Machine Learning-Based Virtual Screening and Identification of the Fourth-Generation EGFR Inhibitors

Abstract

Epidermal growth factor receptor (EGFR) plays a pivotal regulatory role in treating patients with advanced nonsmall cell lung cancer (NSCLC). Following the emergence of the EGFR tertiary CIS C797S mutation, all types of inhibitors lose their inhibitory activity, necessitating the urgent development of new inhibitors. Computer systems employ machine learning methods to process substantial volumes of data and construct models that enable more accurate predictions of the outcomes of new inputs. The purpose of this article is to uncover innovative fourth-generation epidermal growth factor receptor tyrosine kinase inhibitors (EGFR-TKIs) with the aid of machine learning techniques. The paper’s data set was high-dimensional and sparse, encompassing both structured and unstructured descriptors. To address this considerable challenge, we introduced a fusion framework to select critical molecule descriptors by integrating the full quadratic effect model and the Lasso model. Based on structural descriptors obtained from the full quadratic effect model, we conceived and synthesized a variety of small-molecule inhibitors. These inhibitors demonstrated potent inhibitory effects on the two mutated kinases L858R/T790M/C797S and Del19/T790M/C797S. Moreover, we applied our model to virtual screening, successfully identifying four hit compounds. We have evaluated these hit ADME characteristics and look forward to conducting activity evaluations on them in the future to discover a new generation of EGFR-TKI.

1. Introduction

The epidermal growth factor receptor is a kinase within the human epidermal growth factor receptor (HER) family and holds a pivotal regulatory role across multiple malignancies. This protein family serves as an appealing drug target, particularly within the context of nonsmall cell lung cancer.¹⁻⁴ In recent times, with the intensification of research into tumor targets, targeted antitumor medications, typified by epidermal growth factor receptor tyrosine kinase inhibitors (EGFR-TKIs), have emerged as the primary treatment option for patients in the advanced stages of nonsmall cell lung cancer (NSCLC). These drugs have demonstrated notable efficacy, minimal adverse effects, and extended periods of disease-free progression survival post treatment.

The EGFR-TKIs, which include the first generation: erlotinib⁵ and gefitinib,⁶ the second generation: dacotinib⁷ and afatinib,⁸ and the third generation: osimertinib,⁹ are extensively employed in clinical practice, catering to around 90% of patients possessing EGFR-sensitizing mutations. These inhibitors are commonly administered as standalone treatments or in conjunction with monoclonal antibodies to address patient conditions. Unfortunately, prolonged utilization of EGFR inhibitors has encountered substantial challenges due to the emergence of acquired drug resistance. This phenomenon has led to a reduction in clinical effectiveness.¹⁰⁻¹² Recently, approximately 40% of all resistance cases are caused by C797S mutations, making them the primary cause of resistance to third-generation inhibitors.^13,14 Encouraging strides have been made in the development of fourth-generation EGFR-TKIs. Drugs like BLU-945,¹⁵ BLU-701,¹⁶ TBQ3804,¹⁷ BBT-176,¹⁸ BPI-361175,¹⁹ and others have progressed into the realm of clinical research. However, the advancement of most other drugs to clinical studies has faltered, and none have successfully cleared phase III clinical trials. These drugs exhibit diverse structures, including isoindoline-1-one derivatives, 4-amine-pyrimidine derivatives, quinoline derivatives, 2-amino-pyrimidine derivatives, and more. Hence, further investigation remains imperative to uncover the fourth generation of novel EGFR-TKIs.

Computational modeling brings significant advantages to drug discovery and has found extensive application in computer-aided drug discovery.²⁰⁻²² In recent years, numerous QSAR models for EGFR-TKIs have been developed. In the development of wild-type and monomutant EGFR inhibitor models, Chauhan et al.²³ collected 128 quinazoline inhibitors and developed wild-type EGFR and monomutant EGFR (L858R) support vector machine (SVM) models. The coefficients of determination (R²) for the wild-type and mutant models reached 0.83 and 0.71, respectively. Yan et al. collected a data set of 1248 EGFR inhibitors and built two classification models using the methods of Kohonen’s self-organizing map (SOM) and SVM.²⁴ The SOM model had prediction accuracy rates of 98.5 and 96.3% on the training set and test set, respectively, while the SVM model had rates of 99.0 and 97.0%, respectively. They further developed two-dimensional (2D) classification models and three-dimensional (3D) CoMSIA models for wild-type and L858R/T790M double mutant EGFR-TKI.²⁵ For 2D models, the accuracy of each model was greater than 0.87.

