In-Silico discovery of novel cephalosporin antibiotic conformers via ligand-based pharmacophore modelling and de novo molecular design

Highlights

• •Developed a validated pharmacophore model with a high GH score.
• •Identified dug-like candidates through ligand-based virtual screening.
• •Designed 30 novel cephalosporin analogs via de novo fragment-based synthesis.
• •Confirmed synthesis feasibility of top candidate molecules via AI retrosynthesis.

Abstract

Antibiotic resistance poses a significant global challenge as bacteria evolve in response to antibiotic use, leading to prolonged hospitalizations, increased healthcare costs, and higher mortality rates. Cephalosporins, a class of beta-lactam antibiotics, are commonly employed to manage infections; however, their misuse and overuse have contributed to resistance development. In response, in silico methods have emerged as cost-effective and efficient tools for drug discovery. This research aims to predict new compounds using ligand-based pharmacophore models while optimizing existing drugs. We employed a de novo approach to synthesize models of cephalosporin structural motifs, integrating the β-lactam core with potential antibiotic candidates. A shared features pharmacophore (SFP) model was constructed using cephalosporins from PubChem, including cephalothin, ceftriaxone, and cefotaxime. The model comprises hydrogen bond acceptors, hydrogen bond donors, aromatic rings, hydrophobic regions, and negatively ionizable sites. Its robustness was evidenced by a goodness-of-hit (GH) score of 0.739. The generated pharmacophore model, with a score of 0.9268, was utilized to screen a drug library, initially assessing 19 compounds. After the drug-likeness screening, seven promising compounds were identified. These candidates were then fused with the cephalosporin core using genetic algorithms and fragment-based design, resulting in 30 novel synthetic models. Most of these models demonstrated a cephalosporin core, over 70 % average similarity, a TPSA (NO) ≤ 99.85 Å², a drug-likeness (QED) ≥ 0.6, and a Synthetic Accessibility Score (SAScore) ≤ 4.3. Molecular docking and MD simulation evaluations highlighted two candidates—Molecule 23 and Molecule 5, demonstrating superior binding affinities to Penicillin-binding protein 1a (PDB ID: 2V2F) compared to controls. To ensure feasible synthesis, molecular architecture comparison and computational retrosynthesis were performed, confirming the likelihood of successful laboratory synthesis. These findings advance the fight against antimicrobial resistance by establishing a method for designing new, highly effective antibiotic drugs.

1. Introduction

Antibiotics are chemicals that are either synthetic or produced by microorganisms that specifically inhibit the growth of bacteria, viruses, and other microorganisms and can even destroy them with minimal harm to host cells.¹ Although antibiotics have been essential in saving lives and lowering death rates, antimicrobial resistance is making antibiotics less effective, which is making diseases more difficult to treat.² Antibiotic resistance is the ability of microbial cells to resist antibiotic pressure and prevent the bacteriostatic and bactericidal effects of antibiotics. Nowadays, resistance is a big phenomenon in the path of new drug synthesis.⁶ Among many antibiotics, cephalosporin, a critical class of broad-spectrum antibiotics, has played a cornerstone role in the treatment of numerous bacterial infections.³ Their clinical significance extends to the prophylaxis and treatment of infections affecting the skin, ear, bone, respiratory tract, and urinary tract, particularly in hospitalized patients.⁴ Cephalosporin targets peptidase enzymes present in the outer cytoplasmic membrane, referred to as penicillin-binding proteins, and represents the target sites for antimicrobial action.⁵ There are four fundamental mechanisms that are responsible for antibiotic resistance against cephalosporins, which include enzymes produced by bacteria such as beta-lactamase, alterations in penicillin-binding proteins (PBPs), inadequate delivery of antibiotics to the target site, and the development of alternative metabolic pathways.4, 6 An analysis of approximately 10,000 hospital-associated urinary tract infection (UTI) episodes caused by Escherichia coli in the United States revealed an 8.6 % rate of third-generation cephalosporin resistance. In previous studies, it was 7.3 %. According to the studies, the overall rate of resistance to third-generation cephalosporins exceeded 5 to 10 %.⁷ In this study, cephalosporin, including cephalothin, ceftriaxone, and cefotaxime, is mainly focused. Among them, cephalothin is the first-generation antibiotic, and ceftriaxone and cefotaxime are third-generation antibiotics.6, 8These antibiotics are administered to treat complicated urinary tract infections, MRSA (Methicillin-resistant Staphylococcus aureus), and other multidrug-resistant pathogens, such as Neisseria gonorrhoeae.⁵ As bacterial resistance to the established classes of antibiotics used in clinical practice continues to rise, there is an increasing need for new agents capable of treating newly multidrug-resistant organisms.⁹ To deal with the issue, it is necessary to create new generations of antibiotics that are safer and more effective, have a wider range of action against different infections, and can also adapt to the challenges of the modern world.¹⁰.

One promising avenue in the search for new antibiotics is the synthetic development of antimicrobial compounds.¹¹ Traditional methods of drug discovery are often labor-intensive and time-consuming, necessitating a shift towards more efficient strategies.¹² Additionally, clinical trials conducted by the Food and Drug Administration (FDA) are primarily noticeable among the many factors that contribute to the decline in the development of new antibiotics.13, 14 There are several advanced in-silico methods and tools that aid in early drug discovery by reducing the time and cost for drug discovery processes.15, 16, 17 However, pharmacophore modelling is the subfield of computer-aided drug design and one of the major tools in drug discovery,18, 19 and demonstrates promise as an effective way of overcoming bacterial resistance.²⁰ Pharmacophore shows excellent relations between the 3D structure and the activity of molecules. It can also predict three-dimensional quantitative structure-activity relationship (3D QSAR) models and align molecules based on 3D arrangements of chemical features.²¹ Pharmacophore features include a hydrophobic area, aromatic ring system, hydrogen bond donor and acceptor, and negatively and positively ionizable groups²². Various structure- and ligand-based methods have been developed for improved pharmacophore modeling.¹⁸ It has major applications in virtual screening, drug target fishing, ligand profiling, docking, and ADMET prediction.²³ Virtual screening (VS) is the process of identifying potential drug candidates by analysing huge databases of chemical compounds using high-performance computing.24, 25 Two widely used approaches for virtual screening are structure-based methods (docking) and ligand-based methods.²⁶ Ligand-based virtual screening techniques utilize the information found in known active ligands for lead optimization and identification by employing an efficient similarity measure and an accurate scoring system.²⁷ Utilizing the pharmacophore model, we can assess de-novo molecular designs of new cephalosporin conformers that may overcome existing resistance mechanisms by optimizing their interaction with the target molecule.²⁸ Following that, docking is the process of exploring the conformational space of ligands and orientation within a targeted binding site of molecules.²⁹ Molecular docking is one of the most widely used computational methods for repurposing chemical substances toward new therapeutic targets. Docking enables fast, cost-effective virtual screening of databases containing approved drugs, natural products, or prior synthetic entities to target biological macromolecules.30, 31 The computational approaches of drug discovery can then make use of this information to develop new antibiotic drugs that may exhibit increased activity.³².

Our primary goal in this work is to design novel molecular descriptors of cephalosporin antibiotics by fusing them with potential cephalosporin compounds. First, a ligand-based pharmacophore model was generated by utilizing existing drugs of the cephalosporin antibiotic class. This model was further used for screening potential drugs of this class of antibiotics, which were then conjugated with the core ring of cephalosporin antibiotics to develop new cephalosporin conformers. These new compounds were further evaluated using scaffold analysis, molecular docking, and molecular dynamics simulation. As a final and critical component of this research, an algorithmic retrosynthetic strategy was employed to determine the practical chemical synthesis viability of the generated molecules, thereby bridging the gap to real-world applications.

2. Method and materials

2.1. Selection of training set compounds

By conducting searches in the published literature on novel classes of antibiotics,³³ we have identified and selected cephalosporins. We extracted a total of three compounds from the first and third generations of these classes, named cephalothin (PubChem ID: 6024), cefotaxime (PubChem ID: 5742673), and ceftriaxone (PubChem ID: 5479530), and flagged them as training sets. We retrieved their 3D conformer from the PubChem Database (https://pubchem.ncbi.nlm.nih.gov)³⁴ in SDF (Structure Data File) format. For further analysis to create a pharmacophore model, these compounds have been provided as a test. The study overview is represented in Fig. 1, emphasizing the procedures and techniques used to design and evaluate the drug compounds.

