open in a new tab Chemical classification and structures of candidate compounds. (A) Screening workflow of natural product candidates from compound libraries LOTUS. (B) Distribution of chemical types among all compounds suggested by machine learning. (C) Distribution of chemical types within the ‘Alkaloids and derivatives’ category. (D) Chemical structures of the six selected protoberberine alkaloids: phellodendrine ( 1 ), palmatine ( 2 ), coptisine ( 3 ), columbamine ( 4 ), jatrorrhizine ( 5 ), and berberine ( 6 ). Inhibitory effects of protoberberine alkaloids on Th17 cell differentiation in vitro To investigate the potential of the six protoberberine alkaloids as RORγt inhibitors, each compound was tested for its effect on Th17 differentiation in vitro . Naive CD4 + T cells isolated from Il17a EGFP reporter mice were subjected to Th17 polarization with titrated doses of each compound ( Fig. 3 A). Among the six protoberberine alkaloids, berberine ( Ber ) demonstrated the most profound inhibition of IL-17-EGFP expression in CD4 + T cells on day 4 after culture in a dose dependent manner. Coptisine ( Cop ) also significantly inhibited IL-17-EGFP expression at 3.3 μM as well as at 10 μM, while palmatine marginally inhibited IL-17-EGFP expression at 10 μM ( Fig. 3 B and C). To confirm the inhibitory effect of Cop and Ber at the protein level, naive CD4 + T cells isolated from C57BL/6 mice were subjected to Th17 differentiation in the presence of two-fold serially diluted compounds ( Fig. 3 D). Flow cytometric analysis of IL-17 expression in CD4 + T cells confirmed the dose-dependent inhibition of Th17 differentiation by Cop or Ber treatment at the single cell level ( Fig. 3 E and F). When the IL-17 levels in the culture supernatant were measured by ELISA, Cop and Ber resulted in the IC 50 values of 1.859 μM and 0.8561 μM, respectively ( Fig. 3 G and H). To investigate whether the inhibitory effects of Cop and Ber were specific to Th17 differentiation, their effects on Th1 differentiation were tested. Intriguingly, Cop and Ber did not significantly inhibit Th1 differentiation at all tested doses, suggesting that the inhibitory effects of Cop and Ber are Th17-specific ( Supplementary Fig. S3 ). Collectively, these results suggest that Cop and Ber can inhibit Th17 differentiation in vitro by potentially regulating the function of RORγt. Fig. 3. Open in a new tab Initial screening of six protoberberine compounds in Th17 differentiation in vitro . (A-C) Naive CD4 + T cells isolated from Il17a EGFP reporter mice were stimulated with plate-bound anti-CD3ε (1.0 µg/ml) and anti-CD28 (1.0 µg/ml) antibodies in the presence of recombinant human TGF-β (0.5 ng/ml) and recombinant mouse IL-6 (50 ng/ml). Titrated doses of candidate compounds or vehicle control (DMSO) were added from day 0. On day 4, the frequencies of EGFP + cells were analyzed by flow cytometry. (A) Schematic representation of the experimental design. (B) Representative histograms of IL-17-EGFP expression in CD4 + T cells treated with 10 µM of each compound. (C) Quantification of IL-17-EGFP + cell frequencies. (D-H) Naive CD4 + T cells isolated from C57BL/6 mice were cultured under Th17-polarizing conditions as described in (D). On day 4, expression of IL-17 and IFN-γ was assessed by flow cytometry (E), and the frequencies of IL-17 + cells are shown (F). (G, H) IC 50 values for Cop (G) and Ber (H) were calculated using Graphpad Prism 6. phellodendrine ( Phe ), palmatine ( Pal ), coptisine ( Cop ), columbamine ( Col ), jatrorrhizine ( Jat ), and berberine ( Ber ). Data are representative of at least two independent experiments and are presented as mean ± SD. Statistical significance was calculated by one-way ANOVA (n = 3) for Fig. 3 C and 3F, and by two-way ANOVA (n = 3) for Fig. 3 G and 3H. *p < 0.05, **p < 0.01, ***p < 0.005, ****p < 0.001. Coptisine and berberine inhibit expansion of antigen-specific Th17 cells To investigate whether Cop and Ber impact RORγt function in differentiated antigen (Ag)-specific Th17 cells, we benefit from Il17a Cre x Rosa26 YFP Th17 fate mapping mice system, in which Th17 cells are permanently labeled with yellow fluorescent protein (YFP) regardless of current Il17a gene expression if they have turned on Il17a expression once in a lifetime. Eight days after immunization with myelin oligodendrocyte glycoprotein 35–55 (MOG 35-55 ) peptides in CFA, lymphocytes isolated from draining lymph nodes were restimulated with MOG 35-55 peptides and recombinant IL-23 to enrich and expand MOG-specific Th17 cells. Titrated doses of Cop or Ber were added to the culture to investigate their effects on MOG-specific Th17 cell expansion ( Fig. 4 A). After 5 days of restimulation, flow cytometric analysis revealed a significant reduction in the frequencies of IL-17 fate-mapped cells (IL-17 fm-YFP + ) as well as currently IL-17 expressing cells (IL-17 + ) after treatment with either Cop or Ber in a dose dependent manner ( Fig. 4 B–D). Consistent with these results, a dose-dependent reduction in IL-17 production in the culture supernatant was also observed after Cop