Prediction of high-performing spleen-targeted lipid nanoparticles using a deep learning model for robust anticancer immunotherapy

Abstract

Messenger RNA (mRNA) therapeutics hold significant potential across a wide range of medical applications. LNPs are the most clinically advanced mRNA delivery vehicles, but challenges such as off-target effects and liver accumulation limit their broader clinical use. While high-throughput screening is effective for identifying more efficient and selective ionizable lipids, the substantial experimental validation required limits its practical application. In this study, we developed a deep learning model to accelerate ionizable lipid optimization by virtually predicting high-performing ionizable cationic lipids. After applying this model to a series of bis-hydroxyethylamine derived lipids (BDLs), 24 promising candidates were synthesized for delivery efficiency and organ-selectivity validation. Among them, YK-407 exhibited superior in vitro transfection efficiency and in vivo organ-specific mRNA delivery. YK-407 LNPs predominantly targeted the spleen, particularly antigen-presenting cells (APCs). In a mouse OVA tumor model, YK-407 LNPs encapsulating OVA-mRNA almost completely inhibited tumor growth and induced a robust cytotoxic CD8⁺ T cell response in the spleen, outperforming clinically approved SM-102 and Dlin-MC3-DMA. Additionally, we demonstrated that YK-407 LNPs exhibited minimal toxicity for both the liver and spleen, with no significant inflammatory cytokine release. These findings highlight the potential of AI in LNP development and YK-407 holds great promise for applications in mRNA-based treatments.

---

Introduction

Messenger RNA (mRNA) therapeutics have demonstrated significant potential across a variety of medical applications, especially in mRNA vaccines.^1–4 During the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) pandemic, the approval of two SARS-CoV-2 mRNA vaccines^5,6 underscores the transformative impact of this technology. Compared to subunit, killed, live attenuated, and DNA-based vaccines, mRNA vaccines are safer, more efficient, and offer the potential for rapid and scalable manufacturing.^7,8 To date, more than 20 mRNA vaccines have entered clinical trials.^7,9 However, challenges remain in the safety and immunogenicity of mRNA-LNP vaccines, including off-target effects that may result in mRNA translation in non-targeted cells and organs, potentially triggering unwanted immune responses.¹⁰

To induce an effective immune response, mRNA vaccines rely on lipid nanoparticle (LNP) delivery systems to deliver the RNA payload into the cytosol of antigen-presenting cells¹¹ (APCs). Lymphoid organs, such as the spleen and lymph nodes, are rich in APCs, which are located near B and T lymphocytes, creating an optimal environment for immune activation.^12,13 The spleen, as the largest secondary lymphoid organ, plays a critical role in regulating immune responses, with dendritic cells serving as key players in antigen presentation.^14,15 Given that the effectiveness of mRNA vaccines relies on efficient APC uptake and antigen presentation, developing LNPs that target the spleen, particularly its APCs, could significantly enhance mRNA vaccine efficacy, especially for cancer immunotherapy.^16–18

LNPs are currently the most advanced system for mRNA delivery.¹⁹ LNPs typically contain four main components: ionizable lipids, helper lipids, PEG-lipids, and cholesterol.²⁰ Among them, ionizable lipids play a crucial role in RNA encapsulation and endosomal escape through acid-sensitive mechanisms.²¹ These lipids contain pH-sensitive amines that maintain a neutral surface charge under physiological conditions but become protonated in the acidic environment of the endosome, facilitating cargo release.²² While many LNP formulations have enabled significant progress in mRNA therapeutics, a major challenge remains: most LNPs tend to accumulate in the liver, leading to reversible hepatic damage and reduced therapeutic efficacy in other organs.^23–25 In recent years, extensive efforts have been made to achieve organ-selective delivery, such as the development of RNA-LPX,¹⁶ the selective organ targeting (SORT) strategy,^26–28 the optimization of novel ionizable lipids,^29–32 and the incorporation of antibodies.^33–35 However, careful design of lipid components is still essential to ensure both safety and efficacy. Previous studies showed that even subtle structural changes in ionizable lipids can significantly alter the delivery efficiency,^36–38 underscoring the need for the development of diverse ionizable lipids to optimize LNP delivery systems.

Artificial intelligence (AI) has emerged as a powerful tool in accelerating the development of LNPs for mRNA delivery. Recent studies have demonstrated the potential of machine learning (ML) and deep learning (DL) models in predicting LNP transfection efficiency. For example, Li et al. used ML algorithms combined with combinatorial chemistry to speed up the discovery of ionizable lipids for mRNA delivery.³⁹ Furthermore, platforms like AGILE,⁴⁰ TransLNP⁴¹ and LiON⁴² leverage graph neural networks or transformers, to predict transfection efficiency with promising results. Despite these advancements, current AI-based LNP design approaches face significant limitations. Existing models often overlook the unique chemical properties of ionizable lipids, such as their long, structurally similar side chains. Traditional molecular representations, like SMILES, may not effectively capture the subtle differences in atomic environments within the side chains. Moreover, the inherent similarity among ionizable lipids is not adequately considered and can significantly impact transfection efficiency.

