DOI:
10.1039/D5TB00750J
(Paper)
J. Mater. Chem. B, 2025,
13, 7280-7292
GPT-driven generation and biological activity evaluation of novel mRNA trinucleotide Cap1 analogs for mRNA vaccine or immunotherapy†
Received
1st April 2025
, Accepted 13th May 2025
First published on 14th May 2025
Abstract
Analogs of the mRNA 5′-cap are indispensable for mRNA translation, stability, translation efficiency, and immunogenicity, with emerging potential applications in novel preventive and therapeutic interventions. Here, this study presents a novel approach for designing mRNA Cap1 analogs with optimized biological activity. We leveraged the power of generative pre-trained transformer (GPT) architecture to generate novel cap analog sequences. A discriminative model is then employed to select promising candidates based on their predicted expression levels. Our results demonstrate that the GPT-based generative model significantly outperforms a traditional recurrent neural network (RNN) in terms of perplexity, indicating its superior ability to generate diverse and accurate cap analog sequences. Furthermore, the expression screening model achieves high accuracy in identifying potential high-expression candidates. Then, we synthesized a set of designed novel trinucleotide mRNA Cap1 analogs with modified ribose and incorporated it into mRNA using T7 polymerase. A series of experiments revealed that mRNA capped with YK-CAP-01–06 analogs exhibited increased translation efficiency and decapping enzyme stability compared to the commercially available cap-analog-capped mRNA. Finally, the potential application value was explored by constructing OVA, RSV preF- and VZV gE-mRNA vaccines, which resulted in significant (vs. controls) inhibition of tumor growth and an increase in IgG antibody levels in mice.
Introduction
The 5′ cap structure is the core element of eukaryotic mRNA, which is indispensable for pre-mRNA splicing, mRNA translation, intracellular transport and stability.1,2 It consists of 7-methylguanosine linked to the first nucleotide of the mRNA via a 5′–5′ triphosphate bridge, and it has three forms: Cap0, Cap1, and Cap2.3 Cap0 is one of the earliest discovered natural modifications of eukaryotic mRNA, with the structure of m7GpppN, where N represents any nucleotide (A, U, C or G). Cap0 in mammals is accompanied by additional methylations at the 2′-O position of the first one or two transcribed nucleotides, referred to as Cap1 (m7GpppNm) and Cap2 (m7GpppNmNm).4,5
Although the function of these methylations of the mammalian cap structure was not fully understood, it was reported that the methyl of the first transcribed nucleotide in Cap1 played a vital role for distinction between own and foreign mRNAs during viral infection in mammalian cells.6 The distinction may be related to the mechanism of protection of m7Gcapped RNAs with methylation at the 2-O-position of the first transcribed nucleotide from binding with interferon (IFN)-induced proteins with tetratricopeptide repeats (IFITs), which competes with an eukaryotic translation initiation factor 4E (eIF4E).7 Moreover, mRNA vaccines or drugs with a Cap1 structure take a stable protein expression and have low immunogenicity in vivo. Therefore, in vitro transcribed (IVT) mRNAs of medicines and vaccines commonly have a Cap1 structure, whether adopting an enzyme capping or co-transcriptional capping process.8
Cap analogs for the co-transcriptional capping process (the more widely used capping process, compared with the enzyme capping process) have gone through three generations of development. The first generation is m7GpppN-derived dinucleotides, containing Cap0, which have been used for initiating in vitro transcription. However, only about half of the synthesized capped RNA molecules are oriented correctly, and competition from NTPs (especially GTP) further reduces the production yield of effective capped RNAs. To convert them from Cap0 to Cap1, additional enzyme reactions are required, but this process is incomplete, difficult to control, and the production is difficult to purify.9 The second generation is m7G3mpppN-derived dinucleotides, referred to as anti-reverse cap analogs (ARCAs), carrying the modified m7G residue with the blocked 3′ and/or 2′ position on ribose. These ARCA cap analogs direct RNA synthesis only in the “forward” orientation and produce RNA molecules with Cap0 (having 2′ and/or 3′ modifications on the m7G residue). However, the IVT processes have the same disadvantages as those elaborated for the first-generation cap analogs, such as reduced yield caused by competition from NTPs (specifically GTP) and additional enzymatic reactions to convert ARCA Cap0 RNA to ARCA Cap1 RNA, with a non-quantitative, difficult-to-control process and hard-to-separate products.10 The third generation is m7GpppN2mpN-derived trinucleotides, referred to as Cap1 analogs, carrying methyl at the 2-O-position of the first transcribed nucleotide. Cap1 analogs make the mRNA co-transcriptional capping process simple and high-yielding, with a capping efficiency reaching over 95%. At the same time, it is possible to directly obtain the structure of Cap1 (the 2′-O methylation of the first nucleotide), which has become the most widely used cap analog in the industry.11–13
Current approaches to designing mRNA cap analogs often rely on a laborious experimental trial-and-error method, which is time-consuming and resource-intensive. This study aims to develop a high-precision generative and discriminative model for mRNA cap analogues, leveraging advanced deep learning algorithms, specifically the GPT (generative pre-trained transformer)14 architecture for generative modeling and a discriminative model for selecting promising candidates. This innovative approach has the potential to significantly advance the development of cutting-edge fields, including gene therapy, vaccine development, and antiviral drug design, by rapidly and efficiently discovering and designing novel mRNA cap analogs with tailored properties. Then, we synthesized a set of trinucleotide cap analogs with modified ribose enabling manufacturing of RNA featuring Cap1 (m7GpppAmpG) structures based on the designing result. We also preliminarily investigate the structure–activity relationship of these variously capped RNAs. We assess how the cap structure variations influence the quality of IVT transcribed mRNAs, overall protein expression of exogenously delivered mRNA in different mammalian cell lines, susceptibility to decapping, and immune response to different mRNA vaccines for tumor or infectious diseases.
