Development and OECD-Guided Experimental Validation of Eigenvalue-Based QSAR Models for Skin Permeation

In silico methods such as quantitative structure-activity relationships (QSARs) are widely used in drug development for prescreening of 'drug-likeness'. Most of these tools correlate biological properties to two-dimensional (2D) descriptors, which do not account for important structural elements pertaining to three-dimensional (3D) geometry and shape. To address this, we have developed the 'EigenValue Analysis' (EVANS) methodology, a variant of QSAR in which descriptors are computed as a hybrid of 2D physicochemical properties and 3D geometry. Having previously applied EVANS to build QSAR models for several datasets across a range of endpoints relevant to drug discovery and development, we now apply EVANS to build QSAR models for transdermal permeation. A high-quality dataset of 187 compounds with human skin permeability data was carefully curated using strict exclusion criteria. The EVANS methodology was applied and QSAR models were built correlating the calculated hybrid eigenvalues to the logarithmic skin permeability coefficient (log K_p) using multiple linear regression (MLR), random forest (RF) and support vector machine (SVM) algorithms. The SVM-based model built using five descriptors demonstrated the best performance. Further external validation was performed by retrieving 52 chemicals from the SkinPiX database, providing additional insight into the model's generalisability. The model was subsequently used to make predictions for two untested compounds (4-methylumbelliferone and levofloxacin), and experimental validation following the protocol described under OECD test no. 428 confirmed that EVANS model predictions closely matched observed values and outperformed established models including those of Potts and Guy, Frasch and the Modified Robinson model. Thus, we present a robust model for prediction of transdermal permeation, along with a QSAR methodology that has been extensively benchmarked and is experimentally validated. The EVANS methodology is open source, with codes and documentation freely available at https://github.com/anish19292/Eigenvalue_analysis.