To address these challenges, this paper introduces a novel deep learning model and method for predicting LNP transfection efficiency. We utilize the Atom-in-SMILES (AIS) segmentation method to more effectively represent ionizable lipids. AIS captures the chemical environment of branched atoms, providing a richer and more informative molecular description. The model explicitly considers the differential structure between ionizable lipids. By analyzing the differences in molecular structure, we aim to better understand the relationship between subtle structural variations and changes in transfection efficiency. The model is trained within a contrastive learning framework, enabling it to learn both the relationship between individual molecular structures and transfection efficiency. This innovative approach significantly improves the accuracy and generalizability of LNP transfection efficiency prediction, and a library of bis-hydroxyethylamine derived lipids (BDLs) were predicted as high-performing ionizable cationic lipids. Subsequently, these predicted ionizable cationic lipids were optimized and synthesized including 6 different head groups and 10 different lipid tails. The resulting LNPs exhibited favorable physicochemical properties, including optimal particle size and low polydispersity index. In vitro experiments demonstrated that YK-407 shows significantly enhanced in vitro transfection efficiency compared to YK-009.

In vivo, YK-407 almost exclusively delivered mRNA cargo to the spleen, particularly targeting APCs. We further evaluated the therapeutic performance using an MC38-OVA (ovalbumin) tumor model and a B16F10 melanoma model. We found that YK-407 LNP encapsulating OVA-mRNA nearly completely inhibited tumor growth, outperforming clinically used SM-102 and Dlin-MC3-DMA. Further mechanistic studies revealed that YK-407 formulation induced significantly stronger OVA-specific cytotoxic CD8⁺ T cell response in the spleen, which may explain the observed better antitumor efficacy. These findings provide a foundation for further optimization of ionizable lipids with better spleen targeting efficiency. Moreover, the remarkable therapeutic efficacy observed in the mouse tumor model provides promising avenues for future clinical applications.

Results and discussion

Overall model design

There are three modules in our model: encoder, feature process module, and decoder (Fig. 1a and b). The input consists of AIS tokens, which are embedded using a random embedding approach, with positional encoding incorporated into the random embeddings. The encoder consists of multiple encoding layers, each comprising a multi-head self-attention layer, normalization with residual connections, a feed-forward fully connected layer, and another normalization with residual connection layer. In this study, the encoder contains four encoding layers, with four heads in the multi-head self-attention mechanism and a hidden feature dimension of 64. The feature process module is used to compute the difference between the latent vectors of two molecules (Fig. 1c). Specifically, the process involves taking the latent vectors and performing a dot product with a different structure mask (DSM). This operation results in a latent vector that represents the difference in structure of each molecule. Subsequently, the difference between the two latent vectors is computed by subtracting the dot product result. The decoder consists of three linear layers, with a GLUE activation function and a dropout layer applied after the first and second linear layers. The input dimension of the first linear layer is set to the dimension of the latent vector, which is 1024. Both the input and output dimensions of the second linear layer are 1024. The input dimension of the final linear layer is 1024, and its output dimension is 1.

Fig. 1 Architecture of the deep learning model. (a) A high-level overview of the model's workflow. (b) The structure of the encoder module. (c) The feature process module.

The pre-training dataset was collected from published studies containing in vitro experimental data on the transfection efficiency of 1200 ionizable lipids in HeLa and RAW264.7 cells. The model takes a specific ionizable lipid structure as input and predicts its transfection efficiency as the output. This approach can significantly reduce the time and resources required for compound synthesis and experimental evaluation.

Characterization of DL model performance

To evaluate our model's effectiveness, we compared its performance against several established methods: Random Forest (RF), LightGBM, AGILE, and TransLNP. RF and LightGBM are machine learning models utilizing molecular descriptors,⁴³ while AGILE employs a pre-trained graph neural network to encode molecules as graphs.⁴⁰ TransLNP, a state-of-the-art transformer-based model, leverages 3D structural features and achieves a mean squared error (MSE) of 5 on the AGILE dataset through data balancing techniques.⁴¹

We evaluated all models, including ours, on the AGILE dataset using MSE and Pearson correlation coefficient (PCC) as performance metrics. Our model outperforms all other methods, achieving the lowest MSE and highest PCC, indicating a significant improvement in predicting LNP transfection efficiency (Table 1).

Table 1 Performance of different models in predicting LNP transfection efficiency

Model	MSE↓	PCC↑
RF	5.73	0.66
LightGBM	7.10	0.56
AGILE	7.41	0.51
TransLNP	5.17	0.69
Ours	4.81	0.73

---

A primary screen was then conducted by evaluating the transfection efficiency of 624 designed lipid nanoparticles (LNPs) to identify a high-performance subset (Fig. S1b). For fine-tuning, we utilized the best-performing model and a dataset of 199 ionizable cationic lipids to develop a screening model. The high-efficiency LNPs identified in the initial screen were re-evaluated by the fine-tuned model; molecules predicted to exhibit low protein translation in the liver were selected as the final output (Fig. S1c). This model was then used to predict candidate ionizable lipids with superior delivery efficacy.

Optimization and synthesis of the ionizable lipids

Previous studies on spleen-targeted LNPs have highlighted the potential benefits of incorporating multiple tertiary amines.^24,44 Additionally, ethanolamine has been reported to reduce the formation of hydration layers and enhance nucleic acid interactions to increase delivery efficacy.^45,46 Based on these findings, we generated a set of ionizable lipids featuring a symmetric head group with two ethanolamine-like groups connected by 2–4 carbon atoms, along with two lipid tails which could be either the same or different (Fig. 2a).

Fig. 2 Optimization and synthesis of ionizable lipids. (a) The chemical structure of bis-hydroxyethylamine derived lipids (BDLs) used in this study and the formulation to form LNPs. (b) Structures of 6 different head groups and 10 different lipid tails synthesized in this study.

