Method of regulatory network that can explore protein regulations for disease classification

Abstract

Objective: To develop regulatory network to explore and model the regulatory relationships of protein biomarkers and classify different disease groups.

Methods: Regulatory network is constructed to be a hopfield-like network with nodes representing biomarkers and directional connections to be regulations in between. The input to the network is the measured expression levels of biomarkers, and the output is the summation of regulatory strengths from other biomarkers. The network is optimized towards minimizing the energy function that is defined as the measure of the disagreement between the input and output of the network. To simulate more complicated regulations, a sigmoid kernel function is imposed on each node to construct a non-linear regulatory network.

Results: Two datasets have been used as test beds, one dataset includes patients of nasopharyngeal carcinoma with different responses to chemotherapy drug, and the other consists of patients of severe acute respiratory syndrome, influenza, and control normals. The regulatory networks among protein biomarkers were reconstructed for different disease conditions in each dataset. We demonstrated our methods have better classification capability when comparing with conventional methods including Fisher linear discriminant (FLD), K-nearest neighborhood (KNN), linear support vector machines (linSVM) and radial basis function based support vector machines (rbfSVM).

Conclusion: The derived networks can effectively capture the unique regulatory patterns of protein markers associated with different patient groups and hence can be used for disease classification. The discovered regulation relationships can potentially provide insights to revealing the molecular signaling pathways. In this paper, a novel technique of regulatory network is proposed on purpose of modeling biomarker regulations and classifying different disease groups. The network is composed of a certain number of nodes that are directionally connected in between in which nodes denote predictors and connections to be the regulation relationship. The network is optimized towards minimizing its energy function with biomarker expression data acquired from a specific patient group, thus the optimized network can model the regulatory relationship of biomarkers under the same circumstance. To simulate more complicated regulations, a sigmoid kernel function is imposed on each node to construct a non-linear regulatory network. The regulatory network can extract unique features of each disease condition, thus one immediate application of regulatory network is to classifying different diseases. We demonstrated that regulatory network is capable of performing disease classification through comparing with conventional methods including FLD, KNN, linSVM and rbfSVM on two protein datasets. We believe our method is promising in mining knowledge of protein regulations and be powerful for disease classification.

Fig. 1

Structure of regulatory network.

Fig. 2

Classification framework based on regulatory networks.

Fig. 3

Performances of our RN (A) and NRN (B) classifiers for the NPC dataset.

Fig. 4

The expression levels of the selected proteins in chemo-responders (RS) and nonresponders (NR).

Fig. 5

Optimized non-linear regulatory networks with five nodes for the NPC dataset. Subfigure A is for RS; subfigure B is for NR.

Fig. 6

Comparison of the performance of our NRN classifiers with several conventional classification methods on the NPC dataset.

Fig. 7

Performances of our RN and NRN classifiers for the SARS dataset.

Fig. 8

Comparison of the expression levels of the 5 proteins used by the 5-biomarker NRN classifier in normals, IFZ and SARS groups.

Fig. 9

Optimized non-linear regulatory networks with five nodes. Red and green lines represent positive and negative regulations, respectively, and the width of lines indicates the strength of regulations. A is for normal group; B is for influenza-infected group; C is for SARS group. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of the article.)

Fig. 10

Comparison of the performance of our NRN classifiers with several conventional classification methods on the SARS dataset.

Fig. 11

Pearson relations of regulatory matrices by different sigmoid factors. (A) and (B) is for the 5 regulatory matrices of the RS and NR classes, respectively, and (C) is between the regulatory matrices of RS and those of NR.

Fig. 12

Probability distributions of the deviations of regulatory coefficients by different sigmoid factor for the RS and NR class.