Ionizable Lipids in Nanoparticle Formulations

• Ionizable lipids are amphiphilic molecules with pH-dependent charge that facilitate RNA complexation, endosomal escape, and reduced systemic toxicity.
• They consist of a hydrophilic amine-based head and hydrophobic tails designed to optimize membrane packing, spontaneous curvature, and nanoparticle formation.
• Computational and simulation studies use synthesis-aware design and coarse-grained models to predict transfection efficiency and guide formulation improvements.

Ionizable lipids are amphiphilic lipids whose headgroup charge state changes with the chemical environment, most prominently with pH. In contemporary lipid nanoparticle research, the term usually denotes lipids with an ionizable, hydrophilic head and one or more hydrophobic tails, designed to be positively charged under acidic conditions and near-neutral at physiological pH so that they can condense RNA during formulation, promote endosomal escape, and reduce systemic charge burden in circulation (Ou et al., 2024). In membrane biophysics, closely related phenomena are discussed under the language of acidic, charged, protonable, or chemically modified lipids, where protonation and deprotonation alter electrostatic repulsion, hydration, lipid packing, spontaneous curvature, leaflet asymmetry, and in some regimes even bilayer permeability and discrete ion conduction (Angelova et al., 2018, Heimburg, 2010). The modern medicinal-chemistry meaning and the older membrane-physical meaning are not identical, but they are connected by the same core principle: lipid charge state is a state variable that reorganizes lipid assemblies.

1. Definitions and semantic scope

In the LNP delivery literature represented here, an ionizable lipid is defined operationally by a head–tail architecture and by pH-switching behavior. The hydrophilic head typically contains a secondary or tertiary amine, while the hydrophobic region contains multiple tails; the desired behavior is neutrality at physiological pH and positive charge in acidic environments such as endosomes (Yu et al., 2024, Zhao et al., 2024). This design logic is treated as the central reason ionizable lipids are favored over permanently cationic lipids for mRNA delivery, because the same molecule can support nucleic-acid complexation and intracellular release while avoiding a persistent systemic positive charge (Ou et al., 2024).

A broader membrane-biophysical usage predates the LNP field. In that literature, cardiolipin, phosphatidylglycerol, phosphatidic acid, phosphoinositides, ganglioside GM1, phosphatidylserine, and even chemically modified zwitterionic systems are treated as pH-sensitive lipids because their protonation state modulates interfacial charge and collective membrane mechanics (Angelova et al., 2018). This broader usage is directly relevant to ionizable lipids in the modern sense because it establishes that protonation is not merely a molecular descriptor of a headgroup; it is a control parameter for membrane organization.

A persistent source of confusion is the scope of the paper "Lipid Ion Channels" (Heimburg, 2010). That review is not about modern cationic or ionizable LNP lipids, pKa engineering, or synthetic tertiary-amine delivery lipids. Its relevance is more fundamental: it shows that whenever protonation, ionic screening, calcium binding, additive partitioning, or voltage shifts membrane phase behavior, the bilayer’s own permeability can change dramatically and even discretely (Heimburg, 2010). This is a membrane-state perspective on ionizable lipid behavior rather than a medicinal-chemistry perspective.

2. Charge-state thermodynamics and membrane activity

The foundational physical picture is that protonation and deprotonation modify headgroup charge, which in turn changes electrostatic repulsion, hydration, packing, equilibrium area per lipid, and spontaneous curvature. In the bilayer environment, ionization is not a bulk-solution constant. The apparent pH response depends on membrane surface potential, charged-lipid fraction, and ionic strength, so the pH range over which a lipid changes packing in a membrane can differ from its intrinsic solution pKa (Angelova et al., 2018). For phosphatidylglycerol-containing membranes, increasing the PG fraction from 10% to 20% shifts the pH range of protonation-related packing changes upward by 0.5 to 1 pH unit, illustrating that membrane context renormalizes ionization behavior (Angelova et al., 2018).