or Ber treatment ( Fig. 4 E). These results suggest that Cop and Ber inhibit the expansion of already differentiated Ag-specific Th17 cells by presumably regulating the function of RORγt. Fig. 4. Open in a new tab Coptisine ( Cop ) and berberine ( Ber ) suppress reactivation of antigen-specific Th17 cells ex vivo . Single-cell suspensions were prepared from the draining inguinal lymph nodes of Il17a Cre x Rosa26 YFP fate mapping mice immunized with MOG 35-55 in CFA. Cells were restimulated with MOG 35-55 peptides and recombinant mouse IL-23 in the presence of titrated doses of Cop , Ber or vehicle control (DMSO) for 5 days. As a negative control (Nil group), some cells were cultured without MOG 35-55 peptides and IL-23. (A) Illustration of the experimental procedure. (B) On day 5, expression of IL-17 fm-YFP and IL-17, in combination with CD44, was analyzed in gated CD4 + T cells by flow cytometry and the representative contour plots are shown. (C, D) The frequencies of CD44 + YFP + cells (C) and CD44 + IL-17 + cells (D) are shown. (E) IL-17 levels in the culture supernatant were measured by ELISA on day 5. Data are representatives of two independent experiments and are presented as mean ± SD. Statistical significance was calculated by one-way ANOVA (n = 3) for Fig. 4 C–E. *p < 0.05, **p < 0.01, ***p < 0.005, ****p < 0.001. Coptisine and berberine directly bind to RORγt To confirm whether the Th17-modulating activities of Cop and Ber were caused by direct binding to RORγt, surface plasmon resonance (SPR) was employed ( Fig. 5 A and B). SPR results clearly showed that both Cop and Ber were directly bound to RORγt-ligand binding domain (LBD) with equilibrium dissociation constants (K D ) of 59.7 and 108.1 μM, respectively. RORγt-LBD exhibited a slightly stronger (approximately 2-fold) binding affinity for Cop than Ber . The different binding affinity between Cop and Ber for RORγt-LBD was attributed to the association rate constant (k a ) rather than the dissociation rate constant (k d ). Both k d values (1/s) were nearly identical (4.691E-03 for Cop and 4.138E-03 for Ber ), whereas k a values (1/Ms) were distinguished from each other (78.54 for Cop and 38.28 for Ber ). Fig. 5. Open in a new tab Surface plasmon resonance measurement analysis and docking studies of coptisine ( Cop ) and berberine ( Ber ) with RORγt. (A, B) The binding curves of surface plasmon resonance measurement on immobilized CM5 chip at different concentrations of Cop (A) and Ber (B). K d values of Cop and Ber are 59.7 and 108.1 μM, respectively. (C, D) The 3D interactions between RORγt (PDB: 4QM0) and Cop (C) or Ber (D) complex are visualized by Maestro Schrödinger suite. Cop formed stable pi-pi and pi-cation interactions with two key residues (Phe388 and His479). Ber had the pi-bonds with His479. pi–pi stackings are shown as cyan dashed lines, and pi cation interaction is shown as a dark green dashed line. The cutoff distance for key interactions was set at 3 Å. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) Docking study and molecular dynamics simulation We conducted docking studies to investigate whether Cop and Ber effectively bind to the ligand-binding site of RORγt (PDB: 4QM0), which is the crystal structure of the human RORγt-LBD in complex with a tertiary sulfonamide that serves as an inhibitor with lower lipophilicity and higher selectivity than other candidates ( Fig. 5 C and D) [ 54 ]. Cop formed 4 pi interactions with Phe388 and His479, which were the binding residues of other inhibitors [ 5 ]. A stable pi-cation interaction formed between Phe388 and the nitrogen atom within Cop [ 55 , 56 ]. Furthermore, the unique pi-pi stacking interactions of two aromatic rings and the dioxolane ring with His479 were considered to potentially stabilize the interaction network [ 56 ]. In contrast, Ber , possessing dimethoxy residues instead of the dioxolane ring found in Cop , exhibited a pi-pi interaction between the ring with dimethoxy group and His479. However, this conformation resulted in an increased distance between the nitrogen atoms in Ber and Phe388. Next, we performed a molecular dynamics study of Cop to better understand the interaction with RORγt. During the simulation time of 1,000 ns, the protein–ligand complex remained stable ( Supplementary Fig. S4 A). The protein backbone RMSD against Cop showed only minor fluctuations, remaining below 1.5 Å at a steady state. Analysis of the interaction fraction between protein residues and Cop revealed consistent interaction with Phe388 and His479, averaging 31 % and 35 % contact strength, respectively, over the total simulation time ( Supplementary Fig. S4 B and C). Interestingly, we observed a variation in the ligand RMSD around 300 ns. During 0–300 ns, the interaction with Phe388 was dominant with an average contact strength of 75 % ( Supplementary Fig. S4 D and E). However, between 400 and 1,000 ns, the pi-pi interaction with His479 reached an average contact strength of 53 % ( Supplementary Fig. S4 F and G). These results are consistent with the docking study indicating that both Phe388 and His479 