Comparing the performance of quantum descriptors and classical descriptors based ML models for drug like property prediction

Abstract

In drug discovery, the accurate prediction of pharmacokinetic properties plays a crucial role in identifying compounds with favorable absorption, distribution, metabolism, and excretion profiles. Among these, the partition coefficient (logP), blood–brain barrier permeability (logBB), and aqueous solubility (logS) are central to assessing drug–likeness and bioavailability. Traditional descriptor–based models often rely solely on classical cheminformatics features, which may fail to capture the underlying quantum–mechanical phenomena influencing molecular behavior. This work integrates classical, quantum mechanical, and shape/surface descriptors (Quantum Mechanical Polar surface area(QMPSA)) to develop predictive models for logP, logBB, and logS. Quantum descriptors, including frontier orbital energies and Gibbs free energy difference (∆G), were derived from Density Functional Theory (DFT) calculations, while surface and volumetric properties were computed using the Multiwfn program. These were complemented by classical Mordred descriptors, enabling the evaluation of six descriptor set combinations for LogP and LogS, while three descriptor set combinations were used for LogBB. Models were systematically compared using curated datasets comprising 28,930 molecules for log P , 6,460 for log S, and 7,213 (classification) / 976 (regression) for log BB. For log P , the highest performance was achieved with an ExtraTrees regressor on Classical descriptors, attaining R2 = 0.9481 and RMSE = 0.3847. For log S, Gradient Boosting with Classical descriptors produced R2 = 0.8644 and RMSE = 0.8488. The log BB models demonstrated complementary strengths: a classification model with Random Forest on a hybrid model with Classical descriptors and quantum descriptors achieved an accuracy of 0.9411 and an AUC of 0.9974 on hybrid classical and quantum descriptors, while a regression with ExtraTrees on a hybrid Classical+Quantum descriptor reached R2 = 0.665 and RMSE = 0.403. Together, these results highlight the dominance of Classical cheminformatics descriptors for lipophilicity and aqueous solubility, with hybridization yielding improvements in blood-brain barrier permeability tasks. The results demonstrate that incorporating quantum–mechanical information along- side structural and surface descriptors enhances predictive accuracy and offers a generalizable, theory driven framework for early–stage drug property estimation.