Significant strides have been made in the QSAR modeling of EGFR tyrosine kinase inhibitors. However, there are still gaps when it comes to modeling small-molecule inhibitors targeting the EGFR^{L858R/T790M/C797S} triple mutation. This is due to the short duration of inhibitor studies focused on the C797S mutation, leading to a small sample size of compounds. Because of the resulting structural diversity stemming from the relatively low number of inhibitors targeting triple mutations, it becomes imperative to train models using known activities and structures to propel advancements in medicinal chemistry studies. This paper intends to develop an advanced machine learning model capable of identifying small-molecule inhibitors that zero in on the EGFR^{L858R/T790M/C797S} triple mutation. Furthermore, this model will be applied to virtually screen potential inhibitors, and molecular structure descriptors will be extracted to guide the synthesis of inhibitors (Figure 1). A considerable portion of the existing literature combines all of the descriptors into a single feature selection method for screening purposes, thereby rendering it difficult to acquire precise structural descriptors for guiding compound synthesis. In this context, we introduced an integrated framework that initially employs the full quadratic effect model to screen structural descriptors, followed by utilizing Lasso for screening nonstructural descriptors. This approach guarantees the acquisition of essential structural molecular descriptors crucial for predicting activity, thereby furnishing a broader array of avenues for designing molecular structures. To initiate the process, we amassed a data set comprising 221 small-molecule inhibitors specifically targeting the EGFR^{L858R/T790M/C797S} triple mutation. The open-source toolkit RDKit was employed to condense the SMILES representation of each inhibitor, yielding a comprehensive set of descriptors elucidating structural information. Subsequently, an integrated framework combining full quadratic effect modeling and Lasso modeling was employed to meticulously select pivotal descriptors, encompassing both structural and general descriptors. These acquired descriptors were then employed in training the support vector regression (SVR) model. Concurrently, the structural descriptors played a pivotal role in steering the synthesis of small-molecule inhibitors engineered to target the EGFR^{L858R/T790M/C797S} triple mutation. This was corroborated through kinase inhibition activity experiments. Furthermore, our model was extended to virtual screening, wherein molecular simulation docking techniques were harnessed to identify the potential EGFR-TKIs. The safety of these identified hits was subjected to rigorous analysis via an ADME evaluation. Ultimately, by leveraging the structural descriptors as guidance, three inhibitors were meticulously designed and synthesized. In-depth in vitro kinase activity assays were subsequently conducted to validate their efficacy in inhibiting the activity of the triple mutant kinases, with varying degrees of success. Moreover, virtual screening resulted in the identification of four hits with the potential to inhibit triple mutant kinases. These hits are now poised for further in-depth investigation and study.

Figure 1.

Flowchart of molecular descriptor selection, model construction, and virtual screening.

2. Materials and Methods

2.1. Data Preparation

We collected 30 papers published in recent years, which focus on small-molecule inhibitors of EGFR ^{L858R/T790M/C797S} triple mutations and compiled a data set containing 221 valuable active structures and their corresponding kinase inhibitory activities (expressed as IC₅₀).²⁶⁻⁵⁵ For screening efficiency, only structures with IC₅₀ < 5 μM and with accurate numerical representation have been retained in the collation. The activity values were converted to the pIC₅₀(−log₁₀(IC₅₀)) for the sake of the same magnitude on our data.

2.2. Data Preprocessing

In order to validate the performance of the machine learning models, the whole data set was split into a training and a test set at a 3:1 ratio randomly, whose training set included 165 compounds and the test set included 56 compounds.

An important challenge in computational chemistry was to present molecular structures as pieces of information so that computers can process them and use them to train models. The method we used here is to convert the compound SMILES formula into 2D descriptors via RDKit,⁵⁶ which is an open-source toolkit for cheminformatics based on 2D and 3D molecular manipulations of compounds. The generated molecular descriptors capture various features of individual molecules, including molecular properties, connectivity, composition, topological information, and MOE-type (molecular operating environment) information.

In this work, calculated molecule descriptors and the pIC₅₀ values of the whole data set were standardized using the following equation

1

where x^*_i represents the scaled value; x_i represents the original value; x_mean represents the mean value; and x_std represents the standard deviation (SD).

2.3. Feature Selection

There are two types of descriptors, structural and nonstructural descriptors, in the data set. The structural descriptors can help to filter the new drugs, and the nonstructural descriptors can benefit the prediction of activity of new drugs. It is crucial to select both useful descriptors that will be used to train the prediction model. Thus, in this work, we proposed an integrated framework by using the full quadratic effect model and the Lasso model for the selection of important molecular descriptors to develop a machine learning model and guide the subsequent inhibitor synthesis. We used the full quadratic effect model to take the lead in selecting the descriptors with significant effects from the structural molecular descriptors. On the other hand, given that the data set used in this paper was high-dimensional and sparse, where the number of variables (221) and the number of features (208) were close, the Lasso model was used to filter significant nonstructural descriptors from the 208 molecular descriptors calculated by RDKit. Finally, the descriptors screened from the fusion model described above were fitted together to build a machine learning model and guide the synthesis of inhibitors.

2.3.1. Structural Descriptor Selection

The full quadratic effect model is a type of fixed effect model that describes the causal relationship between the factors and the response that persists throughout the experiment. It reveals the causal relationship between the factors and response by employing a regression model to find out the useful first-order effects, second-order effects, and interaction effects between factors (eq 2) that have a significant effect on the dependent variable

2

where Y represents the response, x₁ and x₂ represent the effects of factors, ε represents the random error term, and β₀–β₅ represents the parameters.