The same review gives a continuum description in which a chemically modified lipid fraction $\phi$ in one monolayer changes both equilibrium density and spontaneous curvature. In that framework, protonation-driven deformation cannot be reduced to a single mechanism: a local chemical perturbation alters the preferred neutral-surface density and the preferred curvature of the modified leaflet, and the resulting normal-force contributions are often of comparable magnitude (Angelova et al., 2018). This is the mechanistic basis for pH-driven migration, polarization, tubulation, and invagination in model membranes.

"Lipid Ion Channels" extends the same logic to permeability and conduction near chain-melting transitions. The central quantitative relation is that permeability increases with excess heat capacity,

$P = P_0 + \alpha \Delta c_P,$

because the transition amplifies area fluctuations and compressibility, making pore nucleation more likely (Heimburg, 2010). The paper explicitly identifies temperature, hydrostatic pressure, lateral pressure or tension, electrostatic potential, proton activity, calcium concentration, and chemical potentials of additives as relevant thermodynamic variables (Heimburg, 2010). This suggests that ionizable lipids affect conduction indirectly by changing the thermodynamic distance to the fluctuation-rich transition regime.

Electrostatics enters the membrane-state description directly. At low charge density and high salt, the membrane surface potential is written as

$$\Psi_0=\frac{1}{\epsilon \epsilon_0 \kappa}\sigma \propto \frac{\sigma}{\sqrt{c}},$$

with $\sigma=q/A$, and the associated electrostatic free energy scales as

$F_{el}\propto \frac{q^2}{A\sqrt{c}}.$

Because gel lipids occupy smaller area than fluid lipids, the same total charge yields higher $\sigma$ and larger electrostatic free energy in the gel state; protonation, ionic screening, and charge asymmetry therefore shift the melting point and alter permeability (Heimburg, 2010). The review reports that lowering pH can shift $T_m$ by 10–20 K in charged membranes such as dimyristoyl methylphosphatidic acid, and that Kaufmann and Silman observed proton-induced channel events in soybean phosphatidylcholine near pH $\sim 1$, close to its membrane $pK_a$ (Heimburg, 2010). In this sense, ionizable lipids can render lipid assemblies electrically active without invoking a protein pore.

3. Molecular architecture and design variables

Modern ionizable lipids are described as modular molecules built from an ionizable head and hydrophobic tails. The delivery-oriented generative literature emphasizes the asymmetry between polar, amine-containing heads and lipid-like tails, treating the lipid as a head-centered assembly problem rather than a generic small-molecule graph-generation problem (Ou et al., 2024, Zhao et al., 2024). This modularity is one reason existing small-molecule generators trained on generic medicinal chemistry perform poorly when transferred directly to lipid space.

The structural axes emphasized across the computational papers are head identity, tail identity, number of tails, head–tail connecting atom, branching, ring content, ester incorporation, and nitrogen-containing spacer motifs. LipidBERT’s 10 million virtual lipid library explicitly contains lipids with 2 to 6 tails and highlights clusters associated with branched alkyl chains, alkyl chains connecting two nitrogen atoms, aromatic and cyclic motifs, and benzene or five-membered rings connected to one, two, or three ester groups (Yu et al., 2024). At the same time, that work does not provide a systematic medicinal-chemistry taxonomy of scaffold families, target pKa windows, or biodegradability rules (Yu et al., 2024).

The synthesis-aware generation papers formalize these design variables through explicit head and tail building-block libraries. In the Monte Carlo tree search framework, head building blocks are filtered from ZINC20 by molecular weight $< 500$ g/mol, $P = P_0 + \alpha \Delta c_P,$0, and the presence of amine functional groups with exclusion of ammonium-based molecules; reactive heads are further required to contain carboxyl, hydroxyl, or amine functionality (Zhao et al., 2024). A related DAG-based generator uses heads from ZINC20 with molecular weight $P = P_0 + \alpha \Delta c_P,$1 g/mol, $P = P_0 + \alpha \Delta c_P,$2, amine groups, and 1–3 reactive functional groups, together with 15,302 reactive tails obtained by searching ZINC for synthetically accessible analogs of 8,176 tails extracted from 48,548 LIPID MAPS structures (Ou et al., 2024). These filtering strategies encode a specific view of ionizable lipid chemistry: a protonatable, relatively polar head attached to one to three or more hydrophobic tails through an overview-compatible linkage set.