Fig. 1.

Overview of the study. The investigation includes the retrieval of drug compounds from PubChem and pharmacophore model generation to drug development.

2.2. Common feature pharmacophore model generation

We conducted ligand-based pharmacophore modeling on the training set by using LigandScout 4.5 (https://www.inteligand.com/ligandscout).³⁵ After importing the training set of compounds and performing the “create Ligand-based pharmacophore” process in LigandScout, based on chemical structure alignment, compounds with the highest pharmacophoric features (having the highest pharmacophoric fitting score) generated a 3D shared features pharmacophore (SFP) model. From the ten generated models, we selected the best model, which represents hydrogen bond acceptor (HBA), hydrogen bond donor (HBD), aromatic ring (AR), hydrophobic feature (H), negative ionizable (NI), and excludes the volume sphere, which defines an area for which any compound is not allowed to map.³⁶ Then we incorporated the pharmacophores of each ligand and selected the SFP model from an alignment perspective to align them together for better visualization as well as optimization. Here, our generated model is used for the next virtual screening process.

2.3. Database generation from the benchmark library of ZINCPharmer

In this study, we prepared a compound library of screening compounds based on our generated SFP model. Then we looked through our database in ZINCPharmer³⁷ to identify and retrieve potential compounds that have been matched with the generated pharmacophoric features that would be used in the further virtual screening process. The Zinc database consists of 13,190,317 commercially available chemical compounds integrated with the Pharmit web server (https://pharmit.csb.pitt.edu/). In ZINCPharmer (https://zincpharmer.csb.pitt.edu/), we uploaded the pharmacophore characteristics of each of the selected three ligands in the SDF file. Then, by analysing our generated models circumstantially with feature positions in the map for every ligand, we selected our desired features in a combinatorial way with one or several HBA and HBD and with some fixed features, including 1AR, 1H, and/or 1NI, to submit the query for screening and finding the result of our desired database. Then, all the possible generated hits for every submission were saved as an SDF file format for further analysis.

2.4. Pharmacophore model evaluation

To validate the constructed SFP model, we employed the Goodness-of-Hit (GH) scoring method, a widely recognized approach for assessing model reliability and predictive accuracy.38, 39 The GH scoring method, also known as the Gunner-Henry (GH) method, was utilized to evaluate the model's ability to selectively identify active compounds while minimizing false positives and false negatives within a dataset comprising both active and inactive compounds. Key performance metrics included total hits (Ht), percentage yield of active compounds, active ratio, enrichment factor (E), false negatives, false positives, and the GH score itself.⁴⁰.

Furthermore, additional validation parameters such as the model F1 score, pharmacophore map precision, and recall were examined to ensure a comprehensive assessment of the model's performance.41, 42 The GH score ranges from 0 to 1, where a value of 0 represents a null model with no predictive capability, while a score approaching 1 indicates an ideal model with high accuracy in identifying active compounds.⁴³ These evaluation metrics are critical in determining the pharmacophore model's robustness, ensuring its ability to effectively distinguish between active and inactive compounds while maintaining an optimal balance between precision and recall.

Goodness-of-hit (GH) score:

$$\mathit{GH}=\left(\frac{\mathit{Ha}(3A+Ht)}{4HtA}\right)\times \left(1-\frac{\mathit{Ht}-Ha}{D-A}\right)$$

here,

D = Total molecules in the database.

A = Total number of actives in the database.

Ht = Total hits.

Ha = Active hits.

Additional analysis:

Pharmacophore map Precision:.$\mathit{Precision}=\frac{\mathit{Ha}}{\mathit{Ha}+D}$

Pharmacophore map Recall:.$\mathit{Recall}=\frac{\mathit{Ha}}{A}$

Model F1 score:.$F1=2\times \frac{\mathit{Precision}\times Recall}{\mathit{Precision}+Recall}$

We constructed a small library with a total of 500 compounds obtained from ZINCPharmer in our previous study, with 10 actives collected from ChEMBL.⁴⁴ The database was employed to assess the pharmacophore model’s effectiveness in distinguishing between active and inactive compounds. The database screening was performed using the LigandScout.

2.5. Virtual screening of the generated model to find the potential leads

To select a minimum number of candidate molecules as lead that were well fitted in our generated SFP model, we performed virtual screening. Performing virtual screening reduces the number of compounds for experimental testing by comparing pharmacophoric features in the query pharmacophore to pharmacophoric features present in the molecules of the screened library based on the alignment method.²² For this study, first, we converted all the saved query results from previous analyses into a few SDF file formats by using OpenBabel GUI tools.⁴⁵ In LigandScout 4.5,³⁵ to generate a screening database in IDB file format, we loaded our OpenBabel GUI-converted SDF files one by one. Drug library compatibility with LigandScout is limited to the.idb file format, requiring the application of the software’s built-in ‘generate database’ function for requisite format conversions. In this process, duplicate hits (compounds) were removed, and only authentic compounds were incorporated into the generated IDB file. After creating the IDB file format of the database for screening, we transferred our generated SFP model from a screening perspective, and then we selected all the IDB files one by one to perform screening.

After performing screening, we obtained well-fitted molecules in our SFP model as hit libraries. Molecules were arranged according to their pharmacophoric-fit score. All the resultant hits from each database were selected and retrieved from the ZINC database by substance searching with the mentioned ZINC ID as the SDF file format for further filtration by applying the drug-likeness screening tool.

2.6. Drug-likeness prediction and physicochemical properties screening

In this study, based on Lipinski’s rule of five⁴⁶ and physicochemical properties, training sets and hit compounds were screened. The compounds retrieved from the Zinc database were checked for drug-like properties using the SwissADME tool (https://www.swissadme.ch/), which evaluates required ADME (absorption, distribution, metabolism, and excretion), drug-likeness (rule of five), physicochemical properties, and medicinal chemistry friendliness.⁴⁷ Lipinski’s rule of five is mainly alluded to: hydrogen acceptor ≤10, hydrogen donor ≤5, an octanol/water partition coefficient (iLOGP) value ≤5, molecular weight ≤500 Da, and number of rotatable bonds ≤10.⁴⁶ An orally active drug should not violate more than one of the criteria mentioned. With the aid of canonical SMILE information, by accessing the PubChem server (https://pubchem.ncbi.nlm.nih.gov/) for the training set of compounds and by accessing the substances search in the Zinc database (https://zinc.docking.org/substances/home/) for the hit compounds, we performed this study in SwissADME. The properties that we extracted from the evaluation of the study included molecular weight, number of H-bond acceptors, number of H-bond donors, molar refractivity, the total polar surface area (TPSA), the partition coefficient between n-octanol and water (iLOGP), GI absorption, blood–brain barrier (BBB) permeant, and drug-likeness. After carrying out the assessment, we subjected the drug-like compounds to further studies.

2.7. Automated de novo molecular design to develop novel cephalosporin antibiotics

In this study, we aimed to develop novel chemical entities of cephalosporin antibiotics by integrating the ZINCpharmer-derived hits. Following a drug-likeness screening process, multiple potential compounds were identified that conformed to established drug-likeness criteria. Given the challenge of antibiotic resistance in infectious bacteria, our approach involved incorporating these potential hits into the cephalosporin antibiotic class to explore novel drug candidates. To facilitate de novo molecular design, we employed AlvaBuilder v.1.0.10⁴⁸, a molecular generation tool from Alvascience.⁴⁹.

Alvabuilder utilizes genetic algorithms⁵⁰ for de novo molecular design, enabling the creation of novel molecules with desirable pharmacological properties. The software assembles molecular structures by combining structural fragments from a predefined training set, implementing a graph-based molecular construction approach. Additionally, it integrates QSAR/QSPR modelling, drug energy minimization, stability analysis, and descriptor calculation to assess newly generated compounds.⁵¹ The molecular fragmentation process within AlvaBuilder categorizes molecules into ring systems, linkers, and side chains for systematic recombination.

