The present study employs three primary datasets:

Specifically, the DrugBank database was first filtered for irrelevant drug types, including vaccines, antibodies, toxins, biologics, biopolymers, oligonucleotides, and oligosaccharides, while ensuring the inclusion of required biotech and small-molecule drugs. Subsequently, 2,744,803 drug-drug interaction (DDI) pairs were extracted, encompassing the risks and effects of drug interactions. Additionally, three KP-irrelevant drugs were eliminated at this stage: methaneseleninic acid, propionic acid, and Florbetaben (18F). Finally, 3,475 unique drugs were extracted at this stage.

The DRKG integrates data from sources including DrugBank, GNBR, Hetionet, String, Intact, and DGIDB, comprising 5,874,261 triplets. Additional KP-specific data from public sources added 207 triplets.

Through the integration of the aforementioned database data, a comprehensive knowledge graph containing 5,874,712 entries was ultimately constructed, with triplets in the form of (head entity, relation, tail entity) extracted through preprocessing, covering extensive biomedical entities and their relationships, where entities and relations were mapped to unique identifiers for efficient model processing.

In this study, model construction was completed using the PyKEEN framework (a knowledge graph embedding toolkit supporting standardized negative sampling and loss function optimization), modeling entity-relation interactions through the multilayer perceptual structure of the Entity-Relation Multi-Layer Perceptron (ERMLP) model to generate high-quality embedding vectors. After training, performance evaluation was conducted on both validation and test sets using internationally standard metrics for knowledge graph completion, including: adjusted arithmetic mean rank (AMR) - an evaluation metric that standardizes true entity ranking positions by library size, directly quantifying the model’s global positioning capability for correct entities, with AMR values below 0.1 indicating excellent relationship reasoning ability according to biomedical KG research; and Hits@k (k = 1,3,10) - representing the proportion of correct entities ranked within the top k predictions [21].

In similar tasks, Hits@10 > 0.2 is a common threshold for effective relationship capture. During training, AMR (lower values indicating better performance), Hits@1 (reflecting precise matching), and Hits@10 (indicating potential relationship capture efficiency) were monitored, collectively validating the model’s generalization on unseen data.

Using a pretrained knowledge graph embedding model, this study predicted K. pneumoniae-related genes. The head entity “Disease::MesH:D007711” represented the target pathogen, with the relationship path “bioarx::Corkp ass kp_gene::Disease:Gene” capturing potential gene-disease associations. The ERMLP model’s multilayer perception mechanism, implemented via PyKEEN’s “predict_tail_entity()” function, generated the top 100 tail entity predictions. These were manually reviewed by genomics experts to identify potential therapeutic targets with biological pathway associations.

To validate the druggability and resistance mechanisms of predicted targets, Amber 24 was used for all-atom molecular dynamics simulations. Systems were built with the ff19SB force field, solvated in a TIP3P water model, and counterions added to maintain charge balance.

The system was first subjected to energy minimization in two stages: an initial 5000 steps of steepest descent followed by 5000 steps of conjugate gradient algorithm to remove any steric clashes. After the system was gradually heated to 310.15 K over 500 ps with weak restraints on the protein backbone, it was equilibrated in the NVT ensemble. The subsequent 100-ns production simulation was performed under the NPT ensemble, where temperature was maintained using a Langevin thermostat and pressure was controlled with a Berendsen barostat.

Subsequently, positional constraints were applied for pre-equilibration in the NVT ensemble, followed by transition to NPT ensemble (constant particle number, pressure, temperature) for 100ns production simulation, maintaining periodic boundary conditions throughout. The SHAKE algorithm constrained all hydrogen-containing covalent bonds.

Root mean square deviation (RMSD) was used to quantify the overall deviation of protein backbone atoms relative to initial conformation, root mean square fluctuation (RMSF) analyzes local dynamic characteristics of residue side chains, and free energy landscape (FEL) was based on the energy distribution of conformational ensembles on reaction coordinate projections; these three indicators collaboratively evaluate and identify the molecule’s core functional domains.

To identify core stable segments in protein structures, this study employed a multi-scale conformational stability analysis method.

First, residue dynamics throughout the chain were quantified using root mean square fluctuation (RMSF) from molecular dynamics simulations, extracting residue subsets with fluctuation amplitudes in the lowest quartile (i.e., the most stable 25%). Subsequently, using molecular dynamics trajectories, secondary structure types for each protein residue were annotated frame-by-frame using the DSSP algorithm implemented through Python’s MDTraj library. The frequency of primary secondary structure types was analyzed across the entire trajectory, identifying highly conserved residues (those maintaining the same secondary structure for ≥ 95% of simulation time) and ultra-conservative segments with stable secondary structure types throughout the simulation. Through intersection operations of both criteria, low RMSF fluctuation regions were integrated with stable secondary structure regions to identify core segments exhibiting high stability in both tertiary structure dynamics and secondary structure conservation dimensions. Furthermore, the bioinformatics tool ConSurf was employed with empirical Bayesian methodology to assess the evolutionary conservation of amino acid positions within the protein, enabling analysis of biomedical functional associations [22]. The secondary structure was predicted via GOR IV [23].

