Successful drug development requires not only the optimization of pharmacological specificity and potency, but also a method for efficient drug delivery to the target site. Many drug candidates fail to achieve their therapeutic potentials because of poor bioavailability [1]. Oral drug delivery avoids the pain and discomfort associated with injections and also the risk of accidents and infections caused by misuse of needles. For these reasons, the oral route is by far the easiest and most convenient mode of drug administration, and oral availability is a highly desirable property for candidate drugs under development. However, before an orally administered drug can reach its site of action, it must first cross the intestinal epithelial barrier by passive diffusion, carrier- or receptor-mediated uptake or active transport and enter the systemic circulation [2]. Molecules with low permeability and/or absorption rates are not suitable for oral administration, and there has been great interest in finding ways to avoid producing potent but non-permeating molecules [3]. Several screening paradigms for evaluating drug absorption have been employed to enhance the probability of success through the stages of drug development and a number of methods have been developed to assess oral availability using in vivo, in vitro, in situ or in silico models [4].

The most widely-accepted in vitro absorption model uses Caco-2 cell monolayers. Because Caco-2 cells express several types of transporter proteins, both the passive and active transport potentials of a compound can be investigated [5–7] and several experimental methods have been developed using this model to test the absorption of drugs by the human intestine [8–10]. However, these experimental cell-system methods are rather labor-intensive and not easily applicable to high-throughput screening. As an alternative approach, computational modeling can provide a quick and inexpensive way of evaluating the intestinal permeability of a compound before synthesis. This enables us to prioritize molecules for in vitro and in vivo studies and improve the overall properties of the compounds that proceed along the drug discovery pathway. A number of models for Caco-2 cell permeability or human intestinal absorption have been reported that predict the oral absorption properties of drugs, mostly limited to small organic molecules [11–14].

Rapid developments in biotechnology and peptide synthesis have made it possible to exploit the unique pharmacological activities of peptides; thousands of different peptides have been designed, synthesized and subjected to a range of screening procedures and biological assays. To analyze the vast amounts of biological data on peptides, quantitative structure-activity relationship (QSAR) models have been successfully employed. For example, several QSAR models have been developed to predict the peptide binding activities of target proteins, resulting in good correlations with in vitro data [15–19], and these have proved useful in generating leads through the screening of large peptide libraries. It is surprising that QSAR models have seldom been applied to other pharmacological properties of peptides, especially since failure to comply with pharmacological demands is likely to terminate the development of a candidate peptide drug [20, 21]. Although a few previous QSAR studies have investigated the affinities of peptides to intestinal transport proteins, the machine-learning processes were performed not on the basis of sequence information but of chemical structure [22–24]. There have been a few reports on the prediction of intestinal absorption of non-peptide compounds from molecular structure. Wessel et al. reported a QSAR study on a set of 86 compounds with known percentage human intestinal absorption (%HIA) values [25]. To obtain a predictive model, they used a neural network to map molecular structure descriptors to %HIA. Polley et al. applied Bayesian regularized neural networks to develop a statistically significant QSAR model for human intestinal absorption [26].

In this work, we report the first QSAR models to predict the intestinal permeabilities of peptides on the basis of their sequences. A group of peptides crossing the intestinal barrier were selected from a random phage-peptide library using the 'peroral phage display technique', a newly developed in vivo technique in which a phage-peptide library is administered orally to rats and the intestinal barrier-permeable phages are collected from the internal organs. Using the sequence set of the selected phage-displayed peptides, we constructed an artificial neural network model to evaluate the intestinal permeabilities of peptides using various descriptors of the physicochemical properties and occurrence of the amino acid residues.

First, 852 random heptapeptide sequences were generated as negative control data, keeping the frequency of each amino acid residue the same as in the random phage library. We utilized a feed-forward neural network for our sequence-based permeability prediction. Eight models were derived for training data set by varying the type of peptide descriptor and/or the number of neurons in the (single) hidden layer. The predictive features of the resulting model are illustrated in Figure 1, which clearly shows that our model can distinguish effectively between intestinal barrier-permeable and impermeable peptides.

To test the effect of the number of objects on overtraining, we also constructed neural network models for 1:3 data set in which the negative control data set was three times larger than the positive; Table 2 summarizes the capacity of these models for prediction. Considering models with the same network architectures, the differences in ROC scores between the training and test set were generally smaller in the 1:3 than the 1:1 data set. This result shows that the performance of the model is less affected by overtraining if the size of the data set is increased.

To validate the sequence dependency, we compared the intestinal permeabilities of peptides with identical amino acid compositions but different sequences. We selected three peptides with different prediction scores from among the 852 intestinal barrier-permeable heptapeptides: TQKSGPV, with a high score (1.03), HKGPFQS, with a medium score (0.78), and QPMNSLT, with a low score (0.52). For each of these peptides, we generated a set of peptides with all the possible sequence permutations of the seven amino acids (7! = 5040) and intestinal permeabilities were predicted using the model with network architecture (7 × 20)-0-1. The wide distribution of the prediction scores for the peptide sets (Figure 2) clearly indicates that the intestinal permeability of a peptide depends on its sequence.

