pH-Sensitive Amino Lipid-Driven Pore Formation Enables Endosomal Escape of Lipid Nanoparticles Authors: Akhil Pratap Singh, Kana Shibata, Yusuke Miyazaki, Wataru Shinoda Published: 2026 ABSTRACT AbstractLipid nanoparticles (LNPs) have transformed nucleic acid delivery for vaccines and gene therapies, yet their efficiency remains limited by incomplete endosomal escape. While experimental studies have revealed endosomal membrane disruption facilitating cytosolic release, the molecular basis of this process remains poorly understood. Here, we employ molecular simulations to elucidate how LNPs fuse with the endosomal membrane and release their payloads. We identify multiple fusion pathways, with a dominant stalk-pore mechanism. Our simulations reproduce and rationalize key experimental observations, including the transfer of ionizable lipids from LNPs to the membrane, which promotes nucleic acid reorientation and facilitates stalk formation and expansion. Furthermore, lipid shape, pH sensitivity, membrane tension, and nucleic acid encapsulation emerge as critical molecular determinants of endosomal escape efficiency. These findings advance our mechanistic understanding of intracellular delivery and provide a framework for the rational design of LNPs with improved performance for gene therapies and RNA vaccines.Subject terms: Computational biophysics, Membrane biophysics, Biotechnology Introduction Nucleic acid-based therapies demonstrate significant promise across various applications, including viral vaccines, protein replacement therapies, cancer immunotherapies, cellular reprogramming, and genome editing1–4. However, the effective intracellular delivery of nucleic acid payloads remains a major challenge due to their inherent susceptibility to enzymatic degradation and poor bioavailability, primarily resulting from their large size and strong negative charge3,5. Successful gene therapy further requires not only targeted delivery of mRNA to specific cell types but also efficient cytosolic release, where the mRNA can engage the cellular translation machinery3,4. Among the various delivery platforms developed to address these challenges, lipid nanoparticles (LNPs) have emerged as the most effective non-viral carriers for siRNA and mRNA therapeutics, owing to their high drug-loading capacity, biocompatibility, and tunable properties6–11. The clinical success of LNP-based systems has been proven by the mRNA vaccines developed by BioNTech/Pfizer and Moderna, as well as the siRNA-based drug Patisiran®12–15. Despite these advances, key barriers remain to achieving the LNPs’ full potential16: Although LNPs are readily internalized via endocytosis, they often become trapped within endosomal compartments and are subsequently degraded during endosomal maturation and acidification17,18. Inefficient endosomal escape is widely recognized as one of the principal limitations of LNP-based RNA delivery19,20. For instance, studies have shown that only ~2% of siRNA cargo reaches the cytosol following LNP delivery21. For RNA therapeutics to be effective, endosomal escape must occur prior to degradation. Currently, this process is considered a bottleneck in LNP development22. While viral vectors naturally overcome this challenge by fusing with the cellular membranes through specialized proteins, LNP would offer superior alternatives due to their ability to carry larger payloads, along with improved safety and lower immunogenicity, provided that endosomal escape can be sufficiently enhanced23. Current efforts to improve this process have focused primarily on tuning the lipid composition of LNPs11,24,25. For example, LNP containing cholesterol has been shown to enhance delivery efficiency, likely due to its ability to stabilize high-curvature membrane regions involved in LNP–endosome fusion26. Furthermore, ionizable lipids have proven particularly effective: these lipids remain neutral at physiological pH but become positively charged in the acidic endosomal environment, promoting electrostatic interactions with anionic endosomal membrane (EM) lipids and thereby facilitating membrane destabilization and fusion, leading to endosomal escape8,27–31. Despite advances in LNP formulation, relatively few experimental studies using spectroscopic and microscopic techniques have directly investigated the endosomal escape process21,32,33. A recent study revealed that damage to the EM, indicated by galectin recruitment, facilitates the release of nucleic acid into the cytosol32. This study also reported the enrichment of ionizable lipids in EMs and the disintegration of LNPs, both of which contribute to membrane destabilization and disruption, potentially leading to nucleic acid release. However, these experimental studies have not fully elucidated the precise molecular mechanisms underlying endosomal escape, largely due to inherent challenges of visualizing such dynamic, nanoscale processes at molecular resolution. Molecular dynamics (MD) simulations provide a powerful in silico approach for exploring these mechanisms at the molecular level. While all-atom (AA) MD simulations can deliver highly detailed insights, they are computationally prohibitive for systems as large and complex as LNPs. In contrast, coarse-grained (CG) simulations provide an effective way to simulate millions of particles over microseconds by using larger timesteps and a smoother energy landscape34,35. CG simulations, therefore, have become increasingly important for studying the architecture of large assemblies, such as LNPs and lipoplexes, as well as the fusion events between DNA-lipoplexes and EM35,36. However, many CG approaches rely on transferable, modular “building-block” parameterization strategies and do not explicitly