Application of Support Vector Machine Classification Model to Identification of Vascular Endothelial Growth Factor Receptor Inhibitors

Abstract

Background:

Nowadays, with the increasing prevalence of cancer mortality, finding the best cancer inhibitors is vital. Angiogenesis, which refers to the formation of new blood vessels from existing ones, undergoes abnormal changes in the physiological process of solid tumors. Vascular endothelial growth factor receptor (VEGFR) plays a crucial role in angiogenesis. Hence, one of the suggestions in cancer treatment has been inhibiting VEGFR signaling to prevent angiogenesis. The computational approach as an in vitro alternative method is crucial to reduce time and cost. This study aimed to use classification algorithm to separate potent inhibitors from inactive ones.

Materials and Methods:

In order to apply the machine learning model, biological compounds were extracted from the BindingDB database. Due to the large number of molecular features, the classification model was susceptible to overfitting. To address this issue, a correlation-based feature selection algorithm was proposed as a means of feature reduction. Subsequently, for the classification step, a support vector machine model that utilizes both linear and non-linear kernels was employed.

Results:

The implementation of the support vector machine model with the radial basis function kernel, along with the correlation-based feature selection method, resulted in a higher accuracy (81.8%, P value = 0.008) compared to other feature selection methods used in this study. Finally, two structures were introduced with the highest binding affinity to inhibit the second VEGFR.

Conclusion:

According to the results, the correlation-based feature selection method is more accurate than other methods.

INTRODUCTION

The process of drug discovery includes several steps that aim to identify appropriate drug targets from a large number of drug-like compounds. A crucial stage in this process involves the identification of lead compounds, followed by their optimization, in vitro and in vivo studies, and pre-clinical and clinical trials. As a result, the exploration of new treatments becomes a costly and time-consuming endeavor, with costs exceeding 1 billion dollars and a timeframe ranging from 10 to 15 years. The reported failure rate of this process, which takes into account pharmacokinetic properties such as absorption, distribution, metabolism, excretion, and toxicity, is estimated to be approximately 40–60%, underscoring its significance. Consequently, the integration of computer-aided drug discovery techniques in preliminary investigations by prominent pharmaceutical companies and research groups has played a crucial role in expediting the drug discovery and development process.[1]

Cancer is a global public health concern that has emerged as one of the primary contributors to mortality worldwide, primarily due to its high prevalence. Therefore, the development of novel chemical compounds is imperative. Metastasis, the process by which cancer cells migrate to distant tissues through the bloodstream, leading to the formation of secondary tumors, poses a significant challenge in cancer treatment. By effectively inhibiting metastasis, it becomes possible to restrict tumor growth and impede disease progression.[2]

Angiogenesis plays a crucial role in the development of metastasis by facilitating the formation of new blood vessels from existing ones. This process ensures an adequate oxygen supply to cells and efficient removal of waste products. Vascular endothelial growth factor receptor (VEGFR), a member of the tyrosine kinase receptor family, serves as a key mediator in both physiological and pathological angiogenesis, playing an important role in the early stages of this process.[3]

VEGFR promotes the rapid formation of new blood vessels by stimulating the proliferation and migration of endothelial cells. Due to the significant role of VEGFR in angiogenesis, there is an increasing focus in cancer research on identifying the most potent signaling pathways related to this receptor. Recent studies have revealed promising results in clinical trials involving the use of inhibitors that specifically target VEGFR and its secondary receptors for the treatment of tumors.[4]

The observation that tumor blood vessels exhibit distinct characteristics compared to normal vessels has been well established. Tumor blood vessels are characterized by heightened angiogenesis, increased vascular permeability, and irregular distribution. A study conducted in 2015 demonstrated the clinical success of targeting the VEGFR-2 signaling pathway to inhibit tumor-induced angiogenesis in cancer treatments. By gaining a deeper understanding of the binding modes, molecular interactions, and inhibitory mechanisms of VEGFR-2 tyrosine kinases through binding studies of VEGFR-2 protein structures, drugs with improved selectivity and enhanced safety profiles have been discovered. Consequently, this review aims to showcase the considerable potential of molecular modeling in the identification and optimization of small molecules that specifically target receptor tyrosine kinases. Moreover, it seeks to substantiate the validity of VEGFR tyrosine kinases as promising drug targets for cancer treatment.[5]