In practice, the model can be estimated and extrapolated by statistical software to understand the relationships among the independent variables in order to better understand and interpret the data. In this paper, the OLS model was built using the statsmodels package in Python, containing the first-order and interaction effects of the variables as well as the second-order effects, with the size of the p-value used to determine whether the effect was significant.

2.3.2. Lasso

Lasso regression is a widely used statistical algorithm for variable selection. It is a least-squares with a penalty term (eq 3). Unlike the L2 penalty term in ridge regression, the L1 penalty term in Lasso regression achieves feature selection by compressing the coefficients of insignificant variables to zero without setting any value to zero. Therefore, for high-dimensional sparse data sets, it is more suitable to use Lasso regression to select features

3

2.4. Activity Prediction Model Development and Evaluation

2.4.1. SVR Model

Support vector regression (SVR) model is a widely used supervised learning model for regression analysis, which was first proposed by Drucker et al.⁵⁷ It translates the low-dimensional linearly indivisible data set into a high-dimensional space via a nonlinear transform defined by inner product function and then calculates the optimal hyperplane by minimizing the distance to the furthest data point from the hyperplane. The introduction of SVR is available in ref (58).

In this work, we used the RBF kernel function (eq 4), which transforms the input vector from a low-dimensional space to a high-dimensional space and generates outputs through the weighted summation of its hidden units. As a result, it is suitable for tackling nonlinear problems. The optimization of parameters C, ε (eq 5), and γ was taken out by using the lattice search method with 5-fold cross-validation. In 5-fold cross-validation, the training data were randomly divided into a training set and a test set five times, and we scored each test set by the mean-squared error (MSE), with the best average score indicating the best hyperparameters. We repeated the 5-fold cross-validation 10 times and obtained the average score for each set of hyperparameters. Then, we averaged these MSE values, with the minimal mean MSE value corresponding to the optimal hyperparameter

4

5

2.4.2. Model Evaluation

In this work, two universal statistic metrics (R² and MSE) were applied to assess the model effectiveness. The R² and MSE were calculated by eqs 6 and 7, respectively

6

7

where n represents the total number of compounds; y represents the observed value of a compound; ŷ represents the predicted value of a compound; and y̅ represents the average of y.

2.5. General Procedure for Synthesis

The coupling reaction of 2,4-dichloroquinazoline and tert-butyl-4-(2-aminoethyl)piperazine-1-carboxylate was catalyzed by triethylamine. The product was coupled with phenylboronic acid with different structures under a palladium catalyst. The tertiary butoxycarbonyl protecting group and methyl group of the compound were removed in the presence of boron tribromide to give the final product. The reaction steps are shown in the S1.

2.5.1. tert-Butyl-4-(2-((2-chloroquinazolin-4-yl)amino)ethyl)piperazine-1-carboxylate (2)

1.75 g (7.63 mmol) of tert-butyl-4-(2-aminoethyl)piperazine-1-carboxylate was added to a 100 mL round-bottom flask and 30 mL of tetrahydrofuran was dissolved; 1 mL (7.63 mmol) of triethylamine and 1.51 g (7.63 mmol) of 2,4-dichloroquinazoline were added to the reaction system and stirred at 25 °C for 4 h. After the reaction, it was concentrated under reduced pressure, dissolved in 20 mL of DCM, and extracted three times with a saturated sodium carbonate solution. The organic phase was combined and concentrated under reduced pressure. A yellow oily product of 2.5 g was obtained through silica gel column chromatography and purification, with a yield of 83%.

A white solid; yield: 83%; ¹H NMR (400 MHz, chloroform-d) δ 7.81–7.68 (m, 3H), 7.50 (ddd, J = 8.2, 6.4, 1.8 Hz, 1H), 3.81–3.71 (m, 2H), 3.53 (q, J = 4.7, 4.1 Hz, 4H), 2.78 (t, J = 5.9 Hz, 2H), 2.60–2.49 (m, 4H), 1.49 (s, 9H).

2.5.2. tert-Butyl-4-(2-((2-(2-hydroxyphenyl)quinazolin-4-yl)amino)ethyl)piperazine-1-carboxylate (3a)

0.50 g (1.28 mmol) of compound 2, 0.17 g (1.28 mmol) of (2-hydroxyphenyl) boronic acid, and 0.016 g (0.011 mmol) of tetra (triphenylphosphine) palladium were added in a 100 mL three-necked flask, placing the reaction system in a nitrogen atmosphere. 10 mL of 2 M sodium carbonate solution, 10 mL of n-butanol, and 6 mL of toluene were added in sequence and reacted at 110 °C for 18 h. During this time, TLC monitors the reaction process. After the reaction was completed, heating was stopped and the reaction system was filtered through diatomaceous earth and washed with ethyl acetate; the filtrate was extracted with saturated sodium bicarbonate solution three times, the organic phase was merged, and a white solid product of 0.42 g was obtained by silica gel column chromatography, with a yield of 75%.

