Drug-likeness prioritised selection identifies anti-Escherichia coli candidates confirmed by molecular docking, dynamics simulations, and antibacterial assays

Abstract

To address the imbalance between antibacterial potency and developability in cephalosporin discovery against Escherichia coli, we developed a comprehensive screening strategy guided by the principle of maximum drug-likeness. An integrated evaluation framework was established, consisting of 33 independent predictive submodels across five dimensions: physicochemical properties, pharmacokinetics, safety, efficacy, and stability. This framework was combined with a five-fold property-spectrum scoring mechanism (S_5F) to enable multidimensional and quantitative prioritisation of candidates based on overall developability. Application of this strategy to the eMolecules library yielded 15 high-potential candidates. Experimental results showed compound M3 as the lead molecule, exhibiting notable antibacterial activity against E. coli (minimum inhibitory concentration [MIC] = 16 μg/mL). Molecular analyses further demonstrated that M3 achieved superior binding stability relative to the reference drug Cefaclor through a multimodal, high-affinity interaction network with the target protein. This strategy reduces late-stage attrition risk and provides a robust paradigm for rational antibacterial drug discovery.

1. Introduction

Escherichia coli (E. coli) is one of the most common opportunistic pathogens worldwide and poses a persistent clinical challenge in Gram-negative infections^1,2. Beta-lactam antibiotics, particularly cephalosporins, act by binding penicillin-binding proteins (PBPs, especially PBP1b) to inhibit bacterial cell wall peptidoglycan synthesis and have been central to the treatment of E. coli infections^3–6. However, widespread and occasionally inappropriate antibiotic use has driven resistance through beta-lactamase production, reduced outer-membrane permeability, and PBP mutations, diminishing the efficacy of established agents and creating an urgent need for novel scaffolds with superior developability^7–9.

Drug discovery has evolved from empirical trial-and-error to rational design and, more recently, to systematic, data-driven paradigms^10,11. In early-stage research on cephalosporins targeting E. coli, efficient lead identification strategies are decisive for discovering new beta-lactam cores or side chains^12,13. With advances in computational chemistry and artificial intelligence, high-throughput screening (HTS) has been integrated with rational drug design (RDD) and computer-aided drug design (CADD), yielding multi-tiered,efficient,and cost-effective discovery workflows^14–16. Although HTS enables large-scale, parallel testing with high throughput and minimal prior assumptions, high costs and substantial false-positive rates limit its utility^17,18. As structures of targets including E. coli PBP1b, PBP2, and beta-lactamases have been elucidated, structure-driven optimisation has become a key route to cephalosporin innovation, but it depends critically on high-resolution structures and reliable binding-mode hypotheses^19,20.

Computer-aided drug design (CADD) provides a robust technical foundation for antibacterial discovery^21–23. By combining structure-based drug design (SBDD) and ligand-based drug design (LBDD) with molecular docking, pharmacophore modelling, molecular dynamics (MD) simulations, and quantitative structure-activity relationship (QSAR) analysis, CADD enables rapid pre-screening of large libraries and markedly improves lead identification efficiency^24–29. Prior studies have characterised interaction patterns between cephalosporins and the active pockets of E. coli PBPs, elucidating structural determinants of antibacterial activity and beta-lactam ring stability^30,31. QSAR models quantify relationships between structural features and bioactivity to guide rational optimization^32–34. Molecular dynamics simulations further reveal resistance mechanisms arising from PBP mutations or porin alterations, informing anti-resistance design^35,36. Nevertheless, prevailing CADD pipelines focus primarily on ligand-target interactions and offer limited, integrated assessment of physicochemical, pharmacokinetic, and safety attributes;consequently, activity-centric evaluations often yield candidates that later fail owing to developability liabilities.

Given the high-cost, high-risk nature of drug research and development, the central methodological challenge is to increase screening efficiency while safeguarding overall candidate quality^37–40. To address the difficulty of balancing activity and developability, an innovative maximum drug-likeness concept is proposed, in which a multidimensional, quantitative framework integrates activity predictions with developability parameters to model physicochemical properties, pharmacokinetics, efficacy, safety, and stability, enabling global optimisation and prioritised ranking during virtual screening.

In this study, representative cephalosporins targeting E. coli were used as references to build an ensemble deep learning framework under the maximum drug-likeness paradigm for prioritising large compound libraries.By integrating multiple layers of investigation, including in vitro antibacterial activity assays, enzyme inhibition assay, molecular docking, and molecular dynamics simulations, high-scoring candidate molecules were systematically analysed, with the aim of discovering cephalosporin leads with novel structures and strong development potential. This work provides new candidate molecules for drug-resistant infections and methodological guidance for optimising early antibacterial screening.

2. Materials and methods

2.1. Data sources

2.1.1. Dataset1: Representative marketed cephalosporin drugs

Three commonly used and structurally representative cephalosporins active against E. coli were selected as reference compounds: Cefaclor, Cefprozil, and Cefotaxime. All three belong to the beta-lactam class of antibiotics; their primary targets are penicillin-binding proteins (PBPs), which are key enzymes in bacterial cell-wall biosynthesis. These agents are regarded as important therapeutic options for infections caused by susceptible E. coli, including those of the urinary, gastrointestinal, and reproductive tracts. Collectively, they are considered to typify the structural features and mechanistic determinants underlying cephalosporin activity^41,42. The names and chemical structures of these reference compounds are provided in Table 1.

Table 1.

Name and chemical structure of reference drugs.

No.	Name	Structure
1	Cefaclor
2	Cefprozil
3	Cefotaxime

2.1.2. Dataset2: Candidate molecular library

The screening library used in this study was obtained from eMolecules (San Diego, CA, USA). The database contains approximately 320,000 small molecules with well-defined physicochemical annotations and diverse chemotypes, thereby providing an extensive source for subsequent property-based screening.