are the main interaction residues with Cop . Coptisine treatment ameliorates the induction and progression of psoriasis-like skin inflammation Having confirmed that Cop is a potent inhibitor of RORγt, we set out to investigate whether it can regulate RORγt function in a Th17-driven autoimmune disease. To this end, we evaluated the therapeutic efficacy of Cop in the imiquimod-induced psoriasis-like skin inflammation prevention model ( Fig. 6 A, upper panel). During the induction phase, Cop treatment led to a notable reduction in the development of psoriasis-like skin inflammation in the ear skin, as evidenced by a significant decrease in ear thickness, redness, and scale formation ( Fig. 6 B and C). Flow cytometric analysis of lymphocytes revealed that Cop treatment significantly reduced the absolute cell number of IL-17 producing T cells, including Th17 and γδT17 cells, in the ear skin and draining lymph nodes ( Fig. 6 D–G). Consistent with these findings, IL-17 production in the culture supernatant of draining lymph node cells after overnight restimulation was remarkably diminished in the Cop -treated group compared to the vehicle-treated group ( Fig. 6 H). These results suggest that Cop treatment prevented the induction of psoriasis-like skin inflammation by inhibiting the function of RORγt + IL-17-producing T cells. Fig. 6. Open in a new tab Coptisine ( Cop ) treatment ameliorates the induction and progression of psoriasis-like skin inflammation. Imiquimod was daily applied to both ears of C57BL/6 mice were treated daily with imiquimod applied to both ears for six consecutive days. Cop (1 mg) was administered intraperitoneally once daily, starting from either day 0 (prevention model) or day 2 (therapeutic model) following disease induction. Vehicle control (DMSO) was intraperitoneally injected in the same volume. Changes in ear thickness were monitored daily. Mice were sacrificed for analysis on day 6. (A) Schematic diagram of the imiquimod-induced psoriasis-like skin inflammation models. (B, C) Changes in left ear thickness (B) and representative images of mice on day 6 in the prevention model (C). (D-G) Absolute numbers of γδT17 cells (D) and Th17 cells (E) in the ear skin, and γδT17 cells (F) and Th17 cells (G) in the draining lymph nodes (dLNs), analyzed by flow cytometry. (H) IL-17 levels in the culture supernatant of dLN cells restimulated with PMA and ionomycin for 24 h were measured by ELISA. (I) Changes in left ear thickness in the therapeutic model. (J, K) Absolute numbers of γδT17 cells (J) and Th17 cells (K) in the ear skin were analyzed by flow cytometry. Data are representatives of two independent experiments and are presented as mean ± SD. Statistical significance was calculated by two-way ANOVA (n = 6) for Fig. 6 B and I, and by Student’s t -test for Fig. 6 D–H, J, and K. *p < 0.05, **p < 0.01, ***p < 0.005, ****p < 0.001. To evaluate the therapeutic potential of Cop in the treatment of psoriasis, Cop was administered 2 days after the induction of psoriasis-like skin inflammation ( Fig. 6 A, lower panel). Treatment with Cop significantly delayed the progression of ear thickening ( Fig. 6 I). Consistently, the numbers of γδT17 cells and Th17 cells in the ear were markedly reduced following Cop treatment ( Fig. 6 J and K). Histological examination revealed that Cop treatment markedly attenuated epidermal hyperplasia and dermal inflammatory cell infiltration, accompanied by a significant reduction in ear thickness compared with vehicle group ( Supplementary Fig. S5 A and B). Furthemore, immunohistochemical analysis demonstrated that the Cop -treated group displayed substantially fewer Ki-67 + cells in both the epidermis and dermal infiltrates ( Supplementary Fig. S6 A and B). Taken together, these findings suggest that Cop may inhibit RORγt and modulate the function of IL-17-producing T cells in vivo . Discussion The discovery of novel, and effective RORγt inhibitors remains a critical goal for treating Th17-mediated autoimmune diseases like psoriasis. This study successfully demonstrates an integrated computational and experimental workflow for identifying candidate compounds from the complex chemical space of NPs. By combining ML-based virtual screening, chemotaxonomic analysis, molecular docking, molecular dynamics, and biological validation, we identified two protoberberine alkaloids, Cop and Ber , as direct RORγt inhibitors with therapeutic potential. To ensure a well-curated dataset for the construction of ML model, we collected RORγt-related bioassay datasets from the PubChem repository. The largest dataset (AID: 2551) had an imbalanced ratio of inactive molecules, which we addressed by randomizing chemical similarities to achieve a more balanced dataset with active molecules. Additionally, to exclude compounds that may yield false positive results by binding to the Gal4 DNA binding domain or the UAS luciferase reporter, we employed a dataset (AID 2546) using a vector with the VP16 transactivation domain instead of the RORγt transactivation domain to develop the ML model for selecting candidates that