A white solid; yield: 75%; ¹H NMR (400 MHz, chloroform-d) δ 8.54 (dd, J = 7.9, 1.8 Hz, 1H), 7.77–7.67 (m, 2H), 7.40 (m, J = 23.0, 8.6, 7.0, 1.6 Hz, 2H), 7.04 (dd, J = 8.2, 1.2 Hz, 1H), 6.94 (ddd, J = 8.2, 7.1, 1.2 Hz, 1H), 6.73 (t, J = 4.5 Hz, 1H), 3.84 (q, J = 5.7 Hz, 2H), 3.52 (t, J = 5.1 Hz, 4H), 2.79 (t, J = 6.0 Hz, 2H), 2.54 (t, J = 5.1 Hz, 4H), 1.50 (s, 9H).

2.5.3. tert-Butyl-4-(2-((2-(5-cyano-2-hydroxyphenyl)quinazolin-4-yl)amino)ethyl)piperazine-1-carboxylate (3b)

A white solid; yield: 72%; ¹H NMR (400 MHz, chloroform-d) δ 8.07 (d, J = 2.2 Hz, 1H), 7.93–7.87 (m, 1H), 7.81–7.71 (m, 2H), 7.67 (dd, J = 8.6, 2.2 Hz, 1H), 7.49 (ddd, J = 8.3, 7.1, 1.2 Hz, 1H), 7.06 (d, J = 8.7 Hz, 1H), 3.91 (s, 3H), 3.75 (q, J = 5.5 Hz, 2H), 3.49 (t, J = 5.0 Hz, 4H), 2.75 (t, J = 6.0 Hz, 2H), 2.51 (t, J = 5.0 Hz, 4H), 1.47 (s, 10H).

2.5.4. tert-Butyl-4-(2-((2-(5-chloro-2-hydroxyphenyl)quinazolin-4-yl)amino)ethyl)piperazine-1-carboxylate (3c)

A yellow solid; yield: 72%; ¹H NMR (400 MHz, chloroform-d) δ 7.90 (dd, J = 8.4, 1.2 Hz, 1H), 7.77–7.71 (m, 2H), 7.45 (ddd, J = 8.3, 7.0, 1.2 Hz, 1H), 7.31 (dd, J = 8.8, 2.8 Hz, 1H), 6.93 (d, J = 8.8 Hz, 1H), 6.73 (t, J = 4.7 Hz, 1H), 3.82 (s, 3H), 3.74 (q, J = 5.6 Hz, 2H), 3.48 (t, J = 5.0 Hz, 4H), 2.71 (t, J = 6.0 Hz, 2H), 2.49 (t, J = 5.0 Hz, 4H), 1.47 (s, 9H).

2.5.5. 2-(4-((2-(Piperazin-1-yl)ethyl)amino)quinazolin-2-yl)phenol (4a)

0.20 g (0.44 mmol) of compound 3a was added to a 50 mL Schlenk flask. 15 mL of anhydrous dichloromethane was dissolved in a nitrogen atmosphere and stirred at −30 °C for 15 min; then, 514 μL (5.34 mmol) of boron tribromide was added, and after 1 h, the reaction was moved to room temperature. After the reaction, 15 mL of methanol was added to a low-temperature bath at −30 °C to quench the reaction. The solvent was removed by vacuum distillation and dissolved in 20 mL of ultrapure water. The pH was adjusted to 12 or above, extracted three times using a mixture of 18 mL of dichloromethane and 2 mL of methanol, and the organic phase was merged. Vacuum distillation was used to remove solvents, and reverse phase silica gel column chromatography was performed to obtain 102 mg of green solid product with a yield of 65%.

A white solid; yield: 65%; mp: 170.6–173.3 °C; ¹H NMR (400 MHz, methanol-d₄) δ 8.44 (dd, J = 7.9, 1.6 Hz, 1H), 7.95 (d, J = 8.1 Hz, 1H), 7.73–7.60 (m, 2H), 7.44–7.27 (m, 2H), 6.94–6.83 (m, 2H), 3.81 (t, J = 6.9 Hz, 2H), 2.86 (t, J = 4.9 Hz, 4H), 2.71 (t, J = 6.9 Hz, 2H), 2.57 (s, 4H). ¹³C NMR (101 MHz, methanol-d₄) δ 161.34, 160.99, 159.39, 147.16, 132.88, 131.89, 129.07, 125.90, 125.46, 122.00, 119.55, 117.97, 116.87, 113.61, 57.03, 53.62, 44.74, 37.52. HRMS(ESI) m/z: ([M + H]⁺) calcd For C₂₀H₂₄N₅O⁺: 375.1976, found: 375.1981.