For this study, three antibiotics from the cephalosporin family, Cephalothin, Cefotaxime, and Ceftriaxone, were selected as test set ligands. The core cephalosporin structure was maintained throughout the molecular design process, ensuring the retention of the essential ß-lactam ring, which is a critical component of cephalosporin antibiotics.⁵² This ß-lactam ring consists of a four-membered ring with a carbonyl functional group (Fig. 2), serving as a nucleophilic attack site for hydrolysis and contributing to the reactivity of ß-lactam antibiotics.52, 53 Consequently, all generated drug molecules retained this structural feature, while the remaining molecular framework was remodeled using key features from the ZINCpharmer hits.

Fig. 2.

Visual representation of the conserved Cephalosporin core ring, a key structural feature of Ceftriaxone, Cefotaxime, and Cephalothin antibiotics.

The identified hit compounds were imported into AlvaBuilder as the training dataset. The software was configured to utilize an arithmetic mean as the aggregation method, enabling the crossover of molecular fragments while preserving the integrity of the β-lactam-containing core rings. QSAR/QSPR modeling facilitated the selection of chemically compatible functional groups that could engage in strong covalent or non-covalent binding. Several other selection criteria were imposed, including a topological polar surface area (TPSA) of ≤140 Å², a Synthetic Accessibility Score (SAScore) of ≤4, and a drug-likeness (QED) value of ≥0.6.⁵⁴ The genetic algorithm parameters were configured with a population size of 30 and 30 iterations.

The SAScore served as a key benchmark for assessing the feasibility of synthesizing the generated compounds in a laboratory setting.⁵⁵ This heuristic-based score incorporates multiple factors, including molecular complexity, the presence of challenging functional groups, and the availability of synthetic pathways.⁵⁶ SAScore values range from 1 to 10, where lower scores indicate greater synthetic feasibility, whereas higher scores suggest increased complexity.⁵⁷ To ensure practical synthesis, we set the SAScore threshold at ≤4, prioritizing compounds that were both theoretically effective and synthetically accessible. Applying these parameters, AlvaBuilder generated 30 synthetic cephalosporin conformers, all of which retained the invariant core cephalosporin structure, including the ß-lactam ring. The newly designed synthetic models were obtained in PDB format using the SwissTargetPrediction tool (https://www.swisstargetprediction.ch/). These structures were subsequently subjected to further screening and analyses to identify potential drug candidates.

2.8. Molecular docking studies

Molecular docking is the method of a structure-based drug design that simulates the molecular interaction between receptors and ligands and predicts the binding mode and affinity between them.⁵⁸.

2.8.1. Selection and preparation of proteins

Protein preparation is the main step of molecular docking studies; it involves mainly the deletion of water molecules, the addition of hydrogen atoms, and protein energy minimization.⁵⁹ From DrugBank (https://go.drugbank.com),⁶⁰ we selected our target protein by searching with the name of our selected three ligands or drugs (cephalothin, cefotaxime, and ceftriaxone) of the cephalosporin antibiotic class. We identified Penicillin-binding protein 1a (Uniprot ID: Q8DR59, PDB ID: 2V2F) as our target and obtained it with a resolution of 1.90 Å and an R-value-free score of 0.246⁶¹ from the RCSB Protein Data Bank (https://www.rcsb.org/pdb)⁶² in PDB format. The obtained protein molecules may typically contain heavy atoms and include a co-crystallized ligand, cofactors, metal ions, and water molecules. With the aid of BIOVIA Discovery Studio⁶³, all the mentioned components were eliminated from the protein structure as a refinement process, polar hydrogen was added, and then the protein file was saved as a PDB file. Additionally, we also identified the protein’s active site by utilizing PyMol.⁶⁴ Afterward, Swiss-Pdb Viewer⁶⁵ was utilized for energy minimization and the fixation of deformed protein structural geometries.⁶⁶ After finalizing the process in Swiss-Pdb Viewer, the prepared protein was saved as a PDB file and further used for docking analysis.

2.8.2. Ligand selection

We selected ceftriaxone (PubChem ID: 5479530), cefotaxime (PubChem ID: 5742673), and cephalothin (PubChem ID: 6024) as our control drugs, which were obtained from the PubChem database³⁴ in our previous study. Newly designed synthetic de-novo models of 30 cephalosporin compounds and control drugs that possessed all of the desired physicochemical properties were regarded as ligands for docking studies.

2.8.3. Docking approach

Molecular docking was conducted against the protein molecule for selected standard drugs and de-novo designed synthetic compounds. The molecular docking was carried out using AutoDock Vina in the Python Prescription (PyRx) 0.8 virtual screening tool.⁶⁷ In PyRx, protein molecules were converted to PDBQT format as AutoDock macromolecules. All the SDF files of the ligand molecule were inserted into Open Babel in PyRx and converted to AutoDock Ligand PDBQT format after minimization. The grid points in the X, Y, and Z axes were set in the active site pocket center of the protein with a grid docking center X: 48.7078, Y: −5.8364, Z: 21.0270 and dimensions (Angstrom) X: 15.8011, Y: 14.6859, Z: 18.6309. The lowest binding energy of each cluster was retrieved as exemplary.⁶⁸ The results were graded in the order of highest binding affinity or lowest binding energy. After performing docking, binding interactions of target proteins for respective ligands were inspected by BIOVIA Discovery Studio Visualizer⁶⁹ and recorded carefully. The obtained affinity of synthetic compounds was compared with the used control drugs.

2.9. Molecular dynamics simulations (MDS)

Molecular dynamics (MD) simulations were employed to assess the stability and convergence of interactions of the formed complex between the protein target and the drug candidates.70, 71 The simulations were performed for a period of 200 ns utilizing the Desmond simulation software, a product of Schrödinger LLC.72, 73 In this study, the protein–ligand complex was a key component for determining the static representation of the ligand's binding orientation within the protein's active site. MD simulations track atomic trajectory analysis over time and provide insight into ligand binding behaviour in an in vivo-like environment by integrating Newton's classical equation of motion. Prior to initiating MD simulations, system relaxation was carried out using the default Desmond relaxation protocol.

Protein-ligand complexes underwent preprocessing through the Protein Preparation Wizard in Maestro software, involving optimization, minimization, and system construction via the System Builder tool. The TIP3P (Transferable Intermolecular Interaction Potential 3 Points) solvent model was applied, and the complex was enclosed in an orthorhombic box to mimic realistic environmental conditions.⁷⁴ The OPLS_2005 force field (ff) was used for simulations, with counter ions added as necessary to neutralize the system.⁷⁵ To simulate physiological conditions, a 0.15 M NaCl solution (Na + ions) was introduced via the system builder option. The simulation was conducted under equilibrium conditions, employing NVT (constant number of particles, volume, and temperature) and NPT (constant number of particles, pressure, and temperature) ensembles to maintain a stable temperature of 300 K and pressure of 1 atm, ensuring conservancy of temperature (T), pressure (P), and moles (N).

To assess the reliability of the simulations, several parameters were measured for both control and top-ranked complexes, including root mean square deviation (RMSD),⁷⁶ root mean square fluctuation (RMSF), protein–ligand contact (PL-contact), solvent-accessible surface area (SASA), radius of gyration (RG), and molecular surface area (MolSA). ⁷⁷

The root mean square deviation (RMSD) quantifies the average displacement of a subset of atoms in a given frame relative to a reference frame. It is calculated for each frame in the trajectory as:

$${\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{D}}_{\mathrm{x}}=\sqrt{(\frac{1}{N}}{\sum }_{i=1}^N\left(r^{\prime }i\left(t_x\right)\right)-{\mathrm{r}}_{\mathrm{i}}{({\mathrm{t}}_{\mathrm{r}\mathrm{e}\mathrm{f}}))}^2)$$

where N represents the number of atoms in the selection, t_ref is the reference time (typically the first frame at t = 0), and r' is the position of selected atoms in frame x after superimposing on the reference frame at time t_x. The calculation is iteratively performed across all frames of the simulation trajectory. RMSD provides insight into atomic positional fluctuations over time, with higher values indicating greater structural deviations and lower values suggesting stability.78, 79, 80.