validate cross-interactions involving ionizable amino lipids across diverse chemical environments36,37. Moreover, CG models are often validated using fixed configurations, which limits their broader applicability38,39. Our previous work has shown that accurate modeling of LNPs requires CG parameters that reliably capture both inter- and intramolecular interactions of ionizable amino lipids34. These parameters are crucial for developing reliable CG models tailored to specific LNP systems, enabling unbiased investigations of LNP nanostructures that are initiated from randomly organized components and their interactions with specific targets such as EMs. Building on this foundation, the present study applies a consistent CG framework to investigate lipid-mediated interactions between pH-sensitive lipid 1,2-dilinoleyloxy-3-dimethylaminopropane (DLin-MC3-DMA, or MC3) containing LNPs and endosomal membranes, with the aim of generating mechanistic insight and testable hypotheses rather than exhaustive statistical characterization. Crucially, we explicitly incorporate the lipid's pH-dependent protonation states to mimic the endosomal environment. Nucleic acids are modeled as 12-base-pair double-stranded DNA (dsDNA; PDB: 1BNA) to simplify the complexity of nucleic acid structure and dynamics. While double-stranded RNA and DNA share broadly similar electrostatic properties, differences in flexibility and nonbonded interactions become more pronounced for longer strands. Our CG model (SPICA force field) was developed with parameterizations derived from AA simulations, enabling us to capture specific nonbonded interactions between LNP constituents and realistic EM components, including phosphatidylcholines (PC), cholesterol (CHOL), phosphatidylethanolamine (PE), phosphatidylserine (PS), and phosphatidylinositol (PI). Our simulations reveal that protonated MC3 interacts strongly with anionic lipids on the luminal side of the endosomal membrane under acidic conditions, leading to localized membrane remodeling and the formation of non-bilayer structures such as stalk intermediates. These structural changes drive fusion-pore formation and subsequent membrane rupture, facilitating the release of nucleic acid payloads into the cytosol. Furthermore, our findings reveal that amino lipids are crucial for stalk expansion, ultimately leading to pore formation and the release of dsDNA. Additionally, our results emphasize the importance of lateral membrane tension in EMs, as well as the spatial distribution of nucleic acids within LNPs, in promoting efficient release of the payload. Overall, this study provides molecular-level insights into the endosomal escape mechanisms of LNPs, offering rational design principles for developing next-generation delivery systems with enhanced therapeutic efficacy. Results Lipid-DNA interactions To enable LNPs simulations with a model EM, we first determined the nonbonded interaction parameters between nucleic acids and EM lipids that were not available in the SPICA molecular library. Notably, the SPICA force field (FF) already contains all the inter- and intramolecular interaction parameters for MC3 lipids, CHOL, PC, and nucleic acid (dsDNA) (Supplementary Fig. 1A)34,40–42. However, the nonbonded interactions between MC3 or dsDNA and EM lipids (Supplementary Fig. 1B)—PS and PI—were missing in the SPICA FF. In this study, we explicitly optimized these parameters to reproduce reference data from AA simulations. These data included the surface tension of binary lipid monolayers, density profiles of CG sites, 2D radial distribution functions (RDFs) between the centers of mass of the lipids in the binary monolayers, and the free energy profiles of nucleobases and deoxyribose (DERB) across PS and PI membranes. The optimized CG parameters closely reproduced the AA reference data (Supplementary Figs. 2, 3, and Supplementary Table 1). MC3 lipid programmed LNP-endosome fusion pore formation and expansion To investigate the molecular origin of endosomal escape mechanism, including the puzzling EM damage observed experimentally32, we performed unbiased CG MD simulations of hydrated EM with solvated LNPs (Fig. 1A), using pre-equilibrated structures described in the Methods section. We simulated four systems, each for 20 to 30 μs, under slightly different conditions regarding ion concentrations and varying ratios of charged to neutral MC3 lipids in the LNPs. All simulated systems, along with their compositions and conditions, are summarized in Supplementary Tables 2, 3. The LNP with EM systems in these in silico experiments are denoted as pH4LI, pH5PI-1, pH5PI-2, and pH5PI-3. The pH4LI was simulated at low ionic strength (LI), containing only counterions, and represents a more acidic condition (approximately pH 4), as previous reports have indicated that ionic composition varies along the endocytic pathway43. In contrast, pH5PI-1, pH5PI-2, and pH5PI-3 are independent replicates simulated under physiological ion strength (PI; 150 mM NaCl) at approximately pH 5. The differences observed among the physiological salt conditions arise from stochastic variations in initial configurations rather than from changes in parameters. LNP-endosome fusion pore formation and expansion. A Schematic representation of an LNP entrapped within an endosome. Inset: CG representation of the LNP-EM system. Color scheme: Different lipid types are shown in distinct colors. Lipid headgroups, tails, and dsDNA molecules are depicted using van der Waals (VDW), Licorice, and QuickSurf (iso-surface) representations, respectively. For clarity, water and ions are omitted. B Time evolution of the fraction of migrated MC3lipids in the EM after stalk formation, showing rapid translocation in