Developing Lead Compounds of eEF2K Inhibitors Using Ligand–Receptor Complex Structures

Abstract

The eEF2K, a member of the α-kinase family, plays a crucial role in cellular differentiation and the stability of the nervous system. The development of eEF2K inhibitors has proven to be significantly important in the treatment of diseases such as cancer and Alzheimer’s. With the advancement of big data in pharmaceuticals and the evolution of molecular generation technologies, leveraging artificial intelligence to expedite research on eEF2K inhibitors shows great potential. Based on the recently published structure of eEF2K and known inhibitor molecular structures, a generative model was used to create 1094 candidate inhibitor molecules. Analysis indicates that the model-generated molecules can comprehend the principles of molecular docking. Moreover, some of these molecules can modify the original molecular frameworks. A molecular screening strategy was devised, leading to the identification of five promising eEF2K inhibitor lead compounds. These five compound molecules demonstrated excellent thermodynamic performance when docked with eEF2K, with Vina scores of −12.12, −16.67, −15.07, −15.99, and −10.55 kcal/mol, respectively, showing a 24.27% improvement over known active inhibitor molecules. Additionally, they exhibited favorable drug-likeness. This study used deep generative models to develop eEF2K inhibitors, enabling the treatment of cancer and neurological disorders.

1. Introduction

Phosphorylation and dephosphorylation are crucial for cellular signal transduction and regulation [1,2,3]. Mediated by kinases and phosphatases, these processes are vital for cell growth, differentiation, metabolism, and apoptosis [4,5,6]. Kinases phosphorylate proteins by transferring phosphate groups, leading to conformational changes that affect protein activity, localization, and interactions [7,8]. This modification acts as a rapid, reversible regulatory mechanism in response to external stimuli [2,9]. Phosphatases counteract this by dephosphorylating proteins, restoring them to their original state and adjusting signaling pathway activities [5,6]. The balance between kinase and phosphatase activities is essential for cellular homeostasis and normal physiological functions [2,10]. Imbalances can result in diseases like cancer, diabetes, and neurodegenerative disorders [10]. Therefore, research on drugs targeting phosphatases or kinases is of significant importance.

The α-kinase family represents a distinctive subfamily of protein kinases, playing a pivotal role in various cellular processes, including protein translation, Mg²⁺ homeostasis, intracellular transport, cell migration, adhesion, and proliferation [8,11,12,13,14]. Currently, six α-kinases have been identified in humans [1,8]. Among these, the eukaryotic elongation factor 2 kinase (eEF2K) plays a significant role in cancer [15], exhibiting notable expression in various cancer cells. It regulates crucial processes such as apoptosis, cell survival, and autophagy [16,17]. This discovery holds substantial importance for research and treatment strategies in oncology [18]. Moreover, eEF2K plays a significant role in various biological systems, including neurology, cardiology, myology, and immunology, and in the pathogenesis of several neurological disorders, such as Alzheimer’s disease and depression [14,19,20,21,22]. Interestingly, eEF2K is not essential for the survival or health of mammals under normal conditions. This makes eEF2K inhibitors a critical breakthrough in drug development, offering new avenues for therapeutic intervention [23,24].

The development of pharmaceuticals faces challenges such as high costs and lengthy timelines. With the advancement of artificial intelligence, using AI technology to accelerate drug discovery has become a feasible option, promising significant reductions in drug development time [25,26]. In particular, two main methods for AI-assisted new drug discovery are generating molecules using protein pocket structures [27] and generating molecules based on known drug structures [28,29]. Recently, a technique that simultaneously utilizes both protein and known drug structures for molecule generation—molecule generation based on receptor–ligand complexes—has demonstrated more accurate molecule generation capabilities and is becoming a hot topic in research [30,31,32].