Benchmark simulation studies sharpen the structure–function language further by comparing concrete chemotypes. The coarse-grained MD work focuses on MC3, Lipid5, and ALC-0315, and later applies the same framework to designed lipids L1–L6 (Bai et al., 3 Aug 2025). In that analysis, headgroup hydrophilicity, tail number and bulk, and the resulting cone shape are the principal structural determinants; linker chemistry is present in the coarse-grained mapping but is not developed as an independent mechanistic axis (Bai et al., 3 Aug 2025).

4. LNP self-assembly, internal organization, and formulation dependence

In the coarse-grained simulation framework, ionizable lipids perform three linked functions in LNPs: they drive RNA encapsulation during acidic formulation, control the self-assembly pathway and final internal nanostructure, and enable intracellular release while limiting toxicity at physiological pH (Bai et al., 3 Aug 2025). The decisive variable is not merely whether the lipid is ionizable, but how its protonated and deprotonated states reshape the whole particle.

Assembly from dispersed components proceeds through a stepwise sequence: dispersed lipids plus RNA, formation of small micellar or bilayer-like patches below 10 nm, growth into closed vesicular structures above 10 nm, RNA-mediated adhesion and fusion between vesicles, and emergence of mature LNPs with multiple internal reverse micelles, a near-monolayer envelope, and PEG-lipids retained on the outer surface (Bai et al., 3 Aug 2025). This description rejects a homogeneous-sphere picture and instead treats the mature particle as a compartmentalized object with inverse-curvature aqueous domains.

Chemistry-dependent differences are pronounced. MC3 shows slower fusion, fewer and larger reverse micelles, a rougher and less hydrophilic surface, and weaker propensity for high-curvature inverse organization. Lipid5 and ALC-0315 are described as more strongly cone-shaped than MC3, with faster fusion, smoother surfaces, and more numerous internal reverse micelles; ALC-0315 shows the highest number of reverse micelles and the smallest reverse micelles on average (Bai et al., 3 Aug 2025).

Ionizable lipid	Structural emphasis	Simulated tendencies
MC3	Single hydrophilic headgroup; weaker cone geometry	Fewer, larger reverse micelles; least hydrophilic surface; strongest collapse after deprotonation
Lipid5	Headgroup retains more hydrophilicity after deprotonation; three tails	Faster fusion; more reverse micelles than MC3; better integrity at pH 7.4
ALC-0315	Similar headgroup chemistry to Lipid5; four hydrocarbon tails	Highest reverse-micelle number; smallest reverse micelles; smoother surfaces

Surface composition measurements for small self-assembled particles quantify these differences: MC3 surfaces are 26.3% hydrophilic, 54.1% hydrophobic, and 19.6% neutral; Lipid5 surfaces are 33.8% hydrophilic, 52.1% hydrophobic, and 14.1% neutral; ALC-0315 surfaces are 32.0% hydrophilic, 52.9% hydrophobic, and 15.1% neutral (Bai et al., 3 Aug 2025). The same study varies formulation ratios across 50:40:10, 70:25:5, 30:60:10, and 30:40:30 ionizable lipid:cholesterol:helper lipid conditions and shows that increasing ionizable-lipid fraction generally increases surface hydrophobicity, while helper lipids suppress reverse micelles and increase surface hydrophilicity; cholesterol has lipid-specific effects, increasing reverse micelles in MC3 but decreasing them in Lipid5 and ALC-0315 (Bai et al., 3 Aug 2025). A plausible implication is that no universal formulation ratio can be assumed across ionizable lipid chemotypes.

The pH transition is modeled as a switch from a fully protonated acidic state around pH 4 to a fully deprotonated state around pH 7.4. After deprotonation, ionizable lipids move inward, helper lipids and cholesterol become surface-enriched, reverse micelles decrease, layering weakens, and in unfavorable chemistries aqueous compartments merge or collapse (Bai et al., 3 Aug 2025). MC3 shows the strongest structural rearrangement, visible phase separation, dehydrated interior lipid domains with cholesterol, and partial RNA release in large particles at pH 7.4; Lipid5 retains cargo and integrity better, plausibly because its hydroxyl-bearing headgroup preserves more hydrophilicity after deprotonation (Bai et al., 3 Aug 2025).

