Integrated computational and experimental benchmarking of Bacillus phage endolysins reveals the relationship between peptidoglycan-fragment recognition descriptors and antibacterial performance Authors: Roxana Portieles, Xinmin Ma, Jianjian Hu, Yuxiu Xu, Hongli Xu, Nayanci Portal González, Gabriela Santos-Portal, Rabia Durrani, Ramon Santos-Bermúdez, Orlando Borrás-Hidalgo Published: 2026 ABSTRACT Protein-based antibacterials such as bacteriophage endolysins offer a targeted therapeutic strategy against Gram-positive pathogens. However, prioritizing the most effective candidates from the large sequence diversity available remains a significant challenge. Here we present a standardized computational-experimental benchmarking framework that evaluates seven phage-derived endolysin variants (E1, E2, E3, E7, E10, E12, and E15) identified from Bacillus genomes. We combined molecular docking and residue-level interaction mapping against muramyl dipeptide (MDP), a minimal conserved peptidoglycan motif, with 1000-ns molecular dynamics simulations, MM/PBSA binding free-energy estimation, and matched functional inhibition assays against Staphylococcus aureus and Micrococcus luteus. Computational analyses revealed generally favorable MDP recognition across variants, albeit with notable differences in contact patterns and complex stability profiles. Experimental screening identified E2 as the most potent antibacterial agent against both species, while E7 and E1 performed strongly in selected computational metrics. Integrated analysis showed only modest correlations between computational descriptors of fragment recognition/stability and observed antibacterial performance. This study establishes a practical comparative benchmarking platform for endolysin candidate prioritization, nominates E2 and E7 as promising candidates for further development, and highlights E1 as a potential structural scaffold for rational engineering, while explicitly demonstrating both the utility and the current limitations of using minimal peptidoglycan fragments as proxies for full cell-wall recognition in lysin benchmarking.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1038/s41598-026-57871-x. Introduction Antimicrobial resistance (AMR) is a major global health threat that continues to cause significant and increasing mortality worldwide1. Among resistant pathogens, Staphylococcus aureus remains a persistent clinical challenge because it quickly adapts through various mechanisms, including target modification, drug inactivation, and envelope- or efflux-related strategies that reduce antibiotic exposure at the target site2,3. These facts support the development of antibacterial strategies that act rapidly, can bolster existing antibiotics, and may be less prone to traditional small-molecule resistance pathways. Bacteriophages encode specific antibacterial functions. During the late lytic cycle, phage endolysins (lysins) break down peptidoglycan, causing the bacterial cell wall to rupture and releasing progeny virions. In Gram-positive bacteria, the exposed cell wall allows recombinantly produced lysins to access the peptidoglycan from outside the cell and rapidly kill the bacteria, making lysins a promising class of targeted protein antibacterials4. Despite the growing number of phage-derived endolysin sequences available, the absence of standardized cross-variant benchmarking frameworks has hindered systematic prioritization. Most studies still characterize single lysins using disparate computational or experimental approaches, making direct comparisons and reliable predictions of functional lysis from sequence or structure extremely challenging5–12. A key feature of many lysins is their modular architecture, where one or more catalytic domains are combined with cell wall-binding elements that improve localization and specificity at the bacterial surface. This modularity also enables engineering: domain recombination, chimeric designs, and physicochemical adjustments (such as charge and stability) can change spectrum, potency, and performance in relevant conditions. Accordingly, high-throughput engineering platforms have shown that combinatorial design can speed up hit-to-lead optimization and find variants with better antibacterial activity. Clinical translation is also progressing. Anti-staphylococcal lysins such as exebacase (CF-301) have entered controlled human trials as antibiotic adjuncts for severe S. aureus infections6, highlighting the therapeutic potential of this approach. Beyond systemic infections, lysins are also being investigated for combination therapies and biofilm-related settings, where rapid enzymatic killing could complement other antibacterial strategies7. More broadly, the rationale for narrow, targeted antimicrobials is supported by evidence that broad-spectrum antibiotic use can cause lasting microbiome disruption in humans8–11, which may lead to downstream effects on the host13. Muramyl dipeptide (MDP) was selected as the minimal conserved peptidoglycan motif because it is present in virtually all bacterial cell walls and is directly recognized by the catalytic