inhibit RORγt. To validate the reliability of the GCN model's predictions, we systematically assessed its interpretability using the IG method. This analysis provided strong evidence for the model's robustness and the reproducibility of its predictions, suggesting that it learns meaningful and generalizable chemical representations that support the stability of the GCN framework. Therefore, this interpretability analysis complements the model's predictive performance metrics, offering an additional layer of validation for the trustworthiness of the machine learning workflow applied in this study. A key contribution of our approach was the subsequent filtering and prioritization of ML hits using chemotaxonomy. The structural diversity of NPs can pose challenges for the finding of novel candidates on the specific targets. Many natural products share common structural features derived from representative biosynthetic pathways. Grouping compounds based on similar core structures can serve as an effective strategy for identifying candidates with comparable bioactivities. The chemotaxonomic approach helped navigate the diversity of predicted NP hits and focusing experimental resources efficiently. By classifying the top 1,000 ML candidates, we could strategically deprioritize scaffolds with poor drug-like properties (e.g., lipids) or those containing known RORγt modulators with liabilities (e.g., certain terpenoids/steroids related to digoxin/ursolic acid). Through this rational filtering process, we prioritized protoberberine alkaloids, which are a class with known bioactivities but relatively unexplored potential against RORγt. Considering that the structural diversity of protoberberine derivatives is largely determined by the presence and position of hydroxyl, methoxy groups or 1,3-dioxolane rings on both benzene rings, we selected six representative protoberberine derivatives with these structural modifications for further study. Among six protoberberine derivatives, both Cop and Ber demonstrated potent and selective inhibition of Th17 differentiation in vitro . SPR experiments unequivocally established their direct binding to the RORγt ligand-binding domain, confirming them as RORγt ligands. Cop exhibited a higher binding affinity (K D ∼60 µM) than Ber (K D ∼108 µM), consistent with molecular docking and dynamics simulations suggesting potentially more favorable interactions for Cop , particularly involving pi-pi stacking with His479 and pi-cation interactions with Phe388. Although a transition was observed around 300 ns, the RMSD between the protein and ligand contacts remained stable throughout the 1,000 ns simulation, indicating that this shift does not result from conformational instability. Instead, the temporal transition between Phe388- and His479-dominated binding modes suggests an induced-fit binding mechanism, in which Cop adapts to local conformational changes within the RORγt binding site. This dynamic may contribute to an extended residence time and potential allosteric effects, highlighting the complex interplay between ligand flexibility and protein plasticity. Ber is a well-known natural product with diverse pharmacological activities, including inhibition of Th17 differentiation [ [57] , [58] , [59] , [60] , [61] ]. However, the underlying mechanism has remained unclear. Our study provides strong evidence that Ber directly engages RORγt, identifying a likely mechanism responsible for its effects on Th17 cell differentiation. To our knowledge, this is the first demonstration of direct RORγt binding by Ber , offering mechanistic insight into its selective action on Th17 differentiation [ 62 ]. Nonetheless, we do not rule out the possibility that RORγt binding is not solely responsible for Th17 inhibition, as previous studies have reported that Ber can also interact with TGFBR1 and JAK2, both of which are critical regulators of Th17 differentiation [ 62 ]. Cop , being less extensively studied, represents a more novel RORγt-targeting scaffold identified through our workflow. Based on its stronger binding affinity compared to Ber , along with its relative novelty, Cop was advanced to in vivo testing. In the preclinical imiquimod-induced psoriasis model, Cop demonstrated significant therapeutic efficacy, reducing skin inflammation and suppressing IL-17 production by γδT cells and CD4 + T cells. This in vivo activity strongly supports the potential of Cop as a therapeutic candidate for psoriasis and potentially other RORγt + T cell-driven autoimmune diseases. Pharmacokinetic (PK) studies of Cop have reported that its oral bioavailability is low, mainly due to poor membrane permeability and rapid metabolism [ 63 , 64 ]. Intraperitoneal (i.p.) administration may achieve higher plasma concentrations than oral delivery. Therefore, further PK studies following i.p. administration in an in vivo psoriasis animal model are needed to address these bioavailability limitations. In addition, the development of Cop derivatives or novel formulations will be important to overcome its poor bioavailability and enhance its therapeutic