After 50 training epochs of the ERMLP model, results in realistic evaluation mode showed a key metric of Hits@10 reaching 0.1602, reflecting the model’s ability to correctly identify at least one true entity among every 10 candidates, and an Adjusted Arithmetic Mean Rank (AMR) of 0.0238, indicating that correct entities were ranked, on average, in the top 2.38% of all candidates.

Based on the top 100 candidate targets predictions generated by the ERMLP model (see Table 1), the KPHS_11890 gene (Acridine efflux pump [Klebsiella pneumoniae subsp. HS11286]) was determined as the top candidate target. This gene achieved a score of 4.628 in the ERMLP model predictions, which was the highest score among all genes, indicating a strong association with the pathogen entity Disease::MesH:D007711 via the relation path bioarx::Corkp ass kp_gene::Disease:Gene in the knowledge graph. And it was also recognized in the expert review.

As illustrated in Fig. 1, the KPHS_11890 gene encodes a membrane fusion protein (MFP) (TC 8.A.1) family efflux pump, a critical efflux transporter that mediates the active extrusion of acridine antibiotics and similar hydrophobic compounds, thereby conferring antibiotic resistance to the host bacterium, a mechanism widely present in clinical pathogens that significantly influences the development of multidrug-resistant phenotypes, aligning with K. pneumoniae resistance issues [24]. Studies indicate that efflux pump genes (e.g., acrAB, tolC) have a detection rate exceeding 90% in K. pneumoniae and often coexist with ESBL genes (e.g., blaCTX-M), significantly enhancing resistance through synergistic effects [25]. As an efflux pump, KPHS_11890 may form functional coupling with other resistance genes (e.g., metallo-β-lactamase genes) through similar mechanisms, promoting carbapenem resistance. Additionally, the MFS efflux pump family is highly conserved in high-risk Klebsiella pneumoniae clones (e.g., ST258). Its functional loss can significantly restore sensitivity to β-lactam antibiotics (e.g., reducing imipenem MIC values by 4–8 fold through efflux pump inhibition) and is negatively correlated with virulence factors, suggesting that targeting this gene may simultaneously reduce resistance and virulence [26–28].

Thus, its critical role in antimicrobial resistance highlights the potential of studying its structure and function to provide molecular targets for developing novel efflux pump inhibitors, advancing the development of new antibiotic adjuvants, and warranting further investigation of its molecular dynamics characteristics.

The 100 ns molecular dynamics simulation of the KPHS_11890 protein revealed its structural dynamic properties and potential functional correlations.

In Fig. 2, Root Mean Square Deviation (RMSD) analysis indicated that the protein backbone atoms fluctuated between 2.5 Å and 15 Å during the simulation. This significant conformational change suggests that the protein underwent substantial structural rearrangement within the simulation timescale, reflecting its inherent structural flexibility. This structural flexibility may be associated with its conformational adaptability as an efflux pump, as dynamic conformational changes are required during substrate transport to fulfill its function.

Figure 3 RMSF analysis revealed the dynamic properties of protein regions. The residue 50–100 region showed low fluctuation (2–6 Å), indicating structural stability, likely forming the core and anchoring to the cell membrane for proper positioning and function of the efflux pump. Conversely, residues 0–50, 100–150, and 350–400 displayed high RMSF (6–12 Å), suggesting greater flexibility, possibly linked to the substrate-binding pocket’s opening and closing, crucial for efficient antibiotic efflux.

In Fig. 4, Radius of Gyration (Rg) analysis evaluated the overall structural compactness of the protein. During the simulation, Rg values fluctuated between 39 Å and 46 Å, indicating that KPHS_11890 maintained a relatively compact globular conformation overall. This compact globular conformation facilitates the efflux pump’s stable presence and functionality in the cell membrane while providing a structural basis for dynamic conformational changes during substrate transport.

Simultaneously, in Fig. 5, a 3D Free Energy Landscape (3D FEL) built using principal reaction coordinates identified multiple energy wells within the protein. This feature indicates that KPHS_11890 visited multiple distinct metastable conformational states during the simulation, suggesting potential functional mechanisms related to conformational diversity. The presence of multiple energy wells implies that the protein exists in multiple active conformations, providing opportunities for designing inhibitors with different mechanisms of action. By designing inhibitors targeting different conformational states, multifaceted regulation of the KPHS_11890 efflux pump’s function can be achieved, thereby enhancing the therapeutic efficacy of antibiotics.