2.2. Maximum drug-likeness (MDL)

To address the challenges of prolonged timelines, high costs, and high failure rates in drug research and development, a screening concept termed “Maximum Drug-Likeness” is advanced. The core of this concept is the systematic prediction and evaluation of 33 key properties across five dimensions—physicochemical properties, pharmacokinetics, efficacy, safety, and stability—in order to quantify the maximum multidimensional similarity of candidate compounds to the set of approved drugs. These 33 properties are listed in Table 2. The objective is to directly identify molecules that exhibit the greatest similarity to marketed drugs, thereby improving development efficiency and success rates from the outset.

Table 2.

Classification of 33 properties across five dimensions.

Category	Global No.Range	Properties (Superscript = global serial number)
Physicochemical properties	P1-P11	Aqueous solubility (LogS)1; Lipophilicity (Partition Coefficient, LogP)2; Melting point3; Boiling point4; Surface tension5; Density6; Viscosity 7; Flash point8; Vapor pressure9; Dissociation constant10; Hydrolysis (half-life value)11.
Pharmacokinetics	P12-P18	Bioavailability12; Plasma protein binding rate13; Maximal rate of metabolism14; Biliary excretion rate15; Urinary excretion rate16; Volume of distribution17; Half-life18.
Efficacy	P19-P23	Minimum Inhibitory Concentration (MIC)19; Enzyme inhibition constant(Ki)20; Receptor affinity21; Maximum effect model parameter (Emax)22; 50% effective dose(EC50)23.
Safety	P24-P30	Median lethal dose (LD50) 24; No Observed Adverse Effect Level (NOAEL)25; Tetrahymena pyriformis 50% growth inhibition concentration 26; Median lethal concentration (LC50)27; Developmental toxicity28; Ames mutagenicity29; hERG_risk30.
Stability	P31-P33	Photostability31; Chemical stability32; Thermostability33

2.3. Descriptor calculation based on maximum drug-likeness

To extract multidimensional information reflecting structural differences from the initial set of approximately 320,000 compounds, Dragon version 7.0 (Talete SRL) was employed to compute molecular descriptors for all molecules. Dragon generates numerically encoded features with explicit chemical interpretation from two-dimensional structures and is among the most widely used descriptor platforms for quantitative modelling.

After standardisation in ChemDraw, initial structures for all compounds were exported as Structure Data File (SDF) files. A dataset containing the three marketed cephalosporins (Dataset 1) and a dataset comprising the full set of approximately 320,000 compounds (Dataset 2) were jointly imported into Dragon version 7.0 for molecular characterisation. A uniform set of two-dimensional (2D) descriptors, independent of three-dimensional conformations, was used. These included property descriptors, functional-group counts, topological indices, and matrix-based families such as BCUT, GETAWAY, WHIM, and Burden eigenvalue descriptors, as well as descriptors for ring systems, branching indices, and topological polar surface area. In total, approximately 3,655 2D descriptors were generated.

These descriptors serve as mathematical and topological representations of molecular structure. By quantifying composition, connectivity, and spatial distribution features, they provide foundational data for linking structural characteristics to multidimensional developability attributes, including physicochemical properties, pharmacokinetics, efficacy, safety, and stability. Physicochemical descriptors directly reflect fundamental attributes such as molecular volume, size, lipophilicity, and polarity, which determine baseline properties including solubility and absorption potential⁴³. Pharmacokinetic-related descriptors capture features pertinent to absorption, distribution, metabolism, and excretion (ADME), thereby influencing systemic exposure⁴⁴. Efficacy-related topological and charge descriptors characterise potential modes of interaction with biological targets; polarity indices and polar surface area measures, which have been associated with minimum inhibitory concentration (MIC), are indicative of pharmacological potency⁴⁵. Safety descriptors correlate with potential toxicity liabilities and are used to flag structural motifs associated with risk⁴⁶. Stability descriptors, by evaluating bond types and strengths, molecular rigidity, and the presence of reactive functional groups, support the prediction of chemical and metabolic stability⁴⁷.

2.4. Multi-Criteria screening model based on maximum drug-likeness

Guided by the core concept of maximum drug-likeness, a prediction framework leveraging an ensemble of deep neural networks with transfer learning was developed. The ensemble comprises 33 independent deep neural networks, each dedicated to one key property. The workflow is depicted in Figure 1. The objective is not to optimise any single metric in isolation, but to maximise the concordance of a candidate molecule with the overall distribution of approved cephalosporin drugs in a multidimensional property space—that is, to minimise the candidate’s aggregate deviation from the statistical reference space defined by real drugs.

Figure 1.

Schematic diagram of the MDL-based virtual screening process.

2.4.1. Supervised pretraining of the screening model

Thirty-three task-specific deep neural networks were constructed, each corresponding to a single property prediction. Supervised pretraining was performed on experimentally determined data from large public repositories (ChEMBL and PubChem) to ensure property-specific specialisation and predictive accuracy. For each property, data were randomly partitioned into training, validation, and test sets in proportions of approximately 7:1.5:1.5. All networks shared a common architecture: three fully connected hidden layers with 2,048, 1,024, and 128 neurons, respectively; rectified linear unit (ReLU) activations in the hidden layers; and regularisation via batch normalisation and dropout (dropout rate 0.2) to enhance generalisation. Output layers were matched to task type: a single linear neuron for regression, and a sigmoid-activated neuron for binary classification. Mean squared error was used as the loss function for regression, and binary cross-entropy was used for classification. The generic architecture of a single property-prediction submodel is illustrated in Figure 2.

Figure 2.

General network architecture for predictive sub models.

Regression (mean squared error, MSE):

$$\text{MSE}=\frac{1}{N_{\mathit{\text{batch}}}}{\displaystyle \sum }_{i=1}^{N_{\mathit{\text{batch}}}}{\left(y_i-{\widehat{y}}_i\right)}^2$$ (1)

where $N_{\mathit{\text{batc}}\mathrm{h}}$ denotes the batch size, $y_i$ is the reference value, ${\widehat{y}}_i$ is the predicted value.

