Beta-caryophyllene, a bicyclic sesquiterpene with established anti-inflammatory, analgesic, and anticancer credentials, has previously shown promising xanthine oxidase inhibitory activity. The catch is that its potency falls well short of allopurinol. In a new computational study published in Results in Physics, Arif Setiawansyah, Muhammad Ikhlas Arsul, and Rony Abdi Syahputra of Indonesia set out to close that gap with an unusual strategy: instead of relying on medicinal chemists to sketch analogs by hand, they asked three different artificial intelligence chatbots to redesign the molecule. DeepSeek, ChatGPT, and Claude AI each received an identical standardized prompt describing beta-caryophyllene’s structure, its known but inferior inhibitory activity, and the goal of enhancing binding specificity at the enzyme’s active site.

The output was a library of twelve candidate derivatives, each delivered as a SMILES string ready for computational evaluation. What emerged was not a uniform set of suggestions but a revealing portrait of machine-specific chemical reasoning. DeepSeek consistently targeted the C-4 position of the eight-membered ring, swapping the methyl group for ether linkages, triazole rings, and imidazole rings, often justifying the choices through interactions with the enzyme’s molybdenum center. ChatGPT focused on the C4-C5 diene region, proposing epoxidation, phenol ring incorporation, and combined epoxy-phenol hybrids designed to boost hydrogen bonding and pi-pi stacking. Claude AI took a fundamentally different route, attacking the strained cyclobutane ring and the C-13 side chain with carboxymethyl additions, hydroxylation, ring expansion, and tertiary amine integration.

Before any binding calculations, the team screened the twelve candidates against Lipinski’s Rule of Five, the classic filter for oral drug-likeness. Molecular weights ranged from a compact 220.18 to 364.28 daltons, comfortably below the 500-dalton ceiling, and hydrogen bond donor and acceptor counts stayed well within limits. Ten of the twelve derivatives satisfied the criteria with at most one violation. The exceptions were CD2 and CD4, whose LogP values of 6.84 and 7.38 flagged excessive lipophilicity, a property associated with poor aqueous solubility and nonspecific protein binding. The authors note that formulation strategies such as liposomal encapsulation or nanoparticle delivery could rescue such high-LogP compounds, and they contrast the derivatives with allopurinol’s markedly hydrophilic LogP of minus 0.35.

Quantum chemical calculations at the B3LYP/def2-SVP level of density functional theory then mapped the electronic consequences of each modification. Frontier molecular orbital analysis showed that the DeepSeek series preserved relatively large HOMO-LUMO gaps of 5.691 to 6.769 electron volts, with the pyrimidine-bearing DS4 emerging as the most electronically stable compound in the set. In contrast, GPT3’s polyphenolic architecture drove its gap down to 4.825 electron volts, the smallest of all, signaling enhanced polarizability and potential for pi-pi stacking with aromatic amino acids, but at the cost of oxidative stability. Global reactivity descriptors told a similar story: chemical hardness values spanning 2.413 to 3.385 electron volts placed DS4 and DS1 at the metabolically inert end of the spectrum, while softer molecules like GPT3 and CD1 promised adaptive, induced-fit binding. The electrophilicity index proved especially telling, with CD1’s high value of 2.419 electron volts hinting at possible covalent engagement of nucleophilic residues, while CD5’s minimal 0.969 electron volts pointed to purely reversible, non-covalent interaction modes.

The decisive test came from molecular docking against the crystal structure of xanthine oxidase, using the Protein Data Bank entry 3NVY and a rigorously validated protocol whose redocking of the native ligand quercetin reproduced the crystallographic pose with an RMSD of 1.54 angstroms. Here the Claude AI derivatives dominated. CD4 posted a binding free energy of minus 9.7 kilocalories per mole with a predicted inhibition constant of just 0.08 micromolar, while CD1 followed at minus 9.0 kilocalories per mole and 0.3 micromolar. Both crushed the parent beta-caryophyllene, which scored minus 5.8 kilocalories per mole with an inhibition constant of 55.4 micromolar, and both outperformed oxypurinol, the active metabolite of allopurinol, which registered minus 6.22 kilocalories per mole and 22.41 micromolar. The interaction maps explained why: CD4 anchors itself through hydrogen bonds to the catalytic residues Thr 1010 and Arg 880 while wrapping the pocket in pi-pi and pi-alkyl contacts with Phe 914, Phe 1009, Ala 1079, Leu 1014, Leu 873, Phe 649, and Val 1011.

