Turmeric-Pueraria-Coix Seed Formula Points to Coumestrol Targets in Rheumatoid Arthritis

A computational study has mapped how phytochemicals from a classic Turmeric-Pueraria-Coix seed (TPC) botanical formula, long used in dietary and nutritional management of chronic inflammatory disease, might act on the immune-metabolic networks of rheumatoid arthritis (RA). Rather than relying on the static databases of conventional network pharmacology, the research team assembled a multilayer pipeline that combined synovial transcriptomics from multiple patient cohorts, ensemble machine learning, interpretable feature attribution, molecular docking, explicit-solvent all-atom molecular dynamics simulations, and single-cell in silico perturbation. The result is not a clinical recommendation but a prioritized set of testable molecular hypotheses, centered on the Pueraria-derived phytoestrogen coumestrol and three synovial genes: PTGS2, CD38, and MAPK8.

The biological backdrop is the growing recognition that RA is as much a metabolic disease as an inflammatory one. Under persistent hypoxia and pro-inflammatory stimulation, fibroblast-like synoviocytes (RA-FLS) and macrophages within the synovial microenvironment undergo profound energy-metabolism rearrangement, accompanied by local accumulation of reactive oxygen species and iron-metabolism disorders. These conditions trigger lipid peroxidation and the pathological damage associated with ferroptosis, an iron-dependent form of regulated cell death. Because RA-FLS display tumor-like proliferative behavior and a strong tolerance to ferroptosis, targeting ferroptosis-related pathways in these cells has emerged as a potential strategy for restraining synovial hyperplasia and joint erosion. Dietary polyphenols, with their reported anti-inflammatory, antioxidant, and immunomodulatory activities, are considered promising nutritional agents for modulating exactly these pathways.

To identify the formula’s bioactive constituents, the researchers screened phytochemical profiles from the TCMSP and SymMap databases using strict drug-likeness criteria: oral bioavailability of at least 30 percent, drug-likeness scores of at least 0.18, molecular weight below 500 daltons, and secondary SwissADME filtering across five drug-likeness models. This yielded 85 candidate bioactive phytochemicals, predominantly dietary polyphenols, flavonoids, and allied functional lipids, with 687 predicted human targets. Intersecting these with 4,275 RA-related targets consolidated from GeneCards, DrugBank, PharmGKB, OMIM, and TTD produced 385 shared regulatory nodes. Mapping those shared targets against curated ferroptosis and oxidative-stress gene sets isolated 40 intersecting candidates, and protein-protein interaction topology via STRING and CytoNCA revealed eight hub nodes: NFKB1, BCL2, IL6, MAPK8, AKT1, PTGS2, SRC, and HIF1A. Among the phytochemicals themselves, 2-monoolein from Coix seed, coumestrol from Pueraria, and bisdemethoxycurcumin from turmeric ranked among the highest-degree network nodes.

The transcriptomic arm of the study drew on six GEO synovial datasets. The largest, GSE89408 with 152 RA and 28 healthy-control samples, served as the training cohort, while the remaining datasets were merged into two batch-corrected external validation cohorts totaling 85 samples. The team evaluated 175 distinct machine-learning strategies built from 15 algorithms, spanning regularized regressions, discriminant analyses, ensemble learners, and standard classifiers. A partial least squares-generalized linear model (plsRglm) with 10-fold cross-validation achieved the best balance of stability and performance, with an AUC of 0.837, and retained 15 core feature genes. Intersecting these machine-learning features with the topological hubs, and testing differential expression across both external validation cohorts, narrowed the field to six statistically robust candidates: ALOX5, CD38, CDK1, EZH2, MAPK8, and PTGS2. An exploratory nomogram built on these six genes reached an AUC of 0.998 within the analyzed dataset, with each individual feature scoring above 0.85.

Interpretability analysis using the SHAP framework, which quantifies each feature’s contribution to model output, ranked PTGS2, CD38, and MAPK8 as the three most influential genes. Together they accounted for more than 80 percent of the cumulative absolute SHAP contribution, and permutation feature-importance analysis produced the same ranking. Higher expression of these genes was consistently associated with positive SHAP contributions toward classification into the RA group. PTGS2, the gene encoding COX-2, carried the largest nomogram coefficient, followed by MAPK8. Notably, classical ferroptosis effector genes such as GPX4 and SLC7A11 did not survive the penalized feature selection, a pattern the authors suggest may indicate that the transcriptomic signals captured by the model concentrate in upstream inflammatory and metabolic pathways rather than in terminal ferroptosis machinery.

With PTGS2, CD38, and MAPK8 prioritized as docking receptors, ten TPC phytochemicals passed a secondary ADMET screen requiring high gastrointestinal absorption, non-inhibition of CYP3A4, at most one Lipinski violation, a bioavailability score of at least 0.55, and zero PAINS alerts. Coumestrol then underwent 100-ns all-atom molecular dynamics simulations in explicit solvent with all three proteins. Within the sampled trajectories, coumestrol retained recurrent contacts with CD38 residues ASP217 and LYS214 and with MAPK8 residues ASN156 and ASP151. MM/PBSA binding free-energy calculations yielded estimates of -113.35 and -129.83 kJ/mol for the CD38 and MAPK8 complexes respectively, and free-energy landscapes characterized the conformational regions sampled during production. Strikingly, the PTGS2-coumestrol complex lost its initial docking pose within the 100-ns simulation window, a divergence the authors emphasize as a caution against trusting static docking scores alone when receptor flexibility and aqueous solvation effects are considered.

