Lipid Nanoparticles in Gene Therapy

Updated 7 July 2026

Lipid Nanoparticles (LNPs) are nanoscale delivery vehicles composed of ionizable lipids, helper phospholipids, cholesterol, and PEG-lipids that encapsulate nucleic acids for therapeutic applications.
They self-assemble via rapid mixing and pH-triggered reorganization to facilitate endosomal escape and optimize targeting, critical for mRNA vaccines and siRNA therapeutics.
Advanced computational models and multi-modal characterization techniques, including SAXS and cryo-TEM, guide the scalable design and optimization of LNP formulations.

Lipid nanoparticles (LNPs) are the leading nonviral vehicles for delivering nucleic acids such as mRNA and siRNA. In the canonical therapeutic formulation, an ionizable lipid, a helper phospholipid, cholesterol, and a PEG-lipid self-assemble into nanoscale carriers that stabilize labile cargo, mediate endosomal escape, and can be programmed through composition, structure, and administration route. Their current importance is closely tied to mRNA vaccines and siRNA therapeutics, but the contemporary literature treats LNPs less as a single fixed object than as a family of dynamically reorganizing, formulation-dependent soft-matter systems whose function emerges from coupled lipid chemistry, processing history, and biological transport (Kumar et al., 2024, Liu et al., 2024).

1. Core composition and physicochemical basis

A typical nucleic-acid LNP comprises four constituents: an ionizable lipid, a helper phospholipid, cholesterol, and a PEG-lipid conjugate. The ionizable lipid encapsulates nucleic acids and drives endosomal escape; the helper phospholipid contributes structural integrity and membrane interactions; cholesterol enhances membrane rigidity, particle stability, and intracellular delivery; and the PEG-lipid prolongs circulation by reducing rapid clearance and controlling aggregation (Kumar et al., 2024, Li et al., 24 May 2026).

Component	Primary role	Representative notes
Ionizable lipid	Encapsulation and endosomal escape	Central active component
Helper phospholipid	Structural support and fusion-related behavior	Often DSPC or DOPE
Cholesterol	Rigidity, stability, phase modulation	Improves membrane packing
PEG-lipid	Steric stabilization and circulation control	Reduces aggregation and opsonization

Ionizable lipids are amphiphilic molecules comprising an ionizable hydrophilic head and multiple hydrophobic tails. Their defining property is charge switching across physiological and endosomal conditions: they are designed to be near neutral at physiological pH yet positively charged in acidic environments, balancing RNA complexation, cellular uptake, and reduced toxicity. A widely cited mechanistic guideline is a pKa in the range of approximately 6–7. In the generative-design literature, ionizability is operationalized by verifying net charge neutrality at pH 7.4 and cationicity at pH 5 via predicted pKa values and Henderson–Hasselbalch charge calculations (Zhao et al., 2024, Ou et al., 2024).

For a basic site, the protonation fraction is written as

$\alpha_{\text{protonated}} = \frac{1}{1 + 10^{\mathrm{pH} - \mathrm{p}K_a}}.$

This relation formalizes why LNP assembly and performance are sensitive to the apparent pKa of the ionizable headgroup. At acidic pH, a high protonation fraction supports electrostatic association with polyanionic nucleic acids; at physiological pH, reduced protonation lowers systemic cationic charge. This pH-responsive behavior underlies both formulation in acidic buffer and later neutralization for delivery (Ou et al., 2024, Bai et al., 3 Aug 2025).

2. Self-assembly, internal organization, and pH-triggered restructuring

In therapeutic nucleic-acid LNPs, rapid solvent exchange and electrostatic complexation produce particles with a core rich in ionizable lipids, nucleic acid, and water, and an exterior enriched in PEGylated and helper lipids. Coarse-grained molecular dynamics resolves a characteristic pathway in which dispersed lipids and siRNA first form micelles or bilayer-like patches below 10 nm, then close into vesicles above 10 nm; RNA stranded at vesicle surfaces nucleates adhesion and fusion, yielding mature LNPs. Reverse micelles are central internal motifs, and their prevalence depends strongly on ionizable-lipid geometry: ALC-0315 forms the most numerous and smallest reverse micelles, Lipid5 is intermediate, and MC3 forms fewer but larger reverse micelles (Bai et al., 3 Aug 2025).