domain of most amidase-type endolysins. This choice enables standardized, sequence-agnostic comparisons across variants without the computational intractability of simulating full polymeric peptidoglycan. Nevertheless, MDP lacks the polymeric glycan chain, stem-peptide context, cross-links, and avidity effects provided by the native cell wall. The present study therefore explicitly tests the extent to which fragment-level recognition descriptors correlate with whole-cell lytic performance, thereby revealing both the strengths and the limitations of current in silico proxies for rational endolysin design. Here, we profile seven phage-derived endolysin variants (E1, E2, E3, E7, E10, E12, and E15) predicted from Bacillus genomes using a standardized benchmark framework that combines docking to a conserved peptidoglycan motif (muramyl dipeptide), residue-level interaction mapping, short MD stability descriptors, MM/PBSA binding free-energy estimates, and matched-concentration inhibition assays against S. aureus and Micrococcus luteus. By closing this methodological gap, our integrated pipeline provides a comparative benchmarking framework that evaluates the relationship between computational descriptors of fragment recognition and experimental lytic performance, while explicitly acknowledging the limitations of using minimal peptidoglycan fragments. Methods Endolysin identification and sequence resources Seven putative phage-derived endolysin genes (E1, E2, E3, E7, E10, E12, and E15) were obtained from Bacillus altitudinis/Bacillus aryabhattai genome resources (BioProject PRJNA781678). Coding regions were translated into protein sequences and curated for subsequent analyses. Primer pairs used to amplify each endolysin coding sequence are listed in Supplementary Table 1. Structural modelling and model quality assessment Three-dimensional models for each endolysin were generated by homology modelling using SWISS-MODEL14. The model geometry and stereochemistry were evaluated with MolProbity15, and additional structure-consistency checks were performed with ERRAT and Verify3D via the SAVES server. Only models that satisfied acceptable stereochemical and fold-quality thresholds were used for docking and molecular dynamics (MD) simulations. Global fold-quality metrics (LGscore and MaxSub) were calculated with ProQ and provided as comparative indicators of overall fold plausibility across variants, rather than as MolProbity outputs (Supplementary Table 2)16. Domain architecture and catalytic motifs Domain architecture and catalytic motifs were predicted using InterProScan v5 and Pfam 37.0. Boundaries of the catalytic (EAD) and cell wall-binding (CBD) domains were confirmed by structural modelling and are summarized in Supplementary Table S5 and Supplementary Fig. S3. Full-length models were used for all subsequent docking and molecular dynamics simulations. Ligand definition and preparation Muramyl dipeptide (MDP; PubChem CID: 451714) was used as a conserved minimal peptidoglycan motif to enable standardized comparisons across endolysin variants. MDP was downloaded in SDF format and converted to PDB using Open Babel 3.1.1. Protonation was set to reflect physiological pH (7.4), hydrogens were added, and rotatable bonds were defined consistently across all docking runs. Binding-site prediction and docking Putative ligand-binding pockets were predicted using P2Rank and visualized via PrankWeb17,18. For each endolysin, the top-ranked pocket was selected to ensure consistent site definition across variants. Docking was performed with the HADDOCK2.4 web server following a standardized protocol across all variants (detailed parameters in Supplementary Table 2). To define ambiguous interaction restraints (AIRs) for HADDOCK, active residues were designated as (i) conserved catalytic-motif residues within the predicted catalytic domain (for example, His/Asp/Asn in amidase/glycosidase-like motifs) and (ii) the top pocket residues identified by P2Rank. Passive residues were automatically assigned by HADDOCK as neighboring solvent-accessible residues around the active residues using default settings19–24. Docking calculations were carried out using the standard HADDOCK stages and sampling sizes: (i) rigid-body docking (it0): 1000 structures, (ii) semi-flexible refinement (it1): 200 structures, and (iii) explicit-solvent refinement (water): 200 structures, with default ambiguous interaction restraints (AIRs) settings unless otherwise specified (full run settings in Supplementary Table 2). The normalized docking-confidence score (0–1) was calculated from HADDOCK clustering outputs as: confidence score = (cluster size × cluster Z-score) / max(cluster size × cluster Z-score) across all clusters in the run (Supplementary Table 2). Docking outputs were summarized using (i) the HADDOCK score and cluster-level statistics (cluster size and cluster Z-score), (ii) the normalized docking-confidence index (‘confidence score’), and (iii) ligand RMSD relative to the selected reference pose. Ligand RMSD values (17–27 Å) are reported as a descriptive measure of pose dispersion for this small flexible fragment