The same study links internal order to encapsulation efficiency through

$P = P_0 + \alpha \Delta c_P,$3

L1 and L2 show disordered interiors and low EE of about 7% and 9%, whereas L3 and L4 show well-defined reverse micelles and high EE of about 94% and 84% (Bai et al., 3 Aug 2025). This supports the paper’s central hypothesis that reverse-micelle-forming propensity is a useful in silico indicator for ionizable lipid performance.

5. Generative and synthesis-aware design

A major contemporary direction is to define ionizable lipid discovery as an overview-constrained search problem rather than unconstrained molecular generation. One MCTS-based model formulates the state as the current molecule or partial assembly and the action as the next building block to add, with the search terminating at a “two-tail lipid” (Zhao et al., 2024). The search is guided by a policy network and by two predictors: a Chemprop lipid classifier trained on 180,000 lipid and 180,000 non-lipid samples, and an ionizability pipeline based on MolGpKa plus net charge evaluation at pH 7.4 and pH 5 (Zhao et al., 2024). In this framework, a molecule qualifies as ionizable if its predicted net charge is neutral at physiological pH 7.4 and positive in acidic conditions around pH 5 (Zhao et al., 2024).

The search is synthesis-aware at three levels: building blocks come from a purchasable database, reactive functional groups are filtered up front, and assembly is restricted to predefined reaction templates (Zhao et al., 2024). Quantitatively, the baseline SyntheMol-like approach generated 16,477 two-tail lipids from 10,000 simulations, of which 4,513 were predicted ionizable lipids, for an ionizable lipid rate of 0.2739; random combination gave 0.1547; guided MCTS testing stabilized the unique ionizable-lipid rate around 0.73 to 0.8 (Zhao et al., 2024). Table 1 of that paper reports that Guided MCTS Test produced 545 unique ionizable lipids with unique ionizable lipid rate 0.7372, average SA score 4.24, and retrosynthesis-valid rate 0.2679 (Zhao et al., 2024). The same paper is explicit that the property objective is a weak surrogate for full delivery performance and that the reaction templates are not well matched to lipid-specific chemistry (Zhao et al., 2024).

A related DAG-based generator adapts Synthesis-DAGs to lipid-specific data and replaces Molecular Transformer with Chemformer as the reaction predictor, motivated by copy-paste errors when the older predictor handled large lipid heads and tails (Ou et al., 2024). The dataset contains 70,536 synthesis paths and 43,741 building blocks, consisting of 38,431 unique heads and 5,310 unique tails; the training set contains 5,905 one-tail lipids, 21,301 two-tail lipids, and 36,274 three-tail lipids (Ou et al., 2024). Its best generation regime, DAG+Chem, achieves lipid rate 92.6%, ionizable lipid rate 83.4%, validity 1.000, novelty 0.999, and FCD 3.797 (Ou et al., 2024). The same work also uses AGILE-guided iterative optimization toward HeLa transfection efficiency, constraining output to two-tail lipids with tails of at least 10 carbons (Ou et al., 2024). This suggests a shift from static library generation toward closed-loop, property-biased navigation of synthesizable lipid space.

6. Representation learning and transfection prediction

The scarcity of public ionizable-lipid corpora has motivated representation learning on virtual lipid spaces. LipidBERT pre-trains a BERT-base-like Transformer with 12 encoder blocks, hidden size 768, feed-forward intermediate size 3072, and 12 attention heads on a proprietary corpus of 10 million virtual ionizable lipids represented as canonical RDKit SMILES with a character tokenizer (Yu et al., 2024). Its lipid-specific secondary tasks include number-of-tails prediction for 2–6 tails, two versions of connecting-atom prediction, head/tail token classification, and rearranged or decoy SMILES classification (Yu et al., 2024). In fine-tuning on in-house LNP data, the model reports validation $P = P_0 + \alpha \Delta c_P,$4 values greater than 0.9 in most regression tasks except toxicity, which is 0.79; for ex vivo fluorescence intensity in lung, $P = P_0 + \alpha \Delta c_P,$5 improves from 0.26 with a 0.05M pretraining set to 0.94 with 10M lipids, and for LNP size it reaches 0.91 at 10M (Yu et al., 2024). The same paper reports that on AGILE’s public 1,200-ionizable-lipid dataset, LipidBERT reaches $P = P_0 + \alpha \Delta c_P,$6 after 100 epochs while AGILE fluctuates around 0 (Yu et al., 2024).

