Prediction of the in vitro permeability determined in Caco-2 cells by using artificial neural networks

Abstract

Caco-2 cells are currently the most used in vitro tool for prediction of the potential oral absorption of new drugs. The existence of computational models based on this data may potentiate the early selection process of new drugs, but the current models are based on a limited number of cases or on a reduced molecular space. We present an artificial neural network based only on calculated molecular descriptors for modelling 296 in vitro Caco-2 apparent permeability (P_app) drug values collected in the literature using also a pruning procedure for reducing the descriptors space. Log P_app values were divided into a training group of 192 drugs for network optimization and a testing group of another 59 drugs for early stop and internal validation resulting in correlations of 0.843 and 0.702 and RMSE of 0.546 and 0.791 for the training and testing group, respectively. External validation was made with an additional group of 45 drugs with a correlation of 0.774 and RMSE of 0.601. The selected molecular descriptors encode information related to the lipophilicity, electronegativity, size, shape and flexibility characteristics of the molecules, which are related to drug absorption. This model may be a valuable tool for prediction and simulation in the drug development process, as it allows the in silico estimation of the in vitro Caco-2 apparent permeability.

Introduction

Oral administration of drugs, due to its ease and patient compliance, is the preferred route and a major goal in the development of new drug entities. It is also traditionally one of the reasons for either discontinuation or prolongation of the development time of compounds. These problems lead to a new paradigm, a multivariate approach, in compound selection and optimization (Venkatesh and Lipper, 2000). In this context, and as a consequence of combinatorial chemistry, initial screening of compounds in a number of thousands is done typically by using in silico approaches. In vitro tests are responsible to reduce the number of compounds from hundreds to dozens and in vivo animal models to 1–5 finally potential drugs that are included in clinical trials.

Various in silico models already exist to predict drug absorption potential. One of the most known is the “Lipinski rule of 5” (Lipinski et al., 2001), based on the analysis of successful drugs, in which indications for poor absorbed compounds are based on the number of H-bond donors and acceptors, molecular weight and ClogP. After that initial effort, various models were proposed to quantitatively predict oral absorption. These were based on the search to correlate some molecular descriptors to measures of bioavailability by means of multivariate regression tools. Models were developed based on the in vivo Human Intestinal Absorption (Abraham et al., 2002, Butina, 2004, Klopman et al., 2002, Norinder and Osterberg, 2001, Zhao et al., 2001), in vivo absorption rate constants (Linnankoski et al., 2006), in vivo jejunal effective permeability (Winiwarter et al., 2003, Winiwarter et al., 1998), in vitro apparent permeability in artificial membranes (Fujikawa et al., 2005, Fujikawa et al., 2007, Nakao et al., 2009, Verma et al., 2007) and in vitro apparent permeability (P_app) in Caco-2 cells (Castillo-Garit et al., 2008, Degim, 2005, Di Fenza et al., 2007, Fujiwara et al., 2002, Hou et al., 2004, Nordqvist et al., 2004, Ponce et al., 2004, Santos-Filho and Hopfinger, 2008, Yamashita et al., 2002a).

In vivo based data, although obviously the target for the lead drug, presents some drawbacks. In vivo global bioavailability results from different physical−chemical and biological processes that are difficult to isolate. To be absorbed, a drug needs first to be in its soluble form. Passive diffusion may be the principal driving force for drug absorption, but either active transport or efflux mechanisms may condition the final bioavailability. Instability in the gastro-intestinal fluids and metabolization may also reduce the amount of drug that effectively reaches the systemic circulation. For these reasons, calculation of the absorption rate constants, which uses pharmacokinetic data after intravenous and oral administration, may underestimate the absorption. Collection of data on Human Intestinal Absorption is also experimentally demanding. It requires the evaluation of the fraction of dose, either as parent or metabolites, eliminated by the faeces and urine which may easily result in underestimation of the actual fraction of dose absorbed. Finally, data on jejunal effective permeability, which isolates the absorption process but is experimentally complex, is limited to only a small number of drugs. Another drawback in these three approaches is that prediction confirmation is only possible when the drug is administered to humans in the end stages of drug discovery, leaving no room for “fine-tuning” during the development phase.

In vitro tests, on the contrary, are routinely performed by pharmaceutical companies in an early stage of drug development. For this reason, a large number of P_app values for different molecules are being produced and are/will be available to build databases with a large span of values from high to low bioavailabilities. Parallel artificial membrane permeation assays (PAMPA) are a very promising tool to predict absorption, but there are various experimental variables to be considered and in practice it appears to produce the same outcome that a simple Log D_7.4 assay in predicting high and low absorption (Galinis-Luciani et al., 2007). Caco-2 cells, on the contrary, are a widely preformed in vitro test with interesting properties when extrapolating results to bioavailability. It is a cell system, characterized by easy handling and at the same time resemble morphological and biochemical characteristics of the intestinal cells (Vogel, 2006), that shows a sigmoid relationship between the fraction absorbed in humans and the P_app across the cells (Artursson and Karlsson, 1991).

A number of statistical multivariate methods for developing in silico models are currently used. Artificial neural networks (ANN), however, have shown superior performance to the linear multivariate class of methods in various QSAR models (Fujiwara et al., 2002, Paixao et al., 2009, Sutherland et al., 2004, Votano et al., 2004). To the best of our knowledge, only three studies were previously reported to predict in vitro apparent permeability (P_app) in Caco-2 cells using ANN and molecular descriptors. Of these, two studies (Degim, 2005, Fujiwara et al., 2002) were based on a small number of drugs, did not make any particulate approach to avoid overfitting, nor made a convenient search of optimal molecular descriptors to predict the in vitro Caco-2 permeability. The other study (Di Fenza et al., 2007) resolved some of these problems, but was based on proprietary analogue drugs. Since the proposed model was built with the use of molecules contained in a very similar chemical space, it was proved to be less competent in predicting Log P_app values in a more chemical diverse external dataset.

With the purpose to overcome these limitations, we present a backpropagation ANN model using early stop to predict in vitro Caco-2 Log P_app based on 296 structurally different drugs and drug-like molecules collected in the literature. This large dataset would provide a broad molecular space, ideal to enlarge the applicability of the proposed model. Various molecular descriptors, encoding different molecular characteristics were initially used. Reduction of the number of possible molecular descriptors was made by a simple pruning procedure, allowing to point out some possible factors influencing the drugs absorption process. Use of early stop during the whole modelling procedure also ensured that optimal prediction capabilities were maintained in the final model. This model, and proposed methodology, may be a valuable tool in early drug development.