rather than as an indicator of complex stability25–27. Ligand RMSD was calculated after superimposing the protein backbone of each frame onto the reference docked pose; the relatively high values reflect the intrinsic flexibility of the small MDP ligand, which can occupy multiple low-energy binding orientations while maintaining persistent key contacts with catalytic and pocket residues. Result inspection and figure rendering were performed in PyMOL23,24 (Table 1). Table 1Summary of docking and interaction descriptors across endolysin variants.EndolysinHADDOCK scoreConfidence scoreLigand RMSD (Å)number contactsAvg distance (Å)Strong contacts (< 2.5 Å)Affinity score (Num/Avg)E1-231.700.83619.0172.3452.99E2-226.500.82219.9892.9813.02E3-218.700.79826.5342.6521.51E7-245.280.87019.9352.4632.03E10-236.100.84822.5992.9613.04E12-237.200.85119.2892.8403.17E15-221.700.80717.3563.1201.92Docking score, confidence score and ligand RMSD are reported for the top-ranked pose. “number contacts” and distance-based metrics are derived from polar-contact analysis of endolysin–MDP complexes. The affinity score is defined as (# contacts)/(average contact distance, Å) and used as an internal comparative descriptor. Table 1Summary of docking and interaction descriptors across endolysin variants.EndolysinHADDOCK scoreConfidence scoreLigand RMSD (Å)number contactsAvg distance (Å)Strong contacts (< 2.5 Å)Affinity score (Num/Avg)E1-231.700.83619.0172.3452.99E2-226.500.82219.9892.9813.02E3-218.700.79826.5342.6521.51E7-245.280.87019.9352.4632.03E10-236.100.84822.5992.9613.04E12-237.200.85119.2892.8403.17E15-221.700.80717.3563.1201.92Docking score, confidence score and ligand RMSD are reported for the top-ranked pose. “number contacts” and distance-based metrics are derived from polar-contact analysis of endolysin–MDP complexes. The affinity score is defined as (# contacts)/(average contact distance, Å) and used as an internal comparative descriptor. Summary of docking and interaction descriptors across endolysin variants. Docking score, confidence score and ligand RMSD are reported for the top-ranked pose. “number contacts” and distance-based metrics are derived from polar-contact analysis of endolysin–MDP complexes. The affinity score is defined as (# contacts)/(average contact distance, Å) and used as an internal comparative descriptor. Protein–ligand interaction profiling Non-covalent contacts between each endolysin and MDP were identified from the representative docked complexes using PLIP and confirmed through distance inspection in PyMOL23. Hydrogen bonds and salt bridges were summarized as residue-level interaction fingerprints (Fig. 1, sSupplementary Table 3). Molecular dynamics simulations For each endolysin–MDP complex, three independent production runs of 1000 ns were carried out. Each replica was started from the same energy-minimized and equilibrated structure but with different initial velocity distributions generated using distinct random seeds. MD simulations were conducted in GROMACS28 using the CHARMM36m force field29. Full-length endolysin–PG complexes were solvated in a cubic TIP3P water box and neutralized with counterions. Energy minimization was carried out for 5,000 steps, followed by equilibration in the NVT (100 ps, 300 K) and NPT (100 ps, 1 bar) ensembles. Production runs lasted 1000 ns with a 2 fs timestep, employing periodic boundary conditions and particle-mesh Ewald electrostatics. RMSD values were calculated for the protein backbone atoms relative to the initial frame of each replica. Structural stability was assessed by the mean RMSD ± standard deviation across the three independent replicas. Convergence of the trajectories was evaluated through time-resolved RMSD and radius of gyration (Rg) profiles, persistence of ligand–protein contacts, and block-averaged MM/PBSA calculations from the final 500 ns of each simulation (Fig. 2). MM/PBSA binding free energy calculation Binding free energies (ΔG) were calculated from the MD trajectories using gmx_MMPBSA30. Temperature was kept at 300 K with the velocity-rescale (V-rescale) thermostat, and pressure was maintained at 1 bar with the Parrinello–Rahman barostat. All bonds involving hydrogen were constrained with LINCS, and long-range electrostatics were addressed using particle-mesh Ewald under periodic boundary conditions. Protein backbone RMSD was determined relative to the initial docked complex. For MM/PBSA, snapshots were extracted every 10 ps from the final 500 ns of each trajectory to ensure comparable sampling across complexes. ΔG was decomposed into van der Waals, electrostatic, polar solvation (PB), and nonpolar (SASA) contributions, and per-residue energy decomposition was utilized to identify interaction hotspots. MM/PBSA outputs were interpreted as relative binding-energy estimates within a standardized protocol rather than absolute affinities30,31. All analyses were performed on full-length endolysin–MDP complexes. Cloning, recombinant expression, and purification Endolysin coding sequences (E1, E2, E3, E7, E10, E12, and E15) were amplified by PCR and cloned into the cold-shock expression vector