all systems. The fraction of migrated MC3 lipids was calculated relative to the total number of MC3 molecules. Error bars represent standard errors (SE) estimated by block averaging. Data are presented as mean values +/- SE. C, D Representative snapshots showing the time evolution of the LNP-EM system during MD simulations at C first fusion site and D second fusion site, leading pore formation and expansion. The color scheme is the same as in panel (A). Inset: Representative snapshots of membrane fusion stages: (i) At the first fusion site, two closely apposed bilayers form an HD via an intermediate reverse micellar (RM) phase. (ii) At the second fusion site, radial expansion of the stalk and subsequent HD formation initiate and enlarge a small fusion pore. Color scheme: Pink for CG atoms of lipid headgroups, green for lipid tails. For clarity, water, ions, and dsDNA molecules are omitted. E Cross-sectional view of the LNP-EM complex at 16.3 μs, showing the solvent-accessible surface area (SASA) represented by iceblue color. An inset highlighting a cylindrical region at the narrowest pore for counting water molecules. F The semi-quantitative estimation of the narrowest pore water volume (NPWV) reveals significant pore expansion occurring in both pH4LI and pH5PI-1 systems. The NPWV analysis used snapshots from a 100 ns trajectory, with each data point representing measurements taken at regular intervals. The reported statistics were derived from these values. In the box plot, the center line represents the median, the box spans the 25th to 75th percentiles, the whiskers extend to the most extreme data points within 1.5 times the interquartile range, and open circles indicate outliers beyond the whiskers. System pH4LI and pH5PI-1 underwent successful fusion of the LNPs with the EM, resulting in delivery of the payload into the cytosol. Notably, pH5PI-1, at physiological ion concentrations (150 mM NaCl) and containing LNPs with a mixture of charged and neutral MC3 lipids (representing slightly higher pH~5 conditions), yielded qualitatively similar results to those described here. In contrast, pH5PI-2 and pH5PI-3 exhibited only membrane fusion activity but did not show transfection within the 30 μs simulation time. We will discuss this further below. Analysis of MD trajectories revealed that LNPs approached the EM within tens of nanoseconds at an acidic pH (Fig. 1C). However, control simulations conducted at neutral pH exhibited no binding between LNPs and the EM during the 1 μs simulation time, due to the absence of protonated MC3 lipids (Supplementary Fig. 4). This emphasizes the specific role of MC3 protonation under acidic conditions and is consistent with the expectation that strong electrostatic interactions between the cationic MC3 lipids and anionic PS/PI lipids drive the initial contact. However, fusion did not proceed at this stage due to insufficient interfacial dehydration, a known kinetic barrier to membrane fusion44. To address this, we modified the initial setup to include a dehydrated interface, with one corner of the LNP oriented toward the EM, resulting in successful stalk formation (Fig. 1C). Interestingly, LNP-EM fusion followed two distinct pathways, both culminating in a hemifusion diaphragm (HD) state45, in which the outer leaflet of the LNP and EM merged and exchanged lipids (Fig. 1C, D, and Supplementary Fig. 5). At the initial fusion site, a lipid aggregation such as a reverse micelle (RM) formed and subsequently evolved into the HD structure (inset of Fig. 1C). No further changes occurred at this site. However, it is worth mentioning that we observed the migration of MC3 lipids in the EM during HD formation across all simulated systems (Fig. 1B, and Supplementary Fig. 6), consistent with experimental report32. This is likely due to the unique structure of MC3 lipids, which feature a charged headgroup and polyunsaturated tails, causing greater disruption at the fusion site compared to conventional lipid bilayer or vesicles and typical membrane fusion stalk formation (Supplementary Fig. 7). Importantly, the migration of cationic MC3 lipids into the EM creates an ionic imbalance in the LNPs. This imbalance triggers the rearrangement and accumulation of dsDNA near the fusion site, which we will discuss further in the next section. Consequently, this dsDNA reorganization promoted the formation of an expanded stalk-like intermediate, which rapidly progressed into the formation and expansion of a fusion pore (Fig. 1D, E). This fusion pore enabled the release of the dsDNA payload into the cytosol. Notably, we observed distinct time evolutions for pore formation and expansion in both pH4LI and pH5PI-1 systems (Fig. 1D, Supplementary Figs. 1F, 5). These findings are consistent with the stalk-pore fusion model widely supported in prior membrane fusion studies46–48. All things considered, protonated MC3-lipid maps out an active fusion site for pore creation, presumably causing EM damage, in addition to the LNP and EM interaction, in agreement with experimental reports32,49. Cationic-MC3, PE, PS lipids facilitate membrane fusion LNP transfection efficiency is significantly influenced by the shape and arrangement of lipid components38. To understand how different lipids affect fusion pore formation, we first visually analyzed lipid migration and sorting during fusion (Fig. 2A). We then quantified the number of lipids transferred between the LNP and EM. Lipid transfer from the LNP to the EM occurred in the order: MC3 > CHOL > DSPC. Conversely, lipid transfer from the EM to the LNP followed the trend: PI > PS > PC > PE > CHOL (Supplementary Table 4). These trends were further supported by measurements of lipid orientation relative to the membrane