Green Chemistry PAPER Cite this: Green Chem., 2016, 18, 4468 Received 5th April 2016, Accepted 15th July 2016 DOI: 10.1039/c6gc00945j www.rsc.org/greenchem Predicting skin permeation rate from nuclear magnetic resonance spectra† Nan An, a John-Hanson Machado, b Yuechuan Tang, a Jakub Kostal b and Adelina Voutchkova-Kostal* a In order to systematically approach the design of chemicals with minimal toxicity we are in need of predictive tools that can be applied seamlessly during chemical synthesis and characterization. Our approach is to develop models that utilize spectroscopic data, called Quantitative Spectrometric Data-Activity Relationships, to predict bioavailability and toxicity. Such models do not require knowledge of chemical structure and can be applied to chemical mixtures. Here we report a predictive QSDAR for skin permeation rate (log K p ) of organic chemicals from NMR spectroscopic data and molecular weight. The model is trained on a large data set consisting of structurally diverse chemicals and has been thoroughly externally validated – once with a withheld subset of the original data set, and once with a distinct set of complex biologically active compounds curated by Klimisch scoring (r 2 = 0.838, q ext1 2 = 0.837, q ext2 2 = 0.419). The model performs equally or better than prevailing struc- ture-based methods, and offers a number of advantages for facilitating rational design of safer chemicals. Introduction Estimation of skin permeation rates of chemicals is of para- mount importance to medicinal chemists, formulators and toxicologists. 1,2 Medicinal chemists consider skin permeability of dermal APIs in an effort to deliver the desired dose; formu- lators seek to understand permeability in order to create effective and safe cosmetic products; and toxicologists con- sider permeation when carrying out risk and alternatives assessments. 3 Furthermore, molecular designers seek to mini- mize skin permeation of commercial chemicals in order to reduce the probability of toxicity. Experimental methods for testing skin permeability include in vitro diffusion chamber experiments, 4 biomonitoring experiments 5,6 and excised skin experiments, most commonly from rat and pig. 2 However, these methods are often cost-prohibitive and time-consuming, which makes the development of fast and accurate predictive methods highly desirable. Predictive methods are also required in order to rationally design new commercial chemi- cals with minimal skin permeation. 1 A number of predictive quantitative structure–activity relationships (QSARs) exist that successfully relate skin per- meation rates, log K p , to chemical structure. 2,7–16 If one exam- ines the QSARs that are not constrained to a single functional class (trained on more than 100 compounds), the coefficient of determination of training set (r 2 ) ranges from 0.63 to 0.945. 2,7–10,17–22 The salient properties driving existing QSAR models are hydrophobicity, reflected by octanol–water par- tition coefficient (log P), 23 molecular size (or volume) and intermolecular interactions, such as hydrogen bonding. 24 As a result, many QSARs for skin permeation rate share the generic form, 10,17,22,24 log K p ¼ aðhydrophobicityÞþ bðmolecular sizeÞ þ cðinteractionsÞþ d The majority of these QSARs rely on physicochemical and structural descriptors derived from the 2D or 3D chemical structures, and many of them suffer from limited applicability domains and low external predictivity. 22 The limitations of existing log K p models stem from two factors: first, the in- consistency in the measurement of K p , which depends on the carrier solvent, the mode and location of application to the skin, and biological factors; and second, the existence of numerous mechanisms by which a chemical can permeate through the epidermis, such as intracellular diffusion, inter- cellular transport and follicular transport. 15,25 Herein we attempt to develop a quantitative model for log K p that does not use structure-based descriptors, but rather nuclear magnetic resonance (NMR) spectroscopic data. Such models are referred to as quantitative spectrometric data- activity relationships (QSDARs). 26 A handful of literature reports document successful attempts to integrate NMR spec- troscopic data into predictive models for physicochemical † Electronic supplementary information (ESI) available. See DOI: 10.1039/c6gc00945j a Department of Chemistry, The George Washington University, Washington, DC, USA. E-mail: avoutchkova@email.gwu.edu; Fax: +202-994-5873; Tel: +202-994-6121 b Computational Biology Institute, The George Washington University, Washington, DC, USA 4468 | Green Chem. , 2016, 18, 4468–4474 This journal is © The Royal Society of Chemistry 2016 Published on 25 July 2016. Downloaded by George Washington University on 28/10/2016 19:31:30. View Article Online View Journal | View Issue