pCold I to generate N-terminal His-tagged constructs. Recombinant plasmids were propagated in E. coli DH5α, verified by Sanger sequencing, and transformed into E. coli BL21(DE3) for protein production. Cultures were grown in LB medium supplemented with ampicillin (100 µg mL−1) at 37 °C to an OD600 of approximately 0.6, induced with IPTG (0.5 mM), and expressed at 15–16 °C for 18 h. Cells were harvested by centrifugation and lysed by sonication in binding buffer (20 mM sodium phosphate, 0.5 M NaCl, 30 mM imidazole, pH 7.4). Clarified lysates (after centrifugation followed by 0.22 μm filtration) were purified via immobilized metal affinity chromatography using Ni-NTA resin (HisPur™ Ni-NTA Gravity Column; Thermo Fisher Scientific), following the manufacturer’s instructions. Protein purity was assessed with SDS-PAGE, and concentrations were measured using the bicinchoninic acid (BCA) assay. Protein concentrations were normalized on a molar basis using theoretical molecular weights calculated from the sequences (Supplementary Table S5)32–36. Antibacterial activity assays Antibacterial activity was tested against Staphylococcus aureus (ATCC 29213) and Micrococcus luteus (ATCC 4698) using an agar well diffusion assay37 and a turbidity reduction assay. Agar well diffusion For S. aureus, overnight cultures were adjusted to approximately 10⁸ CFU mL−1 and spread onto Mueller–Hinton agar. Wells (6 mm) were filled with 100 µL of endolysin (1 mg mL−1 in PBS, pH 7.4), incubated at 37 °C for 24 h, and inhibition zones were measured with a digital calliper and reported as the total diameter including the well. For M. luteus, a Micrococcus indicator-lawn diffusion method was used on nutrient-broth agar supplemented with Tween 20. Wells were filled with 100 µL of endolysin or PBS control, and plates were incubated at 30 °C for 24 h before measuring the zones38. Turbidity reduction assay and exponential rate estimation To measure lysis kinetics independently of agar diffusion, turbidity decline was tracked as OD600 over time after adding purified endolysins to standardized bacterial suspensions. Overnight cultures of S. aureus were grown in LB at 37 °C and M. luteus in nutrient broth at 30 °C, then diluted into fresh medium and adjusted to a uniform starting turbidity (OD600 = 1.0). Endolysins were buffer-exchanged into PBS (pH 7.4) and added to bacterial suspensions at a standard protein amount (final concentration 10 µg/mL, unless stated otherwise). Suspensions were incubated under suitable conditions for each species (37 °C for S. aureus, 30 °C for M. luteus) with shaking, and OD600 was measured at 0, 2, 4, 6, 8, and 10 h. OD600 values were normalized to the initial measurement (t = 0) for each condition to enable comparison across variants. PBS-only controls were run in parallel. For diffusion-independent ranking, an apparent lysis constant was estimated using an exponential model39 by fitting ln (OD600) against time during the linear part of the decay curve (a predefined time window applied consistently across variants). The slope of the fit (− kapp) served as a comparative measure of the turbidity-reduction rate, with a higher kapp indicating faster apparent lysis under standard assay conditions. Because OD600 decay over extended incubation can reflect lysis as well as growth suppression or aggregation, turbidity trajectories are interpreted here as a comparative kinetic screen under matched conditions rather than a direct measurement of catalytic turnover. Statistics and reproducibility All statistical analyses were carried out using GraphPad Prism 9. Normality was assessed with the Shapiro–Wilk test. Group differences were analyzed using one-way ANOVA with Tukey’s multiple comparisons test. Statistical significance was set at P < 0.05. Data are presented as mean ± s.e.m. (n = 3 independent biological replicates, each performed in technical triplicate). Biological replicates, technical replicates, and MD trajectory frames are clearly distinguished throughout. Results Structural modelling and domain architecture Homology models of the seven endolysin variants (E1, E2, E3, E7, E10, E12, and E15) were generated using SWISS-MODEL14. All models satisfied acceptable stereochemical and fold-quality thresholds as assessed by MolProbity15, ERRAT, and Verify3D (Supplementary Table 2). Global fold-quality metrics (LGscore and MaxSub) calculated with ProQ indicated good to very good model quality across the panel and were used as comparative indicators of overall fold plausibility (Supplementary Table 2). Domain architecture and catalytic motifs were predicted using InterProScan v5 and Pfam 37.0. All variants share a conserved modular organization consisting of an N-terminal catalytic domain (N-acetylmuramoyl-L-alanine amidase family) and a C-terminal cell wall-binding domain (CBD), although the specific type of CBD differed among variants (Supplementary Table S5 and Supplementary Fig. S3). Full-length models were used for all subsequent docking and molecular dynamics simulations40–42. Contribution of the cell wall-binding domain (CBD) All seven endolysins share a conserved modular architecture