potential. Conclusions In summary, this study validates an efficient, multi-faceted strategy combining advanced ML, chemotaxonomic knowledge, and targeted experimental work for NP-based drug discovery. We successfully identified and characterized Cop as a novel, direct-acting RORγt inhibitor with significant in vivo efficacy against psoriasis-like skin inflammation. Although the binding affinities of Ber and Cop were not sufficiently potent for direct drug development, confirmation of their interaction with RORγt highlights the potential for developing natural product-based RORγt inhibitors through structure-guided optimization and derivative design. This work not only delivers a promising lead compound but also provides a robust approach applicable to discovering modulators for other challenging biological targets by leveraging the rich chemical diversity of natural products. Further investigation of the PK properties and structural optimization of Cop derivatives is warranted to comprehensively evaluate their therapeutic potential. Compliance with ethics requirements All the animal procedures were conducted in accordance with the animal use guidelines approved by the Institutional Animal Care and Use Committee of Seoul National University (SNU-191210-3-11), Lee Gil Ya Cancer and Diabetes Institute (LCDI-2022-0088), and Incheon National University (INU-ANIM-2024-0026-01). Declaration of generative AI and AI-assisted technologies in the writing process During the preparation of this work the authors used chatGPT4 to improve text flow and clarity, and simplify technical language. After using this service, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication. CRediT authorship contribution statement Hojin Yoo: Conceptualization, Data curation, Formal analysis, Software, Visualization, Writing – original draft, Writing – review & editing. Sang-Jun Han: Investigation, Validation, Writing – original draft. Jeong-Eun Lee: Investigation, Validation. Chaeyeon Cho: Formal analysis, Methodology. Donggyun Hong: Formal analysis, Methodology. Birang Jeong: Formal analysis, Methodology. Sijin Kim: Investigation. Go-Yeon Jung: Investigation. Minjeong Ma: Investigation. Soeun Jung: Investigation. Beomjun Park: Data curation, Formal analysis, Software. Namgil Lee: Data curation, Formal analysis, Software. Hee-Seop Yoo: Investigation, Validation. Kwang-Jin Cho: Writing – review & editing. Min-Duk Seo: Conceptualization, Writing – review & editing. Yeonseok Chung: Conceptualization, Writing – review & editing. Byung-Seok Kim: Conceptualization, Writing – original draft, Writing – review & editing. Heejung Yang: Conceptualization, Methodology, Writing – original draft, Writing – review & editing. Declaration of 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. Acknowledgements This work was supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF) grant (Project No. RS-2021-NR061647, RS-2021-NR061691 and 2020R1A3B207889011) funded by the Ministry of Education, Science and Technology, Republic of Korea. This work was also supported by the Incheon National University Research Grant in 2024 (2024-0243). The schematic illustrations in Fig. 3A, 3D, 4A, and 6A were created with BioRender ( http://www.biorender.com ). Footnotes Appendix A Supplementary data to this article can be found online at https://doi.org/10.1016/j.jare.2025.09.004 . Contributor Information Min-Duk Seo, Email: mdseo@ajou.ac.kr. Yeonseok Chung, Email: yeonseok@snu.ac.kr. Byung-Seok Kim, Email: byungseokkim@inu.ac.kr. Heejung Yang, Email: heejyang@kangwon.ac.kr. Appendix A. Supplementary data The following are the Supplementary data to this article: Supplementary Data 1 mmc1.docx (5.3MB, docx) Supplementary Data 2 mmc2.xlsx (1MB, xlsx) Supplementary Data 3 mmc3.xlsx (7.9MB, xlsx) References 1. Skayem C., Taieb C., Halioua B., Baissac C., Saint A.M. Epidemiology of psoriasis: a worldwide global study. Acta Derm Venereol. 2025;105 doi: 10.2340/actadv.v105.42945. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 2. Griffiths C.E.M., Armstrong A.W., Gudjonsson J.E., Barker J.N.W.N. Psoriasis. Lancet. 2021;397:1301–1315. doi: 10.1016/S0140-6736(20)32549-6. [ DOI ] [ PubMed ] [ Google Scholar ] 3. Gupta M.A., Gupta A.K. Psychiatric and psychological co-morbidity in patients with dermatologic disorders: epidemiology and management. Am J Clin Dermatol. 2003;4:833–842. doi: 10.2165/00128071-200304120-00003. [ DOI ] [ PubMed ] [ Google Scholar ] 4. Wang R., Campbell S., Amir M., Mosure S.A., Bassette M.A., Eliason A., et al. Genetic and pharmacological inhibition of the nuclear receptor RORα regulates TH17 driven inflammatory disorders. Nat Commun. 2021;12:76. doi: 10.1038/s41467-020-20385-9. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Zhang Y., Luo X., Wu D., Xu Y. ROR nuclear receptors: structures, related diseases, and drug discovery. Acta Pharmacol Sin. 2015;36:71–87. doi: 10.1038/aps.2014.120. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Ciofani M., Madar A., Galan C., Sellars M., Mace K., Pauli F., et al. A validated regulatory network for Th17 cell specification. Cell. 2012;151:289–303. doi: 10.1016/j.cell.2012.09.016. