Introduction

Peptides have attracted attention for their ability to regulate a variety of cellular events by interacting with receptors on the cell surface. A large number of biomaterials that conjugate functional peptides have been developed for medicine, drug delivery, tissue engineering, bio-imaging, and food additives. However, identifying functional peptides by exhaustive screening that covers all the possible combinations of 20 amino acids is terribly time consuming.

Computationally assisted peptide screening and design have become invaluable tools in overcoming this inefficiency in peptide screening. One of the most advanced areas in the computational analysis of peptides is MHC ligand prediction (1,2,3,4). The peptides that bind to the highly polymorphic molecule MHC were discovered partially by experiment, and all of the data have been accumulating in databases, such as MHCPEP (http://wehih.wehi.edu.au/mhcpep/) and MHCBN (http://www.imtecfh.res.in/reghava/mhcbn/). Using these data as training data, quantitative modeling studies on peptides binding to the MHC have been constructed to have high predictive accuracies for distinguishing unknown peptide ligands by use of various types of algorithms, including a Bayesian neural network (1), artificial neural network (ANN) (2), weight matrix (3), inheritable bi-objective genetic algorithm (4), and support vector machines (5,6). There are also strategies that predict peptide folding (7) or interactive relationships with target receptors (8), which should offer powerful support for drug screening. However, the predictive efficiency of most of these methods is evaluated with their own datasets. Repeated confirmations of the peptide interaction assays to validate the robustness of predictors are not usually performed. In our previous work, we constructed a new computational model for predicting MHC class II binding peptides. A fuzzy neural network (FNN) (9) and hidden Markov model (10) were demonstrated to increase the predictive accuracy of an acquired prediction model. Also, it should be noted that existing computational work is often applied only to restricted prediction studies. For example, the MHC binding peptides possess a core binding region of approximately 9-mer, and important positions that are known to match the structure of the cavity in the receptor have been evaluated for their affinity. These biological restrictions can make the modeling process simpler and increase predictive accuracy. It would be helpful to develop a new computational tool for general predictive studies having no or only light restrictions.

In this study, we focused on cell-peptide interaction, screening for ideal synthetic peptide sequences to apply to the design of cell-culture scaffolds. Since cells express various receptors on the cell surface, the assay output—that is, the cell-adhesion rate—consists of much complex information. However, by repeated evaluation of the model, we found that such complex interactions can be modeled and that by interpreting the constructed model, structural rules for target peptides can be established as “design rules” through the use of our original tool, the FNN (11).

The FNN is a type of ANN that automatically constructs complex model structures by learning the hidden relationship between input and output data, thus functioning as a predictor. It has been shown that high predictive accuracies can be obtained in the analysis of various data derived from complex phenomena, such as bioprocess factors and fermentation efficiency (12), protein structure information and protein enantioselectivity (13), expression profiles of microarray data and the diagnosis of disease development (14), and the prediction of peptide binding to MHC class II (9). As compared with ANN, FNN has an advantage: the “fuzzy layer.” This allows interpretation of the model structure and extraction of the quantified relationship between input and output values as a rule designated as the “if-then” rule. This feature distinguishes FNN from the usual neural network-based models, since the former provides rules for structural design described by combinations of the physical properties of important residues, whereas the latter function only as a black box. We have used this tool to enhance the efficiency of peptide screening when the peptide interaction rules are ambiguous, such as in cell-peptide interactions, which are the focus of this study.