2.5.6. 2-(5-Cyano-2-hydroxyphenyl)-N-(2-(piperazin-1-yl)ethyl)quinazolin-4-aminium (4b)

A yellow solid; yield: 20%; mp: 179.2–181.3 °C; ¹H NMR (400 MHz, methanol-d₄) δ 8.29 (s, 1H), 8.08 (dd, J = 8.1, 1.1 Hz, 1H), 7.78–7.72 (m, 2H), 7.53–7.43 (m, 2H), 6.85 (d, J = 8.6 Hz, 1H), 3.90 (t, J = 6.9 Hz, 2H), 2.90 (s, 4H), 2.80 (t, J = 6.9 Hz, 2H), 2.64 (s, 4H). ¹³C NMR (101 MHz, methanol-d₄) δ: 159.55, 159.43, 144.62, 134.98, 134.07, 133.36, 126.31, 125.47, 122.21, 119.79, 119.44, 118.90, 113.81, 102.78, 100.18, 56.84, 53.47, 44.85, 37.67. HRMS(ESI) m/z: ([M + H]⁺) calcd For C₂₁H₂₃N₆O⁺: 375.1928, found: 375.1929.

2.5.7. 2-(5-Chloro-2-hydroxyphenyl)-N-(2-(piperazin-1-yl)ethyl)quinazolin-4-aminium (4c)

A green solid; yield: 22%; mp: 172.4–174.2 °C; ¹H NMR (400 MHz, methanol-d₄) δ 8.40 (d, J = 2.7 Hz, 1H), 8.02 (dd, J = 8.2, 1.4 Hz, 1H), 7.79–7.66 (m, 2H), 7.48 (ddd, J = 8.3, 6.9, 1.4 Hz, 1H), 7.27 (dd, J = 8.7, 2.8 Hz, 1H), 6.89 (d, J = 8.8 Hz, 1H), 3.87 (dd, J = 8.3, 5.9 Hz, 2H), 2.91 (t, J = 4.9 Hz, 4H), 2.80–2.63 (m, 6H). ¹³C NMR (101 MHz, methanol-d₄) δ: 160.19, 159.69, 159.51, 146.64, 133.02, 131.46, 128.17, 126.05, 125.84, 122.63, 121.98, 120.63, 118.55, 113.70, 56.89, 53.67, 44.74, 37.51. HRMS(ESI) m/z: ([M + H]⁺) calcd. For C₂₀H₂₃ClN₅O⁺: 384.1586, found: 384.1595.

2.6. Kinase Profiling

All enzymatic reactions were performed at 30 °C for 40 min as the assay method reported by Kashem. 50 μL of reaction mixture contained 40 mM tris, pH 7.4, 10 mM MgCl₂, 0.1 mg/mL of bovine serum albumin (BSA), 1 mM of dithiothreitol (DTT), 10 μM of adenosine triphosphate (ATP), 25 ng of kinase, and 0.2 mg/mL of enzyme substrate (Poly (Glu, Tyr)). Compounds were diluted in 10% dimethyl sulfoxide (DMSO), and 5 μL of this dilution was added to 50 μL of the reaction mixture so that the final concentration of DMSO was 1% in all reactions. The assay was performed by using a Kinase-Glo Plus luminescent kinase assay kit. It measures kinase activity by quantifying the amount of ATP remaining in the solution after the kinase reaction. The luminescent signal from the assay correlates with the amount of ATP present and inversely correlates with the increase in kinase activity. The IC₅₀ values were calculated using the nonlinear regression of normalized dose responses using graphpad prism 5.0 software.^59,60

2.7. Virtual Screening

To identify potential tyrosine kinase inhibitors (TKIs) for EGFR^{L858R/T790M/C797S} mutations, we utilized an SVR model to conduct virtual screening experiments on a small-molecule compound data set sourced from the CHEMBL database. Molecular docking was performed by using the CDOCKER module in Discovery Studio. All compounds underwent structure optimization using the MM2 method of ChemDraw 3D and were assigned a Forcefield in the simulation module. The EGFR^{L858R/T790M/C797S} protein (PDB:6LUD)⁶¹ was used as the protein receptor for docking. Before docking, the protein receptor was preprocessed by inserting missing atoms in incomplete residues, modeling the missing loops, and removing cocrystallized water. To generate an aspherical grid with a radius of 10 Å, the binding site was defined around the centroid of the ligand. The resulting protein–drug complexes were then analyzed for their interactions after the docking simulations.

The ADME properties play a critical role in determining the efficacy and safety of drug candidates. Predicting these properties is essential to preventing drug failure during clinical trials. To this end, the Swiss ADME web server, offered by the Swiss Institute of Bioinformatics (SIB), is employed to compute several properties, such as drug-likeness, synthetic accessibility, and PAINS, for the selected compounds.⁶² The drug-likeness can be predicted by several rules, such as Lipinski, Ghose, Veber, Egan, and Muegge. Furthermore, the synthetic accessibility of the compounds is evaluated on a scale of 1–10, where a lower score indicates an easier synthetic route and a higher score indicates a more complex route. Compounds that satisfied the criteria for drug-likeness were free of PAINS, demonstrated favorable ADME properties, and were readily synthesized, which are considered the most promising drug candidates.

