A computational method for prediction of saliva-secretory proteins and its application to identification of head and neck cancer biomarkers for salivary diagnosis.

Human saliva is rich in proteins, which have been used for disease detection such as oral diseases and systematic diseases. In this paper, we present a computational method for predicting secretory proteins in human saliva based on two sets of human proteins from published literatures and public databases. One set contains known proteins which can be secreted into saliva, and the other contains the proteins that are deemed to be not extracellular secretion. The protein features with discerning power between two sets were firstly gathered. Then a classifier was trained based on the identified features to predict whether a protein was saliva-secretory one or not. The average values of the sensitivity, specificity, precision, accuracy, and Matthews correlation coefficient value by 10-fold cross validation repeated 100 times were 80.67%, 90.56%, 90.09%, 85.53%, and 0.7168, respectively. These results indicated that our selected features are informative. We applied the classifier for prediction saliva-secretory proteins out of all human proteins, if a known biomarker was likely to enter into saliva, and the potential salivary biomarkers for head and neck squamous cell carcinoma. We also compared the top 1000 proteins predicted by computational methods in different kind of fluids. This work provided a useful tool for effectively identifying the salivary biomarkers for various human diseases and facilitate the development of salivary diagnosis.