In developing eEF2K inhibitors, several studies have adopted the molecule generation method based on receptor–ligand complex models. For example, eEF2K inhibitors have been generated [33] or silico screening [34] using protein homology modeling [35,36], advancing the research in molecule generation for eEF2K inhibitors [37]. However, these methods, which are not based on precise target structures, need further improvement in molecule fidelity. In an exhilarating development in 2022, the protein structure of eEF2K was reported for the first time [13,38]. This milestone will undoubtedly significantly impact future drug design targeting eEF2K. To our knowledge, no eEF2K inhibitors designed using this crystal structure instead of a homology model have been reported. This breakthrough opens new avenues for more precise and effective inhibitor development against eEF2K. Furthermore, molecule generation and molecular docking models are undergoing iterative upgrades [31,39]. Using the accurate eEF2K protein crystal structure, more inhibitor structures, and the latest state-of-the-art molecule generation models, generating molecules for this target inhibitor is a much-anticipated area of research.

This study utilized the newly disclosed eEF2K protein structure and existing inhibitor structure molecules. We employed the DiffDock [40] to dock the protein with known biologically active compounds. We used the LiGAN [39] molecular generation model, which integrates target protein structures and drugs to generate a series of eEF2K inhibitor molecules. Statistical analysis revealed that these molecules exhibit an optimal Vina score when binding to eEF2K. Furthermore, by comparing their activities with those of reference inhibitors, it was demonstrated that the generative model effectively captures the characteristics of ligand–receptor complex structures. Subsequently, we screened a candidate compound library using metrics such as drug-like properties and synthesizability, identifying five lead compounds with favorable performance and demonstrating their potential research value.

2. Results and Discussion

2.1. Data Collection and Preparation

2.1.1. Construction and Data Analysis of the Known eEF2K Inhibitor Database (eEF2K-bio)

The creation of the known eEF2K inhibitor database was facilitated by the ChEMBL [41], a comprehensive resource of bioactive molecules. By searching for “eEF2K”, structures of of 71 confirmed bioactive inhibitors were identified. These known eEF2K inhibitors were subsequently standardized using RDKit [42], resulting in the establishment of the “eEF2K-bio” database.

Analyzing the physicochemical properties of the data in eEF2K-bio (refer to Figure 1), it became evident that the structures of known inhibitors require further optimization. Looking at the distribution of molecular weights, we observed that compounds with a molecular weight over 500 comprise 15% of the dataset (Figure 1a). In terms of LogP, which is a measure of a compound’s permeability through passive diffusion when orally administered, it is generally believed that the gastrointestinal tract best absorbs compounds with a moderate LogP [43] (range of 0 to 3). However, only 20% of the compounds in eEF2K-bio fall within this range (Figure 1b). Regarding drug-likeness, a QED [44] (Quantitative estimate of drug-likeness) score above 0.67 is considered favorable, but 21% of the compounds scored below this threshold (Figure 1c). As for synthesizability, the SA (Synthetic accessibility) score quantitatively represents the complexity and feasibility of a compound’s chemical synthesis, considering factors such as reference atoms and bonds, functional groups, reaction pathways, and topological structures. The difficulty of synthesizing small molecules is rated on a scale of 1 to 10, with values closer to 1 indicating more accessible synthesis and those nearer to 10 indicating more challenging synthesis. In eEF2K-bio, 49% of the compounds have an SA score above 5 (Figure 1d). From eEF2K-bio, the IC₅₀ values of these compounds range from 6 to 60 μM, with the most active molecule having an IC₅₀ of 100 nM. Forty-two molecules have IC₅₀ values lower than 10 μM, while the remaining twenty-nine are distributed between 10 and 60 μM (Figure 1e).

We examined the relationship between the biological activity of known inhibitors and the changes in binding affinity during the docking process. The binding affinity (represented as the Vina score and measured in kcal/mol) was calculated using AutoDock [45,46]. The distribution of Vina scores in the eEF2K-bio database is detailed in Figure S1, and the distribution of pIC₅₀ is shown in Figure S2. In the correlation plot of Vina scores and pIC₅₀ (Figure 1e), it was evident that most molecules with lower biological activity (IC₅₀ > 10 μM, or pIC₅₀ < 5.0) exhibit significantly higher Vina scores, clustering in the upper left corner of Figure 1e. Conversely, molecules with higher activity (IC₅₀ < 10 μM, or pIC₅₀ > 5.0) demonstrate significantly lower Vina scores, mostly clustering in the bottom right corner. We fitted the relationship between Vina scores and pIC₅₀ using a linear regression equation, resulting in a Pearson correlation coefficient of −0.72. This indicates a strong negative correlation between the biological activity and their Vina scores of molecules in the eEF2K-bio database. Based on this, to facilitate the activity evaluation of candidate eEF2K inhibitor compounds generated by the model, we utilized Vina scores to estimate biological activity.