Simultaneously inhibiting VEGFR-2 and Src (Sarcoma viral oncogene homolog) has the potential to enhance the effectiveness of VEGFR-2-targeted cancer therapies. Src, a member of the non-receptor tyrosine kinase family and a proto-oncogene tyrosine protein kinase, plays a role in promoting colon cancer progression.[6]

A 2017 study employed a multi-stage virtual screening approach to identify new multi-target agents. The screening process involved a ligand-based support vector machine method, a drug similarity rules filter, and structure-based molecular docking. By applying this approach to a large chemical library, compounds that act as dual inhibitors of VEGFR-2 and Src were identified. Experimental validation through kinase inhibition methods and cell viability assays confirmed the findings. Several compounds from six different molecular scaffolds were selected for further biological evaluation.[6] Compound 3, belonging to the 2-amino-3-cyanopyridine scaffold, exhibited a promising anti-proliferative effect and demonstrated dual-target activity against VEGFR-2 and Src.[6]

In a 2018 study, the importance of VEGF and its receptors in both normal and abnormal blood vessel formation (angiogenesis) in tumor progression and metastasis was reaffirmed. Additionally, the study highlighted their significance in the diagnosis and treatment of renal cell cancer, the primary cause of kidney cancer-related deaths worldwide. Through a structure-based virtual screening approach, a comprehensive analysis was conducted on 310,000 molecules from the Protein Data Bank (PDB) database. This analysis resulted in the identification of 23 potential inhibitors, with particular attention given to a compound named SCHEMBL469307, which exhibited high potential in effectively inhibiting the VEGFR-2 protein.[7]

In the same year, a study utilized a combination of ligand-based and structure-based virtual screening methods on a dataset of 1841 entries from the BindingDB database. This comprehensive approach led to the development of an integrated pathway for effectively screening potential inhibitors that target VEGFR-2. Notably, three Food and Drug Administration-approved drugs were identified as novel VEGFR-2 inhibitors, presenting promising outcomes from the study.[8]

In the subsequent research conducted in 2019, a thorough analysis was carried out on a dataset consisting of 2.5 million molecules sourced from the databases Zinc, National Cancer Institute database (Nci), Maybridge, and Asinex. Utilizing virtual screening methods and molecular dynamics simulations, this extensive investigation ultimately resulted in the identification and analysis of 7000 promising compounds. Additionally, a pharmacophore model was developed to facilitate the identification of potential inhibitors targeting VEGFR-2.[3]

In 2020, an investigation was conducted using a ligand-based virtual screening approach on a dataset comprising 2.4 million molecules obtained from the Zinc database. This study resulted in the identification of four compounds, namely, 3, 7, 10, and 13. Among these compounds, compound 10 exhibited promising inhibitory activity against VEGFR-2, with an value of 19.3 µm.[9]

VEGFR-2, being an effective target in the treatment of angiogenesis-related tumors, has seen the approval of small-molecule inhibitors with various scaffolds for diseases such as kidney cancer and non-small-cell lung cancer. However, there remains a high demand and challenge for the development of new VEGFR-2 inhibitors in the market. In a research conducted in 2020, a series of 10 novel VEGFR-2 inhibitors based on the N-methyl-4-oxo-N-propyl-1,4-dihydroquinoline-2-carboxamide scaffold were discovered through structure-based virtual screening. This finding not only expands the chemical space of current VEGFR-2 inhibitors but also contributes to addressing the need for novel inhibitors in the field.[10]

In this study, the ligand-based drug design approach was employed to identify the most suitable inhibitors targeting VEGFR. Various methods, including Quantitative Structure-Activity Relationship (QSAR) modeling, filter-based selection, and embedded feature selection, were utilized to select the optimal inhibitors. Following the development of the QSAR model, it was applied to a dataset consisting of more than 900,000 molecules that had no known effect on the secondary receptor of VEGF. Subsequently, the best conformers were obtained, and docking simulations were conducted on the compounds selected from the QSAR model. Compounds exhibiting the most favorable binding energies with the target protein were identified. Ultimately, two compounds with the potential for yielding promising results in inhibiting the secondary receptor of VEGF were identified and reported [Figure 1].