Prediction accuracy also depends strongly on representation choice and dataset curation. LANTERN re-audits the HeLa dataset from Xu et al. and finds 235 label mismatches and 100 duplicate SMILES entries caused by collapsed geometric isomerism, then curates a final dataset of 1,100 unique ionizable lipid molecules (Mehradfar et al., 3 Jul 2025). In that benchmark, an MLP trained on count-based Morgan fingerprints plus 210 RDKit expert descriptors achieves the best random-split performance with $P = P_0 + \alpha \Delta c_P,$7, $P = P_0 + \alpha \Delta c_P,$8, $P = P_0 + \alpha \Delta c_P,$9, and $$\Psi_0=\frac{1}{\epsilon \epsilon_0 \kappa}\sigma \propto \frac{\sigma}{\sqrt{c}},$$0, compared with AGILE’s $$\Psi_0=\frac{1}{\epsilon \epsilon_0 \kappa}\sigma \propto \frac{\sigma}{\sqrt{c}},$$1, $$\Psi_0=\frac{1}{\epsilon \epsilon_0 \kappa}\sigma \propto \frac{\sigma}{\sqrt{c}},$$2, $$\Psi_0=\frac{1}{\epsilon \epsilon_0 \kappa}\sigma \propto \frac{\sigma}{\sqrt{c}},$$3, and $$\Psi_0=\frac{1}{\epsilon \epsilon_0 \kappa}\sigma \propto \frac{\sigma}{\sqrt{c}},$$4 (Mehradfar et al., 3 Jul 2025). Under Murcko scaffold split, performance drops markedly, with the best reported model being kNN + Morgan at $$\Psi_0=\frac{1}{\epsilon \epsilon_0 \kappa}\sigma \propto \frac{\sigma}{\sqrt{c}},$$5 and AGILE falling to $$\Psi_0=\frac{1}{\epsilon \epsilon_0 \kappa}\sigma \propto \frac{\sigma}{\sqrt{c}},$$6 (Mehradfar et al., 3 Jul 2025). The paper’s central inference is that explicit chemically informative encodings outperform internally learned embeddings on small, specialized ionizable-lipid datasets (Mehradfar et al., 3 Jul 2025).

TransMA addresses a different difficulty: transfection cliffs. It uses a 3D molecular Transformer, a molecule Mamba sequence model, and an atom-level mol-attention fusion block to predict transfection efficiency on the 1,200-ionizable-lipid AGILE dataset in Hela and RAW 264.7 cells (Wu et al., 2024). The labels are log2-transformed, and the paper constructs 4,267 transfection cliff pairs in Hela and 2,104 in RAW 264.7, accounting for 68% and 81% of all data respectively (Wu et al., 2024). On scaffold split, TransMA reports Hela MSE 3.64, MAE 1.57, $$\Psi_0=\frac{1}{\epsilon \epsilon_0 \kappa}\sigma \propto \frac{\sigma}{\sqrt{c}},$$7, and PCC 0.75, and RAW 264.7 MSE 1.63, MAE 0.98, $$\Psi_0=\frac{1}{\epsilon \epsilon_0 \kappa}\sigma \propto \frac{\sigma}{\sqrt{c}},$$8, and PCC 0.54; on cliff split, Hela performance is MSE 4.36, MAE 1.62, $$\Psi_0=\frac{1}{\epsilon \epsilon_0 \kappa}\sigma \propto \frac{\sigma}{\sqrt{c}},$$9, and PCC 0.79 (Wu et al., 2024). The model’s attention maps assign high weights to the specific atoms that differ in cliff pairs, such as a case with structural similarity 0.91, transfection difference 1.69, and differing carbon and nitrogen atoms receiving attention scores of 0.84 and 0.86 (Wu et al., 2024). This does not establish mechanistic causality, but it does make atom-level ranking of subtle structural edits more accessible.

