Machine Learning Screens Malaysian Kelulut Honey Compounds Against Allergy Receptor

Researchers in Malaysia have built a machine learning model that ranks the natural compounds of kelulut honey, a polyfloral honey produced by stingless bees, by their predicted ability to bind the human histamine H1 receptor, the molecular target of classic antihistamine drugs. The study, published in Discover Molecules, combined an ensemble learning workflow built in the open-source KNIME Analytics Platform with conventional structure-based virtual screening, and found that an Extreme Gradient Boosting (XGBoost) model outperformed three rival ensemble methods across training, test and external validation data.

The team first profiled the honey chemically using liquid chromatography coupled with quadrupole time-of-flight mass spectrometry (LC-QTOF-MS). Compounds were classified as acidic, peroxide or non-peroxide, with the peroxide class notably absent, which the authors attribute to the instability of R-O-O-R structures under heat and low water activity. Non-peroxide constituents included flavonoids, phenolic acids and less common molecules such as reptosides, platycogenic acid and kauronic acid. From prior work, ten compounds had been flagged as both absorbable in the gastrointestinal tract (HIA+) and impermeable to the blood-brain barrier (BBB-), a profile the researchers argue could in principle limit the central nervous system side effects, such as sedation and dizziness, associated with first- and second-generation antihistamines.

To train the predictive model, the authors compiled a dataset of 14,000 compounds from PubChem, each annotated with a binding affinity score derived from virtual screening against the H1 receptor and its SMILES structural notation. After curation steps that removed molecules with unusual valences, salts and unconnected structures, 13,816 molecules remained. The KNIME workflow generated 127 two-dimensional descriptors from RDKit, capturing atom types, bonds and topological features, and 85 three-dimensional descriptors from the Weighted Holistic Invariant Molecular (WHIM) scheme, which encodes spatial and conformational information. Dimensionality reduction then whittled this pool down: a low variance filter cut the descriptors from 212 to 154, z-score normalization standardized their scales, and a linear correlation filter with a threshold of 0.8 removed redundant features, leaving 72 descriptors for model construction. Outliers in each descriptor were also removed.

Four ensemble algorithms were trained on 85 percent of the data, with the remaining 15 percent reserved as an internal test set: XGBoost, Random Forest, Gradient Boosted Trees and a Tree Ensemble. Hyperparameters were optimized with a Parameter Optimization Loop node using 10-fold cross-validation. XGBoost emerged as the clear leader, achieving an R-squared of 0.818 on the training set, 0.830 on the test set and 0.799 on an independent external validation set of 2,023 compounds. Its mean absolute errors were 0.476, 0.467 and 0.506, and its root mean square errors 0.677, 0.657 and 0.707 across the three datasets. Random Forest showed more stable but lower R-squared values of roughly 0.77 across datasets with higher error metrics, while Gradient Boosted Trees declined to 0.773 and 0.783 on the test and external sets, which the authors read as a sign of weaker generalization. The Tree Ensemble performed worst overall, with a negative skew in its predicted binding energies indicating systematic overprediction at the low-affinity end.

Reliability on unseen chemicals was checked with an applicability domain analysis based on Euclidean distances between compounds, implemented through the Enalos Domain-Similarity node. The applicability domain threshold was calculated as 2.222, and all evaluated compounds fell within a distance range of 0.042 to 2.188, meaning the external validation chemicals occupy chemical space similar to the training set. The authors also situate their result against earlier work: a previously reported surrogate docking model reached a test R-squared of 0.77 across 11 drug targets, and a random forest binding affinity model reached 0.76, figures comparable to the XGBoost model developed here. They attribute XGBoost’s edge partly to its built-in L1 (Lasso) and L2 (Ridge) regularization, which constrains overfitting arising from bias and variance in the data.

When the trained model was deployed on the ten HIA+, BBB- kelulut honey compounds, one molecule stood out. Compound CID73642 was ranked the top binder by XGBoost with a predicted binding affinity of -9.00 kcal/mol, and the second binder by redocking in PyRx with -7.3 kcal/mol, a relative discrepancy of 20.9 percent. Both methods nonetheless agreed in prioritizing it. Other compounds showed larger divergences in ranking: CID5377686 was rated second by the machine learning model at -8.36 kcal/mol but only eighth by PyRx at -6.70 kcal/mol, a 22.0 percent gap, while CID5315263 ranked third under XGBoost at -8.31 kcal/mol but ninth under PyRx at -6.40 kcal/mol, a 26.0 percent difference. In the opposite direction, CID11057730 received the lowest XGBoost prediction at -6.99 kcal/mol, 7.0 percent below PyRx’s -7.50 kcal/mol.