Binary classification (binary cross-entropy,BCE):

$$\mathit{\text{BCE}}=-\frac{1}{N_{\mathit{\text{batch}}}}{\displaystyle \sum }_{i=1}^{N_{\mathit{\text{batch}}}}\left[y_i\text{log\hspace{0.17em}}\left({\widehat{y}}_i\right)+\left(1-y_i\right)\text{log\hspace{0.17em}}\left(1-{\widehat{y}}_i\right)\right]$$ (2)

where y_i is the reference label, ranging in $\left\{\text{0,1}\right\}$, ${\widehat{y}}_i$ is the predicted probability, ranging in $\left[\text{0,1}\right]$.

2.4.2. Quantification and normalization of molecular properties

To quantify, within a common mathematical space, the overall similarity between a candidate molecule and the reference drugs, a five-domain property spectrum was constructed for each molecule from the outputs of the 33 trained submodels. This spectrum, spanning physicochemical properties, pharmacokinetics, efficacy, safety, and stability, was represented as a 33-dimensional vector $\mathrm{V}\mathrm{\in }{\mathrm{R}}^{33}$.

Because physical units and numerical scales vary substantially across properties, direct distance calculations would otherwise be dominated by properties with larger magnitudes. A unified normalisation strategy was therefore applied to continuous regression outputs. For bioactivity measures spanning several orders of magnitude (e.g. minimum inhibitory concentration and median lethal dose), log-scale transformations were applied during model training. At prediction time, each continuous property was normalised via min-max scaling using training-set statistics, together with a clipping mechanism to avoid extrapolation beyond the chemical space covered by the training data (out-of-distribution control). Define:

$$v_k=\text{clip}\left(\frac{y_k-\mathrm{\text{min}}\hspace{0.17em}\left(Y_{\mathit{\text{train}}}^k\right)}{\mathrm{\text{max}}\hspace{0.17em}\left(Y_{\mathit{\text{train}}}^k\right)-\mathrm{\text{min}}\hspace{0.17em}\left(Y_{\mathit{\text{train}}}^k\right)},\mathrm{ }0,\mathrm{ }1\right)$$ (3)

where $y_k$ is the raw prediction for property k, $Y_{\mathit{\text{train}}}^k$ is the set of reference values for property k, $\text{clip}\left(x\text{,0,1}\right)\text{=}\mathrm{\text{max}}\hspace{0.17em}\left(\text{0,}\mathrm{\text{min}}\hspace{0.17em}\left(x\text{,1}\right)\right)$. This ensures that each normalised feature $v_k$ lies within [0, 1].

This ensures that each normalised feature v_k lies within [0, 1]. For classification properties (e.g. risk of human ether-a-go-go related gene [hERG] channel–mediated cardiotoxicity and chemical stability), the positive-class probability from the model output was taken directly as the corresponding vector component, which naturally lies in [0, 1] and preserves predictive uncertainty and confidence. Through this normalisation and probabilistic treatment, every molecule’s property spectrum was mapped to the 33-dimensional unit hypercube.

2.4.3. Construction of property spectrum and similarity measurement

Given that each dimension has explicit biological meaning and its absolute magnitude reflects property strength, similarity indices that depend only on directional concordance (such as cosine similarity) may underweight meaningful numerical differences. A Euclidean-distance-based measure was therefore adopted to define the composite similarity score S_5F.

$$S_{5F}\left(V_{\mathit{\text{cand}}},V_{\mathit{\text{ref}}}\right)=1-\frac{{‖V_{\mathit{\text{cand}}}-V_{\mathit{\text{ref}}}‖}_2}{\sqrt{N_{\mathit{\text{prop}}}}}$$ (4)

where $V_{\mathit{\text{cand}}}$ is the normalised property vector of the candidate molecule, $V_{\mathit{\text{ref}}}$ is the normalised property vector of the reference drug, $N_{\mathit{\text{prop}}}\text{=33,\hspace{0.17em}}{‖\cdot ‖}_2$ is the Euclidean distance, $\sqrt{N_{\mathit{\text{prop}}}}$ is the maximum theoretical distance within the unit space.

The value of S_5F approaches 1 if and only if the candidate molecule closely matches the reference drug across all 33 property dimensions. The final composite score for a candidate was defined as the maximum S_5F obtained across all reference drugs, thereby capturing the candidate’s greatest overall property similarity to any known effective drug. The screening approach based on maximal drug-likeness was implemented on the commercial virtual screening platform 5FMDL Screener (Real-Drug Technology Co., Ltd., Shanghai, China).

2.5. Comprehensive evaluation strategy

Based on the computed S_5F composite similarity scores, candidate molecules were ranked in descending order, and the top 15 were selected. To systematically assess the drug potential of these high-scoring candidates identified by the maximum drug-likeness strategy, a multi-level, complementary, and integrated evaluation strategy was employed. Under this strategy, in vitro antibacterial assays were performed to determine the inhibitory activity and minimum inhibitory concentrations of the candidate compounds, followed by enzyme inhibition experiments to explore potential molecular targets and modes of action. Molecular docking and molecular dynamics simulations were subsequently conducted to provide structural insights into the experimental observations.

2.5.1. In vitro antibacterial activity assay

For microbiological testing, the reference strain E. coli ATCC 25922 (China General Microbiological Culture Collection Centre, CGMCC, Beijing, China) was used. Bacteria were cultured in Luria–Bertani (LB) broth (Guangdong Huankai Biotechnology Co., Ltd., Guangzhou, China) at 37 °C and 150 rpm to the mid-logarithmic phase (optical density at 600 nm [OD₆₀₀] approximately 0.4–0.6), and then adjusted with sterile normal saline to a 0.5 McFarland standard.