3. Results and Discussion

3.1. Diversity of Data Set

The data set of 221 small-molecule inhibitors of EGFR^{L858R/T790M/C797S} triple mutations with various molecular structures was first visualized by the bioactivity (pIC₅₀) distribution. It was generally accepted that a compound was highly active when its IC₅₀ < 100 nM, i.e., pIC₅₀ > 7, and from Figure 2A, it was found that 58% of the inhibitors were highly active. Then, to measure the similarity of molecular structures, we calculated the Tanimoto coefficients on the basis of MACCS fingerprints. In general, compounds with a Tanimoto coefficient less than 0.7 could be considered to have significant variability. The frequency of distribution histogram (Figure 2B) demonstrated the similarity of each inhibitor. It was obvious that there were 79.5% pairs of inhibitors in the whole data set whose Tanimoto coefficient values were less than 0.70. It can be concluded that the data set employed in this study exhibited a high degree of diversity.

Figure 2.

Histograms of the distribution of pIC₅₀ values (A) and the Tanimoto coefficient (B) for 221 compounds.

3.2. Performances of Machine Learning Models

3.2.1. Feature Selection

In order to select the most pertinent descriptors relating to the bioactivity (pIC₅₀), we initiated the process with 10 molecular structural descriptors extracted from RDKit. Specifically, these descriptors were NOCount, NumHAcceptors, NumAromaticHeterocycles, fr_C_O, fr_Nhpyrrole, fr_amide, fr_aniline, fr_imidazole, fr_piperdine, and fr_piperzine. These descriptors hold substantial influence over EGFR-TKIs and represent fundamental facets of the molecular structure. Both NOCount and NumHAcceptors denote the counts of nitrogen/oxygen and hydrogen bond acceptors, respectively. These counts are pivotal for fostering interactions in the domain structure. The remaining descriptors, which encompass the tally of aromatic heterocycles, ketoaldehydes, pyrroles, amides, anilines, imidazoles, piperidines, and piperazines, are extensively utilized within EGFR-TKIs. Moreover, their structural existence within these compounds has been documented in various reports. A full quadratic effect model was constructed by 10 structural molecule descriptors with an R² of 0.750 and a Prob (F-statistics) of 7.33 × 10^–26, which showed the accuracy and effectiveness of the model. Out of 65 variables, 10 are first-order effects, 45 are interaction effects, and 10 are second-order effects. The p-values of three first-order effects and two interaction effects are less than 0.05 (Table 1), indicating a significant effect between this variable and the dependent variable pIC₅₀. Among the three first-order effects, aromatic heterocycles and aniline structures are commonly used in the structural design of EGFR-TKIs (Figure 3). Aromatic nitrogen heterocycles, in particular, often act as the parent core structure of inhibitors. For example, Gefitinib is the first-generation EGFR-TKI used for the treatment of nonsmall cell lung cancer caused by primary mutations.⁶³ Amino-quinazoline can form hydrogen bond interactions with Met793 and Thr790 of tyrosine kinases. To address the continuous occurrence of acquired resistance, the fourth-generation EGFR-TKI Angew-2017 continues the aminoquinoline structure.⁶⁴ It can efficiently inhibit the Del19/T790M/C797S mutant kinase (IC₅₀ = 17.9 nM). Brigatinib, a fourth-generation EGFR-TKI, has structural similarities to osimertinib. The IC₅₀ value for the L858R/T790M/C797S triple mutant kinase is 38.3 nM. In the molecular simulation of brigatinib and tyrosine kinase, the part replaced by piperidine is defined as the part with low affinity and can be modified. This is consistent with the negative correlation results of the piperidine structure in the full quadratic effect model.⁶⁵

Table 1. Results of the Full Quadratic Effect Model.

variable	coefficient	standard error	p-value
NumAromaticHeterocycles	2.3402	1.030	0.024
fr_piperdine	–3.8413	1.524	0.013
fr_aniline	1.4642	0.552	0.009
NumAromaticHeterocycles* fr_piperdine	–1.2650	0.373	0.001
NumAromaticHeterocycles* fr_aniline	–0.4882	0.186	0.009

Figure 3.

Structural analysis of classical EGFR-TKIs.

Subsequently, we employed the Lasso model to identify significant molecular descriptors from the comprehensive set of descriptors (totaling 208) calculated by using RDKit. The outcomes of this endeavor are presented in Table 2. Ultimately, we amalgamated the 5 variables extracted from the full quadratic effect model with the 7 variables pinpointed by the Lasso model, constituting a holistic collection of pivotal molecule descriptors for the purpose of constructing a support vector regression model.