The root mean square fluctuation (RMSF) serves as a measure of localized changes along the protein chain. It is determined using the equation:

$${\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{F}}_{\mathrm{i}}=\sqrt{(\frac{1}{T}}{\sum }_{t=1}^T<(\mathrm{r}\prime \mathrm{i}(\mathrm{t}))-{\mathrm{r}}_{\mathrm{i}}{(\mathrm{t}\mathrm{r}\mathrm{e}\mathrm{f}))}^2>)$$

where T is the duration of the trajectory, t_ref is the reference time, r_i denotes the position of residue i, and r' represents the atomic position of residue i after superposition on the reference structure. Angle brackets indicate averaging over the selected atoms within the residue. This metric provides insight into the flexibility of individual residues during the simulation, helping to identify regions of structural stability and fluctuation.

2.10. Molecular pattern comparison and computational retrosynthesis route analysis

Following the final screening process, several potential synthetic conformers of cephalosporin antibiotics were identified. The subsequent step involved assessing drug compatibility by comparing these conformers to a control drug, Ceftriaxone. An algorithmic approach was employed to analyze substructure relationships and chemical pattern similarities between the control drug and the newly predicted drug candidates. The SMARTS.plus tool⁸¹ (https://smarts.plus/) from ZBH–Center for Bioinformatics (https://www.zbh.uni-hamburg.de/en.html) was utilized for conducting the comparative study. The tool integrates SMARTSview, SMARTS-miner, and SMARTScompare for advanced molecular pattern analysis. Within SMARTScompare, two molecular expressions (potential drug and control drug) were uploaded, allowing the server to compute substructure relationships and pattern similarities. The results were visually represented through graphical node mapping, facilitating the interpretation of pattern relationships.⁸².

To establish a retrosynthetic route for the modeled drugs, the IBM RXN for Chemistry program⁸³ (https://rxn.app.accelerate.science/rxn/) was utilized. This web-based graphical user interface (GUI) supports the prediction of chemical reactions and the design of organic synthesis pathways. IBM RXN for Chemistry is based on a stepwise transformer network topology, which applies advanced machine- learning techniques to forecast chemical transformations and generate pathways for chemical synthesis.⁸⁴ The program incorporates both template-free and template-based methodologies to deconstruct target molecules by strategically breaking specific bonds, converting them into intermediates, and applying various chemical strategies until the final available compound is determined. To efficiently explore potential synthetic pathways, a probabilistic approach leveraging Monte Carlo Tree Search (MCTS) was employed.⁸⁵.

The retrosynthetic routes generated by the platform were evaluated and assessed using multiple factors, including reaction straightforwardness, the procurement of starting materials, and overall synthetic feasibility. Although the baseline settings provided by IBM RXN were primarily used in the analysis, additional interactive refinements were applied to optimize the proposed routes, ensuring the development of the most effective and practical synthetic strategies.

3. Results

3.1. Pharmacophore model generation

The three selected compounds were used as training sets for developing a pharmacophore model. These compounds were analysed for critical structural and chemical property evaluation by using LigandScout. Based on the alignment of the structural properties of the chosen compounds, the model was generated. With a score of 0.9268, the best shared-features pharmacophore (SFP) model was selected. The model had 14 feature patterns, including 8 hydrogen bond acceptors (HBA), 3 hydrogen bond donors (HBD), 1 aromatic ring (AR), 1 hydrophobic zone (H), and 1 negative ionizable (NI) area (Fig. 3). Through the study, selected antibiotics as a training set were also assessed and had their depicted pharmacophore features distinctly described (Fig. 4). Table 1 summarizes the respective antibiotics (training set) pharmacophore features and their shared pharmacophore features.

Fig. 3.

Schematic representation of the shared pharmacophore model. (A) The spatial arrangement of the generated SFP model-1 with the training set compounds. (B) The features are shown in different colors. Hydrogen bond acceptor (HBA) is shown in red spheres; hydrogen bond donors (HBD) with green spheres and arrows; an aromatic ring (AR) with a small blue sphere; a hydrophobic (H) space with a yellow sphere; and the small red sphere represents a negative ionizable (NI) zone. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 4.

The three training sets contain compounds for developing the pharmacophore model. The model had 10 (A-PubChem ID: 6024), 14 (B-PubChem ID: 5479530), and 13 (C-PubChem ID: 5742673) pharmacophore feature patterns, respectively. The mapping of the respective training set compounds (D) Cephalothin-6024, (E) Ceftriaxone-5479530, and (F) Cefotaxime-5742673 with the generated SFP model.

Table 1.

Pharmacophore features and their shared pharmacophore features of the respective antibiotics.

Training set compound	Pharmacophore features
	H-bond acceptor	H-bond donor	Aromatic	Hydrophobic	Negative ionizable
Cefotaxime	8	2	1	1	1
Ceftriaxone	8	3	1	1	1
Cephalothin	6	1	1	1	1
Shared	8	3	1	1	1

3.2. Library generation of screening compounds

The generated pharmacophore model was utilized as a template model for different pharmacophore features mapping in the respective antibiotics of the training set that were used as a 3D structural search query to perform a virtual database search in ZINCPharmer. In the database screening process, compounds that match the submitted pharmacophore features of interest were gathered as a collection of hits. Throughout the process, from the ZINC compound database, approximately 250k hits (compounds) were obtained as a different query result of the SDF file format.

3.3. Shared feature pharmacophore (SFP) model validation

The validation of the pharmacophore model was a critical step to ensure its reliability and predictive accuracy. To achieve this, we employed an external validation approach utilizing the Goodness-of-Hit (GH) score. The screening database comprised 490 compounds from ZINCPharmer and 10 known active compounds from the ChEMBL database. To assess the model's ability to discriminate between active and inactive compounds, the pharmacophore model was applied as a 3D query for virtual screening. Following the screening process, a total of 13 compounds were identified as hits from the 500-drug compound library. Among these, 4 compounds originated from ZINCPharmer, while 9 were known active drugs from ChEMBL (see Table S1). The validation results, including key performance metrics, are summarized in Table 2, with the GH score serving as a principal measure of model quality.

Table 2.

Validation of the pharmacophore model based on the GH score, F1 score, precision, recall, and enrichment factor (E).

Model parameters	Value
Total molecules in database (D)	500
Total number of actives in the database (A)	10
Total hits (Ht)	13
Active hits (Ha)	9
% Yield of actives[(Ha/Ht) × 100]	69.23 %
% Ratio of actives [(Ha/A) × 100]	90 %
Goodness of hit score (GH)	0.739
False negatives [A − Ha]	1
False positives [Ht – Ha]	4
Enrichment factor (E) [(Ha × D)/(Ht × A)]	34.62
Precision	0.01768
Recall	0.9
F1 score	0.03467

The model demonstrated a high predictive capability, as indicated by a 69.23 % yield of active compounds, which represents the prevalence of active hits across all identified hits. The 90 % active ratio reflects the fraction of active hits relative to the entire set of known active compounds in the database. Additionally, the enrichment factor (E) of 34.62 signifies the model's strong ability to distinguish active compounds from inactive ones, with higher values indicating superior performance in identifying active molecules. Notably, the GH score of 0.739 confirms the high quality of the pharmacophore model, reinforcing its effectiveness for virtual screening and drug discovery applications.

These results validate the robustness of the constructed pharmacophore model, demonstrating its potential for identifying biologically relevant compounds with high precision and efficiency.

3.4. Virtual screening of generated shared feature pharmacophore (SFP) model

By eliminating all the duplicate hits (compounds) from the database library, which were obtained from the ZINC compound database, approximately 200k compounds contained in several IDB file formats were prepared for the virtual screening process of the SFP model. After carrying out the virtual screening process with the prepared compounds database by utilizing the generated SFP model as a template, different hit libraries of 19 well-fit compounds were identified. All the fitted compounds obtained from Hit-library were ranked based on their pharmacophore-fit score, ranging from 107.74 to 128.18 (Table 3). We conducted the study to identify our potential lead molecules that exhibited the corresponding pharmacophore features of our generated SFP model. Fig. 5 illustrates the two-dimensional structures of ZINC compounds alongside their pharmacophore features, while Fig. 6 showcases the three-dimensional representations derived from virtual screening analysis. These compounds are presented based on the significant results obtained, which will lead to a discussion of our subsequent drug-likeness screening method in the following sections.

Table 3.