2.1.2. Gathering and Analysis of eEF2K Protein Structure

The crystal structure of eEF2K, sourced from the Protein Data Bank [47] (PDB), underwent meticulous refinement and analysis. Employing the advanced capabilities of the Maestro software package with the OPLS4 [48] force field, the nucleotide-free form of the eukaryotic elongation factor 2 kinase (PDB ID: 8FO6 [38]) was subjected to a comprehensive cleaning and preparatory process. This rigorous protocol began with the removal of all extraneous water molecules and ions, followed by the addition of hydrogen atoms to achieve accurate atomic configurations. Subsequent stages entailed the thorough removal of non-protein elements within the crystal structure, correction of missing atoms or residues, standardization of nomenclature, and rectification of any significant errors. An energy minimization step was then conducted to secure a stable, low-energy conformation of the protein. In the final phase of preparation, PyMOL [49] (Available at: https://pymol.org (accessed on 13 July 2024)) was utilized to meticulously eliminate any remaining water molecules, thus ensuring the protein was optimally prepped for advanced computational studies.

Given the current lack of precise knowledge regarding the binding sites of inhibitors on eEF2K, we identified a potential binding site within the protein cavity based on the receptor’s structural features. To confirm the accuracy of this predicted binding site, two well-known eEF2K inhibitors, CHEMBL1091075 and CHEMBL34241, were subjected to simulated docking. The results demonstrated that both inhibitors effectively bound to the predicted site (Figure 2a–c), which substantiates the reliability of the identified binding site and lays a solid foundation for subsequent research.

2.2. Generation of eEF2K Inhibitor Lead Compounds

2.2.1. Lead Compound Generation Strategy

The strategy for generating lead compound molecules encompasses three key steps: molecular docking; molecular generation; and molecular screening, as outlined in Figure 3.

(1) Molecular docking. To enable the molecular generation model to grasp the structural intricacies of ligand–receptor complexes, we initially performed docking of known inhibitors from the eEF2K-bio database with eEF2K. This process utilized DiffDock, a model renowned for its state-of-the-art (SOTA) speed and accuracy in deep-learning models for molecular docking.

(2) Molecular generation. In this phase, we employed the LiGAN model, which is a pocket-based molecular generation model, to generate candidate compound molecules for eEF2K inhibitors. This process involved using the ligand–receptor complex structures as inputs. LiGAN employs conditional variational autoencoders, with three-dimensional convolution as the encoder, to create ligand structures that can adapt to changes in protein structure. The model extracts continuous representations of small molecule voxels using 3D convolutions and then utilizes the LSTM-base decoder to reconstruct the molecular SMILES from these continuous representations, thereby completing the molecular generation based on receptor structure.

(3) Molecular evaluation and screening. The molecules generated and optimized above were subjected to molecular docking using AutoDock (v1.2.3), calculating the changes in binding affinity. Based on the previously established relationship between Vina scores and biological activity, we evaluated the top 70 generated molecules in terms of drug-likeness, synthesizability, LogP, and molecular weight. This process led to the selection of lead compounds for eEF2K inhibition.

2.2.2. Generation and Analysis of the eEF2K Inhibitor Candidate Compound Database (eEF2K-Gen)

Utilizing LiGAN for the encoding and decoding process, we generated 20 new eEF2K inhibitor candidates for each reference molecule from the eEF2K-bio. As depicted in Figure 4a, after excluding ineffective structures, we successfully created 1094 valid structures, achieving a generation rate of 78.14%. In the eEF2K-bio dataset, 98.57% of the molecules resulted in at least five inhibitor candidates, demonstrating the robustness and efficacy of this generation strategy. Employing the molecular generation strategy outlined above, we established an eEF2K candidate compound library (eEF2K-gen).

