Natural Product-Likeness Scores Revisited: Machine Learning Models for Predicting Drug-Like Properties from Phytochemical Databases

Abstract

Natural products remain among the richest sources of drug leads, yet the cheminformatic tools used to prioritise molecules were largely developed with synthetic compounds in mind. Natural products occupy a distinct region of chemical space—characteristically higher in molecular weight, oxygen content, fraction of sp³ carbons and stereocentres, and lower in nitrogen, lipophilicity and aromaticity than synthetic drugs—and roughly one in five violate Lipinski's “Rule of Five” while remaining bioactive. This mismatch motivates a fresh look at the two families of scores that guide natural-product-based discovery: drug-likeness metrics (the Rule of Five, Veber, Ghose, Egan, Muegge and the quantitative estimate of drug-likeness, QED) and the natural product-likeness (NP-likeness) score of Ertl and colleagues, a Bayesian fragment-based measure that cleanly separates natural products from synthetic molecules. This review revisits these scores in the machine-learning era.
We summarise the classical rules and their limitations for natural products; describe the NP-likeness score and its open-source and neural-network reimplementations; quantify the systematic physicochemical differences between natural-product and synthetic chemical space; survey the major phytochemical databases (COCONUT, LOTUS, NPASS, KNApSAcK) that now supply training data at scale; and examine how molecular representations (physicochemical descriptors, fingerprints such as ECFP/Morgan, and graph- and SMILES-based encodings) feed machine-learning models—random forests, gradient boosting, deep and graph neural networks—that predict drug-likeness, ADMET and bioactivity. We conclude that data-driven, natural-product-aware models are superseding rigid rule-based filters, but that their reliability depends on data quality, an honestly defined applicability domain, interpretability, and resistance to the historical bias toward synthetic, Rule-of-Five-compliant chemistry.