List of obtained hit compounds from virtual screening.

Serial no.	ZINC ID	Pharmacophore-Fit Score	RMSD	Mol. Index	Number of conformers
1	ZINC03871978	107.74	0.026554	5621	25
2	ZINC03830485	107.76	0.230607	35801	25
3	ZINC21986204	107.85	0.08726	5349	25
4	ZINC35973845	113.87	0.602396	24072	25
5	ZINC03830413	113.9	0.040992	23994	25
6	ZINC04215267	113.92	0.040684	24041	25
7	ZINC04468780	113.94	0.040684	23278	25
8	ZINC04109042	113.94	0.602768	24057	25
9	ZINC04066582	114.12	0.490722	83015	25
10	ZINC95101146	114.69	0.079903	6417	25
11	ZINC40163263	120.64	0.473027	90267	25
12	ZINC38141711	120.71	0.472945	97926	25
13	ZINC03871927	120.76	0.473027	96773	25
14	ZINC13563733	121.07	0.5895844	13	25
15	ZINC28467879	121.26	0.079796	6407	25
16	ZINC38138580	127.51	0.602396	24042	25
17	ZINC72283740	128.02	0.043708	24087	25
18	ZINC65739906	128.10	0.040764	23792	25
19	ZINC72131128	128.18	0.044541	24070	25

Fig. 5.

The 2D chemical structure of ZINC compounds along with their respective pharmacophore features.

Fig. 6.

Respective ZINC compounds in the shared features pharmacophore (SFP) model obtained from virtual screening.

The hit library included 2 compounds with a score of 107.74 and 113.94 that were identified with their recorded names as cefuroxime (ZINC03871978) and cefotaxime (ZINC04468780) in the ZINCPharmer database. These two compounds were also identified as existing cephalosporin antibiotics in the DrugBank database.

3.5. Drug-likeness and physicochemical property analysis

After carrying out the virtual screening process, all the retrieved compounds were subjected to drug-likeness and physicochemical properties screening along with the selected antibiotics (training set compounds) by using SwissADME. Of the retrieved 19 compounds and 3 selected compounds of antibiotics, 7 and 2 compounds had satisfactory drug parameters according to Lipinski’s rule of five. Satisfied compounds are sorted in Table 4 along with some of the assessed physicochemical properties, lipophilicity, pharmacokinetics, and drug-likeness. With the seven selected hit compounds, further studies were implemented.

Table 4.

Physiochemical and pharmacokinetic properties of Hit compounds.

Compound	MW (g/mol)	TPSA (Å2) and Molar Refractivity (MR)	H. Bond	Lipophilicity: Log Po/w (iLOGP)	Water Solubility (Log S (ESOL))	Pharmacokinetics	Drug-likeness (Lipinski)
Cephalothin (standard)	396.44	TPSA: 166.55	Acceptor: 6	1.73	−1.65	GI absorption: Low	Yes
		MR: 98.12	Donor: 2			BBB permeant: No
Cefotaxime (standard)	455.47	TPSA: 227.05	Acceptor: 9	2.05	−1.34	GI absorption: Low	Yes
		MR: 109.90	Donor: 3			BBB permeant: No
ZINC03871978	424.39	TPSA: 199.06	Acceptor: 9	2.00	−1.90	GI absorption: Low	Yes
		MR:100.38	Donor: 3			BBB permeant: No
ZINC04468780	455.47	TPSA: 227.05	Acceptor: 9	1.75	−1.34	GI absorption: Low	Yes
		MR:109.90	Donor: 3			BBB permeant: No
ZINC21986204	424.39	TPSA: 199.06	Acceptor: 9	1.03	−1.89	GI absorption: Low	Yes
		MR: 99.99	Donor: 3			BBB permeant: No
ZINC03830485	424.39	TPSA: 100.38	Acceptor: 9	1.64	−1.90	GI absorption: Low	Yes
		MR: 199.06	Donor: 3			BBB permeant: No
ZINC65739906	486.55	TPSA: 269.14	Acceptor: 8	1.05	−1.27	GI absorption: Low	Yes
		MR: 119.11	Donor: 4			BBB permeant: No
ZINC40163263	479.49	TPSA: 244.35	Acceptor: 10	1.15	−1.98	GI absorption: Low	Yes
		MR: 116.12	Donor: 3			BBB permeant: No
ZINC04066582	454.48	TPSA: 214.16	Acceptor: 8	2.39	−2.17	GI absorption: Low	Yes
		MR: 112.10	Donor: 3			BBB permeant: No

3.6. Conjugation of screened drug-likeness hit compounds with cephalosporin core ring and chemical scaffold analysis

Following the evaluation of screening results based on drug-likeness, the next step involved integrating ZINCPharmer-derived compounds with cephalosporin drugs. Therefore, 3 established antimicrobial agents of the cephalosporin class of antibiotics, including cephalothin, cefotaxime, and ceftriaxone, were selected, as they were also utilized in pharmacophore model generation. As outlined in Section 2.7^, Alvabuilder, the software employed in this study, was utilized to automatically generate molecular models while maintaining electron balance and ensuring seamless fragment integration. This process resulted in the development of 30 novel cephalosporin antibiotic models. During model generation, the cephalosporin core ring (CC(=O)NC1C2SCC(C)=C(N2C1=O)C(O)=O) was preserved to retain its fundamental structure, while additional fragments resulting from ZINCPharmer virtual screened hits were incorporated. The entire set of these molecules, along with their corresponding scores, is provided in Table S2. Additionally, Table 5 presents the 5 leading molecules that exhibited the highest molecular resemblance and the most favorable SAScore, along with an overview of their features. Notably, the similarity score between the newly designed molecules and the reference antibiotic drugs exceeded 70 %, with values ranging from 0.783 to 0.723 (Table 5 and Table S2), indicating a substantial structural resemblance. Approximately 70 % of this similarity was attributed to the cephalosporin core ring containing the ß-lactam structure, while the remaining 30 % resulted from the incorporation of fragments derived from ZINCPharmer hit compounds.

Table 5.

The 5 leading molecules from a pool of 30 synthetically generated model molecules and their SAScore and similarity score.

Molecules no.	Molecules 2D	SMILES	Similarity score (Avg.)	SAScore
Molecule 1		CC1=C(N2C(SC1)C(NC(=O)CCCc1cccs1)C2=O)C(O)=O	0.783	4.101
Molecule 2		CC1=C(N2C(SC1)C(NC(=O)CCc1cccs1)C2=O)C(O)=O	0.781	4.111
Molecule 3		CCC1=C(N2C(SC1)C(NC(=O)CCc1cccs1)C2=O)C(O)=O	0.774	4.15
Molecule 4		CC1=C(N2C(SC1)C(NC(=O)Cc1cccs1)C2=O)C(O)=O	0.767	4.211
Molecule 5		CCCCC(=O)NC1C2SCC(=C(N2C1=O)C(O)=O)Cc1cccs1	0.764	4.186

The newly designed compounds were assessed according to their interaction characteristics and molecular stereochemical configuration. The ß-lactam-containing cephalosporin core ring was incorporated into all generated drug models. The SAScore of these new compounds served as a crucial parameter for evaluating the feasibility of their retrosynthetic pathway. The synthetic models generated by Alvabuilder exhibited an overall SAScore of less than 4.3, suggesting that these drugs are not only hypothetically achievable but also amenable to laboratory procedures, thereby facilitating their progression into the subsequent phases of the pharmaceutical development pipeline. Given the promising performance of these newly designed compounds against the default parameters, all 30 synthetic models were subjected to molecular docking analysis to identify potential drug candidates. To further analyze the chemical scaffolds, three representative compounds from the newly designed conformers were selected for comparison with one of the control drugs, Ceftriaxone, using the SMARTS.plus tool (Fig. 7). The selected compounds-Molecule 23 (Fig. 7A), Molecule 21 (Fig. 7B), and Molecule 29 (Fig. 7C) were chosen based on their significant molecular docking results, which will be elaborated upon in subsequent sections. These compounds function as a typical model for the complete collection of data, enabling a more comprehensive assessment of chemical scaffolds.

Fig. 7.