normal. (Supplementary Fig. 8). Structural relaxations at fusion points and lipid sorting. A Representative snapshots showing the time evolution of the MD simulation of LNP-EM, highlighting lipid sorting. Color scheme: gray for MC3, purple for POPE, cyan for POPS, yellow for POPI, and green for cholesterol. For clarity, water, ions, and dsDNA molecules are omitted. B Cross-sectional view of the fusion points at 12 μs. Lipid compositions at both fusion points are calculated using a cylindrical region, as illustrated. C Comparison of lipid compositions at the fusion points over the time interval from 11.8 μs to 12.1 μs, based on the cylindrical fusion region shown in panel (B). Errors were estimated using block averaging. To quantify lipid sorting at the fusion site, we examined the local lipid composition using a cylindrical region centered on the pore (Fig. 2B, C). PE and PS lipids comprised 11% and 8% of lipids, respectively, at the active fusion site, while PI lipids dominated the blocked fusion site. As expected, negatively charged PI and PS lipids were enriched at fusion sites due to electrostatic attraction to the cationic MC3 lipids. Our results suggest that cone-shaped lipids, such as PE and PS, preferentially accumulate at successful fusion sites, lowering the energy barrier for pore formation, consistent with prior reports50–52. pH-Sensitive MC3 Lipids and EM Curvature Modulate dsDNA Release Next, we examined how the protonation states of MC3 lipids influence fusion pore dynamics in response to pH changes (Fig. 3A). We emphasized the two major scenarios. In the first scenario, all MC3 lipids remained protonated, resulting in slow fusion pore expansion and limited dsDNA release (Supplementary Fig. 8, Supplementary Movies 1, 2), likely due to strong electrostatic interaction among cationic MC3 and negatively charged dsDNA molecules. In the second scenario, as a consequence of pore formation, we mimicked cytosolic influx into the endosome by deprotonating 64% of MC3 lipids at ~17 µs, consistent with prior experimental estimates31. Of note, 64% of MC3 lipids were deprotonated to approximate a pH of ~6.6, based on the experimentally determined pKa of MC3 (6.44) and experimental reports showing that most release events occur upon a pH increase from ~5.6 to ~6.6 following endosomal disruption53. It is also worth noting that the post-rupture state was modeled as a system-level approximation, since spatially resolved pH gradients are beyond the scope of the present coarse-grained framework. Thus, the reduction in charge weakened electrostatic interactions, facilitating pore enlargement and enhanced dsDNA release (Fig. 3B, C, Supplementary Movies 3, 4). dsDNA release is regulated by pH-sensitive MC3 lipids and lateral tension. A Molecular structures of the MC3 lipid headgroup at acidic and neutral pH. B, C Representative snapshots of a single EM-LNP fused complex after randomly replacing 64% of the cationic MC3 lipids with neutral ones at (B) 17.1 μs and (C) 22.6 μs. The color scheme is identical to that in Fig. 1C. The translucent ice blue represents the water phase. For clarity, ions are omitted. D The time evolution of the simulation box length in the X-direction for both systems experiencing pore formation is presented. Error bars represent standard errors (SE) estimated by block averaging. Data are presented as mean values +/- SE. E Representative snapshots showing the final configuration of the simulated system after applying lateral tension, which accelerates dsDNA release into the cytosol and leads to closure of the membrane pore at 27 µs. The lipid color scheme is the same as in Fig. 1C. Cyan indicates unreleased dsDNA molecules adsorbed to the inner leaflet of the EM; purple indicates dsDNA molecules released into the cytosol. F An applied lateral tension to the EM at 22.6 μs in pH4LI and at 14.3 μs in pH5PI-1 resulted in an approximate 40% increase in dsDNA release in the cytosol, rising from 30% to 70%. In a box plot, the central line denotes the median, while the box represents the interquartile range (25th–75th percentiles). Whiskers extend to the most extreme observations within 1.5× the interquartile range, and points beyond this range are plotted as open circles to indicate outliers. To quantify this effect, we monitored simulation cell size in the lateral direction (Fig. 3D), revealing significant EM expansion following deprotonation, consistent with increased lipid migration. These results highlight the critical role of MC3 protonation in regulating pore dynamics and nucleic acid release, in agreement with experimental observations11,33. Despite this, dsDNA release plateaued after 22 µs in the pH4LI system and after 14 µs in the pH5PI-1 system. To further enhance release, we adopted tension-induced fusion, given its crucial role in previous studies54,55. Following pore formation, osmotic swelling of the endosome is expected to generate membrane stress54. In this context, we modeled the post-rupture state by applying lateral tension to the endosomal membrane, implemented as described in previous studies56. Moreover, previous in silico studies have shown that membrane fusion efficiency increases under lateral tension55. We thus applied lateral tension to the EM at 22.6 μs (Fig. 3C) in pH4LI and at 14.3 μs in pH5PI-1 systems, as described in the Methods section, and continued the simulations. It is worth mentioning that the magnitude and timing of the applied tension were chosen to probe the qualitative impact of membrane stress on pore dynamics, rather than to reproduce a specific physiological condition. This intervention dramatically improved dsDNA release, with 70% of encapsulated dsDNA entering the cytosol (Fig. 3E, F). These results suggest