The ionizable lipid candidates were named bis-hydroxyethylamine derived lipids (BDLs), and their high transfection efficiencies and low protein translation in the liver were predicted by the deep learning model. Based on the predictions, 24 BDLs were identified as having high delivery efficiency and were successfully synthesized for further experimental validation (Fig. S2). These 24 BDLs cover six different head groups and ten different lipid tails (Fig. 2b). The synthetic routes are detailed in the SI. Briefly, symmetric BDLs were synthesized through a one-step nucleophilic substitution reaction (Fig. S3a and b). For asymmetric BDLs, the lipid tails were introduced sequentially, starting with Boc-protected head groups (Fig. S3c).

Physicochemical characterization and in vitro screening of BDL-LNPs

We optimized the LNP formulation by an in vitro transfection assay using firefly luciferase (Fluc) mRNA as a reporter to assess the delivery efficiency. Using YK-423 as a model, the molar ratio of ionizable lipid, structural lipid, cholesterol and polymer-conjugated lipid was optimized. The best transfection efficiency was achieved at BDL : DSPC : cholesterol : DMG-PEG2000 = 40 : 10 : 48.5 : 1.5 (Fig. S4), which was selected for further LNP preparation. Using the optimized ratio, we measured the particle size and polydispersity index (PDI) of these BDL-LNPs. The resulting particles ranged in size from 70 to 106 nm (Fig. 3b), suitable for mRNA delivery. Dynamic light scattering (DLS) analysis showed that the PDIs were all below 0.10, demonstrating superior monodispersity and uniformity of the particles (Fig. 3b).

Fig. 3 Physicochemical characterization and in vitro screening of BDL-LNPs. (a) Overview of LNP formulation and preparation using a microfluidic device. (b) The physicochemical properties of LNPs prepared with different BDLs. LNPs comprised BDL : DSPC : cholesterol : DMG-PEG2000 = 40 : 10 : 48.5 : 1.5, and encapsulated Fluc mRNA. (c) Relative bioluminescence intensity in HEK293T cells following delivery of Fluc-mRNA by different BDL-LNPs (n = 5, data represents the mean ± SD.). (d) Chemical structures of YK-407, which was selected for further in vivo study, along with the structures of two control lipids, SM-102⁴⁸ and Dlin-MC3-DMA.⁴⁹

We then employed an in vitro Fluc assay to quantitatively assess the transfection efficiency of different BDL-LNPs in HEK293T cells (Fig. 3c). The results indicated that even a small change of structure significantly impacts transfection efficiency. Specifically, YK-407 exhibited the highest luminescence signal, while the insertion of an additional CH₂ group between the two tertiary amines (YK-413) significantly reduced the transfection efficiency. Notably, a single CH₂ difference in the lipid tail between YK-407 and YK-405 resulted in a three-fold change in transfection efficiency, highlighting the importance of tail structure optimization. Satisfyingly, YK-407 outperformed YK-009, which was previously reported for delivering the COVID-19 Omicron vaccine⁴⁷ (Fig. 3c), with a 1.5-fold increase in transfection efficiency. Based on these in vitro screening results, YK-407 was selected for further in vivo experiments (Fig. 3d).

YK-407 showed superior in vivo spleen specificity

To assess the in vivo delivery efficiency and organ expression tropism of these LNPs, we used a mouse model with Fluc-mRNA as a reporter. SM-102, the lipid used in Moderna's COVID-19 mRNA vaccine⁵ (Fig. 3d), served as the control. The Fluc-mRNA was encapsulated in LNPs and injected intramuscularly into female BALB/c mice. After 6 h, a bioluminescence imaging substrate was injected intraperitoneally. Whole body imaging showed that YK-407 exhibited protein expression levels similar to those of SM-102, with sustained protein expression for up to 72 h (Fig. 4a and Fig. S5a). We also noticed that SM-102 exhibited significant expression in the liver, while YK-407 showed minimal liver uptake (Fig. 4a). After harvesting the organs and measuring the luminescence, we found that SM-102 almost exclusively targeted the liver, whereas YK-407 predominantly delivered mRNA to the spleen, with only minimal protein translation occurring in the liver (Fig. 4b and Fig. S5b). The spleen-targeting delivery ability might be related to the protein corona on the LNPs during circulation,²⁵ but the detailed mechanism requires further investigation. These findings suggest that YK-407 is a promising candidate for spleen-specific delivery, potentially minimizing liver toxicity.

Fig. 4 YK-407 and SM-102 exhibited in vivo spleen selectivity and effectively targeted antigen-presenting cells (APCs). (a) and (b) Representative IVIS images of whole body (a) and major organs (b) with bioluminescence at 6 h post injection of different Fluc-mRNA LNPs injected intramuscularly in female BALB/c mice at a dose of 0.25 mg kg⁻¹. (c) and (d) FACS quantification (c) and representative flow cytometry diagrams (d) of eGFP⁺ cells expressed in different types of splenocytes 24 h after intramuscularly injection of eGFP-mRNA LNPs at a dose of 0.5 mg kg⁻¹ (n = 3). DC: dendritic cells, MΦ: macrophages/monocytes, NK: natural killer cells. Statistics were assessed by two-way ANOVA with Tukey's multiple comparisons test. **p < 0.01 and ****p < 0.0001. Data represents the mean ± SEM.

We further validated the consistent biodistribution of YK-407 across different mouse strains. Using C57BL/6 mice, we replicated the aforementioned experiments and obtained results similar to those in BALB/c mice (Fig. S6). LNPs formulated with SM-102 primarily accumulated in the liver, whereas those formulated with YK-407 exhibited greater accumulation in the spleen. Moreover, both groups of LNPs showed distinct expression in the inguinal lymph nodes, which is consistent with previous studies on the immunization route via intramuscular injection. Additionally, neither group of LNPs exhibited significant expression in other organs (lung, heart and kidneys), indicating that both LNPs possess specific targeting properties and that there are no significant differences between these two different mouse strains. We also examined the biodistribution of YK-407 following systemic administration. The experimental procedure was similar to that of intramuscular injection, with the exception that intravenous injection was performed in C57BL/6 mice. The results are shown in Fig. S7 which indicated that, compared with Dlin-MC3-DMA and SM-102, YK-407 still exhibited a greater propensity to deliver the mRNA into the spleen. The results also demonstrated the superior splenic targeting ability compared to the spleen targeted selective organ targeting (SORT) LNP MC3-18PA.