consisting of an N-terminal enzymatically active domain (EAD; N-acetylmuramoyl-L-alanine amidase family) and a C-terminal cell wall-binding domain (CBD). The specific CBD families differ among variants (SPOR in E2, LysM in E3/E10/E12, PG-binding domain in E7, PlyG-like in E15, and a domain of unknown function in E1; Supplementary Table S5 and Supplementary Fig. S3). Although the present analyses focused on EAD-MDP interactions, the CBD is known to enhance substrate positioning, increase avidity toward polymeric peptidoglycan, and contribute to overall lytic efficiency4,43,44. The observed partial correlation between EAD-centric computational descriptors and experimental antibacterial activity was therefore consistent with the hypothesis that full lytic performance requires cooperative contributions from both domains. Future studies incorporating larger muropeptides or full peptidoglycan fragments will be necessary to quantify CBD contributions quantitatively. Comparative docking of muramyl dipeptide (MDP) Standardized docking of muramyl dipeptide (MDP) to the predicted binding pockets of each endolysin was performed with the HADDOCK2.4 web server. Docking scores ranged from − 218.7 to − 245.3, indicating generally favorable MDP recognition across the seven variants (Supplementary Table 2). The normalized docking-confidence scores (0–1), derived from cluster size and Z-score, were high for all complexes (0.798–0.870), supporting the selection of the top-ranked clusters for downstream analyses. Ligand RMSD values relative to the representative pose ranged from 17.35 to 26.53 Å. These relatively high RMSD values are reported as a descriptive measure of pose dispersion typical when docking small and flexible peptidoglycan fragments, rather than as an indicator of complex stability25–27. Ligand RMSD was calculated after superimposing the protein backbone of each frame onto the reference docked pose; the relatively high values reflect the intrinsic flexibility of the small MDP ligand, which can occupy multiple low-energy binding orientations while maintaining persistent key contacts with catalytic and pocket residues. Residue-level interaction profiling Non-covalent contacts in the top-ranked docked poses were analyzed with PLIP and PyMOL. Variant-specific polar-contact patterns were observed (Supplementary Tables 3 and Supplementary Fig. S1). The number of contacts ranged from 4 to 9, with average contact distances between 2.34 and 3.12 Å (Supplementary Table 4). An internal “affinity score” ((number of contacts) / average distance) was calculated as a simple descriptive metric to summarize contact density. Variants E1, E2, and E10 displayed particularly compact interaction networks involving conserved catalytic residues and additional pocket residues. Molecular dynamics simulations and complex stability To rigorously assess the conformational stability of the endolysin–MDP complexes, three independent 1000 ns molecular dynamics (MD) replicas were performed for each system. Each replica was started from the same energy-minimized and equilibrated structure but assigned different random velocity seeds. The mean backbone RMSD ± standard deviation across the three independent replicas for each complex is shown in Fig. 2. All complexes exhibited stable RMSD profiles after the initial equilibration phase, with low inter-replica variability (narrow shaded regions). This confirms that the observed conformational behavior is reproducible and not the result of stochastic events or entrapment in a local energy minimum of three independent 1000 ns replicas. Additional convergence diagnostics, including time-resolved radius of gyration (Rg) profiles, persistence of ligand–protein contacts, and block-averaged MM/PBSA calculations from the final 500 ns of each replica, further support the robustness of the stability rankings. MM/PBSA binding free energy estimation Binding free energies (ΔG) were calculated from the final 500 ns of each MD trajectory using gmx_MMPBSA. Substantial differences among variants were observed (Fig. 3). E1 exhibited the most favorable (most negative) ΔG value, while E2 showed a comparatively less favorable binding energy despite its strong experimental performance. Per-residue energy decomposition highlighted key residues contributing to the interaction energy in each complex (Fig. 4). These computational metrics provided useful relative rankings of fragment recognition and stability within the standardized simulation protocol. It is important to note that MM/PBSA binding free energies are interpreted here as relative rankings within a standardized protocol rather than absolute affinities. The notably favorable ΔG observed for E1 (around − 70 kcal/mol) likely reflects the absence of entropic (–TΔS) contributions and the approximations inherent to the continuum solvation model; both factors can overestimate binding energies for small, flexible ligands such as MDP30,31. Antibacterial activity against Gram-positive bacteria The lytic activity of the purified endolysins was evaluated against Staphylococcus aureus (ATCC 29213) and Micrococcus luteus (ATCC 