Optimization ACE inhibition activity in hypertension based on random vector functional link and sine-cosine algorithm

Highlights

• Present an alternative method to predict ACE Inhibition Activity.

• Proposed method a modified random vector functional link (RVFL) network.

• Using the sine-cosine algorithm (SCA) to find the optimal configuration of RVFL.

• Results show that performance of proposed model is better than other methods.

Abstract

Bioactive peptides from protein hydrolysates with antihypertensive properties have a great effect in health, which warrants their pharmaceutical use. Nevertheless, the process of their production may affect their efficacy. In this study, we investigate the inhibitory activities of various hydrolysates on angiotensin-converting enzyme (ACE) in relation to the chemical diversity of corresponding bioactive peptides. This depends on the enzyme specificity and process conditions used for the production of hydrolysates. In order to mitigate the uncontrolled chemical alteration in bioactive peptides, we propose a computational approach using the random vector functional link (RVFL) network based on the sine-cosine algorithm (SCA) to find optimal processing parameters, and to predict the ACE inhibition activity. The SCA is used to determine the optimal configuration of RVFL, improving the prediction performance. The experimental results show that the performance measures of the proposed model are better than the state-of-the-art methods.

Introduction

Functional properties of peptides with antihypertensive effect are directly related to their chemical structure [1]. Therefore, the structure-function relationship of peptides from protein hydrolysates needs to be analyzed in depth to allow optimal use of their chemical diversity by the chemical industry. In this study, we measured the concentration of bioactive amino acids in different chemical structures of peptides, and examined the effect of different temperature levels on their ACE inhibition activity, with the aim to control the chemical alteration of hydrolysates. The uncontrolled chemical alteration in hydrolysates may lead to adverse health effects in hypertensive patients [2]. A part of that, ACE inhibition activity is affected by both amino acid type and chain length [3]. Furthermore, the negative chemical charge of acidic amino acids (aspartic and glutamic) reduces ACE inhibition activity [4]. In turn, the presence of arginine, phenylalanine, valine, proline, lysine, tryptophan, leucine, isoleucine and tyrosine in ACE inhibitory peptides increases the strength of ACE binding [5]. Therefore, the drug industry struggles with the challenge of reducing chemical diversity of the production of hydrolysates and evaluating its effects on functional properties of peptides. Thus, it is important to develop fast, and accurate statistical models to predict and account for the influences of chemical alteration on ACE inhibition activity of bioactive peptides.

The relationship between the chemical structure of a peptide and its ACE inhibitory activity still remains to be fully elucidated [6]. Previous studies used the quantitative structure-activity relationship (QSAR) model to predict ACE inhibition activity [7]. However, this model is limited to the structure, and does not consider recently identified features that affect the structure itself. Noteworthy, our model highlights the impact of process conditions on the chemical alteration in the peptide composition, allowing to account for this alteration and further optimize ACE inhibition activity.

Several machine learning methods were applied to a range of regression problems. These include artificial neural networks (ANN) with one or more hidden layers and the backpropagation training [8,9] and the Support Vector Regression (SVR) [10,11]. In an interesting study, an adaptive neuro-fuzzy inference system (ANFIS) model have been used to predict the bioactive peptides in relation to climate change [1].

An interesting variant of neural network paradigm, the random vector functional link (RVFL) network, introduces the direct connection between the input and output neurons [12,13]. The RVFL, is an ANN with a single layer with the input-output connection, and the nodes of its hidden layer are called enhancement nodes [14]. In the RVFL, the weights between input and enhancement nodes are randomly generated from a suitable interval. This interval should constrain the activation functions away from the saturation region. The weights between input and output nodes, and between enhancement and output nodes, are updated during the training process by optimization techniques such as a conjugated backpropagation, or least squares method.

To date, the RVFL has been applied to various applications, including short-term electricity load demand forecasting [15], remote sensing applications [16], forecasting the distribution of the temperature [17], and Big Data processing [18]. In order to improve the performance of the RVFL network, several methods have been applied. Chen and Wan [19] introduced two versions of RVFL to determine the optimal weights and update them on-the-fly. Scientists presented another improvement to RVFL to solve nonlinear dynamic systems [20] by improving the input pattern with nonlinear functional expansion. In another line of research, Scardapane et al. proposed a distributed RVFL algorithm to improve the efficiency of the RVFL [21]. Noteworthy is the RVFL approach, using the Bayesian inference for robust data modeling [22].

However, in the previous studies, the optimization of several parameters in RVFL was a challenge, as it is an NP-hard problem. To overcome this challenge, Zhang and Suganthan [23] presented a comprehensive study to optimize the parameters of RVFL and to improve its performance in classification and regression tasks. But their method is time-consuming as it requires manual permutation [24,25]. This motivated us to propose an automated method for selecting the optimal configuration of RVFL parameters in order to optimize the ACE inhibition activity prediction. In order to determine this configuration, we used the sine-cosine algorithm (SCA) [26].

For the first time, to the best of our knowledge, we propose a descriptive machine learning paradigm, based on RVFL and SCA, to investigate the chemical diversity of bioactive peptides in order to predict their ACE inhibition activity in hypertensive patients. The proposed method, SCA-RVFL, starts by generating a random population where the dimension of each solution represents the number of parameters used for building the RVFL networks. The dataset is divided into training and testing set randomly. The fitness function of each solution is computed by the RVFL built using the parameters of the current solution. The fitness is calculated as root mean square error between the original ACE and predicted ACE. Overfitting of the predicted inhibition activity is evaluated using the testing set. The next step is to determine the best solution, best configuration, and then update the solutions using the sine and cosine functions. These steps are repeated until the stop conditions are met.

The rest of this paper is organized as follows: in Section 2, materials and methods are discussed, including the preparation of the dataset, the RVFL and SCA methods. In Section 3, the proposed method is introduced. Section 4 presents the experimental results and the discussion. Finally, the conclusion and the future works are introduced in Section 5.