Results and discussion
GPT-driven generation of novel mRNA trinucleotide Cap1 analogs with modified ribose
To evaluate the performance of the GPT-based mRNA cap analog generative model, we compared it with a mainstream recurrent neural network (RNN) model, using perplexity as the evaluation metric (Fig. 1). Perplexity measures the uncertainty of a model's predictions on a given dataset, with lower values indicating higher confidence in predicting the next token in a sequence. After training both models on the same dataset, the RNN model achieved a perplexity of 21.1355, while the GPT-based generative model exhibited a significantly lower perplexity of 3.0875, representing an 85.39% reduction. This substantial improvement demonstrates the superior accuracy of the proposed GPT-based model in predicting the next token in the sequence, significantly enhancing the quality and potential diversity of the generated mRNA cap analogs.
 |
| Fig. 1 The architecture of the deep learning models and their performance. (a) The GPT-based mRNA cap analogues generative model. The model processes input SMILES sequences, generating new cap analogue sequences. (b) The expression screening model, which utilizes transformer encoder layers to process SMILES sequences and predict their expression levels. (c) A bar chart comparing the perplexity of the GPT-based model (3.0875) with that of an RNN model (21.1355). The significantly lower perplexity of the GPT-based model indicates its superior performance in generating accurate and diverse mRNA cap analogues. | |
Furthermore, the expression screening model, designed to predict high-expression mRNA cap analogues, achieved an area under the curve (AUC) of 0.93, indicating high accuracy in identifying promising candidates. The ROC curve is shown in Fig. S15a (ESI†). From a pool of 123 generated analogues, YK-CAP-01–07 were predicted to exhibit high expression levels by the screening model and were subsequently selected for further experimental validation (Fig. S15b, ESI†). We visualized the molecules generated by AI along with those in the training set to observe the positions occupied by the structures of the generated molecules within the chemical space. The t-SNE algorithm was employed for dimensionality reduction, and the data used were 2048-bit Morgan fingerprints. Fig. S15c (ESI†) depicts the visualization results, from which we can observe that the generated molecules filled the gaps in the chemical space previously occupied by the original molecules.
Chemical synthesis of YK-CAPs
The GPT-designed seven novel Cap1 analogs feature 2′-O-methyl-modified adenosine on the first nucleotide of the triphosphate cap analogs, with different modifications in the m7G residue. The first series (Fig. 2a), named YK-CAP-01–02, possess the structure of locked nucleic acids (LNAs), with the different substituents on the methylene bridge (C6′) in the ribose ring, and specifically, the dimethylacetamide methyl group is for YK-CAP-01 and the 2,2-difluoroethyl group is for YK-CAP-02. The second series (Fig. 2b), YK-CAP-03–07, possess the different substituents in the normal ribose ring, and specifically, 2-methoxyethyl, difluoromethyl, and diethylaminocarboxy methyl groups at the 3′-O position are for YK-CAP-03–05, respectively, the methoxymethyl group at the C4′ position is for YK-CAP-06, and methylene at the 1′-C position is for YK-CAP-07.
 |
| Fig. 2 Structures and synthetic pathway of trinucleotide Cap1 analogs with modified ribose in this work. (a) Structures of the first series; (b) structures of the second series (note: n = 0 refers to modified ribose directly linking to 7-methylguanosine); (c) chemical synthesis of YK-CAP-01–02, reaction conditions: (i) trimethyl phosphate and POCl3, (ii) 2,2′-dithiodipyridine, imidazole, triphenylphosphine, triethylamine, and DMF, (iii) triethylamine phosphate, ZnCl2, and DMF, (iv) (CH3)2SO4 and 1 M NaOH, (v) ZnCl2 and DMSO. | |
It is important to note that the modified mononucleosides for YK-CAP-01–06 had basically brand-new structures and were successfully synthesized in quite good yields. Specifically, the modified mononucleosides for YK-CAP-01–02 and 06 were synthesized through 10, 11 and 4 steps of reaction using 2-O-(1-methylethylidene)-4-C-[(phenylmethoxy)methyl]-3-O-(phenylmethyl)-L-lyxofuranose as the same starting material, respectively; 1,2-O-isopropyl-α-D-ribofuranose was applied as the starting material for synthesis of modified mononucleosides of YK-CAP-03–05 with 10-, 9-, 10-step reactions; the nucleoside monomer for YK-CAP-07 was obtained through 11-step reactions using 2,3-O-isopropylidene-D-ribofuranoside as the starting material.