Virtual ligand–receptor docking simulations indicate the potential for biological activity among the molecules in the eEF2K-gen library. Using AutoDock, we docked molecules from eEF2K-gen with eEF2K protein and calculated the Vina score. We selected the 70 molecules with the best Vina scores from eEF2K-gen (termed eEF2K-top 70) and compared their scores with 70 molecules from eEF2K-bio. As shown in Figure 4b, the Vina score range for eEF2K-top 70 (blue area) was predominantly between −10.2 kcal/mol and −16.9 kcal/mol, with a few molecules even reaching −20.0 kcal/mol, notably lower than that of the molecules in eEF2K-bio (orange area). Given the correlation curve of pIC₅₀ and Vina score (Figure 1e), this suggests that molecules in eEF2K-gen may have significant potential in terms of biological activity.

The generation of eEF2K inhibitor candidate compounds also demonstrated noteworthy interpretability. We arranged the 1094 molecules generated in the eEF2K-gen library based on the Vina score and then compared them with the pIC₅₀ values of the input molecules from the eEF2K-bio. The results (Figure 4c) indicated that the top 100 molecules with lower Vina scores were generated based on known inhibitors with pIC₅₀ ranges of 4.5 to 7.5. In contrast, the tail-100 molecules with higher Vina scores were derived from known eEF2K inhibitors with a considerably lower pIC₅₀ range of 3 to 8, mainly concentrated around pIC₅₀ = 4 and pIC₅₀ = 5.8. It can be inferred that the generation strategy for eEF2K inhibitor candidate compounds successfully captured the patterns of ligand–receptor interactions, indirectly reflecting the interpretability of the generation strategy.

2.3. Comparative Analysis of eEF2K-Gen and eEF2K-Bio

To compare the characteristics of newly generated candidate compounds with known inhibitors, we utilized the dimensionality reduction tool t-map [50] (tree-map) for visualizing both eEF2K-gen and eEF2K-bio datasets (Figure 5a). In t-map, structurally similar molecules cluster together, while those with significant structural differences diverge. Additionally, interconnected molecules are linked with “tree branches”, revealing their relationships. The plot yields the following insights:

(1) The molecules in eEF2K-bio (represented by orange dots) are broadly categorized into five clusters. We analyzed the cluster at the bottom of Figure 5a, enclosed by a blue dashed line, and found that the upper half of these molecules share a common scaffold of 3-amino-5,6,7,8,9,10-hexahydrocycloocta[b]thieno [3,2-e]pyridine-2-carboxamide. Variations occur at the 4th position with different substituents (Figure 5b). The lower half of these molecules (Figure 5c) shows some changes in the scaffold, such as variations in the number of carbons in the aliphatic ring, but all retain the thieno [2,3-b]pyridine skeleton. From these observations, we inferred that the development of traditional eEF2K inhibitors is largely based on modifying a few fundamental scaffolds; (2) The eEF2K-bio clusters (or dots) are evenly scattered across the dimensionality-reduced map, indicating that the eEF2K-gen (blue dots) effectively learned the structural features of the original inhibitors, encompassing the structural diversity of the eEF2K-bio data in the newly generated candidate compounds; (3) Many branches and clusters formed by the blue dots (eEF2K-gen) lack the presence of orange dots (eEF2K-bio), signifying that eEF2K-gen has generated many new branches and clusters not covered in eEF2K-bio, hence introducing novel scaffolds. These new structures are visible in the molecules from the newly formed clusters. Overall, this strategy for generating eEF2K inhibitor candidates not only captures the characteristics of known eEF2K inhibitors but also innovatively generates new scaffold structures, paving new avenues for the design of eEF2K inhibitors.