These internal structures are not static. In CGMD studies, LNPs assembled with protonated ionizable lipids at approximately pH 4 undergo global reorganization when shifted to pH 7.4. Deprotonation reduces headgroup hydrophilicity, weakens internal layering, decreases reverse-micelle number, enriches helper lipids and cholesterol at the surface, and drives neutral ionizable lipids inward. More than half of deprotonated MC3 lipids were dehydrated in the core and co-clustered with cholesterol into dense lipid phases, whereas Lipid5 retained cargo more effectively, consistent with a neutral headgroup that preserves some hydrophilicity after deprotonation (Bai et al., 3 Aug 2025).

A common misconception is that all lipid nanoparticles can be adequately represented by a uniform spherical core–shell model. That description is useful for many RNA LNP analyses, including SAXS models of MC3-containing systems, but it is not universal. Synchrotron SAXS on solid lipid nanoparticles formulated from cetyl palmitate and polysorbate 80 resolved barrel-like particles built from stacks of coplanar lipid platelets, partially covered by surfactant and bound water, explicitly challenging the classical core–shell picture. This contrast is significant because it shows that “lipid nanoparticles” spans multiple structural classes, from nucleic-acid LNPs with heterogeneous aqueous cores to solid lipid particles with patchy platelet architectures (Spinozzi et al., 2023, Bånkestad et al., 22 May 2026).

3. Manufacturing, mixing physics, and scale-up

Modern nucleic-acid LNPs are commonly formed by rapidly mixing an organic ethanol stream containing lipids with an aqueous acidic buffer stream containing nucleic acids. Mechanistic treatments couple solvent exchange, ionization, complexation, nucleation, growth, aggregation, and transport across reactor and particle scales. A central dimensionless quantity is the Damköhler number, $Da = \tau_{\text{mix}} / \tau_{\text{assembly}}$ : when mixing is faster than assembly, nucleation is more spatially uniform and particles tend to be smaller and more monodisperse; when assembly is faster than mixing, local supersaturation can broaden the size distribution and alter morphology (Inguva et al., 2024).

Population-balance modeling makes these qualitative statements operational. In a microfluidic impinging jet mixer model with lipids dissolved in ethanol at 50 mol% ionizable lipid, 38.5 mol% cholesterol, 10 mol% phospholipid, and 1.5 mol% PEG-lipid, mixed with 0.1 M sodium acetate at pH 5.5, the particle size distribution was governed by the balance of nucleation, growth, and coalescence. The model identified supersaturation and lipid dilution as the key variables controlling that balance. Increasing the flow rate ratio from 1 to 5 shifted the system toward nucleation-dominated kinetics and reduced the Z-average from 752 nm to 84 nm. Reducing the characteristic mixing time from 10 ms to 1 ms decreased the Z-average from 186 nm to 133 nm. By contrast, changing lipid concentration produced a convex size response: the Z-average decreased from 110 nm to 100 nm as concentration increased from 1.5 to 3 mg mL $^{-1}$ , then increased to 115 nm as concentration rose to 7 mg mL $^{-1}$ (Shin et al., 12 Apr 2025).

These results support a broader mechanistic interpretation. Increasing FRR accelerates ethanol dilution and changes solvent composition, while faster micromixing compresses the time window over which local composition gradients persist. Both strategies favor earlier, more distributed nucleation and reduce the time available for growth. Lipid concentration acts less monotonically because it simultaneously raises supersaturation and increases the mass available for growth. This suggests that size control is not reducible to a single “more dilution is smaller particles” heuristic; it depends on how dilution, supersaturation, and assembly timescales intersect in a given mixer and formulation (Shin et al., 12 Apr 2025, Inguva et al., 2024).

Beyond ethanol-based rapid mixing, other lipid nanoparticle classes can be produced by distinct physical routes. A temperature-programmed flow reactor uses lipid polymorphic transitions to fragment micron-scale particles into nanoparticles, producing emulsions or suspensions with particle diameters tunable between 20 and 800 nm. In Precirol ATO 5 systems, the final mean volume diameter correlated with residence time in the pre-burst swelling interval 40–50 °C according to $d_{\text{mean}} = 69 + A \exp(-0.097\, t_r)$ with $R^2 = 0.936$ , and the authors reported performance comparable to high-pressure homogenization at approximately 500 bar or higher without cavitation or bulk overheating (Lesov et al., 2022). This route is not the standard manufacturing paradigm for RNA LNPs, but it underscores the broader engineering diversity of lipid nanoparticle production.

