Improved swarm intelligence algorithms with time-varying modified Sigmoid transfer function for Amphetamine-type stimulants drug classification
Introduction
Amphetamine-type stimulant (ATS) drugs are one of the popular synthetic drugs of abuse. These substances were originally developed for pharmacological research, but, the underground chemist keep modifying the chemical structure of these compounds to evade legal regulation. The novelty of these substances makes them undetectable by traditional drug testing methods [1].
Nowadays, there are different devices and test kits available in the market for use in ATS drug testing [[2], [3], [4]]. However, several drawbacks were discovered within these methods such as long preparation and execution time, costly apparatus, and equipment, complex testing process, requiring well-trained technicians, unreliable and inconsistent outputs from different test kits, and outdated analytical methods. At present, computational methods have been shown as the promising techniques in the cheminformatics field such as in drug design and discovery [5], drug-non drug classification [6], molecular similarity analysis [7], toxicity prediction [8,9], and quantitative structure-activity/property relationships (QSAR/QSPR) analysis [10]. Computational methods also offer much cheaper and faster procedures. Molecular descriptors are an important component that has provided support for many modern computational models. Molecular descriptors are numerical indexes encoded from molecular structure representation of different dimensionalities (0D, 1D, 2D, 3D, or 4D). The higher the dimensionality the more information about the molecular features is stored in the descriptors.
Due to the rapid increment of chemical data, machine learning has become a promising tool to process big data at high volume, veracity, and velocity and with enormous flexibility [11]. In a previous study, Pratama et al. proposed an approach for ATS drug identification by employing a newly developed 3D image pre-processing technique called 3D Exact Legendre Moment Invariants (3D ELMI) as a feature extractor algorithm and several classification algorithms [12,13]. 3D ELMI is a molecular descriptor algorithm that is used to calculate descriptors of 7190 drug compounds (equal size of ATS and non-ATS drugs). It generates 1185 descriptors for each drug compound and is then used as input to the classification algorithms to perform the identification task. The experimental results found that the random forest (RF) classifier is superior in achieving the highest classification accuracy. According to Ref. [14], molecular classification involved three steps: feature extraction, feature selection, and classification. This study aims to improve the ATS drug classification performance by carrying out the feature selection as a data-preprocessing step before executing the classification task.
The existence of new molecular descriptors that generate high-dimensional descriptors has made the feature or descriptor selection step required in computational modeling. Descriptor selection is a well-known non-polynomial (NP) hard combinatorial search problem as the number of possible feature subsets grows exponentially with the increase of dimensionality. The traditional feature selection techniques are inefficient to handle medium or large descriptors or datasets. Therefore, the SI algorithm is one of the core technology to address this issue. SI algorithm is classified as a population-based optimization algorithm having several advantages [15]:
- •
Ease of implementation.
- •
Fewer operators compared to evolutionary approaches.
- •
Fewer parameters to tune.
- •
Retain information about the search space throughout the iteration.
- •
Regularly use memory to save the best solution obtained so far.
Some of the successful works that implemented SI algorithms in the descriptors selection problem are outlined in Table 1.
Descriptor selection has become popular research in the cheminformatics domain where researchers attempt to identify the lowest feasible number of descriptors that can provide good predictive performance [16]. The research area of incorporating and implementing SI algorithms to the feature selection problem is still active to date. According to the No Free Lunch (NFL) theorem, there is no universal algorithm that applies to all optimization problems [17]. Therefore, there is always an opportunity to come up with new metaheuristic-based feature selection algorithms to enhance the process of solving feature selection problems. Numerous new approaches are originated in the literature [[18], [19], [20], [21], [22]]. Motivated by this, this research proposed ten new SI-based feature selection algorithms to obtain the significance and discriminative 3D ELMI descriptors to enhance the classification performance.
The initial version of the SI algorithm produces a continuous solution and is only applicable to solve continuous optimization problems. Therefore, the binary version of the SI algorithm is mandatory to generate a binary solution for addressing binary optimization problems. Examples of binary optimization problems are feature selection [23], and traveling salesman problem (TSP) [24]. The common practice is to use the transfer function as a conversion method [[25], [26], [27]]. The implementation of a transfer function is straightforward and does not increase the complexity of the original algorithm. In addition, the utilization of a suitable transfer function will provide a good balance between the exploration and exploitation phases in the SI algorithm resulting in a better convergence and good classification accuracy. Several popular transfer functions used in the literature are listed in Table 2.