Docking interaction maps of the top three candidates against the H1 receptor, whose crystal structure (PDB ID 7DFL, resolved at 3.30 angstroms by electron microscopy) was preprocessed for the screening, revealed hydrogen bonds between ligand alcohol and ketone groups and residues including ASP107, ARG176, ASP178, LYS179, LYS191, ASN198, TYR431, ASN443, HIS450 and ILE454. ASP107 also formed electrostatic interactions with CID73642 and CID5377686, while HIS450, LYS179 and ASP178 did so with CID5315263. Hydrophobic pi-interactions involving TYR108, TRP428 and PHE432 contacted the stilbene group of CID73642 and the pyrrolidine ring of CID5377686, and pi-pi T-shaped contacts with LYS179, PHE435, HIS450 and ILE454 involved the chromene ring of CID5315263. The docking grid was centered on the histamine binding site, defined by residues including ASP107, TYR108, SER111, THR112, TRP158, ASN198, TRP428, TYR431, PHE432, PHE435 and TYR458.

Feature importance analysis of the XGBoost model, using both weight, the frequency a feature appears across trees, and gain, its impact on predictions, identified scaled molecular refractivity (SMR) as the dominant descriptor, followed by the WHIM descriptors Unit Weight.wlambda3, Unit Weight.wlambda2 and SlogP, an octanol-water partition coefficient. The authors link these descriptors to specific structural traits: CID73642’s bulky glucose moiety raises its wlambda3 value to 23.20, while its glycosidic linkage disrupts molecular planarity and narrows its breadth, giving the lowest wlambda2 among the nine compounds at 2.57. A double bond in the aglycone further constrains flexibility, and the compound’s SMR of 99.54 sits close to the average value of 97 recommended by the Ghose rule. By contrast, CID5315263 and CID5377686 carry SlogP values of 2.91 and 3.17, and the authors suggest that excessive lipophilicity may divert ligands toward hydrophobic regions outside the active site, weakening the hydrogen bonding and ionic contacts needed inside the pocket. CID91017475, despite extreme compactness with a wlambda3 of 27.86, has an SlogP of 4.0852 and an extreme SMR near the upper Lipinski limit, which the authors propose could introduce unfavorable hydration effects at the protein surface.

Drug-likeness screening adds further nuance. Most candidates comply with Lipinski’s Rule of Five, except CID73642 and CID11057730, which exceed the recommended limit for hydrogen bond donors, though the authors note that a single violation does not necessarily preclude bioactivity or oral absorption. All candidates meet Veber’s criteria of fewer than 10 rotatable bonds and a topological polar surface area below 140 square angstroms, while CID5280372 and CID91017475 violate Ghose rules. The authors are explicit about the central limitation of the work: the target variable used to train the model came from PyRx docking scores, not from experimentally measured H1 receptor bioactivity, so the model should be regarded as a docking score-based prioritization method rather than a direct predictor of pharmacological effect. High predicted affinity, they caution, is not synonymous with confirmed antihistamine activity, since docking scores do not capture pharmacokinetic behavior or downstream biological responses. In vitro binding assays and pharmacological studies are therefore required to validate the computational predictions, and the authors propose extending the ensemble framework to other receptor targets and natural product libraries as a next step.

Subject of Research: Machine learning prediction of binding affinity between Malaysian kelulut honey phytochemicals and the histamine H1 receptor

Article Title: Consensus ensemble learning with augmented molecular representation for accelerated virtual screening of kelulut honey phytochemicals targeting histamine H1 receptor

Article References: Tan, W. F., Hassan, S., Edros, R., Awang, S., & Dong, R. (2026). Consensus ensemble learning with augmented molecular representation for accelerated virtual screening of kelulut honey phytochemicals targeting histamine H1 receptor. Discover Molecules, 3(1), Article 27. https://doi.org/10.1007/s44345-026-00065-9

Image Credits: AI Generated

DOI: 10.1007/s44345-026-00065-9

Keywords: kelulut honey, histamine H1 receptor, XGBoost, virtual screening, machine learning, molecular docking, KNIME, LC-QTOF-MS, molecular descriptors, drug discovery, phytochemicals, antihistamines