Mueller–Hinton (MH) agar plates (Guangdong Huankai Biotechnology Co., Ltd., Guangzhou, China) were allowed to dry at room temperature, and a sterile swab was used to evenly seed the bacterial suspension across the plate surface to form a uniform lawn. Test samples were dissolved in dimethyl sulfoxide (DMSO; <1% v/v) (Beijing Solarbio Science & Technology Co., Ltd., Beijing, China) and applied to sterile paper discs at a per-disc loading of 30 µg. The discs were air-dried and placed on the agar surface. A blank control (DMSO only) and a positive control (Cefaclor; Beijing Solarbio Science & Technology Co., Ltd., Beijing, China) were included. Plates were incubated at 37 °C for 22 h, and antibacterial activity was evaluated by measuring the zones of inhibition. All experiments were performed independently in triplicate and conducted in accordance with the Clinical and Laboratory Standards Institute (CLSI) M02 guideline for disc diffusion, including quality control procedures.

2.5.2. Determination of MIC

Compounds exhibiting favourable inhibition zones were advanced to MIC determination by the broth microdilution method in accordance with the CLSI M07-A10 guideline. E. coli ATCC 25922 was adjusted to a 0.5 McFarland standard in saline and further diluted 1:150 in MH broth (Guangdong Huankai Biotechnology Co., Ltd., Guangzhou, China) to yield an inoculum of approximately 5 × 10^{^5} colony-forming units (CFU) per mL. In microtiter plates, twofold serial dilutions of each compound in MH broth were prepared to generate a concentration range from 64 µg/mL to 0.5 µg/mL, and positive growth-control wells contained inoculum and MH broth only. Plates were sealed and incubated statically at 37 °C for 20 h. After incubation, wells were inspected visually for turbidity, and the MIC was recorded as the lowest concentration that completely inhibited visible bacterial growth. All tests were performed independently in triplicate.

2.5.3. Enzyme inhibition assay

Penicillin-binding protein 1b (PBP1b) is a key molecular target involved in the cross-linking of bacterial cell wall peptidoglycan. To elucidate one of the potential mechanisms underlying the antibacterial activity of the candidate cephalosporin compounds, a colorimetric in vitro enzyme inhibition assay was employed to evaluate their inhibitory effects against recombinant E.coli PBP1b (TargetMol Chemicals Inc., Shanghai, China). All reactions were performed in 96-well plates with a total volume of 200 μL. The reaction buffer consisted of 50 mM Tris–HCl (pH7.5) supplemented with 10 mM MgCl₂(Beijing Solarbio Science & Technology Co., Ltd., Beijing, China). Recombinant PBP1b was used at a final concentration of 0.2 μM. p-Nitrophenyl (Shanghai Aladdin Biochemical Technology Co., Ltd., Shanghai, China)acetate was employed as the chromogenic substrate at a final concentration of 1.0 mM. Hydrolysis of p-nitrophenyl acetate catalysed by PBP1b generates p-nitrophenol, which exhibits a characteristic absorbance at 405 nm. The three antibacterial candidate compounds, together with the reference drug Cefaclor, were prepared as 10 mM stock solutions in dimethyl sulfoxide and tested at final concentrations of 1, 5, 10, 20, 50, 100 and 200 μM. The final concentration of dimethyl sulfoxide in the reaction system did not exceed 1% (v/v), a level that had no detectable effect on enzyme activity. Control reactions contained an equivalent volume of dimethyl sulfoxide without inhibitors. Following incubation at 37 °C for 30 min, absorbance at 405 nm was measured using a microplate reader(Molecular Devices,USA). The percentage of enzyme inhibition was calculated according to the following equation:

$$\mathit{\text{In}}\text{h}\mathit{\text{ibition}}\text{\hspace{0.17em}(\%)\hspace{0.17em}=\hspace{0.17em}}\left(1\hspace{0.17em}-\hspace{0.17em}\frac{A_{\mathit{\text{sample}}}\hspace{0.17em}-\hspace{0.17em}{A}_{\mathit{\text{blank}}}}{A_{\mathit{\text{control}}}\hspace{0.17em}-\hspace{0.17em}{A}_{\mathit{\text{blank}}}}\right)\hspace{0.17em}\times \text{\hspace{0.17em}100\%}$$ (5)

where A_sample represents the absorbance of wells containing the test compound, A_control represents the absorbance of wells without inhibitor, and A_blank corresponds to wells lacking enzyme.

All experiments were performed independently in triplicate. Dose–response curves were fitted using nonlinear regression analysis, and the half-maximal inhibitory concentration (IC₅₀) values were determined.

2.5.4. Molecular docking

Molecular docking—a core technique in structure-based drug design used to predict ligand binding modes and affinities to target proteins—was then conducted to assess whether candidates could enter the active site, compare relative binding strengths, and identify key interacting residues^48–50.

PBP1b, a principal target whose inhibition disrupts peptidoglycan cross-linking in the bacterial cell wall, was selected as the docking target (Protein Data Bank [PDB] ID: 3FWL) ⁵¹. The high-resolution crystal structure was obtained from the Research Collaboratory for Structural Bioinformatics Protein Data Bank (RCSB PDB), converted to PDBQT format, and is shown in Figure 1. The molecular structures of the top 15 candidates and the three reference compounds were built in ChemDraw Professional 15.0 and saved as Structure Data File (SDF) format.

Blind docking was performed with the CB-Dock2 platform⁵², which identified five potential binding cavities with volumes ranging from 378 to 3751 Å^3, each exhibiting distinct geometric features and interaction potential. Docking outputs were downloaded as text files, and three-dimensional visualisation and interaction analysis of protein–ligand complexes (PDB format) were performed in PyMOL 3.1.0 and BIOVIA Discovery Studio 2021.