Comparative analysis of the drug structure and chemical scaffold of the newly designed compounds. (A) Molecule 23 – Ceftriaxone; (B) Molecule 21 – Ceftriaxone; (C) Molecule 29 – Ceftriaxone.

As depicted in Fig. 7, most of the newly designed drugs retained the cephalosporin core ring. However, Molecules 23 (Fig. 7A) and 29 (Fig. 7C) exhibited a distinctive ring structure comprising aromatic carbon, nitrogen, and sulfur atoms at specific positions, as indicated by the gray, blue, and yellow dots. Each compound displayed a unique chemical formula and structural arrangement. The β-lactam-containing cephalosporin core ring (Fig. 7) was the exclusive shared structural motif between the designed molecules and the control drugs, while the newly incorporated fragments exhibited no prior interactions. The application of Alvabuilder provided an algorithmic and streamlined process for compound design, ensuring the retention of the cephalosporin core while strategically introducing new fragments. This approach aimed to enhance antibiotic efficacy while minimizing the risk of resistance development.

3.7. Molecular docking analysis

Molecular docking was carried out for the ligands Ceftriaxone (PubChem ID: 5479530), Cefotaxime (PubChem ID: 5742673), and Cephalothin (PubChem ID: 6024) as control and synthetically generated potential therapeutic molecules (a set of 30 molecules) against the energy-minimized crystallographic structure of Penicillin-binding protein 1a (2V2F). The interaction with RMSD zero was considered to be the best (Table S3). Table 6 presents a ranking of the binding affinities of the respective ligand molecules in comparison to control drugs, along with their synthetic accessibility score and interaction profiles.

Table 6.

Binding affinity results of designed compounds and their interacted residues with amino acids.

Compounds name	Docking energy (kcal/mol)	SAScore	Hydrogen Bond Interaction-AA	Hydrophobic Bond Interaction-AA	Electrostatic-AA and Other
Ceftriaxone (control)	−6.2	5.06	GLN427, THR560, GLY542	TYR577, TRP411, ARG545	TRP411
Cefotaxime (control)	−6.6	4.79	GLY546, THR560	TYR577, PHE611	PHE611
Cephalothin (control)	−6.9	4.48	GLY546, THR560	TYR577, PHE611	PHE611
Molecule 9	−7.1	4.231	GLY546	TYR577, PHE611	N/A
Molecule 7	−7.2	4.209	THR560, GLY546	ARG545	N/A
Molecule 5	−7.2	4.186	GLY546, THR560	TYR577, ALA614, ALA615, PHE584, PHE611	N/A
Molecule 26	−7.2	4.318	N/A	TYR577, PHE611, LEU612	TRP411, TYR577
Molecule 22	−7.3	4.127	LYS373, GLY608, TRP411, ASN430	TRP411, PHE611	TYR577, PHE611
Molecule 29	−7.5	4.083	GLY546, THR560, THR543	TRP411, TYR577, PHE611	N/A
Molecule 21	−8.0	3.985	N/A	TRP411, LEU612	GLU582
Molecule 23	−8.6	4.213	LYS373, THR560	TRP411, LEU612	GLU582

The results of docking complexes (Fig. 8) for control drugs (ceftriaxone, cefotaxime, and cephalothin) and synthetically modeled drugs were visualized for respective protein–ligand interactions using the Discovery Studio program. The program provided a depiction of representative three-dimensional molecular interactions of protein–ligand docking complexes, as well as a two-dimensional schematic representation of the complex bond between the ligand and the receptor (Fig. 9). From the obtained result, we further thoroughly analyzed the interactions of control drugs and the three molecules that were more potent than control (most potent binding affinity score against the protein).

Fig. 8.

Three-dimensional 2V2F protein–ligand interaction. (A) Ceftriaxone (control), (B) Cefotaxime (control), (C) Cephalothin (control), (D) Molecule 5, (E) Molecule 29, (F) Molecule 23.

Fig. 9.

Two-dimensional 2V2F protein–ligand interaction. (A) Ceftriaxone (control), (B) Cefotaxime (control), (C) Cephalothin (control), (D) Molecule 5, (E) Molecule 29, (F) Molecule 23.

Ceftriaxone exhibited a docking energy of −6.2 kcal mol⁻¹. It formed four (4) hydrophobic interactions (Pi-Alkyl, Pi-Pi stacked, and Pi-Sigma) with three amino acid residues (ARG545, TRP411, and TYR577) in the active pocket of the target. Among the four (4) hydrophobic interactions, two (2) formed interactions with the residue of TRP411 at different distances of 3.87 and 4.25 Å, while the other two (2) formed interactions with amino acid residues of ARG545 and TYR577 at different distances of 5.23 and 3.53 Å, respectively. In addition, it formed four (4) H-bond interactions with THR560 (two bonds), GLY542, and GLN427 at bond distances of 3.06, 2.15, 2.67, and 2.56 Å, respectively. Other interactions include one (1) Pi-Sulfur and one (1) unfavorable donor–donor interaction with TRP411 at 3.83 Å and SER370 at 1.13 Å, as shown in Fig. 9A.

Cefotaxime exhibited a docking energy of −6.6 kcal mol⁻¹. It interacts with the active pocket of the receptor to form three (3) H-bonds with GLY546 (two bonds) and THR560 residues at bond distances of 2.37, 3.01, and 2.17 Å, respectively. Other noticeable interactions include one (1) Pi-Sulfur with PHE611 (3.72 Å) and two (2) hydrophobic interactions with amino acid residues of TYR577 and PHE611 at bond distances of 3.69 and 4.80 Å, respectively (Fig. 9B).

Cephalothin exhibited a docking energy of −6.9 kcal mol⁻¹. It formed two (2) hydrophobic interactions with two (2) amino acid residues (PHE611 and TYR577) in the active site of the target protein at respective distances (3.69 Å and 4.80 Å). Furthermore, cephalothin also formed three (3) hydrogen bonds with GLY546 (two bonds) and THR560 at 2.37, 3.01, and 2.17 Å bond distances and one (1) Pi-sulfur with PHE611 at 3.72 Å distance (Fig. 9C).

Molecule 5 revealed a docking energy of −7.2 kcal mol⁻¹. It formed two (2) hydrogen bond interactions with GLY546 at 2.22 Å and THR560 at 2.10 Å distances. It also formed five (5) hydrophobic interactions with five amino acid residues: THR577 (3.88 Å), PHE584 (5.03 Å), PHE611 (5.43 Å), ALA614 (3.72 Å), and ALA615 (3.68 Å) (Fig. 9E).

Molecule 29 revealed a docking energy of −7.5 kcal mol⁻¹. It formed four (4) hydrogen bonds with three amino acid residues: THR560 (2.49 Å), GLY546 (2.25 & 3.27 Å), and THR543 (2.56 Å). It also formed three (3) hydrophobic interactions with TRP411, TYR577, and PHE611 at 5.48 Å, 3.58 Å, and 4.57 Å distances, respectively (Fig. 9D).

Molecule 23 revealed a docking energy of −8.6 kcal.mol⁻¹. It formed five (5) hydrophobic interactions with two amino acid residues: TRP411 (4.98, 4.30, 4.00, & 4.53 Å), and LEU612 (4.29 Å). It also formed two (2) hydrogen bonds with LYS373 and THR at 2.94 Å and 3.41 Å distances, respectively. Other interactions include one electrostatic interaction with GLU582 at 4.90 Å (Fig. 9F).

Docking studies reveal that molecules 21, 23, and 29 exhibit a greater affinity compared to other modelled compounds and the control drugs, whereas molecule 5 exhibits moderate affinity. Their binding affinities, coupled with feasible synthetic accessibility, position them as promising candidates for further exploration in the design of their retrosynthesis pathways.

3.8. Molecular dynamics simulations

To validate the ligand interaction patterns and assess the protein–ligand complex integrity, molecular dynamics (MD) simulations were performed using the Desmond program. Based on the binding mode analysis from the molecular docking study, two consensus compounds, Molecule-5 and Molecule-23, along with the control compound (cephalothin), were subjected to 200 ns MD simulations. Post-simulation analyses, including root mean square deviation (RMSD), root mean square fluctuation (RMSF), protein–ligand contact (PL-contact), solvent-accessible surface area (SASA), radius of gyration (RG), and molecular surface area (MolSA), were conducted to gain deeper insights into the stability and interactions of the protein–ligand complexes.