Docking, however, captures only a frozen snapshot. To test whether the complexes survive real thermal motion, the researchers ran 250-nanosecond molecular dynamics simulations in GROMACS with the CHARMM36m force field, explicit TIP3P water, and physiological salt at 310 kelvin. All three ligand-enzyme complexes equilibrated within the first 10 to 15 nanoseconds and remained stable for the rest of the trajectory. CD4 fluctuated around 0.45 to 0.55 nanometers of backbone RMSD, oxypurinol held steady near 0.30 to 0.40 nanometers, and CD1 drifted higher to roughly 0.8 to 1.0 nanometers, though without any progressive destabilization. Residue-level fluctuation analysis confirmed that most of the enzyme stayed rigid, with deviations confined to loop regions and the flexible C-terminus. Radius of gyration and solvent-accessible surface area remained stable across all systems, indicating that neither AI-designed derivative unfolds or globally distorts the enzyme.

Binding free energy calculations using the MM-PBSA method on snapshots drawn from the equilibrated 10-to-250-nanosecond window delivered the study’s headline numbers. CD4 achieved a central binding free energy of minus 23.4 kilocalories per mole, CD1 reached minus 21.7, and oxypurinol trailed at minus 16.3, with all pairwise differences statistically significant. Energy decomposition revealed that the advantage came overwhelmingly from van der Waals contacts: CD4 accumulated minus 35.1 kilocalories per mole of dispersion-driven stabilization compared with minus 23.6 for oxypurinol. Intriguingly, oxypurinol actually won the electrostatic category at minus 28.7 kilocalories per mole, yet still lost overall, demonstrating that hydrophobic pocket complementarity, not polar bonding, drives the AI-designed compounds’ energetic profile. The authors caution, however, that oxypurinol inhibits the enzyme through metal coordination and redox chemistry at the molybdenum center, mechanisms that conventional docking and MM-PBSA cannot fully capture, so the superior calculated energies should be read as evidence of strong non-covalent binding rather than definitive proof of clinical superiority.

The study’s limitations are candidly acknowledged. The AI component provided scaffold-based analog generation rather than fully de novo molecular design, the post-simulation analyses omitted advanced techniques such as principal component analysis and free energy landscape mapping, and no experimental validation was performed. Still, the work stands as a compelling proof of concept that conversational AI systems, each with its own distinct chemical biases, can serve as productive molecular design partners when coupled to rigorous quantum chemistry, docking, and dynamics pipelines. CD4, with its phenolic extension, exceptional van der Waals stabilization, and favorable polar contribution, now stands as the leading candidate for synthesis and enzymatic testing. If laboratory assays confirm what the simulations predict, a humble essential-oil terpene, reimagined by chatbots, could become the template for a safer new generation of uric-acid-lowering drugs.

Subject of Research: AI-assisted computational design of beta-caryophyllene derivatives as xanthine oxidase inhibitors for hyperuricemia and gout

Article Title: Molecular modeling of AI-assisted structural modification of β-caryophyllene toward improved binding to xanthine oxidase: DFT and molecular dynamics studies

Article References: Setiawansyah, A., Arsul, M. I., & Syahputra, R. A. (2026). Molecular modeling of AI-assisted structural modification of β-caryophyllene toward improved binding to xanthine oxidase: DFT and molecular dynamics studies. Results in Physics, Article 108762. https://doi.org/10.1016/j.rinp.2026.108762

Image Credits: AI Generated

DOI: 10.1016/j.rinp.2026.108762

Keywords: beta-caryophyllene, xanthine oxidase, gout, hyperuricemia, AI drug design, molecular docking, molecular dynamics, DFT, MM-PBSA, natural products, drug discovery, allopurinol