2.4. Selection of Lead Compounds from eEF2K-Gen

Screening Strategy and Results

This study primarily focused on the generation of molecules based on ligand–receptor complex structures while not comprehensively considering molecular drug-likeness (such as LogP, molecular weight, synthesizability, polarity, ADMET properties, etc.). To obtain lead compounds with research and development value, we screened molecules from eEF2K-gen. The primary screening strategy involved (1) sorting 1094 compounds according to their Vina scores from low to high and selecting the top 70 molecules; (2) individually assessing the drug-likeness of these 70 molecules, which included parameters like lipophilicity (XLOGP3 between −0.7 and +5.0), size (MW between 150 and 500 g/mol), polarity (TPSA between 20 and 130 Ų), solubility (log S not higher than 6), saturation (fraction of carbons in sp3 hybridization not less than 0.25), and flexibility (no more than nine rotatable bonds). Molecules with poor drug-likeness were eliminated, resulting in five lead compounds (Figure 6).

To analyze the research and development potential of the generated eEF2K inhibitor lead compounds and evaluate the effectiveness of this study’s molecular generation and screening design, we analyzed the molecular structure information of the generated lead compounds. Comparing them with reference molecules in eEF2K-bio led to interesting insights.

(1) From the perspective of similarity between generated and reference molecules, we used RDKit to calculate the Tanimoto coefficient for easier comparison between lead compounds generated in eEF2K-gen and reference molecules in eEF2K-bio. The generated molecules included both functional group modifications of existing skeletons (Figure 6a,b, with similarities of 0.79 and 0.67, respectively) and the creation of new skeletons (Figure 6c–e, with similarities of 0.42, 0.40, 0.44). We visualized the relationship between inhibitor candidates and original bioactive molecules using molecular pair similarity maps, where contour lines are denser in regions of high similarity and sparser in low-similarity areas. Except for areas like aromatic rings and conjugated structures, which have higher similarities, the sp3-sp3 bond similarities were lower, indicating that the generated model can capture key information about the interaction between reference molecules and protein targets and simulate similar molecular structures.

(2) From the perspective of drug-likeness between generated and reference molecules, we used tools like Swiss-ADME for visualization to intuitively assess the drug-likeness of lead compounds, evaluating parameters such as solubility, synthesizability, molecular size, flexibility, and polarity. Molecules within the pink area indicate good drug-likeness; those outside suggest poorer drug-likeness, with closer proximity to the hexagonal shape indicating better drug-likeness. As shown in the right-hand comparison in Figure 6, generally, the generated molecules’ drug-likeness ratings are all within the pink area, whereas three of the five reference molecules (CHEMBL109038, CHEMBL1091074, CHEMBL34241) exceed the pink area. In terms of shape area, except for entry d (CHEMBL1089016-3), which has a lower drug-likeness index than the reference molecule, the drug-likeness of the other four generated molecules is superior to that of the reference molecules. The overall drug-likeness quality of eEF2K-gen is not superior to eEF2K-bio (Figure S3), but the high drug-likeness rating of the lead compounds is mainly due to the scale of eEF2K-gen, which resulted from eliminating many molecules with poor drug-likeness. This method might offer insights into the development of AI in the pharmaceutical field.

(3) The lead compounds generated boasted highly optimized molecules in comparison to the reference compounds, as evidenced by molecular binding affinity. Notably, CHEMBL1091074-18, CHEMBL1090401-7, and CHEMBL1089016-3 stand out with their exceptional optimization, resulting in a noteworthy reduction in Vina score change in docking with eEF2K by 3.97 kcal/mol, 2.90 kcal/mol, and 4.59 kcal/mol, respectively (Figure 6b,c).

To reveal the changes in Vina scores, we visualized the docking of generated lead compounds with eEF2K protein using PyMol, comparing it with the docking situations of reference molecules (Figure 7). The docking results showed that the generated molecules bind well with the eEF2K pocket. The CHEMBL1090389-18 molecule (Figure 7a), which underwent functional group fine-tuning on the same framework, omitted the amino group and replaced the reference molecule’s thiazole ring with a pyridine ring. The pyridine ring created an aromatic-aromatic interaction with the benzene ring of TYR-236, a conclusion further supported by the two-dimensional diagram (the bottom of Figure 7a). CHEMBL1091074-18 (Figure 7b), compared to the reference molecule, increased steric hindrance with a substituent at the 4th position, but the carbonyl group on the amide bond formed hydrogen bonds with the amino of LYS-170, likely contributing to the significant reduction in Vina score. The skeletons of CHEMBL1090401-7 (Figure 7c), CHEMBL1089016-3 (Figure 7d), and CHEMBL34241-15 (Figure 7e) changed, but they bound more closely to the protein pocket. From the trends in Vina scores and biological activity curves, we believe these lead compounds offer a promising basis for further biological activity evaluation in capable laboratories.

