Polymer Communication Prediction of biological response for large combinatorial libraries of biodegradable polymers: Polymethacrylates as a test case Vladyslav Kholodovych a, * , 1 , Anna V. Gubskaya b,1 , Michael Bohrer b , Nicole Harris b , Doyle Knight c , Joachim Kohn b , William J. Welsh a, * a Department of Pharmacology, Robert Wood Johnson Medical School (RWJMS), University of Medicine and Dentistry of New Jersey (UMDNJ), Piscataway, NJ 08854, United States b New Jersey Center for Biomaterials, Rutgers, The State University of New Jersey, New Brunswick, NJ 08854-8087, United States c Department of Mechanical and Aerospace Engineering, Rutgers, The State University of New Jersey, New Brunswick, NJ 08854-8058, United States article info Article history: Received 29 November 2007 Received in revised form 11 March 2008 Accepted 19 March 2008 Available online 27 March 2008 Keywords: Combinatorial chemistry Biomaterials Computer modeling abstract A large virtual combinatorial library of polymethacrylates was, for the first time, designed for computer- aided prediction of biorelevant and material properties and focused polymer synthesis. The dis- tinguishing features of this virtual library include its size (about 40 000 compounds), its explicit representation of relatively long polymer chains, and its accounting for different compositions in the case of copolymers and terpolymers. A subset of 79 polymers taken from a representative sub-library of 2000 polymethacrylates was employed to build initial QSPR-based polynomial neural network models, which were then deployed to predict cell attachment, cell growth, and fibrinogen adsorption on polymer surfaces for these 2000 polymethacrylates. The agreement between predicted and experimentally measured property values for the 50 polymethacrylate copolymers within this virtual polymer space encourages further pursuit of polymethacrylate-based biomaterials, and justifies more extensive de- ployment of computational models derived from larger experimental data sets for the rational design of biorelevant polymers endowed with targeted performance properties. Ó 2008 Elsevier Ltd. All rights reserved. 1. Introduction Utilization of computational molecular modeling methodolo- gies has led to significant accomplishments in a wide variety of fields of research and development. Computational approaches and tools employed successfully in pharmaceutical drug discovery, such as molecular diversity/similarity, virtual screening, and quantita- tive structure–performance relationship (QSPR) models, have be- come essential technologies over the past decade due to their scalability, robustness and predictability. Recent advances in QSPR models derived from machine-learning algorithms, together with the advent of affordable high-performance computer hardware and software, now invites extension of these methodologies to larger and more complex systems such as biomaterials. Nevertheless, computational modeling and prediction of bioresponse phenom- ena for polymeric biomaterials remains a significant challenge. The first combinatorial library of polymers [1] consisted of 112 strictly alternating polyarylate copolymers, and beginning in 1999 [2], has served as the source of data for building an array of pre- dictive QSPR (surrogate or semi-empirical) models that differed with respect to their theoretical basis and computational com- plexity. Statistically robust QSPR models were constructed and validated to predict diverse physicochemical and biological prop- erties such as glass transition temperature (T g ), air–water contact angle (AWCA), fibrinogen adsorption onto polymer surfaces [3,4], and cellular attachment and growth of fibroblast-like cells from mouse embryo on flat surfaces of the selected polyarylates [5]. Several computational techniques, such as multiple linear re- gression (MLR), partial least square analysis (PLS) [5] and artificial neural networks (ANN) [3,6], have been employed and compared with respect to their statistical quality and predictability. Despite some qualitative differences in their predictability, all of these models were reasonably successful in predicting the properties of the ‘‘external’’ set of polymers, i.e., the polymers, which were not involved in generating the QSPR models. These previous compu- tational efforts employed various strategies to select molecular structure-based descriptors. Among these approaches, the decision tree algorithm, the Monte Carlo variation procedure, and principal component analysis (PCA) [4,6] were successful in selecting mo- lecular descriptors that correlated most strongly with experimen- tally determined performance properties. While early models included experimental values of T g and AWCA as descriptors for * Corresponding authors. Fax: þ1 732 325 3475. E-mail addresses: kholodvl@umdnj.edu (V. Kholodovych), welshwj@umdnj.edu (W.J. Welsh). 1 These authors contributed equally to the present work. Contents lists available at ScienceDirect Polymer journal homepage: www.elsevier.com/locate/polymer Polymer 49 (2008) 2435–2439 Contents lists available at ScienceDirect Polymer journal homepage: www.elsevier.com/locate/polymer 0032-3861/$ – see front matter Ó 2008 Elsevier Ltd. All rights reserved. doi:10.1016/j.polymer.2008.03.032