7. Limitations, controversies, and unresolved questions

A first limitation is terminological. Not every paper on pH-sensitive or electrically active lipids is about modern ionizable LNP design. "Lipid Ion Channels" (Heimburg, 2010) and "pH Sensing by Lipids in Membranes" (Angelova et al., 2018) are foundational for understanding state-dependent lipid charge, permeability, and mechanics, but they do not discuss cationic or ionizable lipid nanoparticles, pKa engineering for endosomal escape, or delivery-lipid medicinal chemistry. Treating all of these literatures as interchangeable obscures the distinction between membrane-state theory and LNP-specific design.

A second limitation concerns surrogate objectives. Several computational pipelines classify ionizable lipids by neutrality at pH 7.4 and positive charge at pH 5, or by lipid-likeness plus binary ionizability, rather than by experimentally measured transfection, toxicity, biodegradability, or in vivo efficacy (Zhao et al., 2024, Ou et al., 2024). This means that “good ionizable lipid” is often reduced to a computational proxy. A plausible implication is that generative enrichment for ionizable-lipid-like molecules does not by itself establish formulation utility.

A third limitation is synthetic realism. The MCTS framework constrains generation by reaction templates and purchasable building blocks, but its independent retrosynthesis-valid rates remain limited; the highest value reported in its comparison table is 0.4881 for Guided MCTS Train, and published ionizable lipids score only 0.0433 under the same Syntheseus check (Zhao et al., 2024). The DAG+Chem generator provides explicit synthesis paths, yet it also states that the validity of those proposed routes has not been evaluated experimentally (Ou et al., 2024). Synthesis-aware generation is therefore stronger than unconstrained SMILES generation, but it is not equivalent to laboratory validation.

A fourth limitation is domain specificity. LipidBERT’s fine-tuned predictor is acknowledged to work best for ionizable lipids with scaffolds similar to those in the wet-lab data and may be less accurate for “entirely new scaffolds” (Yu et al., 2024). LANTERN’s scaffold-split degradation shows the same problem from a different angle (Mehradfar et al., 3 Jul 2025). TransMA likewise predicts transfection only within a specific formulation context, using fixed helper lipids and assay conditions, and does not model toxicity, immunogenicity, encapsulation efficiency, serum stability, biodistribution, or manufacturability (Wu et al., 2024). These are not minor omissions; they define the applicability domain of current ionizable-lipid ML.

A fifth limitation is physicochemical resolution. The coarse-grained MD study models pH response as a binary switch between fully protonated and fully deprotonated states rather than a continuous Henderson–Hasselbalch equilibrium, and large-particle simulations can become sensitive to dense initial packing (Bai et al., 3 Aug 2025). This is sufficient to reveal strong design rules—reverse-micelle propensity, surface hydrophobicity, headgroup hydrophilicity after neutralization, and lipid-specific formulation optimization—but it does not resolve fine hydrogen-bonding, detailed ion pairing, or continuous local ionization effects (Bai et al., 3 Aug 2025).

Taken together, the current literature presents ionizable lipids as a multiscale subject. At the membrane-physical scale, protonation changes equilibrium density, spontaneous curvature, phase behavior, and even discrete pore formation (Angelova et al., 2018, Heimburg, 2010). At the nanoparticle scale, headgroup hydrophilicity, tail multiplicity, cone shape, and formulation ratio determine reverse-micelle organization, surface properties, and pH-triggered restructuring (Bai et al., 3 Aug 2025). At the computational-design scale, synthesis-aware generation, virtual-corpus pretraining, and explicit feature-based prediction now define much of the field’s practical methodology, but the strongest models remain bounded by surrogate objectives, dataset bias, and limited prospective validation (Zhao et al., 2024, Ou et al., 2024, Yu et al., 2024, Mehradfar et al., 3 Jul 2025, Wu et al., 2024).