In the present research, we introduced a time-varying modified Sigmoid transfer function with a linear time-varying updating strategy. We also adopted the transfer function that we have proposed in Ref. [28]. To evaluate the efficiency of the particular transfer functions, we integrate them into five continuous SI algorithms: particle swarm optimization algorithm (PSO), whale optimization algorithm (WOA), grey wolf optimization algorithm (GWO), harris hawk optimization algorithm (HHO), and manta-ray foraging optimization algorithm (MRFO). The characteristics of these SI algorithms are summarized in Table 3 while Table 4, Table 5 present the transfer functions used to produce their binary versions.
In a few words, this paper introduces a new approach for tackling descriptors selection problem based on the PSO, BWO, GWO, HHO, and MRFO algorithms, and its main contributions can be summarized as follows:
- 1.
A novel time-varying modified Sigmoid transfer function with a linear (TV1) time-varying updating scheme is introduced as a binarization technique for the metaheuristic algorithm.
- 2.
Ten new binary variants of PSO, BWOA, GWO, HHO, and MRFO algorithms are developed by employing TV1 and our recently proposed transfer function (TV2) in Ref. [28]: BPSOTV1, BPSOTV2, BWOATV1, BWOATV2, BGWOTV1, BGWOTV2, BHHOTV1, BHHOTV2, BMRFOTV1, BMRFOTV2.
- 3.
These proposed SI algorithms are adapted as a feature search for wrapper feature selection in a supervised binary classification task that differentiates ATS and non-ATS drugs.
- 4.
The final results were assessed based on different performance metrics, including the average fitness, average classification accuracy, average fitness, average number of selected features, as well as the respective standard deviation values.
- 5.
The significance of the proposed algorithms was validated against competitive algorithms using a Wilcoxon's rank-sum non-parametric statistical test at a significance level of α = 0.05.
There are several limitations of this study which include:
- 1.
The transfer functions are validated on five SI-based optimization algorithms only.
- 2.
The algorithms are used to solve the descriptors selection problem in the drug analysis domain.
- 3.
Only one chemical dataset is used for algorithms evaluation.
- 4.
This study does not apply any data preprocessing to the dataset.
The remainder of this paper is structured as follows. Section 2 explains the concepts of the proposed new transfer functions and their implementation in PSO, GWO, WOA, HHO, and MRFO algorithms. Section 3 describes the necessary material and methods used in the experiments. Results and discussions are presented in Section 4. Finally, Section 5 concludes and highlights some future works.
Section snippets
The binary version of PSO, GWO, WOA, HHO, and MRFO
[23,24,50]. In BPSO, BWOA, BGWOA, BHHO, and BMRFO, the search agents (solutions) update their positions continuously to any point in the search space based on the best search agent discovered so far. Then the real position of search agents is converted to binary values using the proposed time-varying modified Sigmoid transfer function. This technique forces search agents to move in a binary space by probability definition which updates each element (feature) in the solution (features subset) to
Dataset
In this study, a special chemical dataset was adopted in the experiments. The dataset contains 1187 descriptors to represent the 3D molecular structure of 3595 ATS dan 3595 non-ATS drugs compounds. The descriptors were generated by using a novel 3D Exact Legendre Moment Invariant molecular descriptors algorithm that was developed by Pratama in Refs. [12,13]. The descriptors comprise molecule id, 1185 moment invariants values, and a binary class label, 0 (non-ATS drug) and 1 (ATS drug). During
Experimental results and discussion
Table 8 present the detailed experimental results. The average fitness value achieved by the proposed algorithms is much lower when compared with their existing binary version except for BGWO TV1. It is confirmed by the convergence curves illustrated in Figs. 2–6. This disclosed the potential of TV2 in improving the convergence of BPSO, BWOA, BGWO, and BHHO. Meanwhile, TV1 is capable to help the BMRFO algorithm to escape from local optima more efficiently. BMRFOTV1 is seen to provide the lowest
Conclusions and future work
In this paper, the time-varying modified Sigmoid transfer function with two time-varying update strategies are introduced as binarization methods for PSO, GWO, WOA, HHO, and MRFO algorithms. These proposed binary SI algorithms are used in the descriptors selection problem to improve ATS drug classification accuracy. Among all the best proposed binary algorithms, BMRFOTV1 was seen to be superior and significant in selecting relevant and discriminative descriptors and obtained high classification
Author statement
Norfadzlia Mohd Yusof: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Writing – Original draft, Visualization. Azah Kamilah Muda: Conceptualization, Supervison, Project administration, Funding acquisition. Satrya Fajri Pratama: Conceptualization, Data Curation, Resource, Supervison. Ramon Carbo-Dorca: Wrting – Review & Editing. Ajith Abraham: Wrting – Review & Editing.
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
This work was supported by Universiti Teknikal Malaysia Melaka through the Fundamental Research Grant Scheme [FRGS/1/2020/FTMK-CACT/F00461] from the Ministry of Higher Education, Malaysia.