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. Hirota K., Hashimoto M., Yoshitomi H., Tanaka S., Nomura T., Yamaguchi T., et al. T cell self-reactivity forms a cytokine milieu for spontaneous development of IL-17+ Th cells that cause autoimmune arthritis. J Exp Med. 2007;204:41–47. doi: 10.1084/jem.20062259. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. Sun N., Guo H., Wang Y. Retinoic acid receptor-related orphan receptor gamma-t (RORγt) inhibitors in clinical development for the treatment of autoimmune diseases: a patent review (2016-present) Expert Opin Ther Pat. 2019;29:663–674. doi: 10.1080/13543776.2019.1655541. [ DOI ] [ PubMed ] [ Google Scholar ] 9. Tan J., Liu H., Huang M., Li N., Tang S., Meng J., et al. Small molecules targeting RORγt inhibit autoimmune disease by suppressing Th17 cell differentiation. Cell Death Dis. 2020;11:697. doi: 10.1038/s41419-020-02891-2. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. Gege C. RORγt inhibitors as potential back-ups for the phase II candidate VTP-43742 from vitae pharmaceuticals: patent evaluation of WO2016061160 and US20160122345. Expert Opin Ther Pat. 2017;27:1–8. doi: 10.1080/13543776.2017.1262350. [ DOI ] [ PubMed ] [ Google Scholar ] 11. Nakajima R., Oono H., Sugiyama S., Matsueda Y., Ida T., Kakuda S., et al. Discovery of [1,2,4]triazolo[1,5- a ]pyridine derivatives as potent and orally bioavailable RORγt inverse agonists. ACS Med Chem Lett. 2020;11:528–534. doi: 10.1021/acsmedchemlett.9b00649. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 12. Cherney R.J., Cornelius L.A.M., Srivastava A., Weigelt C.A., Marcoux D., Duan J.-J.-W., et al. Discovery of BMS-986251: a clinically viable, potent, and selective RORγt inverse agonist. ACS Med Chem Lett. 2020;11:1221–1227. doi: 10.1021/acsmedchemlett.0c00063. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Huh J.R., Englund E.E., Wang H., Huang R., Huang P., Rastinejad F., et al. Identification of potent and selective diphenylpropanamide RORγ inhibitors. ACS Med Chem Lett. 2013;4:79–84. doi: 10.1021/ml300286h. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Solt L.A., Kumar N., Nuhant P., Wang Y., Lauer J.L., Liu J., et al. Suppression of TH17 differentiation and autoimmunity by a synthetic ROR ligand. Nature. 2011;472:491–494. doi: 10.1038/nature10075. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 15. Atanasov A.G., Zotchev S.B., Dirsch V.M. International natural product sciences taskforce, Supuran CT. Natural products in drug discovery: advances and opportunities. Nat Rev Drug Discov. 2021 doi: 10.1038/s41573-020-00114-z. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 16. Rodrigues T., Reker D., Schneider P., Schneider G. Counting on natural products for drug design. Nat Chem. 2016;8:531–541. doi: 10.1038/nchem.2479. [ DOI ] [ PubMed ] [ Google Scholar ] 17. Newman D.J., Cragg G.M. Natural products as sources of new drugs over the nearly four decades from 01/1981 to 09/2019. J Nat Prod. 2020;83:770–803. doi: 10.1021/acs.jnatprod.9b01285. [ DOI ] [ PubMed ] [ Google Scholar ] 18. Harvey A.L., Edrada-Ebel R., Quinn R.J. The re-emergence of natural products for drug discovery in the genomics era. Nat Rev Drug Discov. 2015;14:111–129. doi: 10.1038/nrd4510. [ DOI ] [ PubMed ] [ Google Scholar ] 19. Jiménez-Luna J., Grisoni F., Schneider G. Drug discovery with explainable artificial intelligence. Nat Mach Intell. 2020;2:573–584. doi: 10.1038/s42256-020-00236-4. [ DOI ] [ Google Scholar ] 20. Vamathevan J., Clark D., Czodrowski P., Dunham I., Ferran E., Lee G., et al. Applications of machine learning in drug discovery and development. Nat Rev Drug Discov. 2019;18:463–477. doi: 10.1038/s41573-019-0024-5. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 21. Lo Y.-C., Rensi S.E., Torng W., Altman R.B. Machine learning in chemoinformatics and drug discovery. Drug Discov Today. 2018;23:1538–1546. doi: 10.1016/j.drudis.2018.05.010. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 22. Rodríguez-Pérez R., Miljković F., Bajorath J. Machine learning in chemoinformatics and medicinal chemistry. Annu Rev Biomed Data Sci. 2022;5:43–65. doi: 10.1146/annurev-biodatasci-122120-124216. [ DOI ] [ PubMed ] [ Google Scholar ] 23. Bertoni M., Duran-Frigola M., Badia-i-Mompel P., Pauls E., Orozco-Ruiz M., Guitart-Pla O., et al. Bioactivity descriptors for uncharacterized chemical compounds. Nat Commun. 2021;12:3932. doi: 10.1038/s41467-021-24150-4. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 24. Kim S., Chen J., Cheng T., Gindulyte A., He J., He S., et al. PubChem 2025 update. Nucleic Acids Res. 2025;53:D1516–D1525. doi: 10.1093/nar/gkae1059. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 25. Greg Landrum, Paolo Tosco, Brian Kelley, Ricardo Rodriguez, David Cosgrove, Riccardo Vianello, et al. rdkit/rdkit: 2025_03_1 (Q1 2025) Release 2025. https://doi.org/10.5281/ZENODO.591637. 26. Li M., Zhou J., Hu J., Fan W., Zhang Y., Gu Y., et al. DGL-LifeSci: an open-source toolkit for deep learning on graphs in life science. ACS Omega. 