that membrane curvature and associated tension significantly enhance transfection efficiency. Nucleic acid distribution within LNPs influences release efficiency Finally, our investigation focused on how the spatial distribution of payloads within LNPs influences endosomal escape. It is important to note that LNP formulation is governed by a kinetically driven assembly process, which can lead to heterogeneous payload distributions, as demonstrated in previous experimental studies57,58. Visual analysis of MD trajectories revealed that prior to fusion, several dsDNA molecules protruded from the LNP surface (Fig. 4A). After fusion, most dsDNA molecules accumulated beneath the LNP’s outermost bilayer (Fig. 4B). Fusion pore formation occurred at regions where the outer bilayer of the LNP contacted the outer leaflet of the EM, leading to cytosolic release of dsDNA (Fig. 4C). In contrast, dsDNA located at the LNP surface failed to escape, as this region fused with the inner leaflet of the EM (Fig. 4E). It is also worth mentioning that the distribution of dsDNA molecules in the LNP-EM fused complexes of pH5PI-2 and pH5PI-3, which did not undergo pore formation, differed somewhat from the systems where pore formation occurred (Fig. 4F–I, Supplementary Fig. 10). In the systems without pore formation (Supplementary Fig. 11), most dsDNA remained linked to the molecules protruding from the LNP surface. Subsequently, these molecules aligned along the principal axes of the dsDNA and migrated toward the aqueous phase on the luminal side (Supplementary Fig. 11). However, no DNA was released from the complex into the cytosol during the 30μs time interval, presumably resulting in a more relaxed system. The LNP-EM complex highlights released and unreleased dsDNA molecules, as well as distinct dsDNA distributions in systems with and without pore formation. A–E Time evolution of the MD simulation of the LNP-EM complex, highlighting released and unreleased dsDNA molecules in the systems with pore formation (pH4LI). Gray and green spheres represent the lipid headgroups of the LNP and EM, respectively. The dsDNA color scheme is the same as in Fig. 3E. F Representative snapshots of a single EM-LNP fused complex in systems with and without pore formation at ~9 μs. The color scheme is identical to that in Fig. 1C. The arrows in pH5PI-2 indicate the migration of dsDNA into the aqueous phase on the luminal side. For clarity, results from pH5PI-1 (with pore formation) and pH5PI-3 (without pore formation) are omitted, as they are similar to those shown here. G Cross-sectional view of the LNP-EM complex with the solvent-accessible surface area (SASA). Light gray and green transparent regions represent the SASA of the LNP and EM, respectively. H Probability distributions of dsDNA in the LNP-EM fused complex at ~8 μs from the LNP surface toward its interior. The distance from the LNP surface, R1, is defined as the inward projection of R1=Ri−rSASA along the LNP surface normal, shifted by the radius of the SASA probe particle. I The dsDNA amount within 6 nm (1/4th the diameter of the LNP) of the LNP surface, at around 8 μs, shows an increase by about 20% in pH5PI-2 and pH5PI-4 systems, which did not show pore formation, unlike others. Error bars represent standard errors (SE) estimated by block averaging. Data are presented as mean values +/−SE. These results indicate that full encapsulation of nucleic acids beneath the LNP bilayer enhances release efficiency, whereas DNA associated with the surface inhibits transfection. Additionally, successful pore formation and transfection require proper alignment of dsDNA molecules towards the fusion site, rather than protruding outward. Overall, our findings suggest that optimizing the internal localization of nucleic acids is crucial for maximizing delivery efficiency. Discussion In this study, we employed CG MD simulations to investigate the endosomal escape mechanism of LNPs and elucidate the molecular details of their interactions with endosomal membranes, thereby extending beyond the limitations of current experimental methods. While the molecular mechanisms governing nucleic acid transfer from lipoplexes have been extensively studied35, our work distinguishes itself by explicitly modeling the endosomal membrane and a double-stranded nucleic acid-LNP therapeutic analogous to Patisiran®, using a consistent CG framework rather than the modular “building-block” approach often used in multiscale modeling. This approach was motivated by two key considerations in addition to exploring EM damage as observed experimentally32: first, the importance of detailed endosomal escape mechanisms for the clinical efficacy of Patisiran®, and second, the potential to leverage these insights for the rational design of more efficient nucleic acid delivery systems. Our simulations successfully captured LNP-endosomal membrane fusion events, including multiple fusion sites leading to dsDNA transfection. We observed that the fusion pore formation represents a plausible route for endosomal escape, consistent with experimental findings21,32. However, not all fusion sites developed into functional pores. This suggests a two-step free energy landscape: (1) lipid rearrangement during stalk formation, and (2) transition from HD to pore formation—both consistent with prior studies48. At the initial fusion site, the process stalled at an amorphous HD structure, where multiple membranes from the LNP and the endosomal membrane remained entangled, likely due to the high energetic cost of PI-rich regions. In contrast, the second fusion event, involving stalk expansion, proceeded through the stalk–HD–pore pathway, facilitating dsDNA release. Notably, our results revealed that the charged