4. Structural characterization and the inverse problem

LNP characterization is necessarily multimodal. DLS, zeta potential, cryo-TEM, SAXS, and SANS are used to interrogate size, dispersity, surface charge, morphology, encapsulation, and internal organization. In one MC3/cholesterol/DSPC/DMPE-PEG system at 50:38.5:10:1.5 mol%, SAXS was combined with cryo-TEM and DLS; the preparation had 98% encapsulation efficiency by RiboGreen, a cryo-TEM equivalent radius of 346 ± 62 Å, and a DLS hydrodynamic radius of 323 ± 79 Å (Bånkestad et al., 22 May 2026).

SAXS is especially powerful for LNPs because it probes nanometer-to-hundreds-of-nanometers structure, but it is an inverse problem with non-unique solutions. A differentiable machine-learning-accelerated SAXS framework addressed this explicitly by fitting heterogeneous, polydisperse MC3 LNPs with a core–shell particle model coupled to a Gaussian random-field interior. The surrogate reduced prediction cost by four orders of magnitude, enabling 8000 multi-start fits. Applied to experimental data, it found that shell thickness was broadly distributed in the 40–70 Å range and that five HDBSCAN clusters of parameter modes produced near-identical fits. The dominant ambiguity was a trade-off between size-distribution parameters and interior-structure parameters, implying that near-identical SAXS curves can correspond to distinct physical interpretations (Bånkestad et al., 22 May 2026).

This non-uniqueness has practical consequences. It argues against reporting a single best-fit structure without alternative modes, and it strengthens the case for contrast variation, independent size priors, and explicit identifiability analysis. The same lesson appears, in a different structural class, in synchrotron SAXS studies of solid lipid nanoparticles, where detailed modeling resolved lamellar spacings, patchy surfactant coverage, and bound water, revealing a much richer architecture than the classical core–shell simplification would imply (Spinozzi et al., 2023). The broader implication is that LNP structure should be understood as model-conditional and measurement-dependent rather than directly observable from one modality alone.

5. Therapeutic behavior, programmability, and administration routes

Programmable LNP design has been organized into four coupled domains: Architecture, Interface, Payload, and Dispersal. The Architecture domain concerns structural phases, encapsulation, and phase transitions; the Interface domain governs targeting ligands, PEG behavior, and environmental responsiveness; the Payload domain covers nucleic acids, small molecules, proteins, and co-payloads; and the Dispersal domain addresses administration route, biodistribution, and tissue barriers. This framework emphasizes that LNP performance is not solely a matter of composition at formulation but of how formulation, transport, and trigger-responsive behavior interact throughout the particle lifecycle (Liu et al., 2024).

Route dependence is particularly pronounced. Intravenous administration yields immediate systemic exposure and rapid liver uptake. Intramuscular dosing shows strong local retention and lymphatic drainage, with expression at the site within approximately 1 h and in the liver by approximately 3 h. Subcutaneous dosing produces predominantly local expression, approximately 99% at the injection site, and 20-fold lower plasma levels than IV at the same dose. Intranasal and pulmonary delivery make mucus penetration and aerosol stability critical, while ophthalmic and intracerebral routes face distinct anatomical barriers (Liu et al., 2024).

Interface programming can alter these route-dependent outcomes. Active targeting examples include folic acid, mannose, adenosine, peptides, antibodies, and aptamers, while environmental targeting uses pH-responsive lipids, redox-sensitive disulfides and thioketals, enzyme-cleavable groups, and light-triggered lipids. The review literature also highlights lipidomic engineering via SORT molecules such as DOTAP, 18PA, and DODAP, which reprogram organ tropism, and reports that cholesterol removal can reduce hepatic adsorption and redirect mRNA translation toward the lung or other organs (Liu et al., 2024).

These developments should not obscure a persistent mechanistic bottleneck: endosomal escape remains difficult. Multiple sources identify protonation-driven membrane destabilization and curvature-generating internal structures as decisive, but translation from uptake to functional cytosolic delivery remains inefficient. One quantitative example reported a yield of approximately 0.15 protein per delivered mRNA at 24 h, underscoring that uptake, escape, release, and translation are serially coupled efficiencies rather than interchangeable proxies (Liu et al., 2024). This suggests that particle size, biodistribution, and expression should not be interpreted as isolated endpoints.