3. Conclusions

This study marks a pioneering use of the recently unveiled crystal structure of the eEF2K protein, leveraging the structural information from known inhibitor-structure protein receptor complexes to generate an innovative library of inhibitor candidates, dubbed eEF2K-gen. Impressively, these candidates demonstrated superior thermodynamic performance in docking with eEF2K, outshining the inhibitors used in our training process. Our analysis revealed that the molecular generation model employed in this study adeptly captures the crucial ligand–receptor structural information, offering excellent interpretability. Among the top 5 lead compounds identified, a comprehensive analysis of their physicochemical properties suggested they have significant potential biological activity, warranting further investigation. This work, harnessing artificial intelligence tools, not only propels the development of eEF2K inhibitors but also promises to accelerate the discovery of treatments for diseases such as cancer and Alzheimer’s while paving the way for the development of inhibitors targeting other molecular sites.

4. Methods

4.1. Dataset

4.1.1. eEF2K-Bio

A search within the ChEMBL database using “eEF2K” as a keyword yielded 71 unique entries, with IC₅₀ values ranging from 60 nM to 60,000 nM. One molecule containing an unusual “Si” atom, unrecognized by the LiGAN model, was subsequently excluded. The eEF2K inhibitors were standardized using RDKit. There are 70 molecules known to inhibit the eEF2K protein, forming the bioactive dataset named “eEF2K-bio”.

4.1.2. The Crystal Structure of eEF2K

The crystal structure was sourced from the Protein Data Bank. Utilizing the Maestro software package and the OPLS4 force field, the nucleotide-free protein structure of eEF2K (PDB ID: 8FO6) was cleaned and preprocessed. The initial steps involved removing all water molecules and ions and adding hydrogen atoms to achieve accurate atomic configurations. Further processing included the removal of non-protein elements, rectification of missing atoms or residues, standardization of naming conventions, and correction of observable errors. An energy minimization procedure was applied to attain a stable, low-energy conformation of the protein. Finally, PyMOL was meticulously employed to remove any remaining water molecules, ensuring the protein’s pristine condition for subsequent computational analysis.

4.1.3. PyMOL

PyMOL was originally developed by Warren L. DeLano and is currently maintained by Schrödinger, LLC (New York, NY, USA). It is used for the visualization of molecular structures, allowing scientists to see and manipulate 3D models of proteins, nucleic acids, and other biomolecules. PyMOL provides tools for molecular modeling, including measuring distances, angles, and dihedrals, as well as generating surfaces and computing electrostatic potentials.

4.2. Molecular Docking

4.2.1. DiffDock

In this study, molecular docking served a dual purpose. First, we employed the deep learning-based docking tool DiffDock (v1.1.2) to initially position protein-small molecule complexes for reference molecules within the eEF2K-bio dataset. The use of DiffDock was pivotal in accurately positioning the molecules, setting the stage for subsequent processes with LiGAN. The parameters for DiffDock docking were in line with the standard settings outlined in the DiffDock publication, with further details accessible in the code availability section. This step involved blind docking of the small molecule into the appropriate pocket of the protein. Subsequently, these complexes were utilized as input for LiGAN, which requires the docked complex structure as its input format.

4.2.2. AutoDock

Following the generation of new molecules, AutoDock was employed to assess the Vina score of both the newly generated and reference molecules for further analysis. AutoDock facilitates the evaluation of binding affinity for both reference and generated molecules, a critical step in determining whether the newly synthesized molecules retain inhibitory properties. This is because binding affinity is a plausible indicator of inhibitory potential. The docking results were then utilized to gauge the binding affinity of these molecules, aiding in the assessment of their effectiveness as eEF2K inhibitors. Practically, we adhered to the standard protocols provided in the AutoDock official tutorial to compute binding affinity, employing a docking box generated by PyMOL plugins. The comprehensive process is detailed in our code protocol.