state of MC3 lipids and their polyunsaturated tails disrupted the fusion site, promoting MC3 migration from the site of fusion into the EM. This migration drove nucleic acid reorientation within the LNP, leading to dsDNA accumulation at the fusion site, which facilitated stalk expansion and ultimately pore formation. We further demonstrated that the lipid structure and composition at the fusion site critically influence pore formation and transfection efficiency, supporting previous studies59. Cone-shaped lipids such as PS, PE, and cationic MC3 preferentially accumulated at successful fusion sites, lowering the energy barrier for pore formation. In contrast, PI-rich regions, despite facilitating initial fusion through strong electrostatic interactions, failed to complete pore formation. Next, we examined the effect of pH-sensitive lipid protonation states on fusion dynamics. Initially, under acidic conditions, protonated MC3 lipids promoted LNP-endosomal membrane interactions via electrostatic attraction. Upon pore formation, partial deprotonation of MC3 lipids, mimicking cytosolic pH, reduced these interactions, facilitating rapid pore expansion and nucleic acid release. These observations highlight the significance of dynamic charge regulation in MC3 lipids for regulating endosomal escape efficiency. Although we could not simulate a fully enclosed LNP within a curved endosome due to computational limitations, we investigated the role of membrane curvature by applying lateral pressure to the endosomal membrane. This intervention, which mimics curvature-induced stress, substantially enhanced dsDNA release. Our findings suggest that membrane curvature and local tension may be critical for achieving complete nucleic acid delivery. Moreover, our results highlight the significant influence of the spatial distribution of dsDNA molecules within LNP-EM fused complexes, primarily caused by the migration of MC3 into the EM. In our simulations, the two systems that contained dsDNA molecules beneath the LNP bilayer successfully transfected most of the dsDNA, except for those located at the LNP surface. In contrast, the other two systems, which did not undergo pore formation, showed rapid leakage of dsDNA during rearrangements within the LNP-EM complex. This leakage prevented the stalk expansion and pore formation. This issue likely stems from the small size of the LNPs, which increases the risk of nucleic acid leakage60. Therefore, LNP size indirectly influences transfection success. The present study focuses on small LNPs encapsulating dsDNA; however, future studies should extend this analysis to larger LNPs or other nucleic acid cargos, such as mRNA, as both cargo size and structure are expected to play key roles in determining payload distribution. Importantly, these simulation-based insights align with experimental observations from microscopic study11,21,32,33 which have revealed transient pore or EM damage and stepwise fusion pathways similar to those observed here. Moreover, LNP disintegration, MC3-lipids migration from LNP to EM, also impact the endosomal escape process, consistent with experimental observations32. Previous simulation studies also support the strong dependence of membrane fusion behavior on local lipid composition and mechanical stresses48,51. Nonetheless, our study has limitations. As with any CG-MD approach, discrepancies exist between simulation and experimental conditions, including differences in RNA type (siRNA vs. mRNA), RNA-to-lipid ratios, salt identity and concentration, and the absence of natural endosomal curvature and proteins. The sampling remains limited relative to the complexity of biological membranes. These factors could influence the kinetics and pathways of fusion. Future studies should systematically vary these parameters and incorporate diverse lipid compositions to better understand the energetics and topological constraints governing LNP-endosome fusion. Additionally, the present EM model simplifies lipid-mediated interactions and intentionally excludes several relevant components, such as membrane-associated proteins (e.g., Rab, EEA1), PI(3)P, glycosphingolipids, and galectins, which play roles in endosomal maturation and cellular responses61,62. The observed pore-like membrane perturbations should not be interpreted as a universal mechanism for endosomal escape, but rather as a lipid-driven pathway for localized cargo release under specific conditions. These perturbations involve localized lipid mixing and transient aqueous connectivity, not a full lumen-to-cytosol pore. Such localized damage is nevertheless compatible with experimentally reported focal recruitment of galectins to perturbed membrane regions32, as opposed to requiring complete membrane rupture. Future studies could incorporate more endosomal components to better link lipid-driven perturbations to biological outcomes. In the context of pH, a simplified protonation model was used for MC3 lipids, adjusting the protonation fraction based on the neutral-to-cationic ratio and the pKa of MC3 (6.44)31, yielding an ionization level representative of mildly acidic endosomal conditions. While constant-pH molecular dynamics (CpHMD) can capture the coupling between protonation states and conformational changes under uniform pH63–65, it does not account for the spatially heterogeneous pH environments found in endosomes, where a membrane separates an acidic luminal region from a near-neutral cytosolic region. Additionally, CpHMD can exhibit slow pKa convergence and is dependent on specific force fields66. Therefore, more rigorous methods could be employed to address the pH gradient in LNP-EM simulations more accurately. Further, we emphasize that our analysis