Comparative QSPR analysis of novel neighborhood distance-based and classical descriptors for anticancer drugs

Cancer remains one of the leading causes of mortality worldwide and continues to pose a major challenge to public health. Although a large number of anticancer drugs have been developed, their synthesis and experimental evaluation are expensive and time-consuming. As a result, computational approaches have become increasingly important for accelerating drug discovery by establishing relationships between molecular structure and physicochemical properties that govern absorption, distribution, metabolism, and excretion (ADME). In this work, three novel neighborhood distance-based topological descriptors are introduced and employed to investigate eight important physicochemical properties of 50 FDA-approved anticancer drugs, namely molecular weight, LogP, topological polar surface area, hydrogen bond acceptor count, hydrogen bond donor count, rotatable bond count, molecular complexity, and heavy atom count. Initially, the predictive capabilities of the proposed descriptors were examined through polynomial QSPR models and compared with several classical distance-based topological descriptors. The results showed that the newly developed descriptors provide comparable and, in some cases, better predictive performance than existing descriptors. Subsequently, the proposed descriptors were combined with classical topological descriptors and RDKit descriptors to develop Ridge regression, Random Forest, and XGBoost models. Among these models, Ridge regression achieved the best predictive performance for all physicochemical properties. Feature importance analysis based on SHAP further revealed that the proposed descriptors consistently ranked among the most influential features, highlighting their contribution to property prediction. Y-randomization and applicability domain analyses confirmed the robustness and reliability of the developed models. The obtained results demonstrate that the proposed neighborhood distance-based descriptors effectively capture important structural characteristics of anticancer drugs and provide valuable information for QSPR modeling. The developed framework may serve as a useful computational approach for estimating physicochemical properties that are relevant to drug-likeness and ADME-related assessment.