We also conducted in vitro incubations of mouse plasma with SM-102, YK-407, and MC3-18PA, respectively, and subsequently employed differential centrifugation to isolate the plasma proteins bound to the surfaces of distinct LNPs. Upon qualitative analysis using sodium dodecyl sulfate (SDS)–polyacrylamide gel electrophoresis (PAGE) (SDS-PAGE) (Fig. S8), we observed that YK-407 and MC3-18PA exhibited highly enriched bands at the similar M_w. This indicates that the types of proteins adsorbed by YK-407 and MC3-18PA are consistent, primarily comprising actin, myosin, and immunoglobulin lambda-like protein as reported.^23,50 However, these proteins display low correlation and are typically nonspecific. Therefore, the underlying mechanism for spleen targeting remains to be elucidated.⁵⁰

YK-407 effectively targeted major APCs

Next, we investigated the specific cell types targeted by SM-102 or YK-407 using eGFP-mRNA as a reporter. Briefly, mice were intramuscularly injected with eGFP-mRNA encapsulated in SM-102 or YK-407 LNPs. After 24 h, the mice were sacrificed, and their spleens were collected for analysis. Flow cytometry data revealed that YK-407 LNPs primarily delivered eGFP-mRNA to dendritic cells (pDC, ∼2.83% and cDC, ∼0.86% for YK-407 LNPs) and macrophage/monocytes (MΦ) cells (∼0.94% for YK-407 LNPs) (Fig. 4c and d), which constitute central APCs in the spleen. The values were 2.8-, 1.8-, and 1.8-fold higher than that of SM-102, respectively. These results confirmed that YK-407 successfully delivers mRNA into APCs, which makes it suitable for efficient antigen-presenting and T cell activation. Given the superior in vitro transfection efficiency and in vivo expression tropism of YK-407, we next focused on exploring its therapeutic potential.

Therapeutic anti-tumor potential of the YK-407 mRNA vaccine

We next assessed the therapeutic potential of YK-407 in delivering an mRNA vaccine for tumor treatment using the MC38-OVA colorectal carcinoma mouse model or B16F10 melanoma model. In this study, C57BL/6 mice were subcutaneously inoculated with 1.5 × 10⁶ MC38-OVA or B16F10 cells on day 0. Subsequently, the OVA-mRNA LNPs formulated by YK-407, SM-102, or Dlin-MC3-DMA (Fig. 3d) were injected intramuscularly on days 3, 6, 9, 12, 15 and 18, after the tumors reached an average volume of 54 mm³ for the MC38 model (Fig. 5a). For the B16F10 model, YK-407, and SM-102 encapsulated Trp2 (tyrosinase-related protein 2)-mRNA were injected on days 3, 6 and 10 via an intramuscular or intravenous route, respectively. Trp2 is known as a tumor-associated antigen which is important in the B16F10 melanoma model.⁵¹ In the MC38 model, all three LNP formulations demonstrated significant inhibition of tumor growth (Fig. 5b), demonstrating successful OVA-mRNA delivery and immune response induction. More importantly, the YK-407 formulation nearly completely inhibited tumor growth (Fig. 5b), likely due to its selective spleen targeting and efficient antigen delivery to APCs. Furthermore, mice immunized with YK-407 exhibited a 100% survival rate during the experiment (Fig. 5c). Notably, no significant body weight loss was observed in the SM-102 and Dlin-MC3-DMA groups compared to the PBS group (Fig. 5d). The YK-407 group showed significantly lower body weight at day 15 compared to the control group (Fig. 5d), which is likely attributed to the significant reduction of tumor weight.

Fig. 5 The therapeutic effect of YK-407 OVA-mRNA vaccine on the MC38-OVA tumor model. (a) Schematic representation of therapeutic study design. Mice were implanted with MC38-OVA cells and vaccinated with OVA-mRNA/LNPs, formulated with YK-407, SM-102, or Dlin-MC3-DMA(MC3) at a dose of 0.15 mg kg⁻¹ on days 3, 6, 9, 12, 15, and 18. (b)–(d) Average tumor volumes (b), survival curves (c), and body weight change (d) following treatment (n = 8). Statistics were assessed by two-way ANOVA with Dunnett's multiple comparison tests. *p < 0.05, **p < 0.01, and ***p < 0.001. ns means no significant difference. Data represents the mean ± SEM.

In the B16F10 mouse model, C57BL/6 mice received mRNA vaccines formulated with either SM-102 or YK-407 LNPs at a dose of 0.25 mg kg⁻¹ on days 3, 6, and 10 post-tumor inoculation. Both intramuscular and intravenous administration routes were employed to evaluate the influence of delivery modality on antitumor efficacy (Fig. S9). Notably, intramuscular delivery conferred a more pronounced antitumor effect than intravenous administration. Irrespective of the injection route, YK-407 LNP elicited superior tumor suppression compared with SM-102 LNP.