Figure 1.

Schematic illustration of this study

MATERIALS AND METHODS

Materials

This study consists of a combination of ligand-based and structure-based approaches. In the ligand-based phase, compounds related to VEGFR-2 were extracted from the BindingDB database, and feature selection methods and classifications were applied. Blind modeling was performed using the dataset obtained from the ligand-based phase. In the structure-based phase, the crystallographic structure of VEGFR-2 was extracted from the PDB database. Analysis of the compounds from the ligand-based phase was conducted using Py-Rx software. The analysis focused on the protein 6xvj, which was published in 2020 and has a resolution of 1.7 Angstrom.[11]

Data pre-processing

In this step, all chemical compounds were individually isolated and optimized using Openbabel software, and the resulting structures were stored in the Structure-Data File (SDF) format. Subsequently, molecules with similar structures were identified and removed from the database, resulting in a reduction in the total number of compounds to 9,271.

Molecular descriptors

The dataset consisting of over 3,224 compounds was subjected to analysis using the Dragon software (2007). This software generated a comprehensive set of descriptors, categorized into 22 groups ranging from 0D to 3D descriptors. The resulting output from the Dragon software is presented in the form of a numerical matrix, as illustrated below (Equation 1).

Equation 1: Numerical Matrix (Insert the numerical matrix here)

In the above equation, “m” refers to the number of molecules in the dataset, while “n” represents the number of features or descriptors obtained from the Dragon software. It is important to note that in order to minimize redundancy within the input data, any correlated features were eliminated from the input matrix. This process ensured that the resulting dataset contained non-redundant and informative features for further analysis.

The primary objective of the proposed model was to identify the most potent VEGFR inhibitors from a pool of over 900,000 drug-like compounds. The classification model was developed based on the biological activity values of these small molecules. In order to create this model, a separation threshold of 1 µm was set for each ligand. Accordingly, ligands with an value greater than 1 µm were classified as non-inhibitors, while those with lower values were categorized as VEGFR inhibitors. Consequently, a total of 5,896 compounds were identified as inhibitory compounds, while 2,573 compounds were categorized as neutral based on their activity values.

Machine learning model

It is well established that one of the key strategies in computational drug design involves the development of machine learning models for predicting or classifying inhibitors/activators from neutral compounds. However, in many cases, the number of compounds is significantly smaller than the number of molecular descriptors available. Consequently, it is recommended to apply feature selection algorithms to identify the optimal descriptors before the classification process. The inclusion of a feature selection step offers several advantages, such as reducing model complexity, enhancing model performance, and addressing the issue of redundancy commonly encountered in QSAR models.[8,11] The implementation of this model was performed utilizing Python programming language version 3.7.0 and the Scikit-Learn package version 1.0.2.

Feature selection methods

In this study, two types of feature selection algorithms were employed: a filter-based model, which included correlation-based feature selection (CFS), Fisher score, and mutual information, and a wrapper-based model, which utilized Least Absolute Shrinkage and Selection Operator (LASSO) Ftaand Elastic Net methods.

Classification algorithm

The Support Vector Machine (SVM) method is widely recognized as a powerful and popular classification model in supervised machine learning methods. SVM algorithms belong to the category of supervised classification techniques that aim to separate samples into two or more classes using defined kernels. Over time, various types of kernels have been proposed, including non-linear, polynomial, radial basis function (RBF), and sigmoid. Among these, the radial basis function (RBF) kernel is the most commonly utilized type of kernel function.

RESULTS

The primary objective of this study was to identify potent VEGFR inhibitors from an enormous collection of more than 900,000 drug-like compounds. To achieve this, a machine learning model consisting of two main steps, namely, feature selection and classification, was proposed. Prior to conducting any feature selection and classification model construction, the data were initially partitioned into two distinct groups: a training set comprising 70% of the total data and a separate test set comprising the remaining 30%. This division was achieved using a technique known as “test and train split.”