2.5.5. Molecular dynamics simulation

Molecular dynamics simulation, a computational approach based on Newtonian mechanics, was employed to assess the stability of protein–ligand complexes under near-physiological conditions and to characterise the dynamic features of intermolecular interactions⁵³–55.Simulations were conducted using GROMACS 2022 for the most active complex identified in in vitro assays (M2–PBP1b) and for the reference complex (Cefaclor–PBP1b).

Ligand topologies and force-field parameters were generated with the AnteChamber PYthon Parser interface (ACPYPE) using the General AMBER Force Field 2 (GAFF2), ensuring compatibility with the AMBER ff14SB protein force field, which was used to describe the protein. Each system was placed in a cubic periodic box with at least 1.2 nm between the protein surface and the box boundary, solvated with the TIP3P water model, neutralised with Na + and Cl − ions, and adjusted to a salt concentration of 0.15 M.

The simulation protocol comprised energy minimisation and equilibration stages, followed by production dynamics. Steepest-descent energy minimisation was performed for 5000 steps to remove unfavourable contacts. Equilibration proceeded under a constant-number, constant-volume, constant-temperature (NVT) ensemble for 100 ps using a velocity-rescaling thermostat (V-rescale) to raise the temperature to 310 K, followed by a constant-number, constant-pressure, constant-temperature (NPT) ensemble for 200 ps using a Parrinello–Rahman barostat to maintain 1 bar, allowing temperature and density to stabilise. Production runs were conducted for 100 ns under NPT conditions with a 2 fs integration time step. All covalent bond lengths were constrained with the Linear Constraint Solver (LINCS). A 1.2 nm cut-off was applied to van der Waals and short-range electrostatic interactions, and long-range electrostatics were treated with the particle mesh Ewald (PME) method. Atomic coordinates were saved every 10 ps to balance file size with adequate sampling of conformational dynamics.

3. Results and analysis

3.1. Performance evaluation of property-prediction submodels

To evaluate the reliability of the deep learning–based screening framework, the performance of all 33 submodels was assessed on independent test sets. Among the 27 regression submodels, most achieved satisfactory predictive accuracy, with coefficients of determination (R²) between 0.85 and 0.91. The six classification submodels exhibited robust discriminative ability, with areas under the receiver operating characteristic curve (AUC) all exceeding 0.85. Similar performance trends were observed across the training, validation, and test sets, indicating the absence of overfitting and demonstrating good generalisation. Test‑set statistics are summarised in Table 3.

Table 3.

Predictive performance evaluation of the 33 property-specific submodels on the independent test set.

Number	Task type	Evaluation metric	Performance
P1	Regression	R2	0.89
P2	Regression	R2	0.88
P3	Regression	R2	0.86
P4	Regression	R2	0.91
P5	Regression	R2	0.90
P6	Regression	R2	0.89
P7	Regression	R2	0.87
P8	Regression	R2	0.90
P9	Regression	R2	0.88
P10	Regression	R2	0.86
P11	Regression	R2	0.85
P12	Regression	R2	0.86
P13	Regression	R2	0.88
P14	Regression	R2	0.89
P15	Regression	R2	0.86
P16	Regression	R2	0.86
P17	Regression	R2	0.87
P18	Regression	R2	0.89
P19	Regression	R2	0.90
P20	Regression	R2	0.91
P21	Regression	R2	0.89
P22	Regression	R2	0.85
P23	Regression	R2	0.89
P24	Regression	R2	0.87
P25	Regression	R2	0.90
P26	Regression	R2	0.89
P27	Regression	R2	0.91
P28	Classification	AUC	0.90
P29	Classification	AUC	0.89
P30	Classification	AUC	0.90
P31	Classification	AUC	0.91
P32	Classification	AUC	0.92
P33	Classification	AUC	0.90

3.2. Consistency of True - Drug ranking

Three reference drugs were spiked into a library of 320,000 candidate compounds, and the entire collection was scored and ranked using the S_5F composite similarity metric. The ranks and percentile ranges of the reference drugs were recorded. All three consistently fell within the top 1% of the ranking (Table 4), indicating that the scoring strategy effectively captures drug‑like characteristics. In large‑scale libraries, bona fide drugs were accurately identified and prioritised, thereby supporting the reliability and specificity of the scoring scheme.

Table 4.

Ranking positions of reference drugs within the S_5F framework.

Molecule	S5F Ranking	Percentile
Cefaclor	128	Top 0.04%
Cefprozil	65	Top 0.02%
Cefotaxime	155	Top 0.05%

3.3. Analysis of property spectra for candidate molecules

Using the trained submodels, Cefaclor, Cefprozil, and Cefotaxime were selected as reference drugs, and property spectra covering 33 key properties were constructed for these references and for the screened candidates. Normalised property scores were converted into property fingerprint spectra (Figure 3), providing an intuitive depiction of property matching between candidates and references. In panels A–C, the x‑axis denotes indices of the 33 core properties and the y‑axis the normalised property scores. Red curves represent the property profiles of the reference cephalosporins, whereas distinct blue curves correspond to individual candidates. The candidate curves closely track their respective references, with small inter‑profile deviations. This visualisation indicates that the candidates do not merely match a single physicochemical or biological metric; rather, they exhibit high similarity to the reference drugs across multiple dimensions, including physicochemical properties and pharmacokinetic attributes, thereby supporting the effectiveness of the screening method.

Figure 3.

Panoramic comparison of property spectra between three reference drugs and their corresponding top five candidate molecules in the 33-dimensional property space. (A) Cefaclor; (B) Cefprozil; (C) Cefotaxime.

3.4. Virtual screening results

The maximum drug‑likeness–based composite screening model was applied to score and rank 320,000 candidates across five property domains. Based on descending composite scores, the top 15 candidates were selected, their molecular structures and scores are reported in Table 5. All selected molecules achieved five‑domain composite scores greater than 0.94, indicating high concordance with true drugs across the five dimensions. The efficiency of screening candidate molecules for drug-likeness was substantially improved, thereby narrowing the scope of subsequent experimental investigations and providing a high-quality set of candidates for further systematic evaluation of their druggability from both molecular feature and biological activity perspectives.