The RMSD analysis evaluated the structural stability of the protein–ligand complexes over the simulation period. RMSD values calculated across all trajectory frames provided insights into the conformational stability of the protein. Slight variations in RMSD values, typically within the range of 1–3 Å, indicated the attainment of equilibrium, suggesting the stability of the docked complexes. The RMSD plot revealed that the protein–ligand complex remained stable, with fluctuations observed up to 90 ns. The protein–ligand (control) complex recorded the highest RMSD of 3.90 Å at 154 ns and the lowest value of 1.45 Å at 1 ns, with an average RMSD of 3.28 Å. Protein RMSD remained between 2.5 and 2.8 Å, while the ligand RMSD fluctuated between 3 and 11 Å from 85 to 140 ns, suggesting transient instability. RMSF values were calculated to determine the average atomic fluctuations for each amino acid, representing the mean positional changes of individual residues during the simulation period. RMSF values for residues SER-355, SER-472, GLY-628, SER-629, and PRO-631 exceeded 4.5 Å. No ligand interactions were observed with these residues during the simulation, suggesting that these regions were more flexible (Fig. 10A).

Fig. 10.

RMSD and RMSF plots from MD simulations. The C-α backbone (light blue) and ligand (magenta) RMSD are displayed for (A) Control, (B) Molecule 5, and (C) Molecule 23. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

For the protein-Molecule-5 complex, RMSD analysis of the C-a backbone during the 200 ns simulation revealed a highest RMSD of 3.27 Å at 188 ns and a lowest value of 1.49 Å at 1 ns, with an average RMSD of 2.85 Å. The system showed fluctuations up to 80 ns, after which it stabilized for the remainder of the simulation. Ligand RMSD ranged between 5 and 6 Å at 80 ns, indicating stability and equilibrium within the binding pocket. RMSF analysis revealed that residues GLY-628 and SER-629 exceeded 4.0 Å, suggesting flexible regions with no ligand interactions (Fig. 10B).

The protein-Molecule-23 complex exhibited an RMSD peak of 3.43 Å at 194 ns and a minimum of 1.43 Å at 1 ns, with an average RMSD of 2.93 Å. The complex demonstrated minimal fluctuations, indicating stability throughout the simulation. RMSF values for residues SER-355, GLY-628, and SER-629 were approximately 4.0 Å, suggesting these regions experienced flexibility without direct ligand interactions (Fig. 10C). The RMSD and RMSF plots confirmed the overall stability of both Molecule-5 and Molecule-23 throughout the MD trajectory. Ligand RMSD graphs showed no major fluctuations, and the RMSF data suggested that individual amino acid residues maintained their stable conformations. The RMSD values indicated that ligand binding did not induce significant conformational changes in the protein structure, demonstrating the stability of the docked complexes and the effectiveness of ligand–protein interactions.

Protein-ligand contact analysis identified key interactions, including hydrogen bonding, hydrophobic interactions, water bridges, and ionic interactions. In the protein-cephalothin (control) complex, ASP-412, GLN-427, SER-428, and GLU-528 formed stable hydrogen bonds, exhibiting interaction fractions of over 60 %, 30 %, 90 %, and 160 %, respectively. Hydrophobic interactions were observed with TYR-577 and PHE-611, with interaction fractions exceeding 100 % and 50 %, respectively. These interactions remained persistent throughout the simulation (Fig. 11A).

Fig. 11.

The protein–ligand contact fractions during the entire molecular dynamic simulation. Green bars indicate H-bonds; blue bars are for water-mediated H-bonds; purple bars are for hydrophobic interactions, and pink bars indicate ionic interactions. (A) Control, (B) Molecule 5, and (C) Molecule 23. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

For the protein-Molecule-5 complex, hydrogen bond interactions were observed at THR-543, ARG-545, SER-561, and GLU-582, with interaction fractions of 90 %, 35 %, 20 %, and 100 %, respectively. Hydrophobic interactions were noted at TYR-577 and PHE-611, with interaction fractions of approximately 10 % and 110 % (Fig. 11B). The protein-Molecule-23 complex exhibited hydrogen bonding, hydrophobic interactions, and ionic interactions. TYR-577 demonstrated a total interaction fraction of approximately 123 %, while GLU-582 exhibited a fraction of around 100 %. Other interactions with various amino acid residues were consistently detected throughout the trajectory, ensuring stable interactions (Fig. 11C).

Additional structural properties, including SASA, RG, and MoISA, were analysed to assess ligand conformational behaviour, as shown in Supplementary File 4 (Fig. S1). The analysis confirmed that both selected compounds exhibited optimal SASA, RG, and MolSA values, highlighting their suitability as prospective medicinal agents for subsequent therapeutic innovation.

3.9. Retrosynthesis analysis of significant compounds

One of the key challenges in this research was the development of a viable retrosynthetic pathway for the most significant compounds. To ensure the feasibility of the synthesis, the SAScore was carefully maintained using computational approaches. The retrosynthetic pathway for Molecule 23, formulated as “CCCC(=O)NC1C2SCC(=C(N2C1=O)C(O)=O)CCCc1nccs1,” was systematically developed using the IBM RXN for Chemistry program. This system utilizes sophisticated artificial intelligence (AI) techniques, including stepwise transformer network topology, to predict chemical reactions with high accuracy.

The synthesis process for Molecule 23 comprised multiple critical steps, incorporating key reagents such as tetrahydrofuran (THF), triphenylphosphine, water, butyryl chloride, triethylamine, dichloromethane, anisole, and trifluoroacetic acid (TFA) (Fig. 12). The initial steps involved the Staudinger reduction process, which employed THF, water, and triphenylphosphine to convert diphenylmethyl 7-azido-8-oxo-3-[3-(1,3-thiazol-2-yl)propyl]-5-thia-1-azabicyclo[4.2.0]oct-2-ene-2-carboxylate into the desired intermediate containing the cephalosporin core ring. Subsequently, Schotten-Baumann amide synthesis was carried out in the presence of butyryl chloride, triethylamine, and dichloromethane to facilitate the next structural modification. The final step involved ester hydrolysis of the Schotten-Baumann amide using anisole and trifluoroacetic acid (TFA), ultimately leading to the formation of Molecule 23.

Fig. 12.

Illustration of the sequence-to-sequence retrosynthetic pathway for the assembly of Molecule 23, detailing the required chemical reagents and the final product.

The IBM RXN program greatly improved the effectiveness and precision of retrosynthetic pathway development. By employing a probabilistic approach, it predicts reaction outcomes and explores potential routes by leveraging Monte Carlo Tree Search (MCTS), an advanced system for resolving intricate choices.

In addition to Molecule 23, the retrosynthetic pathways for the other two leading molecules, Molecule 29 and Molecule 5, were similarly developed using the IBM RXN platform. The complete reaction sequences for these molecules were documented, with detailed reaction files available as supplementary materials. The full retrosynthesis route for Molecule 23 is provided in Supplementary File 5, while Supplementary Files 6 and 7 contain the reaction details for Molecules 29 and 5, respectively. The supplemental resources offer comprehensive incremental reaction pathway information and the specific chemical inputs involved, ensuring clarity and reproducibility in synthetic planning.

Throughout the retrosynthesis process, the proposed pathways were continuously optimized according to key considerations such as ease of reaction, accessibility of reactants, and comprehensive achievability for laboratory synthesis. This incremental optimization guaranteed the ultimate synthetic pathway was both hypothetically robust and practically executable.

The integration of computational tools in this study underscores the transformative role of AI-powered techniques across modern chemical design and medicinal innovation. Through the use of advanced algorithms and cheminformatics, the retrosynthesis of Molecule 23, Molecule 29, and Molecule 5 was achieved with a high level of precision and reliability, demonstrating the potential of AI-assisted retrosynthetic strategies in accelerating drug development.