OVA-mRNA LNPs induced robust OVA-specific cytotoxic CD8⁺ T cell response in the spleens

To gain mechanistic insight into the observed anti-tumor effects, spleens were collected on day 24 for further analysis (Fig. 6a). First, we assessed the levels of IFN-γ-secreting T cells in splenocytes using an ELISpot assay. The results showed that no response was detected in the non-vaccinated group upon SIINFEKL stimulation, while all vaccinated groups generated significantly increased IFN-γ-secreting T cells (Fig. 6b and c), indicating a robust T cell response induced by the mRNA vaccine. Next, we measured the generation of cytotoxic CD8⁺ T cells. Compared to the untreated control, the percentage of SIINFEKL-specific CD8⁺ T cells increased 21.5-fold, 17.4-fold, and 11.7-fold in the YK-407, SM-102, and Dlin-MC3-DMA groups, respectively (Fig. 6d, e and Fig. S11). These findings suggest that the YK-407 formulation effectively promoted the proliferation of antigen-specific T cells in the spleens, leading to stronger T cell responses than SM-102 and Dlin-MC3-DMA.

Fig. 6 OVA-mRNA LNPs induced a robust OVA-specific cytotoxic CD8⁺ T cell response in spleens. (a) Schematic representation of therapeutic study design. Mice were implanted with MC38-OVA cells and vaccinated with OVA-mRNA/LNPs, formulated with YK-407, SM-102, or Dlin-MC3-DMA (MC3) at a dose of 0.15 mg kg⁻¹ on days 3, 6, 9, 12, 15, and 18. Spleens were collected on day 24 for subsequent experiments. (b) and (c) ELISpot images (b) and spot numbers (c) of IFN-γ-secreting T cells within the spleens of vaccinated mice (n = 8). (d) and (e) FACS quantification (d) and representative flow cytometry diagrams (e) of CD8⁺ T cells bearing T cell receptors binding to H-2Kb OVA tetramer-SIINFEKL within splenocytes of vaccinated mice (n = 5). Statistics were assessed by one-way ANOVA with Tukey's multiple comparison tests. *p < 0.05, **p < 0.01, and ***p < 0.001. Data represents the mean ± SEM.

Safety of YK-407 LNPs in vivo

To evaluate the safety of YK-407 LNPs for potential clinical use, BALB/c mice were subcutaneously injected with 0.5 mg kg⁻¹ of OVA-mRNA encapsulated in YK-407 LNPs. Several biochemical blood parameters were analyzed to assess the potential liver and kidney toxicity. We observed only a slight decrease in serum aspartate transaminase (AST) levels following YK-407 treatment (Fig. 7a), with no significant changes in alanine transaminase (ALT), uric acid (UREA), or creatinine (CREA), indicating minimal liver and kidney toxicity (Fig. 7a). Additionally, the major organs, including the heart, liver, spleen, lungs, and kidneys, were collected from the mice for histopathological examination. The results showed that no obvious pathological changes were observed in any of the organs (Fig. 7b).

Fig. 7 Safety evaluation of YK-407 LNP constructs in vivo. (a) Serum levels of aspartate transaminase (AST), alanine transaminase (ALT), creatinine (CRE), and urea (URE) in BALB/c mice vaccinated with OVA-mRNA-LNPs formulated with YK-407, measured on day 12 after three vaccinations on days 0, 3, and 7 (n = 3). Control refers to treatment. (b) Histopathological examination (H&E) of the heart, liver, spleen, lungs, and kidneys in BALB/c mice on day 12 after three vaccinations on days 0, 3, and 7 with YK-407 OVA-mRNA-LNPs and DPBS (n = 3). (c) Serum cytokines 6 h after the first vaccination (n = 3). Statistics were assessed by unpaired one-tailed Student's t-test. *p < 0.05, ns means no significant difference. Data represents the mean ± SEM.

Finally, serum levels of key cytokines, including IFN-γ, IFN-α, TNF-α, and IL-12, were measured after YK-407 LNP treatment. The serum levels of IFN-α, TNF-α, and IL-12 showed no significant differences compared to the control group (Fig. 7c). Although an increase in serum IFN-γ levels was observed in one mouse, the overall difference was not statistically significant (Fig. 7c). Collectively, these findings support the safety of YK-407 LNPs for further clinical development.

Conclusions

In conclusion, we developed a deep learning model to aid in screening for more efficient ionizable lipids for mRNA delivery. By applying this model to a series of bis-hydroxyethylamine derived lipids (BDLs), 24 lipids were predicted to have high transfection efficiency, and were subsequently synthesized and evaluated for their mRNA delivery performance and organ-specific targeting. After optimizing the lipid formulation, the resulting LNPs showed favorable particle size and polydispersity index. Among these, YK-407 demonstrated superior in vitro transfection efficiency and in vivo spleen-specific targeting, delivering mRNAs to major APCs within the spleen. In a mouse OVA tumor model, OVA-mRNA delivered by YK-407 LNP nearly completely suppressed tumor growth, outperforming both SM-102 and Dlin-MC3-DMA, underscoring its potential for future applications. Further mechanistic studies showed that the YK-407 mRNA vaccine induced stronger OVA-specific cytotoxic CD8⁺ T cell responses in the spleen, which likely contributed to its strong anti-tumor effects. We also demonstrated that the YK-407 mRNA vaccine exhibited promising in vivo safety profiles, supporting its potential for future clinical use.

A recent preprint study also reported the design and synthesis of ionizable lipids with dual ethanolamine head groups, which were then incorporated into a three-component formulation.⁴⁶ In their work, the lipid LQ-3, which shares the same structure as our YK-401, demonstrated promising delivery capabilities, which serves as a good validation of our findings. However, their research primarily focused on the development of symmetric ionizable lipids and new formulations without phospholipids. Recent findings suggest that asymmetric lipids tend to have higher transfection efficacy than symmetric lipids, which may help explain the better performance of YK-407 compared to symmetric ones such as YK-401.⁵² Additionally, their study did not observe any organ-specific cargo delivery, which is a key aspect of our work.

