Ancient Chinese Medicine Compounds Turned Into Next-Generation Antibiotic Candidates by AI

Antimicrobial resistance has become one of the most urgent threats in modern medicine, and researchers are increasingly turning to unconventional sources for answers. In a new computational study published in Discover Chemistry, a team led by Peter Adeolu Adedibu of Florida State University set out to mine the vast chemical library of Traditional Chinese Medicine (TCM) for molecules capable of disabling Escherichia coli, one of the world’s most troublesome multidrug-resistant pathogens. Rather than testing thousands of compounds in the laboratory, the researchers built a multi-stage digital pipeline that combined machine learning, pharmacophore filtering, molecular docking, and molecular dynamics simulations to hunt for molecules that could jam a critical bacterial enzyme called DNA gyrase B.

The choice of target is deliberate and strategically important. DNA gyrase is a type II topoisomerase, an enzyme bacteria need to unwind and rewind their DNA during replication, transcription, and chromosome segregation. Its GyrB subunit hydrolyzes ATP to power this supercoiling process, and blocking that energy supply kills the cell. Crucially, DNA gyrase is absent from humans and other higher eukaryotes, which means drugs aimed at GyrB should, in principle, spare human cells and reduce off-target toxicity. This matters because most current antibiotics, including the widely used fluoroquinolones such as ciprofloxacin, attack the GyrA subunit instead, and resistance mutations in GyrA are spreading rapidly. Targeting GyrB offers a complementary route that resistant strains have not yet fully learned to defeat.

The stakes are enormous. Drug-resistant bacterial infections were responsible for an estimated 1.27 million deaths in 2019, with nearly five million deaths associated with resistance in that year alone. E. coli, a leading cause of urinary tract and bloodstream infections, has accumulated an arsenal of defenses, including extended-spectrum beta-lactamases and AmpC enzymes that destroy penicillins and cephalosporins. The economic toll is projected to reach trillions of dollars in healthcare costs and lost GDP by 2030. Against this backdrop, the researchers hypothesized that TCM compounds, which have been used for centuries and include proven drugs such as artemisinin, could yield chemically distinct scaffolds that conventional synthetic libraries have overlooked.

The team’s workflow began with machine learning. They retrieved 562 compounds with experimentally measured activity against E. coli GyrB from the ChEMBL database and labeled them as strong or weak inhibitors based on their pChEMBL values, using a cutoff of 6.8. After generating 210 molecular descriptors with RDKit and pruning redundant or low-variance features, they used the SelectKBest algorithm to retain the ten most informative descriptors, which included molecular weight, topological polar surface area, and hydrogen bond donor and acceptor counts. Four classifiers were trained and compared: RandomForest, AdaBoost, GradientBoosting, and ExtraTrees. The ExtraTreesClassifier emerged as the clear winner, achieving a test area under the ROC curve of 0.941, an accuracy of 0.858, and a Matthews correlation coefficient of 0.716, outperforming the alternatives and demonstrating reliable discrimination between potent and feeble inhibitors.

With this classifier in hand, the researchers screened a curated TCM library of 2,892 unique compounds, prepared from the MedChemExpress TCM Active Compound Library by removing salts, duplicates, and structurally incomplete entries. The model flagged 537 compounds, roughly 18.6 percent of the library, as predicted strong GyrB inhibitors. These were then passed through an energy-optimized pharmacophore model built from the crystal structure of GyrB bound to a known ATP-competitive inhibitor, using the Protein Data Bank entry 6F86 resolved at 1.9 angstroms. The pharmacophore demanded four features: two hydrogen bond donors and two aromatic ring centers. Only 25 compounds matched all four features with a fitness score above 1.0, and notably, known ATP-competitive inhibitors such as novobiocin also satisfied the hypothesis, providing qualitative confidence that the model captured the essential interaction motifs of the ATP pocket.

The next stage was high-precision molecular docking with Glide XP. To validate the protocol, the researchers redocked the co-crystallized ligand into the binding site and obtained a root-mean-square deviation of just 1.34 angstroms from its crystallographic pose, confirming the docking procedure could reproduce experimentally observed binding modes. Fourteen of the 25 pharmacophore hits scored below minus 5.0 kilocalories per mole and advanced to MM/GBSA rescoring, a more rigorous physics-based method that incorporates molecular mechanics, solvation, and desolvation penalties. Five leads rose to the top: lithospermic acid with a binding free energy of minus 54.94 kilocalories per mole, folic acid at minus 49.76, taxifolin at minus 42.35, brazilin at minus 41.47, and dihydromyricetin at minus 40.17. All outperformed novobiocin, which scored minus 5.49 in docking and minus 36.71 in MM/GBSA, as well as ciprofloxacin, which docked at minus 4.51 under the identical protocol.