4698) using two complementary functional assays. In the agar well diffusion assay, all variants produced clear inhibition zones, with E2 consistently generating the largest diameters against both species (Fig. 5a and Supplementary Fig. S2). The diffusion-independent turbidity reduction assay confirmed the same ranking: E2 produced the most rapid decline in OD600 over 10 h, indicating the strongest apparent lytic kinetics (Fig. 5b). An exponential decay model was fitted to the linear phase of the turbidity curves to derive an apparent lysis rate constant, further supporting E2 as the most potent variant under the assay conditions. Protein concentrations were normalized on a molar basis using theoretical molecular weights (Supplementary Table S5) to ensure fair comparison across constructs of different lengths. Data are presented as mean ± s.e.m. (n = 3 independent biological replicates, each performed in technical triplicate)45–52. Integrated computational-experimental benchmarking To facilitate candidate prioritization, the computational descriptors (docking score, confidence score, affinity score, MD stability metrics, and MM/PBSA ΔG) were integrated with the experimental antibacterial activity data and visualized in a multi-metric heatmap (Fig. 6). The analysis revealed only modest correlations between the computational descriptors of MDP recognition and stability and the observed lytic performance (Pearson r ≈ 0.3–0.4, p > 0.3; Supplementary Table S6). E2 emerged as the most potent inhibitor in both functional assays, while E7 and E1 performed strongly in selected computational metrics. These results support the utility of the standardized benchmarking pipeline for comparative evaluation of endolysin variants and nominate specific candidates for further development. Fig. 1Representative binding interfaces of selected endolysin variants with muramyl dipeptide (MDP). (a) E1, (b) E2, (c) E7. Left panels show the overall surface representation of the endolysin (green) with MDP in the binding pocket (magenta sticks). Right panels show detailed stick representation of key interacting residues (orange) and hydrogen bonds (yellow dashes) with distances in Å. The full set of seven complexes is shown in Supplementary Fig. S1. Representative binding interfaces of selected endolysin variants with muramyl dipeptide (MDP). (a) E1, (b) E2, (c) E7. Left panels show the overall surface representation of the endolysin (green) with MDP in the binding pocket (magenta sticks). Right panels show detailed stick representation of key interacting residues (orange) and hydrogen bonds (yellow dashes) with distances in Å. The full set of seven complexes is shown in Supplementary Fig. S1. Fig. 2Root-mean-square deviation (RMSD) of backbone atoms in the endolysin–MDP complexes during 1000 ns MD simulations. Data represent the mean RMSD ± standard deviation from three independent replicas for each complex (E1-MDP, E2-MDP, E3-MDP, E7-MDP, E10-MDP, E12-MDP, and E15-MDP). Shaded regions indicate the standard deviation. (a) Time evolution of the mean RMSD for each endolysin–MDP complex. (b) Overlay of all systems for direct comparison. Root-mean-square deviation (RMSD) of backbone atoms in the endolysin–MDP complexes during 1000 ns MD simulations. Data represent the mean RMSD ± standard deviation from three independent replicas for each complex (E1-MDP, E2-MDP, E3-MDP, E7-MDP, E10-MDP, E12-MDP, and E15-MDP). Shaded regions indicate the standard deviation. (a) Time evolution of the mean RMSD for each endolysin–MDP complex. (b) Overlay of all systems for direct comparison. Fig. 3MM/PBSA binding free energies (ΔG) of endolysin–MDP complexes. Box plots show the distribution of ΔG values calculated from the final 500 ns of each of the three independent 1000-ns MD replicas. More negative values indicate more favorable binding. MM/PBSA binding free energies (ΔG) of endolysin–MDP complexes. Box plots show the distribution of ΔG values calculated from the final 500 ns of each of the three independent 1000-ns MD replicas. More negative values indicate more favorable binding. Fig. 4Per-residue energy decomposition from MM/PBSA calculations. Key residues contributing most favorably to the binding energy are highlighted for each endolysin variant. Per-residue energy decomposition from MM/PBSA calculations. Key residues contributing most favorably to the binding energy are highlighted for each endolysin variant. Fig. 5Antibacterial activity and lysis kinetics of the seven Bacillus phage endolysins. (a) Inhibition-zone diameters (mm) determined by agar well diffusion assay against Staphylococcus aureus (ATCC 29213) and Micrococcus luteus (ATCC 4698). Bars represent means ± standard error of three independent replicates. Different letters indicate statistically significant differences (one-way ANOVA with Tukey’s post-hoc test, P < 0.0001). (b) Representative photographs of inhibition zones. (c) Turbidity reduction kinetics (OD600) showing exponential decay of bacterial suspensions after addition of purified endolysins (10 µg/mL). Data are means of three independent experiments. E2 consistently exhibited the