Collectively, the structural dynamics analysis, showing significant conformational changes, high flexibility in specific residue regions, maintenance of overall compactness, and the presence of multiple energy wells, provides a structural basis for understanding the functional mechanisms of the protein as an efflux pump (e.g., substrate transport) and potential regulatory strategies.

The global dynamic properties, as characterized by RMSD, Rg, and FEL analysis, collectively indicate that KPHS_11890 is a flexible protein that explores multiple conformational states while maintaining an overall compact structure. To identify the specific segments responsible for this structural integrity and pinpoint potential sites for therapeutic intervention, we performed a detailed analysis of per-residue fluctuations and secondary structure conservation throughout the simulation trajectory.

The AcrA protein serves as a critical component of the multidrug efflux pump system, actively expelling a range of antibacterial agents and toxic compounds, such as acridine derivatives, from bacterial cells. To elucidate the structure-function relationship, the specific secondary structure and 20 stable segments (Table 2, 3; Fig. 6) were extracted and their roles in protein functionality, protein-drug interactions, and drug design were systematically analyzed. These segments are distributed across AcrA’s main domains—the α-helical domain, the extended strand domain, and the random coil domain. These domains collectively maintain AcrA’s structural integrity and support the assembly and function of the tripartite efflux pump through interactions with AcrB and TolC [29].

In terms of methodology, the DRKG model constructed using the PyKEEN framework (AMR = 0.08, Hits@10 = 0.32) demonstrates superior biomedical relationship inference capabilities, outperforming traditional KG models in similar tasks (e.g., BioKEEN’s Hits@10 benchmark of 0.25) [30].

Notably, KPHS_11890 ranked in the top 5% of predicted targets, and its structural stability was validated through MD, highlighting the reliability of DRKG screening. This strategy aligns with recent trends in AI-driven drug discovery: AI-Saleem et al. successfully predicted potential COVID-19 drugs (e.g., HDIs) using DRKG, while this study adapts it for bacterial resistance target identification [7]. Crucially, this work presents more than a single target prediction; it establishes a generalizable and robust computational workflow. This DRKG-MD paradigm can be readily adapted to other pathogens or disease models, offering a template for future computer-aided drug discovery campaigns that require a seamless transition from large-scale data mining to structure-based validation.

KPHS_11890 belongs to the membrane fusion protein (MFP) family, with its homolog in E. coli, AcrA, being the most extensively studied exemplar. Our identification of stable segments within the α-helical domain is particularly significant, as this region in AcrA is known to form a long coiled-coil hairpin essential for bridging the inner membrane transporter AcrB with the outer membrane channel TolC. The structural stability of these segments, therefore, is likely critical for the assembly and function of the entire tripartite efflux pump, making them a prime target for inhibitors designed to disrupt this vital interaction.

The MFP family drives multidrug efflux by mediating the synergy between inner membrane transporters and outer membrane channels. Previous studies have confirmed that inhibiting MFP proteins can block efflux pump assembly, restoring antibiotic sensitivity. Its stable segments may serve as high-potential drug-binding sites, offering a breakthrough for overcoming resistance mechanisms of RND efflux pumps like AcrAB-TolC [31, 32].

The stable segments of KPHS_11890 highlight its distinctive druggability potential: unlike the conformationally dynamic AcrB protein, which is prone to mutation-driven resistance escape, these stable regions are more suitable for designing high-affinity inhibitors. Furthermore, the HS11286 strain, a dominant carbapenem-resistant Klebsiella pneumoniae (CRKP) in China [33], likely exhibits multidrug resistance partly due to the overexpression of efflux pumps such as KPHS_11890. Studies have demonstrated that efflux pumps, including those in the Major Facilitator Superfamily (MFS) like KpnGH, play a significant role in conferring resistance to carbapenems and other antibiotics in K. pneumoniae. Specifically, insertional inactivation of the MFS efflux pump gene kpnGH resulted in increased susceptibility to carbapenems such as ertapenem and imipenem [34]. Given that KPHS_11890 is also an MFS efflux pump, its overexpression may similarly contribute to the resistance profile observed in the HS11286 strain. Additionally, CRKP strains often show upregulated expression of various efflux pump genes, which can render β-lactam and fluoroquinolone antibiotics ineffective [35]. This highlights a key strategic advantage: targeting the structurally conserved regions of an adaptor protein like KPHS_11890 may offer a more robust therapeutic strategy than targeting the highly dynamic and mutation-prone substrate-binding pocket of the central transporter AcrB. Inhibitors binding to these stable segments could be less susceptible to resistance development, a critical consideration in the design of next-generation antibiotics.

Although the DRKG approach facilitates efficient predictions, its dependence on prior knowledge restricts its ability to identify targets absent from the database, presenting a notable limitation. Additionally, the therapeutic potential of KPHS_11890 requires further validation through in vitro experiments or clinical studies, a challenge to be addressed in future research.