2021;6:27233–27238. doi: 10.1021/acsomega.1c04017. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 27. Wang M, Zheng D, Ye Z, Gan Q, Li M, Song X, et al. Deep Graph Library: A Graph-Centric, Highly-Performant Package for Graph Neural Networks 2019. https://doi.org/10.48550/ARXIV.1909.01315. 28. Pedregosa F., Varoquaux G., Gramfort A., Michel V., Thirion B., Grisel O., et al. Scikit-learn: machine learning in python. J Mach Learn Res. 2011;12:2825–2830. doi: 10.5555/1953048.2078195. [ DOI ] [ Google Scholar ] 29. Wu Z., Ramsundar B., Feinberg E.N., Gomes J., Geniesse C., Pappu A.S., et al. MoleculeNet: a benchmark for molecular machine learning. Chem Sci. 2018;9:513–530. doi: 10.1039/C7SC02664A. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 30. Rutz A., Sorokina M., Galgonek J., Mietchen D., Willighagen E., Gaudry A., et al. The LOTUS initiative for open knowledge management in natural products research. Elife. 2022;11 doi: 10.7554/eLife.70780. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 31. Djoumbou Feunang Y., Eisner R., Knox C., Chepelev L., Hastings J., Owen G., et al. ClassyFire: automated chemical classification with a comprehensive, computable taxonomy. J Cheminf. 2016;8 doi: 10.1186/s13321-016-0174-y. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 32. Bowers KJ, Chow DE, Xu H, Dror RO, Eastwood MP, Gregersen BA, et al. Scalable Algorithms for Molecular Dynamics Simulations on Commodity Clusters. ACM/IEEE SC 2006 Conference (SC’06), Tampa, FL: IEEE; 2006, p. 43–43. https://doi.org/10.1109/SC.2006.54. 33. Choi G., Park Y.-J., Cho M., Moon H., Kim D., Kang C.-Y., et al. A critical role for Th17 cell-derived TGF-β1 in regulating the stability and pathogenicity of autoimmune Th17 cells. Exp Mol Med. 2021;53:993–1004. doi: 10.1038/s12276-021-00632-9. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 34. Martyna G.J., Tobias D.J., Klein M.L. Constant pressure molecular dynamics algorithms. J Chem Phys. 1994;101:4177–4189. doi: 10.1063/1.467468. [ DOI ] [ Google Scholar ] 35. Nosé S. A unified formulation of the constant temperature molecular dynamics methods. J Chem Phys. 1984;81:511–519. doi: 10.1063/1.447334. [ DOI ] [ Google Scholar ] 36. Rogers D., Hahn M. Extended-connectivity fingerprints. J Chem Inf Model. 2010;50:742–754. doi: 10.1021/ci100050t. [ DOI ] [ PubMed ] [ Google Scholar ] 37. Kipf TN, Welling M. Semi-Supervised Classification with Graph Convolutional Networks. arXiv:160902907 [Cs, Stat] 2017. 38. Veličković P, Cucurull G, Casanova A, Romero A, Liò P, Bengio Y. Graph Attention Networks 2017. https://doi.org/10.48550/ARXIV.1710.10903. 39. Xiong Z., Wang D., Liu X., Zhong F., Wan X., Li X., et al. Pushing the boundaries of molecular representation for drug discovery with the graph attention mechanism. J Med Chem. 2020;12 doi: 10.1021/acs.jmedchem.9b00959. [ DOI ] [ PubMed ] [ Google Scholar ] 40. Gilmer J, Schoenholz SS, Riley PF, Vinyals O, Dahl GE. Neural Message Passing for Quantum Chemistry. arXiv:170401212 [Cs] 2017. 41. Duvenaud D, Maclaurin D, Aguilera-Iparraguirre J, Gómez-Bombarelli R, Hirzel T, Aspuru-Guzik A, et al. Convolutional Networks on Graphs for Learning Molecular Fingerprints. arXiv:150909292 [Cs, Stat] 2015. 42. Kearnes S., McCloskey K., Berndl M., Pande V., Riley P. Molecular graph convolutions: moving beyond fingerprints. J Comput Aided Mol Des. 2016;30:595–608. doi: 10.1007/s10822-016-9938-8. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 43. Lundstrom D, Huang T, Razaviyayn M. A Rigorous Study of Integrated Gradients Method and Extensions to Internal Neuron Attributions 2022. https://doi.org/10.48550/ARXIV.2202.11912. 44. Sundararajan M, Taly A, Yan Q. Axiomatic Attribution for Deep Networks 2017. https://doi.org/10.48550/ARXIV.1703.01365. 45. Stehlin-Gaon C., Willmann D., Zeyer D., Sanglier S., Van Dorsselaer A., Renaud J.-P., et al. All-trans retinoic acid is a ligand for the orphan nuclear receptor RORβ. Nat Struct Mol Biol. 2003;10:820–825. doi: 10.1038/nsb979. [ DOI ] [ PubMed ] [ Google Scholar ] 46. Stehlin C. X-ray structure of the orphan nuclear receptor RORbeta ligand-binding domain in the active conformation. EMBO J. 2001;20:5822–5831. doi: 10.1093/emboj/20.21.5822. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 47. Kallen J., Schlaeppi J.-M., Bitsch F., Delhon I., Fournier B. Crystal structure of the human RORα ligand binding domain in complex with cholesterol sulfate at 2.2 Å. J Biol Chem. 2004;279:14033–14038. doi: 10.1074/jbc.M400302200. [ DOI ] [ PubMed ] [ Google Scholar ] 48. Kallen J.A., Schlaeppi J.-M., Bitsch F., Geisse S., Geiser M., Delhon I., et al. X-Ray structure of the hRORα LBD at 1.63 Å. Structure. 2002;10:1697–1707. doi: 10.1016/S0969-2126(02)00912-7. [ DOI ] [ PubMed ] [ Google Scholar ] 49. Jin L., Martynowski D., Zheng S., Wada T., Xie W., Li Y. Structural basis for hydroxycholesterols as natural ligands of orphan nuclear receptor RORγ. Mol Endocrinol. 2010;24:923–929. doi: 10.1210/me.2009-0507. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 50. Grycová L., Dostál J., Marek R. Quaternary protoberberine alkaloids. Phytochemistry. 