In the subsequent analysis, we employed K-fold cross-validation on the 70% training dataset, where we set K to a value of 5. Following the identification of the optimal feature selection method, we utilized the entire 70% training dataset for modeling purposes, repeating this process 100 times. It is important to highlight that the 30% test dataset, which was initially set aside, was not utilized in any of the modeling procedures, thus mitigating the risk of overfitting on the test data.

For internal validation, the k-fold cross-validation method was employed. This method is widely used for estimating the performance of an algorithm or machine learning configuration on a given dataset. In k-fold cross-validation, the input data are divided into k different subsets or folds. Each fold is then used as a test set, while the remaining folds are used as the training dataset. This process is repeated k times, and the average performance is reported and for external validation; a randomly selected 30% of the data was set aside as a test set, while the remaining data were used as the training set. This allowed for an independent evaluation of the model’s performance on unseen data.

To assess the performance of the model, four evaluation indexes were utilized: sensitivity (SE), specificity (SP), accuracy (Q), and the Matthews correlation coefficient (MCC). The MCC, which ranges from -1 to 1, is considered the most crucial indicator of binary classification quality. Positive values refer to inhibitory compounds, while negative values represent non-inhibitory compounds.[12]

The equation for calculating the indicators is as follows:

In these equations, TP represents true positives (correctly predicted inhibitory compounds), TN represents true negatives (correctly predicted non-inhibitory compounds), FP represents false positives (incorrectly predicted inhibitory compounds), and FN represents false negatives (incorrectly predicted non-inhibitory compounds).

Setting up the proposed classification parameters

To establish the SVM model for classification, several parameters were configured, including the selection of an appropriate kernel. In this study, four different kernels were examined, and the outcomes were compared to determine the most effective kernel. The evaluation criteria used to assess the performance of the model are presented in Figure 2.

Figure 2.

The result of support vector machine model

Based on the four main evaluation criteria, it was determined that the RBF kernel outperformed the other methods. Consequently, the RBF kernel was selected as the default kernel for the classification step. Moving on to the feature selection step, two approaches, namely, filter-based and wrapper-based, were employed. To identify the best filter-based model, three different methods were applied: correlation-based (CFS), mutual information, and Fisher score. The results obtained from each method were compared with each other as well as with the classification without feature selection, as shown in Table 1. It can be observed that the CFS method yielded superior results compared to the other methods.

Table 1:

Results of support vector machine model with RBF kernel on some filter-based feature selection methods

Filter-based feature selection methods	Number of features	Q (%)	MCC (%)	SE (%)	SP (%)
RBF kernel without feature selection	3224	81.8	57.7	92.3	82.4
Correlation-based feature selection	1431	82.4	59.1	92.2	83.1
Mutual Information	1859	81.5	57	91.9	82.3
Fisher Score	1117	81.2	56.2	91.4	81.8

In the second step, the performance of the CFS-RBF model was compared with those of two different embedded-based models. Figure 3 illustrates the results, indicating that both the Elastic Net and LASSO methods yielded similar outcomes. However, the Elastic Net method exhibited slightly higher sensitivity and accuracy compared to the LASSO method.

Figure 3.

Results of support vector machine model with RBF kernel on embedding-based feature selection methods

To identify the best potent VEGFR inhibitors from the large pool of more than 900,000 chemical compounds, the CFS-Elastic Net-RBF model, as mentioned in the results section, was deemed reliable. The model was repeated 100 times, and all molecules from the BindingDB database were analyzed using this model. As a result, 3300 inhibitory compounds targeting VEGFR-2 were obtained.

Subsequently, all the extracted molecules underwent evaluation using the Py-Rx method, and the results were extracted using Python software. Following this, the receptor molecules were subjected to clustering using the Similarity Search approach. Molecules exhibiting more than 80% similarity were grouped together in respective clusters.