To test the reliability of the peptide sequences from the phage-display experiment as the positive control set and to validate the strength of our model in predicting the intestinal permeabilities of peptides, a separate decoy set was generated as positive control in the training and test set. We constructed supplementary model trained with the decoy set and compared that model with the model trained with the real data set for ability to discriminate between intestinal barrier-permeable and impermeable peptides. The predictive features presented in Figure 3 indicate that model constructed with the decoy set do not discriminate between these two permeability classes of peptides. Also, the validation result for model with network architecture (7 × 20)-0-1 (Table 3) suggest that the predictive power for the test set is considerably greater when the model is constructed with the real set than with the decoy. This result confirms that model trained with the real set is robust discriminator between intestinal barrier-permeable and impermeable peptides, and that the positive real data sequences were collected sufficiently well to allow efficient model construction.

We have developed models for predicting the intestinal permeabilities of peptides. Our models produced nearly identical statistics for multiple training runs and efficiently discriminated among peptides on the basis of intestinal permeability. As shown in the decoy set analysis, models trained with random sequences had no prediction capacity, but the peptide sequences collected from the in vivo experiment served well as positive control sets for the QSAR models.

Although we tried to optimize the network architecture and to minimize overtraining and other related problems during the course of development, some factors in our model might cause prediction errors. We assumed that randomly-selected heptapeptide sequences can be used as negative controls. This assumption can be rationalized on the grounds that heptapeptides with random sequences are very likely to be intestinal barrier-impermeable because the sequences obtained from the in vivo experiment only covered a very small portion of the entire 'heptapeptide space'. Thus, our model correctly predicts permeable rather than impermeable peptides. This indicates that a model based on 1:3 data set is preferable for eliminating intestinal barrier-impermeable peptides if the random sequences chosen as negative controls do indeed show negligible intestinal absorption; as shown in Table 4, the specificity of 1:3 data set models is superior to that of 1:1 data set models. Consequently, our model has the best predictive power for the selection of the permeable peptide in relation to the reliability of the data set.

In this work, we developed models for prediction of intestinal permeabilities of peptides using the feed-forward neural network as an algorithm for training. As shown in Table 2, some problems could be detected in the feed-forward neural network, which are including overfitting, network architecture optimization, and selection of the best QSAR model. To avoid shortcomings such as overtraining which appears to have happened in our models with larger network complexity, robust QSAR models using the Bayesian regularized neural network would be more desirable [26, 29]. The development of QSAR models using the more robust methods like the Bayesian neural network would be a fruitful approach of future work in terms of predictive ability and robustness of model for intestinal permeability of peptide.

Burden et al. [17] noted that property-based descriptors require a more flexible modeling method than binary descriptors to take account of larger contributions from cross terms or nonlinearity. However, our models produced very similar results on the discrimination of intestinal permeability using binary and VHSE descriptors; no statistically significant difference was observed in their ROC scores for the test sets.

The models most widely used for predicting passive intestinal absorption are drug-likeness prediction models such as the Rule of 5 model introduced by Lipinski et al. [30]; they have the advantages of being simple, easy to interpret and quick to compute. In general, such approaches are formulated on the basis of group additive methods, so the predicted intestinal permeability is similar for peptides consisting of the same numbers and types of amino acids, even though they may have different sequences. However, our analysis showed that the intestinal permeability of a peptide depends on its sequence (Figure 2) and cannot be explained simply by using the drug-likeness prediction models of passive transport. Because of its large size, the peptide-phage complex is expected to be transported across the intestinal barrier by other mechanisms such as transcytosis.

Systemic delivery of macromolecules via the oral pathway remains one of the most challenging problems in the drug delivery field, and transcytosis may be a mechanism for transporting therapeutic agents across the intestinal barrier. If a carrier molecule, either a natural ligand or an antibody binding to a transcytotic receptor on the intestinal epithelium, is covalently bound to a therapeutic agent by a short linker, the conjugate can bind to the cognate receptor and undergo vesicular trafficking across the intestinal barrier [31]. This 'carrier-drug conjugate' approach has been tried using drugs conjugated to immunoglobulin G (IgG), lactoferrin, transferrin or folic acid, all of which have cognate transcytosis receptors in enterocytes [32]. To utilize the transcytosis mechanism for an oral drug delivery system, it is essential to identify ligands that can bind to the receptors and facilitate efficient transcytosis across the intestinal barrier. Our QSAR study on the selection of intestinal barrier-permeable peptides should be applicable to the development of peptide 'carriers' for delivering large molecules such as proteins and drugs.

We used artificial neural networks to develop the first models for predicting the intestinal permeabilities of peptides on the basis of sequence information. The high quality models obtained were capable of making reliable predictions. These models are expected to find applications in the selection of intestinal barrier-permeable peptides from large peptide libraries, and the selected peptides might be used to facilitate the transport of large molecules across the intestinal barrier.