Table 2. Results of the Lasso Model.

descriptor	description
BCUT2D_MWHI	similarity calculations of a property at a certain distance from each atom
BCUT2D_CHGHI	charge distribution calculations for the atoms
Chi2v	the second-order valence bonding index of a molecule
HallKierAlpha	calculation of the similarity of the charge distribution and arrangement of atoms
PEOE_VSA10	atomic partial charges on the van der Waals surface area of a molecule in the range [0.10, 0.15].
EState_VSA7	EState VSA Descriptor 7 (1.81≤ x < 2.05)
VSA_EState2	VSA EState Descriptor 2 (4.78≤ x < 5.00)

3.2.2. SVR Model

For constructing the SVR model with the RBF kernel, we selected three optimal parameters (C, γ, ε) by using the lattice search method with 5-fold cross-validation. As a result (Table 3), the best parameters C = 1.7, γ = 0.11, ε = 0.3 for the SVR model built with 12 important molecule descriptors and an R² of 0.745 and an MSE of 0.255 were obtained for test set. The calculated and experimental values of bioactivity (pIC₅₀) on the SVR model are shown in Figure 4.

Table 3. Performance of the SVR Model.

	training set		test set
modeling method	MSE	R2	MSE	R2
SVR	0.147	0.853	0.255	0.745

Figure 4.

Calculated vs experimental values of bioactivity (pIC₅₀) on the SVR model.

3.3. Synthesis of Inhibitors Based on Structural Descriptors

Following the utilization of a comprehensive quadratic effect model, structural descriptors that offered more robust guidance for inhibitor synthesis were preserved. Grounded in the insight gleaned from these descriptors, we judiciously formulated and synthesized inhibitors aligned with their structural attributes. Subsequently, we validated the inhibitory potential of these inhibitors through a series of in vitro experiments.

In the first-order effect, due to the negative correlation between piperidine and the dependent variable pIC₅₀, the presence of piperidine structures should be avoided. At the same time, the coefficient of aniline and aromatic heterocycles in the interaction is negative, indicating that the number of aniline or aromatic heterocycles present will affect the positive correlation of the activity of other moieties. Therefore, in the design, efforts should be made to avoid an overwhelming increase in the number of aniline or aromatic heterocycles. We used 2-aryl-4-aminoquinazole as the starting point. It is not only a classic mother nucleus structure but also contains two relatively balanced structures: aniline and aromatic heterocycles. The regulation of the structure–activity relationship was achieved by replacing different substituents in the aromatic group. To satisfy the interaction between inhibitors and the solvent region outside the tyrosine domain, piperazine structures were connected from the perspective of medicinal chemistry design. A total of 3 inhibitors were synthesized through the above design shown in Figure 5 (4a–c).

Figure 5.

Structure of compounds based on structural descriptors.

3.4. In Vitro Enzymatic Activity Assay

Angew-2017 containing 2-aryl-4-aminoquinazole structure showed excellent inhibitory activity against Del19/T790M/C797S.⁶⁴ Therefore, when in vitro enzyme inhibition experiments on inhibitors with 2-aryl-4-aminoquinazole as the mother nucleus were conducted, two triple mutated kinases, L858R/T790M/C797S and Del19/T790M/C797S, were considered. The results are listed in Table 4.

Table 4. In Vitro Enzyme Inhibitory Activity of the Synthesized Compound on the Triple Mutant EGFR and the Number of Structural Descriptors Contained in the Compound^a.

compound	EGFRL858R/T790M/C797S IC50 (μM)	EGFRDel19/T790M/C797SIC50 (μM)	fr_aniline	NumAromaticHeterocycles
4a	1.221 ± 0.031	1.432 ± 0.022	1	1
4b	0.736 ± 0.044	0.220 ± 0.002	1	1
4c	0.277 ± 0.028	0.089 ± 0.004	1	1
Angew-2017	Ndb	0.018	1	2

^aThe data are the mean ± SD of at least three independent experiments.

^bNd: no data.

The findings demonstrated that the synthesized compound exhibited concurrent inhibitory activity against both triple mutated kinases L858R/T790M/C797S and Del19/T790M/C797S. Notably, the presence of potent electron-withdrawing substituents within the phenolic hydroxyl group emerged as a critical determinant for the inhibitor’s activity. Compound 4c, which was substituted with chlorine, exhibited the strongest inhibitory effect on both triple mutated kinases (IC₅₀^{L858R/T790M/C797S} = 0.277 μM, IC₅₀^{Del19/T790M/C797S} = 0.089 μM). Compound Angew-2017 demonstrated a stronger inhibitory effect on enzyme activity compared to that of the synthesized compound. This compound contained more aromatic heterocyclic structures in its composition. These observations suggest that the number of aromatic heterocycles could have a positive correlation with activity, which aligns with the findings of the full quadratic effect model. In order to determine the binding effect between the synthesized compound and the active site of the enzyme, molecular docking was performed between 4a–c and EGFR^{L858R/T790M/C797S} triple mutated kinase (Figures S17–S19). The results showed that 4a–c can bind to kinase active sites in different forms. The experimental results indicated that the synthesized compounds could be further studied as excellent lead compounds for discovering fourth-generation EGFR-TKIs

3.5. Virtual Screening

In this study, a data set of 50,000 compound molecular structures, sourced from the ChMEBL database, was utilized for virtual screening. A pIC₅₀ value threshold of above 7.5 was set to ensure empirical significance. Consequently, 35 compounds, exhibiting pIC₅₀ values surpassing 7.5 in the SVR model, were pinpointed as potential inhibitors.