Table 5.

Top 15 candidate molecules and their scores.

No.	Specs ID	Structure	S5F	Nearest drug
M1	AS-871/42057286		0.961	Cefprozil
M2	AS-871/42057269		0.960	Cefotaxime
M3	AO-079/14332007		0.960	Cefaclor
M4	AF-399/07629014		0.958	Cefotaxime
M5	AG-205/07689035		0.957	Cefaclor
M6	AE-848/08814049		0.957	Cefprozil
M7	AO-022/43454074		0.956	Cefaclor
M8	AF-399/14662028		0.954	Cefotaxime
M9	AK-968/11197020		0.951	Cefprozil
M10	AN-989/14323005		0.950	Cefaclor
M11	AB-323/13887102		0.948	Cefotaxime
M12	AG-690/11096058		0.947	Cefprozil
M13	AS-871/42056970		0.945	Cefaclor
M14	AG-690/12249326		0.944	Cefotaxime
M15	AJ-292/14408025		0.941	Cefprozil

3.5. Antibacterial activity by disk diffusion

In disc diffusion assays against E. coli ATCC 25922, 15 candidate small molecules and three approved cephalosporin antibiotics were evaluated. The blank control (dimethyl sulfoxide ≤1%) produced no zones of inhibition, indicating no solvent effect on bacterial growth, whereas the positive control (Cefaclor) yielded clear, regular zones, confirming assay robustness. Most candidates showed no obvious antibacterial activity; however, molecules M3, M5, and M7 produced well‑defined zones of inhibition under the assay conditions, indicating inhibition of E.coli ATCC 25922.

The inhibition‑zone diameter of M3 was comparable to that of the cephalosporin control (Figure 4). At a per‑disc loading of 30 µg, M3 produced an average zone of 20.2 mm, whereas Cefaclor produced 23.1 mm, representing the strongest activity. M7 exhibited moderate activity (18.1 mm), and M5 showed weaker activity (14.6 mm), revealing a graded difference in antibacterial potency among the candidates.

Figure 4.

In vitro antibacterial activity of selected candidate molecules against E. coli.

3.6. Minimum inhibitory concentration (MIC) results

To quantify antibacterial potency, the three candidates that exhibited inhibition zones in the disc diffusion assay were subjected to broth microdilution testing to determine minimum inhibitory concentrations (Table 6).

Table 6.

MIC results of candidate molecules against E. coli.

Compd. no	MIC
M3	16µg/mL
M5	32µg/mL
M7	32µg/mL
Cefaclor	4µg/mL

Molecule M3 demonstrated the greatest potency, with an MIC of 16 µg/mL; M5 and M7 each had MICs of 32 µg/mL; the control drug Cefaclor had an MIC of 4 µg/mL. These quantitative results correlated well with the disc diffusion data: larger inhibition zones corresponded to lower MICs and stronger activity. The findings confirm that M3 exhibits the best antibacterial efficacy among the candidates. Although M3 did not surpass Cefaclor, it was markedly more potent than the other candidates and is therefore the most promising antibacterial molecule identified in this screening.

3.7. Enzyme inhibition results

After establishing the in vitro antibacterial activity of the candidate compounds against E.coli through inhibition zone assays and minimum inhibitory concentration determinations, their inhibitory effects on recombinant PBP1b were further evaluated to gain mechanistic insight. As shown in Figure 5, the positive control Cefaclor, as well as all lead compounds, exhibited clear and dose-dependent inhibition of PBP1b activity, thereby confirming the reliability of the enzymatic assay system.

Figure 5.

Inhibition rate of compounds at different concentrations.

Notably, a strong positive correlation was observed between enzymatic inhibition potency and antibacterial activity (Table 7). The lead compound M3, which displayed the strongest antibacterial activity, also showed the most potent inhibition of PBP1b, with an IC₅₀ value (23.16 μM) comparable to that of Cefaclor (14.32 μM). In contrast, compounds M3 and M7, which exhibited relatively weaker antibacterial activity, showed progressively higher IC₅₀ values of 42.77 μM and 58.30 μM, respectively, indicating reduced inhibitory potency.

Table 7.

IC₅₀ results of candidate molecules.

Compd. no	IC50
M3	23.16 μM
M5	42.77 μM
M7	58.30 μM
Cefaclor	14.32 μM

This highly consistent trend strongly suggests that inhibition of the transpeptidase activity of PBP1b represents one of the core molecular mechanisms underlying the antibacterial effects of these lead compounds. The enzyme inhibition results provide direct biochemical evidence supporting the observed antibacterial phenotypes. Based on this experimentally supported target engagement, subsequent molecular docking and molecular dynamics simulations are expected to more reliably and specifically elucidate the precise interaction patterns between the lead compounds—particularly compound M3—and the active site of PBP1b, thereby offering critical structural insights into their mechanisms of action.

3.8. Molecular docking results and interaction analysis

Molecular docking was conducted for the top 15 candidate molecules together with three approved reference drugs to evaluate their potential affinity for PBP1b. For nearly all of the 18 molecules, the most favourable (lowest) docking scores were obtained in the cavity with a volume of 3610 ų (cavity 1); docking results are summarised in Table 8. This observation suggests that this cavity is likely the active site or the principal binding pocket of PBP1b, whose geometry and physicochemical environment appear to provide optimal positioning for cephalosporin engagement and acylation, consistent with covalent inhibition of the protein.

Table 8.

Docking scores of 3 genuine drugs and TOP15 candidate molecules in PBP1b.