References (64)
Synthetic drugs of abuse
Adv. Clin. Chem.
(2021)- et al.
A novel hybrid Harris hawks optimization and support vector machines for drug design and discovery
Comput. Chem. Eng.
(2020) - et al.
Drug/nondrug classification using Support Vector Machines with various feature selection strategies
Comput. Methods Progr. Biomed.
(2014) - et al.
Machine learning in chemoinformatics and drug discovery
Drug Discov. Today
(2018) - et al.
Grey wolf optimizer
Adv. Eng. Software
(2014) - et al.
High-dimensional QSAR/QSPR classification modeling based on improving pigeon optimization algorithm
Chemometr. Intell. Lab. Syst.
(2020) - et al.
The whale optimization algorithm
Adv. Eng. Software
(2016) - et al.
Harris hawks optimization: algorithm and applications
Future Generat. Comput. Syst.
(2019) - et al.
Manta ray foraging optimization: an effective bio-inspired optimizer for engineering applications
Eng. Appl. Artif. Intell.
(2020) - et al.
Binary grey wolf optimization approaches for feature selection
Neurocomputing
(2016)
Binary dragonfly optimization for feature selection using time-varying transfer functions
Knowl. Base Syst.
S-shaped versus V-shaped transfer functions for binary Particle Swarm Optimization
Swarm Evol. Comput.
A time-varying mirrored S-shaped transfer function for binary particle swarm optimization
Inf. Sci.
Whale optimization approaches for wrapper feature selection
Appl. Soft Comput.
Binary grasshopper optimisation algorithm approaches for feature selection problems
Expert Syst. Appl.
Binary ant lion approaches for feature selection
Neurocomputing
Binary artificial algae algorithm for feature selection
Appl. Soft Comput.
Improved binary adaptive wind driven optimization algorithm-based dimensionality reduction for supervised classification
Comput. Ind. Eng.
An overview of forensic drug testing methods and their suitability for harm reduction point-of-care services
Harm Reduct. J.
An LC-MS/MS methodological approach to the analysis of hair for amphetamine-type-stimulant (ATS) drugs, including selected synthetic cathinones and piperazines
Drug Test. Anal.
Amphetamine-type stimulants in drug testing
Mass Spectrom Lett
Using cheminformatics to predict cross reactivity of “designer drugs” to their currently available immunoassays
J. Cheminf.
Efficient toxicity prediction via simple features using shallow neural networks and decision trees
ACS Omega
A review on machine learning methods for in silico toxicity prediction
J. Environ. Sci. Health Part C Environ. Carcinog. Ecotoxicol. Rev.
Incorporating PLS model information into particle swarm optimization for descriptor selection in QSAR/QSPR
J. Chemom.
Three-Dimensional Exact Legendre Moment Invariants for Amphetamine-type Stimulants Molecular Structure Representation
Preparation of ATS drugs 3D molecular structure for 3D moment invariants-based molecular descriptors
A hybridised feature selection approach in molecular classification using CSO and GA
Int. J. Comput. Appl. Technol.
QSAR model for predicting neuraminidase inhibitors of influenza A viruses (H1N1) based on adaptive grasshopper optimization algorithm
SAR QSAR Environ. Res.
No free lunch theorems
IEEE Trans. Evol. Comput.
Chaotic Vortex Search Algorithm: Metaheuristic Algorithm for Feature Selection
A novel hybrid whale optimization algorithm with flower pollination algorithm for feature selection: case study Email spam detection
Comput. Intell.
Cited by (10)
Copula entropy-based golden jackal optimization algorithm for high-dimensional feature selection problems
2024, Expert Systems with ApplicationsOptimizing microarray cancer gene selection using swarm intelligence: Recent developments and an exploratory study
2023, Egyptian Informatics JournalHybrid binary grey wolf naked mole-rat algorithm for fragment-type UWB antenna optimization using time-varying transfer functions[Formula presented]
2023, Expert Systems with ApplicationsImproving Amphetamine-type Stimulants drug classification using chaotic-based time-varying binary whale optimization algorithm
2022, Chemometrics and Intelligent Laboratory SystemsCitation Excerpt :Besides that, the transfer function can improve the exploration and exploitation of the SI algorithm therefore the selection of an appropriate transfer function is important. Various transfer functions have been utilized and examined by several researchers in BWOA for feature selection problems in supervised data classification [22–27]. Another improvement method in the SI domain is the application of chaos theory.
Advances in Manta Ray Foraging Optimization: A Comprehensive Survey
2024, Journal of Bionic EngineeringSigmoid activation function generation by photonic artificial neuron (PAN)
2024, Optical and Quantum Electronics