These findings underscore the therapeutic potential of YK-407 in tumor immunotherapy and provide a solid foundation for further optimization of ionizable lipids to improve spleen-specific mRNA delivery, thereby increasing the therapeutic efficacy and minimizing off-target effects.

Author contributions

G. Song provided project design, administration and supervision; H. Zhang performed conceptualization, investigation, data curation, formal analysis and writing – original draft; J. Ma performed the formal analysis, writing – original draft and writing – review and editing; F. Pan performed the software, writing – original draft; Y. Liu, C. Zhang, H. Huang, W. Zhang, D. Xiu performed methodology; M. Zhang and W. Zhang performed validation.

Conflicts of interest

All authors were employed by Beijing Youcare Kechuang Pharmaceutical Technology Co., Ltd. All authors declare that the research was conducted without any commercial or financial relationships that could be viewed as a potential conflict of interest.

Data availability

The data supporting this article have been included as part of the SI. The content of the SI: Fig. S1–S11, chemical synthesis procedures, materials and methods, Tables S1–S2, SI references. See DOI: https://doi.org/10.1039/d5tb01217a

Acknowledgements

We thank Dr Kai Dong, Dr Jinyu Zhang and Huanyu Wang for helpful advice and discussion.

Notes and references

1. L. Miao, L. Li, Y. Huang, D. Delcassian, J. Chahal, J. Han, Y. Shi, K. Sadtler, W. Gao, J. Lin, J. C. Doloff, R. Langer and D. G. Anderson, Nat. Biotechnol., 2019, 37, 1174.
2. U. Sahin, P. Oehm, E. Derhovanessian, R. A. Jabulowsky, M. Vormehr, M. Gold, D. Maurus, D. Schwarck-Kokarakis, A. N. Kuhn, T. Omokoko, L. M. Kranz, M. Diken, S. Kreiter, H. Haas, S. Attig, R. Rae, K. Cuk, A. Kemmer-Brueck, A. Breitkreuz, C. Tolliver, J. Caspar, J. Quinkhardt, L. Hebich, M. Stein, A. Hohberger, I. Vogler, I. Liebig, S. Renken, J. Sikorski, M. Leierer, V. Mueller, H. Mitzel-Rink, M. Miederer, C. Huber, S. Grabbe, J. Utikal, A. Pinter, R. Kaufmann, J. C. Hassel, C. Loquai and O. Tuereci, Nature, 2020, 585, 107.
3. J. D. Beck, D. Reidenbach, N. Salomon, U. Sahin, Ö. Türeci, M. Vormehr and L. M. Kranz, Mol. Cancer, 2021, 20, 69.
4. H. Zhang, X. You, X. Wang, L. Cui, Z. Wang, F. Xu, M. Li, Z. Yang, J. Liu, P. Huang, Y. Kang, J. Wu and X. Xia, Proc. Natl. Acad. Sci. U. S. A., 2021, 118, e2005191118.
5. L. R. Baden, H. M. El Sahly, B. Essink, K. Kotloff, S. Frey, R. Novak, D. Diemert, S. A. Spector, N. Rouphael, C. B. Creech, J. McGettigan, S. Khetan, N. Segall, J. Solis, A. Brosz, C. Fierro, H. Schwartz, K. Neuzil, L. Corey, P. Gilbert, H. Janes, D. Follmann, M. Marovich, J. Mascola, L. Polakowski, J. Ledgerwood, B. S. Graham, H. Bennett, R. Pajon, C. Knightly, B. Leav, W. Deng, H. Zhou, S. Han, M. Ivarsson, J. Miller, T. Zaks and The COVE Study Group, N. Engl. J. Med., 2021, 384, 403.
6. F. P. Polack, S. J. Thomas, N. Kitchin, J. Absalon, A. Gurtman, S. Lockhart, J. L. Perez, G. Pérez Marc, E. D. Moreira, C. Zerbini, R. Bailey, K. A. Swanson, S. Roychoudhury, K. Koury, P. Li, W. V. Kalina, D. Cooper, R. W. Jr. Frenck, L. L. Hammitt, O. Tureci, H. Nell, A. Schaefer, S. Unal, D. B. Tresnan, S. Mather, P. R. Dormitzer, U. Sahin, K. U. Jansen, W. C. Gruber and The C4591001 Clinical Trial Group, N. Engl. J. Med., 2020, 383, 2603.
7. A. J. Barbier, A. Y. Jiang, P. Zhang, R. Wooster and D. G. Anderson, Nat. Biotechnol., 2022, 40, 840.
8. N. Pardi, M. J. Hogan, F. W. Porter and D. Weissman, Nat. Rev. Drug Discovery, 2018, 17, 261.
9. C. Liu, Q. Shi, X. Huang, S. Koo, N. Kong and W. Tao, Nat. Rev. Cancer, 2023, 23, 526.