6. Computational design, optimization frameworks, and unresolved constraints

The growth of LNP datasets has turned design into a statistical and machine-learning problem as much as a synthetic one. In an early demonstration, a curated dataset of 622 LNPs from published studies was used to predict transfection efficiency of unseen LNPs, with a multilayer perceptron reaching 98% classification accuracy on the test set (Ding et al., 2023). A larger literature-curated framework assembled 6,398 LNP formulations from 16 studies, including constituent SMILES, molar ratios, nucleic-acid type, nucleic-acid-to-lipid ratio, dose, activity, and viability; binary models achieved over 90% accuracy and multiclass models over 95%, and including full composition improved random-forest accuracy from 82.6% for ionizable lipid only to 90.2% (Kumar et al., 2024).

As datasets and benchmarks matured, curation quality became a central issue. LANTERN revisited the HeLa transfection dataset used for AGILE, identified 235 label inconsistencies and 100 duplicate SMILES entries, and retained 1,100 unique ionizable lipids after manual correction and duplicate removal. In that benchmark, an MLP trained on count-based Morgan fingerprints and RDKit Expert descriptors reached $R^2 = 0.8161$ and $r = 0.9053$ , substantially exceeding AGILE at $R^2 = 0.2655$ and $r = 0.5488$ . The same study argued that simpler models with chemically informative features outperformed more complex models relying on internally learned embeddings (Mehradfar et al., 3 Jul 2025).

Other studies reached a different conclusion under different data regimes. LipidBERT was pre-trained on 10 million virtual ionizable lipids and then fine-tuned on proprietary LNP property tasks and the public AGILE dataset. Validation $Da = \tau_{\text{mix}} / \tau_{\text{assembly}}$ 0 exceeded 0.9 for most proprietary regression tasks, with toxicity as an exception at approximately 0.79; on AGILE, Pearson correlation reached 0.98 in HeLa and 0.97 in Raw 264.7, with $Da = \tau_{\text{mix}} / \tau_{\text{assembly}}$ 1 approximately 0.95 after 100 epochs. However, the wet-lab tasks held non-ionizable formulation ratios constant and the authors cautioned that accuracy was strongest for lipids with scaffolds similar to those present in the wet-lab set (Yu et al., 2024). This suggests that apparent disagreement between descriptor-based and foundation-model approaches is partly a question of dataset scope, target definition, and generalization regime.

Predictive modeling has been complemented by generative and safety-aware frameworks. A synthesis-aware Synthesis-DAG approach trained on 70,536 ionizable-lipid synthesis paths generated lipids together with synthesis routes; its DAG+Chem variant achieved a lipid rate of 92.6% and an ionizable lipid rate of 83.4% (Ou et al., 2024). A policy-network-guided Monte Carlo tree search framework operated on purchasable head and tail building blocks and reaction templates to generate ionizable lipids with plausible synthesis pathways (Zhao et al., 2024). LipoAgent then reframed screening around toxicity as a decision-level prerequisite for efficiency: on TransLipid, it reported an average 32% relative improvement in mRNA transfection efficiency prediction over reported lipid-design baselines and screened 10,024 synthetically feasible lipids in approximately 23 hours before wet-lab validation of four candidates (Li et al., 24 May 2026).

Statistical design-of-experiments remains important alongside ML. A Quality by Design workflow for LNP optimization used space-filling mixture–process designs and self-validated ensemble models under the constraint that molar fractions sum to 100%. In the illustrative study, ionizable lipid, helper lipid, and cholesterol each varied from 10% to 60%, PEG from 1% to 5%, N:P ratio from 6 to 14, and flow rate from 1 to 3. The workflow emphasized desirability-based optimization, confirmation runs, and graphical summaries rather than one-shot point prediction (Karl et al., 2022).

Several unresolved constraints recur across these computational literatures. Literature-derived datasets are heterogeneous, often small relative to the chemical and process space, and may omit crucial physical attributes such as size, PDI, zeta potential, and helper-lipid composition. Domain shift across cell types, tissues, routes, and scaffolds remains substantial. Toxicity labels are typically in vitro and incomplete. Structural characterization itself is non-unique, which limits how cleanly mechanistic descriptors can be extracted from measurement. The aggregate lesson is not that LNP design is intractable, but that accurate prediction increasingly depends on integrating chemistry, full composition, process metadata, structural measurements, and safety endpoints rather than treating ionizable-lipid structure as the sole explanatory variable (Kumar et al., 2024, Bånkestad et al., 22 May 2026).