3.5.1. Molecular Docking

Molecular docking is a computational method to identify the bioactive conformations of small molecules within protein binding sites. We employed this method to dock 35 compounds selected through virtual screening with the EGFR^{L858R/T790M/C797S} protein (PDB:6LUD). After the ligands were prepared, a total of 665 conformations were generated. The docking results were sorted by the size of the -CDOCKER ENERGY score, and four hits with a score greater than 30 were obtained.

The visual analysis of the docking results revealed that all four hits were able to interact with the structural domains of the EGFR ^{L858R/T790M/C797S} triple mutant proteins in different ways, as shown in Figure 6. Hydrogen-bonding interactions were observed in all compounds, involving amino acids, such as MET793, SER797, LYS745, and PRO794, among others. Aromatic structures also played an important role in the interaction of compounds with tyrosine kinases such as Pi-Alkyl and Alkyl interactions. The potential hit structures discovered above will provide new possibilities for the design of a new generation of EGFR inhibitors.

Figure 6.

2D interaction diagram of EGFR^{L858R/T790MC797S} in complex with (A) CHEMBL602257, (B) CHEMBL3473990, (C) CHEMBL3718174, (D) CHEMBL4116547, and (E) the corresponding structures of the compounds.

3.5.2. ADME Analysis

ADME parameters are crucial in determining the potential of a chemical compound to function as an active drug. Solubility, lipophilicity, and bioavailability, among other physicochemical properties, are critical factors in evaluating the drug-likeness of a compound. All four hits demonstrated excellent physicochemical properties, as highlighted in Table 5. Each hit satisfied specific criteria, including solubility (−6 < Log S < 0), lipophilicity (0.7 < Log P < 5), molecular weight (150 < MW < 500 g/mol), polarity (20 < TPSA < 130), saturation (0.25 < fraction Csp3 < 1), hydrogen bond acceptors (0 < HBA < 10), and hydrogen bond donors (0 < HBD < 5). Furthermore, drug-likeness was evaluated using important rule-based approaches such as Lipinski’s rule of five and Ghose, Veber, Egan, and Muegge (Table 6). All four hits showed a bioavailability score of 0.55, with no PAINS alerts and a plausible synthetic accessibility score ranging from 3.68 to 4.79, as indicated in Table 5.

Table 5. Physicochemical Properties of 4 Hits.

molecule	MW	Csp3	HBA	HBD	TPSA	Log P	Log S
CHEMBL602257	370.44	0.38	5	4	107.74	2.77	–3.44
CHEMBL3473990	360.4	0.47	6	2	94.58	3.55	–3.11
CHEMBL3718174	386.44	0.52	6	1	104.37	3.14	–3.06
CHEMBL4116547	414.49	0.67	7	3	109.3	3.46	–3.59

Table 6. Predicted Drug-Likeness and Synthetic Accessibility of the 4 Hits.

molecule	drug-likeness violationsa	bioavailability score	PAINS alerts	synthetic accessibility
CHEMBL602257	0	0.55	0	3.68
CHEMBL3473990	0	0.55	0	3.79
CHEMBL3718174	0	0.55	0	4.34
CHEMBL4116547	0	0.55	0	4.76

^aLipinski’s rule of five, Ghose, Veber, Egan, and Muegge.

4. Conclusions

In this study, we assembled a data set encompassing 221 small-molecule inhibitors targeting EGFR^{L858R/T790M/C797S}, a valuable resource for cancer therapy. We employed RDKit to compute 208 molecular descriptors for these compounds. Subsequently, we introduced a fusion framework comprising a full quadratic effect model and a Lasso model. This framework facilitated the selection of both structural and nonstructural molecular descriptors. Utilizing these descriptors, we constructed an SVR model that integrated 12 crucial descriptors, resulting in an R² value of 0.745 and an MSE of 0.255 on the test set. Guided by the structural descriptors employed in the modeling process, we synthesized a range of fourth-generation EGFR inhibitors and conducted in vitro kinase activity assays. The IC₅₀ values of the synthesized compounds reached the nanomolar level. A virtual screening was then conducted on a data set containing around 50,000 compounds, leveraging the SVR model. This screening, combined with molecular simulation docking and ADME analysis, yielded 4 hit compounds. It is expected that activity evaluation on the screened compounds will be conducted to discover a new generation of EGFR-TKI.

Supporting Information Available

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsomega.3c06225

• General synthetic route and available copies of ¹H NMR, ¹³C NMR, HRMS, and HPLC chromatograms for compounds 2, 3a, 3b, 3c, 4a, 4b, and 4c; and molecular docking results of compounds 4a, 4b, and 4c (PDF)

• HLPC analysis of all tested compounds (XLSX)

Author Contributions

^# Hao Chang and Zeyu Zhang contributed equally.

The authors declare no competing financial interest.