10. C. L. Lorentzen, J. B. Haanen, Ö. Met and I. M. Svane, Lancet Oncol., 2022, 23, e450.
11. Y. Qu, J. Xu, T. Zhang, Q. Chen, T. Sun and C. Jiang, Acta Pharm. Sin. B, 2024, 14, 170.
12. M. Wu, Z. Luo, Z. Cai, Q. Mao, Z. Li, H. Li, C. Zhang, Y. Zhang, A. Zhong, L. Wu and X. Liu, EMBO Mol. Med., 2023, 15, e16836.
13. S. Kimura, I. A. Khalil, Y. H. A. Elewa and H. Harashima, J. Controlled Release, 2021, 330, 753.
14. J. D. Gibbins, L. R. Ancelet, T. L. Osmond, T. R. Petersen and I. F. Hermans, J. Immunother., 2015, 38, 321.
15. S. Bevers, S. A. A. Kooijmans, E. van de Velde, M. J. W. Evers, S. Seghers, J. J. J. M. Gitz-Francois, N. C. H. van Kronenburg, M. H. A. M. Fens, E. Mastrobattista, L. Hassler, H. Sork, T. Lehto, K. E. Ahmed, S. El Andaloussi, K. Fiedler, K. Breckpot, M. Maes, D. van Hoorick, T. Bastogne, R. M. Schiffelers and S. de Koker, Mol. Ther., 2022, 30, 3078.
16. L. M. Kranz, M. Diken, H. Haas, S. Kreiter, C. Loquai, K. C. Reuter, M. Meng, D. Fritz, F. Vascotto, H. Hefesha, C. Grunwitz, M. Vormehr, Y. Hüsemann, A. Selmi, A. N. Kuhn, J. Buck, E. Derhovanessian, R. Rae, S. Attig, J. Diekmann, R. A. Jabulowsky, S. Heesch, J. Hassel, P. Langguth, S. Grabbe, C. Huber, Ö. Türeci and U. Sahin, Nature, 2016, 534, 396.
17. F. Wang, J. Lou, X. Gao, L. Zhang, F. Sun, Z. Wang, T. Ji and Z. Qin, Nano Today, 2023, 52, 101943.
18. R. Shimosakai, I. A. Khalil, S. Kimura and H. Harashima, Pharmaceuticals, 2022, 15, 1017.
19. E. Rohner, R. Yang, K. S. Foo, A. Goedel and K. R. Chien, Nat. Biotechnol., 2022, 40, 1586.
20. L. Schoenmaker, D. Witzigmann, J. A. Kulkarni, R. Verbeke, G. Kersten, W. Jiskoot and D. J. A. Crommelin, Int. J. Pharm., 2021, 601, 120586.
21. X. Han, H. Zhang, K. Butowska, K. L. Swingle, M.-G. Alameh, D. Weissman and M. J. Mitchell, Nat. Commun., 2021, 12, 7233.
22. S. C. Semple, A. Akinc, J. Chen, A. P. Sandhu, B. L. Mui, C. K. Cho, D. W. Y. Sah, D. Stebbing, E. J. Crosley, E. Yaworski, I. M. Hafez, J. R. Dorkin, J. Qin, K. Lam, K. G. Rajeev, K. F. Wong, L. B. Jeffs, L. Nechev, M. L. Eisenhardt, M. Jayaraman, M. Kazem, M. A. Maier, M. Srinivasulu, M. J. Weinstein, Q. Chen, R. Alvarez, S. A. Barros, S. De, S. K. Klimuk, T. Borland, V. Kosovrasti, W. L. Cantley, Y. K. Tam, M. Manoharan, M. A. Ciufolini, M. A. Tracy, A. de Fougerolles, I. MacLachlan, P. R. Cullis, T. D. Madden and M. J. Hope, Nat. Biotechnol., 2010, 28, 172.
23. S. A. Dilliard, Q. Cheng and D. J. Siegwart, Proc. Natl. Acad. Sci. U. S. A., 2021, 118, e2109256118.
24. J. Chen, Z. Ye, C. Huang, M. Qiu, D. Song, Y. Li and Q. Xu, Proc. Natl. Acad. Sci. U. S. A., 2022, 119, e2207841119.
25. S. A. Dilliard and D. J. Siegwart, Nat. Rev. Mater., 2023, 8, 282.
26. Q. Cheng, T. Wei, L. Farbiak, L. T. Johnson, S. A. Dilliard and D. J. Siegwart, Nat. Nanotechnol., 2020, 15, 313.
27. X. Wang, S. Liu, Y. Sun, X. Yu, S. M. Lee, Q. Cheng, T. Wei, J. Gong, J. Robinson, D. Zhang, X. Lian, P. Basak and D. J. Siegwart, Nat. Protoc., 2023, 18, 265.
28. E. Álvarez-Benedicto, Z. Tian, S. Chatterjee, D. Orlando, M. Kim, E. D. Guerrero, X. Wang and D. J. Siegwart, Angew. Chem., Int. Ed., 2023, 62, e202310395.
29. Y. Xu, Y. Hu, H. Xia, S. Zhang, H. Lei, B. Yan, Z. X. Xiao, J. Chen, J. Pang and G. F. Zha, Adv. Healthcare Mater., 2023, 13, 2302691.
30. K. Lv, Z. Yu, J. Wang, N. Li, A. Wang, T. Xue, Q. Wang, Y. Shi, L. Han, W. Qin, J. Gong, H. Song, T. Zhang, C. Chang, H. Chen, X. Zhong, J. Ding, R. Chen, M. Liu, W. Zhang, S. Cen and Y. Dong, Adv. Sci., 2024, 11, 2404684.
31. S. Ramishetti, I. Hazan-Halevy, R. Palakuri, S. Chatterjee, S. Naidu Gonna, N. Dammes, I. Freilich, L. Kolik Shmuel, D. Danino and D. A. Peer, Adv. Mater., 2020, 32, 1906128.