Lithospermic acid, a polyphenolic compound long used in Chinese herbal medicine, became the focus of lead optimization. Using bioisosteric replacement, a medicinal chemistry technique that swaps functional groups for alternatives with similar shape and electronic properties, the team generated 191 analogs and pushed them through the same screening cascade. Twelve analogs survived high-throughput virtual screening and extra-precision docking with scores below minus 9.0 kilocalories per mole, and the top five were selected for detailed analysis. The standout, designated Analog-1 or lithospermic acid-103, achieved a Glide XP score of minus 10.382 kilocalories per mole and an MM/GBSA binding free energy of minus 76.41 kilocalories per mole, substantially better than the parent compound and both reference drugs. Structure-activity relationship analysis revealed why: Analog-1 forms a distinctive pi-cation interaction with ARG76, a pocket-shaping residue that neither the parent compound, novobiocin, nor ciprofloxacin engages, while also strengthening hydrogen bonds with ASP49 and ASN46 and expanding hydrophobic contacts with ALA90 and VAL93.

Molecular dynamics simulations over 150 nanoseconds provided a dynamic view of these interactions. Analog-1 contacted 33 protein residues, with ASN46, ASP49, GLY81, ILE94, VAL97, and VAL120 showing interaction occupancies above 40 percent, and it maintained more persistent contact with key ATP-site residues such as ASP73, ASN46, and ARG76 than ciprofloxacin, which touched only 25 residues. The ligand itself displayed considerable flexibility, with its RMSD stabilizing between 6.8 and 9 angstroms after 50 nanoseconds, suggesting it explores multiple orientations within the pocket rather than locking into a single rigid pose. Ciprofloxacin, by contrast, remained relatively stable for most of the trajectory before diffusing in the final 10 nanoseconds, behavior consistent with its preference for the GyrA-DNA cleavage complex rather than the GyrB ATP pocket. The authors caution that these single-trajectory simulations are qualitative and were not replicated.

The ADMET and drug-likeness analysis delivered a sobering reality check. All five bioisosteres violate Lipinski’s Rule of Five three times over, with molecular weights exceeding 590 grams per mole, topological polar surface areas above 207 square angstroms, and numerous hydrogen bond donors and acceptors, predicting poor gastrointestinal absorption and low solubility. Several also triggered PAINS structural alerts. Ciprofloxacin, in contrast, passes all classical drug-likeness criteria cleanly. Yet the picture is not entirely bleak: none of the analogs are predicted to be P-glycoprotein substrates, which could enhance intracellular accumulation under non-oral delivery, none are expected to cross the blood-brain barrier, reducing neurotoxicity risk, and none show predicted hepatotoxicity, hERG liability, genotoxicity, or nephrotoxicity. Some analogs, such as Analog-2 and Analog-4, are predicted to avoid inhibiting major cytochrome P450 enzymes, unlike ciprofloxacin, which is predicted to inhibit CYP1A2.

The authors are careful to frame the entire study as hypothesis-generating rather than definitive. The machine learning model was trained on a modest dataset of 562 compounds without external validation or applicability domain analysis, the MM/GBSA values were computed from single docking poses without entropic corrections, and no selectivity modeling against human topoisomerases was performed. All findings remain computational predictions awaiting experimental confirmation. The proposed next steps are concrete: synthesize Analog-1 and Analog-3, run GyrB ATPase enzyme assays, test minimum inhibitory concentrations against multidrug-resistant E. coli, and profile selectivity against human topoisomerase. If those experiments succeed, a centuries-old herbal scaffold, refined by algorithms rather than intuition, could become the seed of a genuinely new class of antibiotics, one aimed at a target that resistant bacteria have so far left vulnerable.

Subject of Research: Computational screening and bioisosteric optimization of Traditional Chinese Medicine compounds as inhibitors of Escherichia coli DNA gyrase B

Article Title: Computational design and optimization of Traditional Chinese Medicine compounds as potential inhibitors of Escherichia coli gyraseB

Article References: Adedibu, P. A., Abdulwahab, A. A., Nkemehule, F. E., Fatoki, O. O., Duru, I. A., & Bodun, D. (2026). Computational design and optimization of Traditional Chinese Medicine compounds as potential inhibitors of Escherichia coli gyraseB. Discover Chemistry, 3(1), Article 521. https://doi.org/10.1007/s44371-026-00980-3

Image Credits: AI Generated

DOI: 10.1007/s44371-026-00980-3

Keywords: antimicrobial resistance, Escherichia coli, DNA gyrase B, Traditional Chinese Medicine, machine learning, molecular docking, MM/GBSA, molecular dynamics, bioisosteric replacement, lithospermic acid, drug discovery, ADMET