revealed consistent qualitative features across both successful and unsuccessful replicas, supporting the robustness of the proposed mechanism. While additional replicas would further refine statistical confidence, the primary mechanistic insights into lipid-mediated endosomal escape are expected to remain unchanged, with only the relative frequencies of observed outcomes varying due to the inherently out-of-equilibrium nature of the simulations. In conclusion, we present a detailed in silico investigation of LNP fusion with a model endosomal membrane, offering mechanistic insights into the endosomal escape process. Our results pave the way for the rational optimization of LNP formulations. Furthermore, our simulations provide interesting insights beyond the specifics of pore formation, which presumably cause EM damage. It also reproduces and rationalizes experimental observation, which validates a posteriori the quality of our force fields and simulation parameters. Thus, future work could extend these findings by incorporating cell-type-specific endosomal compositions and maturation states to develop task-specific delivery strategies. Importantly, such efforts do not necessarily require new CG FF development but rather a more comprehensive application of existing models. Taken together, our findings provide a mechanistic framework to improve the endosomal escape efficiency of RNA-LNP therapeutics and expand their clinical potential. Methods AA-MD Simulations AA-MD simulations were conducted to generate reference data for SPICA CG modeling of nonbonded interactions between MC3 lipids, dsDNA, and EM components—specifically PS and PI lipids. Simulations were performed using NAMD version 2.1267 and GROMACS 201868 with the CHARMM 36 force field69,70 and TIP3P water model71. All systems were simulated in the NPT ensemble using a 2 fs integration time step. Electrostatic interactions were computed using the particle-mesh Ewald (PME) method72. Lennard-Jones (LJ) interactions were smoothly truncated using a force-switching function between 10 and 12 Å. NAMD was specifically used for free energy calculations with the Colvars module as detailed below73,74. A Langevin thermostat (damping coefficient: 5 ps−1) controlled the system temperature at 310 K. Pressure was maintained at 1 atm using the Langevin piston algorithm75 with a piston period of 200 and a decay time of 100 fs. Semi-isotropic pressure coupling was applied for simulations of lipid bilayer systems. In GROMACS simulations, temperature and pressure were maintained at 310 K and 1 atm using the Nosé-Hoover thermostat76,77 and Parrinello-Rahman barostat78, respectively. CG-MD simulations CG-MD simulations of LNP, EM, and combined LNP-EM systems were performed using either LAMMPS79,80 or GROMACS68 with the SPICA force field34,41,81. LJ interactions were truncated at 1.5 nm. Electrostatics were calculated using the particle-particle particle-mesh method in LAMMPS and PME72 in GROMACS. System temperature was controlled via the Nosé-Hoover thermostat76,77 or a velocity-rescaling method, and pressure was maintained with the Parrinello-Rahman barostat78 using semi-isotropic coupling. A 10-fs integration time step was used. For the LNP-EM system, simulations included a 1 ns equilibration (NVT ensemble) and a 30 μs production run (NPxy PzT ensemble) at 310 K and 1 atm. Surface tension of monolayers To obtain reference data for MC3–PS/PI nonbonded interactions, AA-MD simulations were conducted for binary monolayer systems. MC3 lipids were arranged at a molecular area of 0.80 nm2, representing the liquid-phase monolayer. The assumed area per lipid is 0.70 nm2 for both POPS and POPI. Surface tension convergence in AA-MD was carefully monitored. Subsequently, CG-MD simulations of equivalent monolayer systems were conducted to reproduce the AA-derived surface tension by optimizing CG nonbonded interaction parameters. The resulting CG surface tension values (listed in Supplementary Table 1) show good agreement with AA-MD results. Density profile of the CG segments AA trajectories of POPS/MC3 and POPI/MC3 systems were mapped to CG representations using SPICA tools (https://github.com/SPICA-group/spica-tools). Density profiles of individual CG segments along the monolayer normal were calculated to further guide parameterization. As shown in Supplementary Fig. 1, the segment distributions from CG-MD closely match those obtained by mapping AA trajectories. Free energy profile of nucleobase permeation To refine lipid-DNA nonbond interactions, potential of mean force (PMF) profiles were calculated for nucleobases translocation across POPS and POPI bilayers (Supplementary Fig. 2). For obtaining the PMF, the adaptive biasing force (ABF) method was used, with the reaction coordinate defined along the membrane normal (z-axis), where the center of the membrane was set as z = 0. The coordinate space was divided into 5 Å windows up to 30 Å from the membrane center. Each window was simulated for at least 300 ns to observe the convergence of the free energy profiles. Both the AA and CG simulations utilized the same system setup. NAMD and LAMMPS were used exclusively for free energy calculations, employing the Colvars module for the AA and CG simulations, respectively. Notably, AA free energy profiles for nucleobase permeation, derived using the CHARMM force field, were used as reference data to optimize the CG nonbonded interactions between nucleobases and PS/PI lipids. Following this, the CG nonbonded interactions were refined through iterative CG -MD simulations to reproduce the AA free energy profiles accurately (Supplementary Fig. 3). Preparation of the EM-LNP System A simple EM model composed of 90 POPC, 60 cholesterol, 22 POPE, 14 