4. Discussion

In this research work, we have covered the ligand-based pharmacophore assimilation and screening of potential compounds that have analogous pharmacophore features. We identified the pharmacophore model of the cephalosporin drug family, which is the second major group of beta (β)-lactam antibiotics.⁸⁶ As a result of the acquired resistance of microorganisms and among the most effective therapeutic approaches in medicine against bacterial infection, antibiotics of these classes have drawn major concern in drug design as well as in drug discovery.7, 87, 88 Throughout the study, our main objective is to uncover the potential drug candidate that has more efficiency than the already feasible antibiotics in this family. By beginning with the preliminary analysis of both the two- and three-dimensional structures of the selected antibiotics (Cephalothin, Ceftriaxone, and Cefotaxime) as a training set compound, we performed the pharmacophore model generation process using LigandScout software. The antibiotics were retrieved from PubChem, which is an easily accessible database for the search of depository data widely, their study, and structural retrieval.³⁴ The ligand-based pharmacophore model was built to sort new compounds with similar features.³⁵ Through the spatial arrangement of drug features that offer activity in the vicinity of selected receptor molecules, a shared feature pharmacophore (SFP) model was generated. A pharmacophoric feature is elucidated as a group of atoms that are available in a molecule with a physicochemical property important for binding.⁸⁹ Particularly, a feature is a hydrogen bond acceptor and donor atom; atoms of an aromatic ring; hydrophobic atoms; or a positive and negative atom.⁹⁰ The identification of pharmacophoric features is directly involved in the optimization of the binding of the molecule of interest to a specific receptor.⁹¹ The preferred pharmacophore model was used as a template for a 3D query to search the database of ligand-based virtual screening compounds and their retrieval in ZINCPharmer to generate a compound library.⁹² An essential criterion for assessing the constructed pharmacophoric model is its capability to pinpoint prospective therapeutic agents that exhibit analogous structural and functional characteristics.⁹¹ Among the various independent model testing techniques, the goodness-of-hit (GH) score has emerged as a critical metric for confirming model reliability.⁴³ The GH score is calculated by incorporating an additional dataset of active compounds alongside the query dataset.³⁸ The validity of the pharmacophore model is determined by the number of active compounds it successfully engages. In this study, we achieved a GH score of 0.739, a noteworthy value given that an ideal GH score ranges between 0 and 1.⁹³ Furthermore, additional metrics such as enrichment factor, precision, recall, and F1 score further substantiated the robustness of our constructed pharmacophore model. The constructed structure-based feature pharmacophore (SFP) effectively interacted with 9 out of 10 active compounds, which were selected from a database library comprising 500 compounds.

To identify novel compounds, ligand-based pharmacophore virtual screening was conducted using the developed SFP model against a ligand library containing approximately 200,000 compounds. Virtual screening has become an integral part of the drug discovery process due to its efficiency in identifying new scaffolds with therapeutic potential.⁹⁴ Compared to traditional in vitro and in vivo screening methods, virtual screening offers a rapid and cost-effective approach to discovering lead compounds.²² Following the screening process, 19 compounds exhibited significant compatibility with the pharmacophore model. To refine the selection and eliminate non-viable candidates, drug-likeness screening was performed based on Lipinski's Rule of Five.⁴⁶ The primary objective of drug-likeness screening was to exclude compounds unlikely to exhibit the necessary potency against the intended drug target, while also addressing pharmacokinetics and toxicity concerns that could impact clinical trial outcomes. Ultimately, 7 compounds that adhered to Lipinski's rule were shortlisted for de novo molecular design.

Computational biology has dramatically reshaped drug development by enabling the discovery and optimization of potential drug molecules using advanced algorithms and predictive models⁹⁵. One such tool, Alvabuilder,⁴⁸ facilitates the combination of test-set antibiotics with training-set molecules using diverse parameters to generate novel synthetically modeled drug candidates. The generated models are subsequently screened to extract the most promising agents for further investigation. Our investigation yielded 30 synthetic models of cephalosporin antibiotics by incorporating the essential β-lactam-containing cephalosporin core ring⁵² with cephalosporin-based drug scaffolds obtained from ZINCpharmer. Molecular docking analysis identified eight compounds- Molecule 23, Molecule 21, Molecule 29, Molecule 22, Molecule 26, Molecule 5, Molecule 7, and Molecule 9—as exhibiting superior binding affinities compared to the control drugs (see Table 6). These findings suggest that these compounds hold significant potential as therapeutic agents.

To further analyse of the chemical properties of these compounds, Gaussian computational methods were employed to determine their stoichiometry and structural analogs (see Fig. 7).⁸¹ For comparative evaluation, Ceftriaxone, a well-established cephalosporin antibiotic, was selected as the control drug. Molecular docking studies allowed us to identify the most promising compounds, Molecule 5 and Molecule 23, which were then subjected to 200 ns molecular dynamics (MD) simulations along with the control compound, cephalothin. Key analyses, including root mean square deviation (RMSD), root mean square fluctuation (RMSF), protein–ligand contact (PL-contact), solvent-accessible surface area (SASA), radius of gyration (RG), and molecular surface area (MolSA), confirmed the stability and reliability of these selected compounds as potential drug candidates.⁹⁶.

The retrosynthetic pathway design for the most promising synthetic cephalosporin antibiotics, specifically Molecules 23, 29, and 5, highlighted the transformative role of computational retrosynthesis in modern drug development. Utilizing the IBM RXN program,⁸⁴ we rigorously developed and validated sophisticated compound synthesis pathways with a high degree of precision. This approach enabled the efficient integration of diverse chemical groups while guaranteeing laboratory implementation viability. In-silico retrosynthetic route evaluation plays a pivotal function in drug development by establishing a reliable framework for the synthesis and assessment of novel compounds. A drug candidate that lacks a feasible retrosynthesis model is unlikely to be synthesized successfully in a laboratory setting, emphasizing the importance of this analytical approach.⁹⁷ Integrating computational retrosynthesis into the drug development pipeline enhances efficiency and establishes an innovative approach to assessing the lab-based implementation of newly designed therapeutic agents in real-world applications.

As these computational techniques continue to evolve, they will undoubtedly shape the future of drug development, facilitating the invention of highly efficient and synthetically accessible therapeutic agents. The combined application of pharmacophore modelling, molecular docking, molecular dynamics simulations, and retrosynthetic analysis underscores the power of computational methodologies in accelerating drug development and optimizing lead identification strategies.

5. Conclusion

This study highlights the significant role of computational methods in drug discovery, particularly in developing new cephalosporin antibiotics. Pharmacophore evaluation is a critical component of drug design and the identification of new lead compounds. By employing a ligand-based pharmacophore model, we discovered bioactive compounds through virtual screening, de-novo molecular design, molecular docking, and retrosynthesis route design. The integration of advanced tools enabled the creation of promising candidates with improved binding affinities and favourable pharmacokinetic profiles. Computational retrosynthesis facilitated the meticulous design of synthetic routes, ensuring practical applications for the proposed drugs. Our analyses revealed that Molecule 23 and Molecule 5 emerged as strong therapeutic candidates within the cephalosporin family. However, it is essential to acknowledge the limitations of our research. While our computational findings suggest the feasibility of synthesizing these compounds, further experimental validation through wet lab techniques and in vitro and in vivo studies is crucial. These additional experiments will validate the activity and efficacy of the candidates in biological systems, enhancing global health outcomes. Ultimately, this research lays a foundation for understanding antibiotic resistance and controlling bacterial infections while showcasing the potential of in silico strategies in drug discovery.

6. Ethics, Consent to Participate, and Consent to Publish Declarations

Not applicable.

Funding statement

None. There was no funding source for the study.

CRediT authorship contribution statement

Rayhan Chowdhury: Conceptualization, Data curation, Formal analysis, Methodology, Writing – original draft. Samia Akter Saima: Data curation, Formal analysis, Methodology, Validation. Md. Al Amin: Conceptualization, Data curation, Formal analysis, Methodology, Validation. Md. Kawsar Habib: Formal analysis, Methodology, Validation. Ramisa Binti Mohiuddin: Formal analysis, Methodology, Validation. Ali Mohamod Wasaf Hasan: Formal analysis, Methodology, Validation. Roksana Khanam: Data curation, Formal analysis, Methodology, Validation, Writing – review & editing. Shahin Mahmud: Conceptualization, Formal analysis, Investigation, Methodology, Supervision, Writing – review & editing.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary material

The following are the Supplementary data to this article:

Data availability

All data included in the study will be available for everyone upon request to the authors.