Last, compounds with significant structural similarity and high binding affinity were carefully chosen, and their corresponding CID¹ codes were obtained through the PubChem website. It is worth noting that numerous inhibitors for VEGFR-2 have been proposed based on X-ray crystal structures. However, many of these molecules are still undergoing clinical trials.[13] In our study, the model and analysis yielded two compounds with inhibitory properties against the second receptor of VEGFR (refer to [Figure 4]).

Figure 4.

(a) cid_ 5330860, binding affinity = -14.4; (b) cid_ 53374320, binding affinity = -13.5

DISCUSSION

Upon comparing the structures of our two inhibitors with commercially available inhibitor drugs targeting this receptor, we observed structural similarities between these molecules.[13,14,15]

However, further calculations are necessary to categorize these identified molecules into one of the three classes of VEGFR-2 inhibitors.

The three classes of VEGFR-2 inhibitors are as follows:

ATP²-competitive inhibitors: These inhibitors compete with ATP for binding to the receptor’s active site.

Class II inhibitors: These inhibitors have the ability to bind near the receptor’s hydrophobic pocket.

Class III inhibitors: These inhibitors are covalent inhibitors that bind to specific cysteine amino acid residues.

To determine the class to which our identified molecules belong, additional calculations and analyses are required.[3,16,17]

It is true that the structural similarity among all VEGF receptors allows VEGFR-2 inhibitors to potentially target other receptors within the VEGF family. Additionally, the catalytic domain of VEGFR-2 bears similarity to other receptors like PDGFRs and c-kit. As a result, these inhibitors may not only be specific to VEGFR-2 but can also inhibit other tyrosine kinase receptors.[18,19]

Furthermore, it can be hypothesized that these molecules, given their negative binding affinities and structural similarities with receptor inhibitor drugs, have the potential to bind effectively to the receptor, thereby impeding its interaction with VEGF. By inhibiting the activity of VEGFR and interrupting the progression of its signaling pathway, subsequent events such as angiogenesis, cell growth, and proliferation can be halted, ultimately preventing tumor growth.

These identified molecules may possess anti-tumor and anti-cancer properties, which warrant further investigation in future research endeavors.

Indeed, the results obtained from calculations and mathematical modeling provide valuable insights and predictions. However, to validate and confirm the accuracy and reliability of these findings, it is crucial to conduct experimental studies in controlled laboratory environments. Experimental validation helps to assess the actual biological activity, efficacy, and safety of the identified molecules as VEGFR inhibitors. These studies involve in vitro assays, in vivo animal models, and eventually clinical trials to evaluate the potential of these molecules as therapeutic agents. Experimental validation is essential for translating computational predictions into tangible outcomes and advancing toward the development of effective treatments for cancer and other related conditions.

It can be argued that inhibitors developed using computational drug design approaches offer targeted and precise solutions for the treatment of complex diseases, such as cancer. The accurate calculations and time and cost savings associated with these approaches make them highly valuable in drug development. Therefore, it is recommended that this project and similar studies serve as a foundation for future research. Validating the findings in laboratory experiments will further establish the reliability and efficacy of these computational designs. Ultimately, this approach has the potential to lead us toward the discovery and development of more suitable and personalized drugs for a broad spectrum of diseases, including complex conditions.

CONCLUSION

The findings of this study highlight the potential of VEGFR inhibitory compounds as promising candidates for effective treatments against solid tumor growth. Moreover, these compounds hold promise for addressing other angiogenesis-dependent diseases such as rheumatoid arthritis, diabetic retinopathy, and macular degeneration. The utilization of bioinformatics and computational drug design approaches enables the identification of more accurate and targeted therapeutic goals, while also offering significant time and cost savings. This is particularly crucial for countries with limited resources as it alleviates the burden of exorbitant medical expenses. Additionally, targeted and precise drugs for diseases like cancer can improve patient tolerance and enhance treatment efficacy. Embracing the advancements in computational drug design and bioinformatics opens new avenues for the development of personalized and effective therapies, ultimately benefiting patients and improving healthcare outcomes.

Financial support and sponsorship

Isfahan University of Medical Sciences Research Assistant.

Conflicts of interest

There are no conflicts of interest.