POPS, and 14 POPI lipids was built using CHARMM-GUI, based on previous reports82,83. This model was converted to CG resolution using SPICA tools and replicated seven times in both x and y dimensions to construct a large planar EM. The specifics of each component in the large EM are provided in Supplementary Table 2. The LNP contained 6400 MC3 lipids, 4992 cholesterols, 1284 DSPC lipids, and 64 dsDNA molecules (PDB ID: 1BNA), based on prior studies (Supplementary Table 3)34. The details regarding the construction of LNPs at various pH levels can be found in the Supplementary Methods section. The equilibrated LNPs were placed above the equilibrated EM for the subsequent simulations. To achieve a low ionic strength while including only counterions and maintaining a physiological NaCl concentration, we added additional water and sodium chloride (NaCl) to the system with the LNP and EM. After this adjustment, we further equilibrated the LNP-EM systems for several more nanoseconds before conducting the production run. Dehydration of the LNP-membrane interface To initiate fusion, water molecules were removed from the contact region between LNP and EM by applying a spherical wall potential: where ri is the distance of CG water particle i from the center of the sphere, k is the force constant (10000 kcal/mol/nm3), and R is the sphere radius (5 nm). The sphere center was positioned at the interface between the LNP and EM. This potential was applied for 100 ns, starting from 700 ns into the simulation. Angular distribution of lipids To analyze lipid accumulation in the fusion zone (Supplementary Fig. 6), the tilt angles of EM lipids relative to the membrane normal were computed. Tilt angles were defined as the angle between the lipid head-tail bond vector and the membrane normal, using an in-house Python script. Trajectory data from 1.3 to 1.4 μs were used for this analysis. Lateral tension To accelerate the membrane fusion, lateral tension was applied to the EM. The surface tension γ was computed as: where 〈Pαα〉 (α = x, y, z) denotes the ensemble-averaged pressure tensor and Lz is the simulation box length in the z-direction. A target surface tension of 4 mN/m was set, consistent with experimental conditions84. At 23.2 μs, Lz was 40.7 nm, requiring Pxx and Pyy, to be maintained at - bar. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Description of Additional Supplementary Files Supplementary Movie 1 Supplementary Movie 2 Supplementary Movie 3 Supplementary Movie 4 Reporting Summary Transparent Peer Review file Source data Source Data Acknowledgements This work was supported by the supercomputer facilities of the Institute for Solid State Physics, the University of Tokyo (ISSPkyodo-SC-2023-Ea-0002, 2023-Eb-0004, 2024-Ea-0001, 2024-Eb-0003, 2025-Ea-0002), and the Research Center for Computational Science, Okazaki, Japan (Project: 25-IMS-C094, 24-IMS-C090, 23-IMS-C095). Author contributions A.P.S. led the data analysis, figure preparation, and interpretation of results, and wrote the main draft of the manuscript. K.S. performed the molecular dynamics simulations and initial data analysis. Y.M. contributed to the revision and improvement of the figures. W.S. conceived and designed the study, supervised the project, provided critical input on data interpretation, and substantially revised the manuscript. All authors discussed the results and contributed to the final version of the manuscript. Peer review Peer review information Nature Communications thanks anonymous reviewers for their contribution to the peer review of this work. A peer review file is available. Funding This project has received funding from JSPS KAKENHI Grant Numbers JP21H01880, JP24H00038, JP24H00843, and JP25K01729. Data availability The data generated and analyzed in this study are publicly available via Zenodo at https://zenodo.org/records/17489583. These datasets are provided to facilitate further research and enable independent verification of the results reported in the main and Supplementary Figs. Source data underlying Figs. 1B, 1F, 3D, 3F, 4H, and 4I and Supplementary Figs. 2B–E, 3B–E, 4B, 9, 10, and 13 are provided with this paper in the associated Source Data file. Further information on the SPICA force field, including tutorials and parameter files, is available at https://www.spica-ff.org. Source data are provided with this paper. Code availability The custom Tcl and Python scripts used for data analysis in this study are publicly available via Zenodo at https://doi.org/10.5281/zenodo.21414460. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Akhil Pratap Singh, Kana Shibata. Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-76810-y. References Associated Data Supplementary Materials Supplementary Information Description of Additional Supplementary Files Supplementary Movie 1 Supplementary Movie 2 Supplementary Movie 3 Supplementary Movie 4 Reporting Summary Transparent Peer Review file Source Data Data Availability Statement The data generated and analyzed in this study are publicly available via Zenodo at https://zenodo.org/records/17489583. These datasets are provided to facilitate further research and enable independent verification of the results reported in the main and Supplementary Figs. Source data underlying Figs. 1B, 1F, 3D, 3F, 4H, and 4I and Supplementary Figs. 2B–E, 3B–E, 4B, 9, 10, and 13 are provided with this paper in the associated Source Data file. Further information on the SPICA force field, including tutorials and parameter files, is available at https://www.spica-ff.org. Source data are provided with this paper. The custom Tcl and Python scripts used for data analysis in this study are publicly available via Zenodo at https://doi.org/10.5281/zenodo.21414460.