The mechanistic rationale for these targets is grounded in prior experimental literature. CD38 is a major NAD+-consuming ectoenzyme whose activity during inflammation depletes the cellular NAD+ pool, suppressing NAD+-dependent sirtuin activity and altering metabolic-inflammatory signaling. This depletion increases ferroptosis susceptibility by disrupting NADPH and glutathione homeostasis, thereby impairing antioxidant capacity and GPX4-dependent control of lipid peroxides. Persistent coumestrol association with the CD38 binding region therefore offers a structural rationale for coumestrol-mediated modulation of CD38 activity and downstream ferroptotic networks. MAPK8, a stress-signaling hub involved in oxidative-stress responses, has been linked to MMP regulation and joint damage in RA. The authors are careful to note, however, that their simulations do not demonstrate direct inhibition of MAPK8 activity or establish that coumestrol alters MAPK8-dependent signaling in RA cells.

Single-cell in silico perturbation added a network-level dimension. Using scTenifold and scTenifoldKnk on RA synovial single-cell RNA-sequencing data comprising 30,057 cells, the team simulated virtual overexpression and knockout of the prioritized targets. Overexpression of MAPK8 and PTGS2 converged on significant regulatory displacement of the mitochondrial-encoded genes MT-ATP6 and MT-ND4, implicating these targets in modulating mitochondrial network organization. Knockout of both targets displaced AES and RARRES3, indicating partially convergent downstream regulatory states from distinct upstream perturbations. CD38 deletion, by contrast, triggered a broader and distinct response involving EEF1A1, RARRES3, RPL19, GIMAP7, CSF2RA, RPS2, and RPL7A. The convergence between MAPK8 and PTGS2 suggests that different upstream signals can propagate toward shared downstream regulatory states within the synovial single-cell network.

The study’s functional lipids may operate through pathways entirely distinct from the protein interactions prioritized for coumestrol. In the docking analysis, 2-monoolein showed weaker predicted affinities for the selected targets, but previous work has implicated Coix seed-derived lipids in modulation of the gut microbiota, butyrate production, and the gut-joint axis in inflammatory arthritis. Because microbiome and metabolomic data were not examined in this study, the authors frame these lipid mechanisms as contextual hypotheses rather than direct findings. Similarly, curcumin’s reported suppression of NF-kB-associated inflammatory signaling and enhancement of Nrf2-dependent antioxidant responses, along with experimental evidence that coumestrol attenuates matrix-degrading enzymes and inflammatory cytokine responses in chondrocytes and macrophages, substantiates the computational prioritization of these polyphenols as modulators of oxidative stress and ferroptosis-related networks.

The authors are explicit about the limitations inherent to a purely computational approach. In silico ADMET predictions cannot fully account for the absorption, metabolism, and first-pass disposition of dietary polyphenols in vivo, and future LC-MS/MS studies will be needed to determine systemic and joint-tissue exposure of the prioritized compounds. Each molecular dynamics assessment relied on a single 100-ns trajectory, which may miss rare dissociation events or the full conformational landscape, so the observed CD38-coumestrol and MAPK8-coumestrol interactions do not preclude late-stage dissociation or metastability. The training cohort exhibited marked class imbalance, and although SMOTE was restricted to the training phase, AUC-PR declined from 0.998 in the training set to 0.690 in Validation Set 2, reflecting attenuated cross-cohort generalizability. Reliance on synovial bulk-transcriptomic data also restricts the findings to the synovial compartment, meaning the six-gene signature functions as a molecular prioritization tool rather than a clinical diagnostic classifier. Most importantly, the predicted compound-target interactions lack direct experimental validation, and future studies must integrate target-specific perturbations with direct binding assays, alongside complementary ferroptosis readouts such as lipid ROS, GPX4, and SLC7A11, to confirm these relationships.

Within those constraints, the study delivers a coherent multilayer model of how a polyphenol-rich botanical formula might engage RA pathology. Coumestrol emerges as the leading bioactive constituent, with CD38 and MAPK8 receiving robust structural support as principal interaction targets, while PTGS2 retains prominent transcriptomic and network-level relevance despite dynamic instability during simulation. The single-cell perturbation results demonstrate that these targets drive distinct yet partially overlapping regulatory states, and the recurrent displacement of mitochondrial-encoded genes following overexpression links the prioritized targets to mitochondrial network organization. By decoupling physical binding predictions from transcriptional and network-level responses, the framework offers a disciplined template for mechanistic discovery in complex botanical mixtures, and a concrete, experimentally addressable agenda for validating how turmeric, Pueraria, and Coix seed phytochemicals might modulate the inflammatory, metabolic, and ferroptosis-susceptibility landscape of rheumatoid arthritis.

Subject of Research: Computational biology and rheumatoid arthritis research

Article Title: Computational Prediction of the Therapeutic Mechanisms From the Turmeric‐Pueraria‐Coix Seed Botanical Formula in Rheumatoid Arthritis

Article References: Lan, W., Lu, Q., Rui, X., Li, K., Su, C., Su, Q., Huang, Y., Shi, R., Chen, W., & Tang, F. (2026). Computational Prediction of the Therapeutic Mechanisms From the Turmeric‐Pueraria‐Coix Seed Botanical Formula in Rheumatoid Arthritis. Food Science & Nutrition, 14(10), Article e72503. https://doi.org/10.1002/fsn3.72503

Image Credits: AI Generated

DOI: 10.1002/fsn3.72503

Keywords: rheumatoid arthritis, coumestrol, machine learning, molecular dynamics, ferroptosis, network pharmacology, single-cell RNA sequencing, dietary polyphenols, systems biology, SHAP, Computational, Prediction