strongest lytic activity against both species. Antibacterial activity and lysis kinetics of the seven Bacillus phage endolysins. (a) Inhibition-zone diameters (mm) determined by agar well diffusion assay against Staphylococcus aureus (ATCC 29213) and Micrococcus luteus (ATCC 4698). Bars represent means ± standard error of three independent replicates. Different letters indicate statistically significant differences (one-way ANOVA with Tukey’s post-hoc test, P < 0.0001). (b) Representative photographs of inhibition zones. (c) Turbidity reduction kinetics (OD600) showing exponential decay of bacterial suspensions after addition of purified endolysins (10 µg/mL). Data are means of three independent experiments. E2 consistently exhibited the strongest lytic activity against both species. Fig. 6Integrated heatmap of computational peptidoglycan-fragment recognition descriptors and experimental antibacterial performance. Rows represent the seven endolysin variants; columns show key computational metrics (docking score, normalized confidence score, ligand RMSD, average MM/PBSA ΔG, number of polar contacts, and strong contacts < 2.5 Å) alongside experimental inhibition-zone diameters against S. aureus and M. luteus. Values are z-score normalized within each column for visualization. Modest correlations between computational descriptors and lytic activity are reported in Supplementary Table S6. The heatmap highlights both the relative strengths of individual variants (E2 in experimental activity, E1 and E7 in computational metrics) and the partial concordance between in silico proxies and whole-cell performance. Integrated heatmap of computational peptidoglycan-fragment recognition descriptors and experimental antibacterial performance. Rows represent the seven endolysin variants; columns show key computational metrics (docking score, normalized confidence score, ligand RMSD, average MM/PBSA ΔG, number of polar contacts, and strong contacts < 2.5 Å) alongside experimental inhibition-zone diameters against S. aureus and M. luteus. Values are z-score normalized within each column for visualization. Modest correlations between computational descriptors and lytic activity are reported in Supplementary Table S6. The heatmap highlights both the relative strengths of individual variants (E2 in experimental activity, E1 and E7 in computational metrics) and the partial concordance between in silico proxies and whole-cell performance. Discussion We present an integrated computational-experimental benchmarking framework that systematically evaluates seven phage-derived endolysin variants (E1, E2, E3, E7, E10, E12, and E15) identified from Bacillus genomes. By combining molecular docking and residue-level interaction mapping against muramyl dipeptide (MDP) a minimal conserved peptidoglycan motif with three independent 1000-ns molecular dynamics replicas per complex, MM/PBSA binding free-energy estimation, and matched antibacterial inhibition assays against Staphylococcus aureus and Micrococcus luteus, the pipeline provides a standardized platform for cross-variant comparison. Computational analyses revealed generally favorable MDP recognition and complex stability across most variants, with E7 and E1 standing out in docking score and MM/PBSA binding energy, respectively. Functional assays identified E2 as the most potent antibacterial agent against both Gram-positive species (Fig. 5). Integrated analysis showed only modest correlations between individual peptidoglycan-fragment recognition descriptors and observed lytic performance (Pearson r ≈ 0.3–0.4, p > 0.3; Supplementary Table S6 and Fig. 6). These findings distinguish fragment-level recognition potential from effective whole-cell lysis and underscore the value of multi-metric benchmarking rather than reliance on any single in silico predictor. All seven endolysins share a conserved modular architecture consisting of an N-terminal enzymatically active domain (EAD; N-acetylmuramoyl-L-alanine amidase family) and a C-terminal cell wall-binding domain (CBD), although the specific CBD families differ markedly among variants (SPOR in E2, LysM in E3/E10/E12, PG-binding domain in E7, PlyG-like in E15, and a domain of unknown function in E1; Supplementary Table S5 and Supplementary Fig. S3). The present computational pipeline, which focused on EAD-MDP interactions, necessarily emphasizes catalytic-domain recognition. The CBD is known to mediate substrate positioning, enhance avidity toward the polymeric, cross-linked peptidoglycan lattice of live bacteria, and contribute substantially to overall lytic efficiency. The observed partial concordance between EAD-centric descriptors and experimental activity is therefore mechanistically expected and consistent with the hypothesis that full antibacterial performance requires cooperative contributions from both domains. This study thus highlights the importance of considering domain cooperativity in future rational design efforts. The three independent 1000-ns MD simulations per complex (different random velocity seeds) provided robust descriptors of structural stability. Protein backbone RMSD remained