2007;68:150–175. doi: 10.1016/j.phytochem.2006.10.004. [ DOI ] [ PubMed ] [ Google Scholar ] 51. Vennerstrom J.L., Klayman D.L. Protoberberine alkaloids as antimalarials. J Med Chem. 1988;31:1084–1087. doi: 10.1021/jm00401a006. [ DOI ] [ PubMed ] [ Google Scholar ] 52. Zhang Y., Yu Y., Yan X., Wang W., Tian X., Wang L., et al. Different structures of berberine and five other protoberberine alkaloids that affect P-glycoprotein-mediated efflux capacity. Acta Pharmacol Sin. 2019;40:133–142. doi: 10.1038/s41401-018-0183-7. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 53. Inoue N., Terabayashi T., Takiguchi-Kawashima Y., Fujinami D., Matsuoka S., Kawano M., et al. The benzylisoquinoline alkaloids, berberine and coptisine, act against camptothecin-resistant topoisomerase I mutants. Sci Rep. 2021;11:7718. doi: 10.1038/s41598-021-87344-2. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 54. Fauber B.P., René O., De Leon B.G., Burton B., Deng Y., Eidenschenk C., et al. Reduction in lipophilicity improved the solubility, plasma–protein binding, and permeability of tertiary sulfonamide RORc inverse agonists. Bioorg Med Chem Lett. 2014;24:3891–3897. doi: 10.1016/j.bmcl.2014.06.048. [ DOI ] [ PubMed ] [ Google Scholar ] 55. Yang T., Liu Q., Cheng Y., Cai W., Ma Y., Yang L., et al. Discovery of tertiary amine and indole derivatives as potent RORγt inverse agonists. ACS Med Chem Lett. 2014;5:65–68. doi: 10.1021/ml4003875. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 56. Fauber B.P., de Leon B.G., Burton B., Eidenschenk C., Everett C., Gobbi A., et al. Structure-based design of substituted hexafluoroisopropanol-arylsulfonamides as modulators of RORc. Bioorg Med Chem Lett. 2013;23:6604–6609. doi: 10.1016/j.bmcl.2013.10.054. [ DOI ] [ PubMed ] [ Google Scholar ] 57. Cui H., Cai Y., Wang L., Jia B., Li J., Zhao S., et al. Berberine regulates Treg/Th17 balance to treat ulcerative colitis through modulating the gut microbiota in the colon. Front Pharmacol. 2018;9:571. doi: 10.3389/fphar.2018.00571. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 58. Li Y., Xiao H., Hu D., Fatima S., Lin C., Mu H., et al. Berberine ameliorates chronic relapsing dextran sulfate sodium-induced colitis in C57BL/6 mice by suppressing Th17 responses. Pharmacol Res. 2016;110:227–239. doi: 10.1016/j.phrs.2016.02.010. [ DOI ] [ PubMed ] [ Google Scholar ] 59. Yue M., Xia Y., Shi C., Guan C., Li Y., Liu R., et al. Berberine ameliorates collagen‐induced arthritis in rats by suppressing Th17 cell responses via inducing cortistatin in the gut. FEBS J. 2017;284:2786–2801. doi: 10.1111/febs.14147. [ DOI ] [ PubMed ] [ Google Scholar ] 60. Qin X., Guo B.T., Wan B., Fang L., Lu L., Wu L., et al. Regulation of Th1 and Th17 cell differentiation and amelioration of experimental autoimmune encephalomyelitis by natural product compound berberine. J Immunol. 2010;185:1855–1863. doi: 10.4049/jimmunol.0903853. [ DOI ] [ PubMed ] [ Google Scholar ] 61. Cui G., Qin X., Zhang Y., Gong Z., Ge B., Zang Y.Q. Berberine differentially modulates the activities of ERK, p38 MAPK, and JNK to suppress Th17 and Th1 T cell differentiation in type 1 diabetic mice. J Biol Chem. 2009;284:28420–28429. doi: 10.1074/jbc.M109.012674. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 62. Sun P., Wang Z., Ma Y., Liu Y., Xue Y., Li Y., et al. Advance in identified targets of berberine. Front Pharmacol. 2025;16 doi: 10.3389/fphar.2025.1500511. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 63. Su J., Miao Q., Miao P., Zhao Y., Zhang Y., Chen N., et al. Pharmacokinetics and brain distribution and metabolite identification of coptisine, a protoberberine alkaloid with therapeutic potential for CNS disorders, in rats. Biol Pharm Bull. 2015;38:1518–1528. doi: 10.1248/bpb.b15-00293. [ DOI ] [ PubMed ] [ Google Scholar ] 64. Yan Y., Zhang H., Zhang Z., Song J., Chen Y., Wang X., et al. Pharmacokinetics and tissue distribution of coptisine in rats after oral administration by liquid chromatography–mass spectrometry. Biomed Chromatogr. 2017;31:e3918. doi: 10.1002/bmc.3918. [ DOI ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary Data 1 mmc1.docx (5.3MB, docx) Supplementary Data 2 mmc2.xlsx (1MB, xlsx) Supplementary Data 3 mmc3.xlsx (7.9MB, xlsx) Articles from Journal of Advanced Research are provided here courtesy of Elsevier ACTIONS View on publisher site PDF (2.9 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Evaluation of Peppermint Leaf Flavonoids as SARS-CoV-2 Spike Receptor-Binding Domain Attachment Inhibitors to the Human ACE2 Receptor: A Molecular Docking Study#974119
• FLT3 inhibitors in AML: from classic scaffolds to next-generation approaches.#378403
• A Dual-targeting Tracer TATE-RGD for the Diagnosis of SSTR- and RGD- Positive Tumors#845516
• Nucleo-amino acid derived AMO as a potential JAK inhibitor: machine learning screening, docking, molecular dynamics, ADMET/toxicity, and MM-PBSA analysis.#330811
• PREDICT Therapy Selection for JAK, T-cell, or IL-6 Inhibitor Therapies Using a Molecular Signature Response Classifier (PREDICT)#871643