4.3. Deep Generative Model

The centerpiece of this framework is LiGAN, an innovative deep generative model which plays a pivotal role in the synthesis of new molecular entities. The essence of LiGAN’s methodology lies in its adept utilization of information about docking sites and the structural intricacies of the eEF2K pocket. This model, specifically honed for molecular generation, operates on a conditional variational autoencoder (CVAE) architecture and is optimized for CUDA-enabled GPUs, reflecting its alignment with cutting-edge computational techniques.

The dataset employed in this model adheres to the “oadc-1.0” atom typing scheme for both receptors and ligands. A meticulous 0.5 Å grid resolution was employed, supplemented by randomized grid rotations and translations, a methodological choice aimed at enhancing the model’s predictive accuracy. Structurally, LiGAN’s generative model commenced with 28 filters within its convolution layer, which incrementally doubled in width per factor, and was composed of four convolution blocks, each containing four layers. The layers utilize a 3 × 3 convolution kernel size paired with Leaky ReLU activation, optimizing the model for efficient molecular processing.

During the generative phase, for every 70 input instances, the model yielded 20 samples, each demonstrating an accurate alignment of atoms to densities. This systematic approach not only underscores the precision of the molecular generation process but also exemplifies the efficiency and efficacy of the framework. The outcomes of this study highlighted the potential of this framework as a significant tool in the discovery and development of new eEF2K inhibitors, thereby contributing to the evolving landscape of targeted therapeutic interventions.

4.4. Metrics

4.4.1. Drug-Likeness

This study utilized a comprehensive approach to evaluate the potential of eEF2K inhibitor candidates. The approach involved a combination of molecular assessment metrics and online tools to thoroughly assess drug-likeness, synthetic feasibility, pharmacokinetics, and toxicology. The molecular characteristics were analyzed using various metrics, such as QED, SA score, LogP, Molecular Weight (MW), and Vina score. MW and LogP metrics helped in understanding the ADME profiles. An ideal MW below 500 was considered optimal for drug absorption and distribution, and the optimal LogP range was 0 to 3 for balancing water solubility and cell membrane permeation. Additionally, drug-likeness and synthetic complexity were quantified using the QED and SA score, respectively. The QED score, ranging from 0 to 1, indicated drug potential, with a favorable score above 0.67. The SA score, ranging from 1 to 10, assessed the ease of synthesis. The Vina score was also used and noted for its significant correlation with binding affinity. For pharmacokinetics and toxicological analyses, online tools Swiss-ADME (version 2023) [51,52] (http://www.swissadme.ch, accessed on 29 March 2024) and eMolTox (v1.2) [53] (http://xundrug.cn/moltox, accessed on 29 March 2024) were used to predict the properties of selected eEF2K inhibitor hits.

4.4.2. Tanimoto Similarity

In this study, to assess the structural similarity between compounds, we employed the Tanimoto similarity coefficient. Initially, the structures of the compounds were converted into molecular fingerprints, a method that encodes chemical structural information into binary sequences. For this process, the Extended-Connectivity Fingerprints (ECFP) algorithm in RDKit was chosen due to its efficiency in capturing essential chemical features of molecules. Subsequently, the similarity between each pair of compounds was calculated using the Tanimoto formula. Specifically, the Tanimoto similarity is determined by comparing the binary bits at corresponding positions in two molecular fingerprints. The similarity is calculated using the formula:

$$(\mathrm{T}={\displaystyle \frac{\mathrm{A}\cap \mathrm{B}}{\mathrm{A}\cup \mathrm{B}}})$$

where $\mathrm{A}$ and $\mathrm{B}$ represent the molecular fingerprints of the two compounds. Here, $\mathrm{A}\cap \mathrm{B}$ denotes the count of bits that are 1 in both fingerprints, while $\mathrm{A}\cup \mathrm{B}$ represents the count of bits that are 1 in at least one of the fingerprints. The resulting similarity value ranges from 0 (no similarity) to 1 (identical). A threshold was set to determine significant structural similarity between compounds. This method is widely applied in molecular similarity assessments, particularly in the fields of drug design and cheminformatics.