low and reproducible across all 21 replicas (< 2.5 Å after equilibration; Fig. 2), confirming overall complex stability. The relatively high ligand RMSD values (17–27 Å) reflect the intrinsic flexibility of the small MDP fragment, which can occupy multiple low-energy binding orientations while maintaining persistent key contacts with catalytic and pocket residues (Supplementary Table 4). Likewise, MM/PBSA binding free energies were interpreted as relative rankings within the standardized protocol rather than absolute affinities. The notably favorable ΔG observed for E1 likely arises from the absence of explicit entropic contributions (–TΔS) and the approximations inherent to the continuum solvation model both factors known to overestimate binding energies for small, flexible ligands such as MDP52,53. These limitations are particularly relevant for endolysins, where interactions with polymeric peptidoglycan further exacerbate issues related to conformational sampling and the accuracy of implicit solvation models52,53. These computational metrics therefore serve as useful comparative tools for candidate screening and structural insight but cannot fully replace experimental functional validation. Functional assays confirmed E2 as the most potent antibacterial agent against both S. aureus and M. luteus, with E7 showing intermediate potency and E3 consistently the weakest performer (Fig. 5). Protein concentrations were normalized on a molar basis to ensure fair comparisons across variants of different lengths. The integrated heatmap (Fig. 6) and correlation analysis (Supplementary Table S6) revealed only modest associations between any individual computational descriptor (number of contacts, average/min/max distances, strong contacts, or affinity score) and inhibition-zone diameters. Far from representing a shortcoming of the benchmarking framework, this modest correlation constitutes one of its most important scientific contributions. By explicitly demonstrating the limited predictive power of current EAD-centric, minimal-fragment proxies, the study reveals a critical gap in in silico approaches for lysin engineering: the necessity of incorporating CBD-polymeric peptidoglycan interactions, higher-order muropeptide context, and avidity effects. Such “negative results” are essential for the maturation of rational design pipelines, as they prevent over-reliance on oversimplified metrics and direct future efforts toward more physiologically relevant models. The present study has several important limitations. First, muramyl dipeptide (MDP) represents only the minimal conserved disaccharide-peptide motif recognized by the enzymatically active domain (EAD) and does not recapitulate the polymeric glycan chains, stem-peptide cross-links, or multivalent avidity effects of the native bacterial cell wall. Consequently, the computational pipeline emphasizes EAD-centric interactions while under-representing the contribution of the cell wall-binding domain (CBD). Second, although three independent 1000-ns MD replicas were performed for each complex, the MM/PBSA calculations do not include explicit entropic contributions (–TΔS) and rely on a continuum solvation model; both approximations can overestimate binding energies for small, flexible ligands52. Moreover, conformational sampling remains a major challenge in large, flexible protein–ligand systems such as endolysins53. These issues are further aggravated when studying interactions with polymeric peptidoglycan substrates52. Third, antibacterial assays were conducted using standardized non-native Gram-positive targets (S. aureus and M. luteus) chosen for reproducibility and benchmarking purposes; activity against native Bacillus hosts or clinical isolates may differ. Finally, the long assay timescale and single-concentration format provide a practical but simplified readout of lytic performance. Future extensions incorporating larger muropeptides, coarse-grained simulations of polymeric peptidoglycan, direct catalytic turnover measurements, and additional clinical strains will further strengthen the predictive power of the framework. Despite these limitations, the standardized computational-experimental pipeline established here offers a reproducible and practical platform for cross-variant comparison and candidate prioritization in endolysin engineering. E2 and E7 emerge as the most promising lead candidates for further development and therapeutic optimization, while E1 represents a valuable “binding-forward” structural scaffold whose favorable computational profile can be leveraged through targeted domain-swapping or CBD engineering. By transparently documenting both the strengths and the current limitations of minimal-peptidoglycan-fragment proxies, this study provides a constructive roadmap for the next generation of in silico tools in lysin design and contributes to the broader effort to develop precision protein antibacterials against antimicrobial-resistant pathogens. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1 Supplementary Material 2 Supplementary Material 2 Supplementary Material 3 Supplementary Material 3