Cavity Volumes
3610Å3		1131Å3		3751Å3		1734 Å3		378Å3
Sub.	kcal/mol	Sub.	kcal/mol	Sub.	kcal/mol	Sub.	kcal/mol	Sub.	kcal/mol
Cefaclor	−8.5	Cefaclor	−8.2	Cefaclor	−7.8	Cefaclor	−7.7	Cefaclor	−7.7
Cefprozil	−8.4	Cefprozil	−8.4	Cefprozil	−7.6	Cefprozil	−7.0	Cefprozil	−6.7
Cefotaxime	−8.3	Cefotaxime	−7.4	Cefotaxime	−7.0	Cefotaxime	−6.4	Cefotaxime	−6.2
M1	−9.2	M1	−8.0	M1	−7.3	M1	−8.5	M1	−7.5
M2	−8.8	M2	−7.7	M2	−8.1	M2	−7.2	M2	−6.9
M3	−10.1	M3	−9.3	M3	−8.7	M3	−8.0	M3	−7.6
M4	−9.0	M4	−8.3	M4	−7.8	M4	−8.2	M4	−6.6
M5	−10.3	M5	−9.0	M5	−7.6	M5	−7.9	M5	−7.6
M6	−8.9	M6	−7.8	M6	−7.6	M6	−7.1	M6	−7.6
M7	−10.0	M7	−8.2	M7	−7.9	M7	−8.8	M7	−7.2
M8	−9.6	M8	−7.9	M8	−8.2	M8	−6.8	M8	−7.5
M9	−9.0	M9	−8.2	M9	−7.7	M9	−7.9	M9	−6.7
M10	−9.5	M10	−8.9	M10	−7.8	M10	−7.0	M10	−6.7
M11	−9.0	M11	−8.3	M11	−7.6	M11	−8.0	M11	−6.9
M12	−8.9	M12	−8.2	M12	−7.9	M12	−8.1	M12	−6.7
M13	−9.1	M13	−8.5	M13	−7.9	M13	−7.4	M13	−7.7
M14	−8.9	M14	−8.0	M14	−8.2	M14	−7.0	M14	−6.9
M15	−9.3	M15	−8.0	M15	−7.9	M15	−7.1	M15	−7.3

Across the set, the mean docking score for the top 15 candidates (−9.3 kcal/mol) was more favourable than that for the three approved drugs (−8.4 kcal/mol), indicating that the maximum drug‑likeness–based composite framework is advantageous for identifying molecules with potentially improved binding. Per‑compound docking scores for the reference drugs and the 15 candidates are provided in Table 8. The highest‑scoring candidate, M5 (−10.2 kcal/mol), outperformed the best‑scoring reference drug, Cefaclor (−8.5 kcal/mol), in docking; however, it did not exhibit superior antibacterial activity in disc diffusion assays. This discrepancy indicates that target binding is only one determinant of antibacterial efficacy; suboptimal transmembrane transport or non-productive, peripheral binding that does not perturb catalysis may underlie the weak activity of M5.

Protein–ligand interactions provide the principal driving forces for complex stabilisation and for binding specificity and affinity. In this study, both shared and distinct interaction features were observed for the reference drug and the candidates (Figure 6). Hydrogen bonding—the most prevalent interaction in ligand–target recognition—was observed for Cefaclor and for all screened candidates. For Cefaclor, conventional hydrogen bonds (e.g. with GLN159) and hydrophobic contacts with TYR280 predominated. The number of contact residues was moderate, supporting stable occupancy of the pocket and blockade of protein function, and serving as a benchmark for comparison.

Figure 6.

Molecular docking binding modes of the reference drug and top candidate molecules at the PBP1b target: (A) Cefaclor; (B) M3; (C) M5; (D) M7.

Among the candidates, M3—the most active molecule in vitro—formed a multimodal, high‑affinity interaction network with PBP1b. Conventional hydrogen bonds with LYS112 and ASP107 provided strong polar stabilisation of the bound conformation. An attractive electrostatic interaction with ASP107 further strengthened binding, and multiple hydrophobic contacts enhanced complex stability. The carbon chain and heteroaromatic scaffold of M3 exhibited close topological complementarity to the active‑site pocket, collectively consistent with its favourable antibacterial activity.

By contrast, M5—the best‑scoring molecule by docking—formed hydrogen bonds primarily with noncatalytic residues such as PHE162 and TYR54. Although a cooperative pattern comprising a hydrogen‑bond network, a hydrophobic core, and charge anchoring was present, the focus on peripheral residues would not directly interfere with the catalytic machinery of PBP1b, resulting in weaker inhibition than M3. For M7, the binding mode displayed foundational antibacterial features: conventional hydrogen bonds afforded site anchoring, hydrophobic contacts supported embedding within a hydrophobic pocket, and an attractive electrostatic interaction with ASP107 was present. However, the absence of key functional contacts, a disrupted cooperative interaction network, and suboptimal shape complementarity likely prevented effective translation into antibacterial efficacy, as reflected by zones of inhibition markedly smaller than those of the reference drug.

3.9. Molecular dynamics simulation results

To investigate binding stability and conformational behaviour for the optimal candidate, a 100 ns molecular dynamics simulation was performed for the M3–PBP1b complex, with the cefaclor–PBP1b complex used as a comparator. Root‑mean‑square deviation (RMSD) and root‑mean‑square fluctuation (RMSF) analyses were conducted to evaluate global and residue‑level dynamics, respectively. Simulation trajectories are shown in Figure 7.

Figure 7.

Molecular dynamics simulation trajectories of the Cefaclor–PBP1b and M3–PBP1b complexes: (A) RMSD; (B) RMSF.

To assess binding stability at atomic resolution, conformational changes were compared over the 100 ns simulations. RMSD analysis of protein Cα atoms indicated that both systems reached equilibrium at approximately 20 ns (Figure 7(A)), supporting the plausibility of the initial docked poses. During the equilibrated phase (20–100 ns), the mean RMSD for M3–PBP1b was (2.8 ± 0.3) Å, significantly lower than that for Cefaclor–PBP1b at (3.4 ± 0.3) Å. Over the full trajectory, the M3 complex exhibited narrower fluctuations (2.6–3.2 Å), whereas the reference complex fluctuated more widely (3.0–3.8 Å). These results indicate higher conformational stability for the M3–PBP1b complex and suggest stronger effective binding.