32. O. S. Fenton, K. J. Kauffman, R. L. McClellan, J. C. Kaczmarek, M. D. Zeng, J. L. Andresen, L. H. Rhym, M. W. Heartlein, F. DeRosa and D. G. Anderson, Angew. Chem., Int. Ed., 2018, 57, 13582.
33. I. Tombácz, D. Laczkó, H. Shahnawaz, H. Muramatsu, A. Natesan, A. Yadegari, T. E. Papp, M.-G. Alameh, V. Shuvaev, B. L. Mui, Y. K. Tam, V. Muzykantov, N. Pardi, D. Weissman and H. Parhiz, Mol. Ther., 2021, 29, 3293.
34. T. T. Smith, S. B. Stephan, H. F. Moffett, L. E. McKnight, W. Ji, D. Reiman, E. Bonagofski, M. E. Wohlfahrt, S. P. S. Pillai and M. T. Stephan, Nat. Nanotechnol., 2017, 12, 813.
35. N. N. Parayath, S. B. Stephan, A. L. Koehne, P. S. Nelson and M. T. Stephan, Nat. Commun., 2020, 11, 6080.
36. M. Qiu, Y. Tang, J. Chen, R. Muriph, Z. Ye, C. Huang, J. Evans, E. P. Henske and Q. Xu, Proc. Natl. Acad. Sci. U. S. A., 2022, 119, e2116271119.
37. S. Wu, K. Su, X. Yan, L. Shi, L. Lin, E. Ren, J. Zhou, C. Zhang, Y. Song and S. Liu, J. Controlled Release, 2024, 376, 395.
38. Y. Zhang, C. Sun, C. Wang, K. E. Jankovic and Y. Dong, Chem. Rev., 2021, 121, 12181.
39. B. Li, I. O. Raji, A. G. R. Gordon, L. Sun, T. M. Raimondo, F. A. Oladimeji, A. Y. Jiang, A. Varley, R. S. Langer and D. G. Anderson, Nat. Mater., 2024, 23, 1002.
40. Y. Xu, S. Ma, H. Cui, J. Chen, S. Xu, F. Gong, A. Golubovic, M. Zhou, K. C. Wang, A. Varley, R. X. Z. Lu, B. Wang and B. Li, Nat. Commun., 2024, 15, 6305.
41. K. Wu, X. Yang, Z. Wang, N. Li, J. Zhang and L. Liu, Briefings Bioinf., 2024, 25, bbae186.
42. J. Witten, I. Raji, R. S. Manan, E. Beyer, S. Bartlett, Y. Tang, M. Ebadi, J. Lei, D. Nguyen, F. Oladimeji, A. Y. Jiang, E. MacDonald, Y. Hu, H. Mughal, A. Self, E. Collins, Z. Yan, J. F. Engelhardt, R. Langer and D. G. Anderson, Nat. Biotechnol., 2024 DOI:10.1038/s41587-024-02490-y.
43. D. Y. Ding, Y. Zhang, Y. Jia and J. Sun, arXiv, 2023, preprint, arXiv:2308.01402 DOI:10.48550/arXiv.2308.01402.
44. M. M. Billingsley, N. Gong, A. J. Mukalel, A. S. Thatte, R. El-Mayta, S. K. Patel, A. E. Metzloff, K. L. Swingle, X. Han, L. Xue, A. G. Hamilton, H. C. Safford, M.-G. Alameh, T. E. Papp, H. Parhiz, D. Weissman and M. J. Mitchell, Small, 2024, 20, e2304378.
45. M. J. Bennett, A. M. Aberle, R. P. Balasubramaniam, J. G. Malone, R. W. Malone and M. H. Nantz, J. Med. Chem., 1997, 40, 4069.
46. G. Tilstra, J. Couture-Senécal, Y. M. A. Lau, A. M. Manning, D. S. M. Wong, W. W. Janaeska, T. A. Wuraola, J. Pang and O. F. Khan, J. Am. Chem. Soc., 2023, 145, 2294.
47. J. Long, C. Yu, H. Zhang, Y. Cao, Y. Sang, H. Lu, Z. Zhang, X. Wang, H. Wang, G. Song, J. Yang and S. Wang, Adv. Healthcare Mater., 2023, 12, e2202590.
48. S. Sabnis, E. S. Kumarasinghe, T. Salerno, C. Mihai, T. Ketova, J. J. Senn, A. Lynn, A. Bulychev, I. McFadyen, J. Chan, O. Almarsson, M. G. Stanton and K. E. Benenato, Mol. Ther., 2018, 26, 1509.
49. J. Lu, T. Jiang, X. Li, S. Tan, Y. Zhang, H. Gu, L. Chen, J. Guo, R. Yu, J. Zang, D. Ouyang, H. Yu, H. Yao, M. Qiu and J. Lin, bioRxiv, 2023 DOI:10.1101/2023.10.13.562139.
50. J. Heredero, Á. Peña, E. Broset, B. Blandín, D. De Miguel, T. Alejo, A. Toro, E. Mata, A. López-Gavín, A. Gallego-Lleyda, D. Casabona, V. Lampaya, A. Larraga, E. Pérez-Herrán, D. Luna, I. Orera, E. Romanos, A. García, J. Martínez-Oliván and J. Giménez-Warren, Pharmaceutics, 2025, 17, 459.
51. M. B. Bloom, D. Perry-Lalley, P. F. Robbins, Y. Li, M. El-Gamil, S. A. Rosenberg and J. C. Yang, J. Exp. Med., 1997, 185, 453.
52. X. Han, M.-G. Alameh, Y. Xu, R. Palanki, R. El-Mayta, G. Dwivedi, K. L. Swingle, J. Xu, N. Gong, L. Xue, Q. Shi, I.-C. Yoon, C. C. Warzecha, J. M. Wilson, D. Weissman and M. J. Mitchell, Nat. Biomed. Eng., 2024, 8, 1412.

---

Footnote

† These authors contributed equally.

---