Root mean square fluctuation (RMSF) analysis showed that the active site residues of PBP1b in the M3–PBP1b complex had a significantly lower mean RMSF than those in the Cefaclor–PBP1b complex. This suggests that M3 binding specifically restricts the flexibility of the active site, thereby increasing its conformational rigidity and stability. In contrast, Cefaclor binding resulted in higher residue fluctuations, indicating a weaker constraint on the active site conformation. No significant difference in RMSF was observed in non-active regions between the two complexes, confirming that M3 selectively stabilises the active site without imposing non-specific rigidity on the overall protein structure.Moreover, key interacting residues in the M3–PBP1b complex(e.g. LYS 112 and ASP 107)exhibited low RMSF values (1.0–1.5 Å). This supports the conclusion that specific interactions, including hydrogen bonds and hydrophobic contacts, effectively restrain these residues and contribute to the stability of the complex.

Taken together, the molecular dynamics simulations indicate that M3 matches or exceeds the reference drug with respect to global stability (lower RMSD) and local binding strength (lower RMSF). The trends in dynamic stability observed at the molecular level were found to be in good agreement with both the activity levels exhibited in the in vitro antibacterial assays and the inhibitory potency determined from the in vitro enzyme inhibition experiments, thereby providing supportive structural and dynamical evidence for understanding the antibacterial behaviour and underlying mechanism of action of the compounds. M3 produced a clear inhibition zone with a diameter of 21.2 mm and a minimum inhibitory concentration of 16 µg/mL, comparable to Cefaclor, which yielded a 23.1 mm zone and a 4 µg/mL minimum inhibitory concentration. From a computational perspective, these results strongly support M3 as the most active molecule identified in this virtual screen: it not only engages the target effectively but also forms a stable, persistent complex, indicating substantial potential as a lead compound.

4. Conclusion

A longstanding imbalance between antibacterial activity and developability in the screening of cephalosporin candidates against E. coli was addressed through the proposal and implementation of an innovative “maximum drug-likeness” screening paradigm. The traditional activity-centric screening approach was redesigned to yield a systematic, composite evaluation framework integrating five dimensions—physicochemical properties, pharmacokinetics, safety, efficacy, and stability—across 33 key properties. On this basis, an ensemble deep learning model comprising 33 independent, rigorously validated submodels was established. Abstract property similarity was rendered as visual property spectra and embedded in a 33-dimensional space. A Euclidean distance–based similarity scoring scheme was introduced, with the S_5F composite score adopted as the primary metric, thereby enabling multidimensional quantitative assessment and integrated ranking of candidate molecules.

The experimental results demonstrated that this screening strategy was capable of effectively enriching candidate molecules with antibacterial potential. The 15 highest-scoring candidates by the composite metric were subjected to antibacterial testing and determination of the minimum inhibitory concentration (MIC). Three molecules (M3, M5, and M7) displayed activity against E. coli.Among them, M3 yielded an inhibition zone of 21.2 mm and an MIC of 16 µg/mL, second only to the positive control, Cefaclor (23.1 mm; MIC 4 µg/mL), and exhibited the strongest antibacterial activity among the candidates. The results of the enzymatic inhibition assays, including both inhibition rates and IC₅₀ values obtained for the antibacterial molecules, provide preliminary experimental evidence for elucidating the interaction between the compounds and the target enzyme, as well as their potential antibacterial mechanisms.These results confirm that the maximal drug-likeness framework can accurately identify candidates with bona fide biological activity and establish M3 as a core candidate for subsequent cephalosporin anti–E. coli research.

A mechanistic rationale was provided by molecular docking and molecular dynamics simulation. Docking analyses indicated that catalytic core residues of E. coli penicillin-binding protein 1b (PBP1b) were engaged by M3, with a cooperative network of hydrogen bonds, hydrophobic contacts, and electrostatic interactions formed to effect efficient inhibition of the target. In a 100-nanosecond molecular dynamics simulation, the M3–PBP1b complex exhibited lower conformational fluctuations than the reference complex, consistent with stronger binding stability. By contrast, although M5 and M7 achieved favourable docking scores, the interactions of M5—comprising a hydrogen-bond network, a hydrophobic core, and charge anchoring—were concentrated on peripheral, noncatalytic residues at the edge of the active site, leading to weaker antibacterial activity than M3; for M7, insufficient shape complementarity hindered translation into effective antibacterial efficacy. Accordingly, M3 is nominated as the lead compound with the most favourable developability profile in this study.

In summary, the maximum drug-likeness strategy prospectively integrates multidimensional developability assessment into early-stage virtual screening, effecting a shift from an activity-oriented to a quality-oriented discovery paradigm. This approach can markedly improve conversion from virtual screening to active lead identification and prospectively avert unfavourable developability liabilities, thereby reducing downstream attrition risk. The strategy provides a new avenue for the design and discovery of cephalosporin antibacterials. Future research may be directed towards expanding the screened compound libraries, extending in vivo evaluations of efficacy and toxicity, optimising the computational modelling algorithms, and conducting targeted congener design along with systematic structure–activity relationship studies focused on the core lead compounds.

Funding Statement

This work was supported by the National Natural Science Foundation of China (grant numbers 82260896).

Disclosure statement

No potential conflict of interest was reported by the author(s). Xiangdong Zhao is employed by Beijing Honscent Zhiye Technology Co., Ltd. The company has no financial or non-financial interests in the submitted work, and this employment does not constitute a competing interest. The company also declares no competing interests related to this study.

Data availability statement

The data are available upon reasonable request.