The synthetic pathway leading to trinucleotide Cap1 analogs included monophosphorylation of the above-mentioned mononucleosides, followed by their activation into p-imidazolide (B), and then undergoing diphosphorylation and methylation, ZnCl2-mediated coupling reaction with the imidazole derivative of pAmpG (E) in solution (Fig. 2c, taking series 1 as an example). The intermediates of monophosphates (A), diphosphates (C) and methylates (D) were isolated by ion-exchange chromatography on DEAE Sephadex to give a triethylammonium salt suitable for further reaction. The synthesis details are provided in Schemes S1–S7 in the ESI.† The final Cap1 analogs were purified by preparative high-performance liquid chromatography to give ammonium salts in good yields, and the structures were verified by proton NMR and 31P NMR (Fig. S1–S14, ESI†).
In vitro transcription reaction of 5′-capped mRNAs with YK-CAPs
mRNAs encoding firefly luciferase and capping with YK-CAP-01–07 or the commercially available CleanCap® Reagent AG (N-7113, m7GpppAmpG) and CleanCap® Reagent AG (3′OMe) (N-7413, m7G (3′OMe)pppAmpG) were synthesized by in vitro transcription using T7 Polymerase. The results (Fig. 3a and b) show that the modification of ribose in the m7G residue does not substantially impair the ability of trinucleotide cap analogs to prime in vitro transcription reaction (up to 90% of capping efficiency on 5′-capped Fluc-mRNAs obtained with YK-CAP-01–06, which was at a similar level for those obtained with CleanCap® Reagent AG, and analyzed by LC-MS to assess the ratio of capped and uncapped RNA). On the other hand, the IVT unit template yields with YK-CAP-01–06 were also at a similar level of yields with CleanCap® Reagent AG, which suggested that these novel cap analogs (except YK-CAP-07) are effective substrates for the T7 RNA polymerase.
 |
| Fig. 3 Transcription results of capped mRNAs. (a) IVT yields were determined with a spectrophotometer after initial purification. IVT reactions were incubated at 37 °C for 2 h containing 2 μg template, 2 μL 100 mM ATP, GTP, CTP, and 1-methyl-pseudouridine-5′-triphosphate, 2 μL 100 mM various cap analogues in a 20 μL reaction volume. (b) RNA capping efficiency was calculated as a fraction of capped RNA versus total RNA. (c) Fraction of decapped RNA after 45 min of incubation with hDcp2. (d–f) Integrity, purity and dsRNA content of mRNA after IVT. Statistical significance was calculated using a one-way ANOVA (analysis of variance) with Dunnett's multiple comparisons test, and data are shown as mean ± SD (ns: no significant difference, *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001). | |
Surprisingly, the experiments provided IVT unit template yields with YK-CAP-01–02 of 151.0–163.2 μg versus the yields with YK-CAP-03–06 of 153.2–173.5 μg, which indicated that the configuration of LNA allowed bulkier substituents on the methylene bridge (C6′) in the ribose ring with no significant interference for IVT. Moreover, in terms of IVT yield, capping efficiency, integrity and susceptibility to hDcp2, these hydrophobic substituents (dimethylacetamide methyl group for YK-CAP-01 and the 2,2-difluoroethyl group for YK-CAP-02) demonstrated superior performance compared to the unmodified LNA (Fig. S16a–d, ESI†). In both in vitro and in vivo firefly luciferase expression experiments, YK-CAP-01 exhibited expression levels comparable to those of unmodified LNP, whereas YK-CAP-02 demonstrated significantly higher expression levels (Fig. S16e and f, ESI†).
Susceptibility to the decapping enzyme hDcp2
The stability of 5-capped mRNAs in complex biological mixtures in vivo is quite often challenged from hydrolysis by a cellular decapping enzyme. In eukaryotes, there are two main mRNA degradation pathways (5′-to-3′ and 3′-to-5′), and decapping is degradation pathways (5′-to-3′ and 3′-to-5′), and decapping is the first irreversible event in the 5′-to-3′ mRNA degradation pathway, which is executed by the Dcp1/2 complex as the main decapping enzyme.4 Dcp2 is a catalytic subunit belonging to the Nudix family of hydrolases, whereas Dcp1 is a regulatory subunit, which can interact with additional decapping enhancers. The hydrolysis of 5′-capped mRNAs between the α and β phosphates in the 5′–5′ triphosphate chain catalyzed by Dcp2 yield 7-methylguanosine diphosphate and 5′ monophosphorylated mRNA, which are no longer available for translational machinery and will undergo further degradation by 5′-exonucleases. Therefore, the stability of mRNA can be improved by increasing its resistance to Dcp2, to prolong and increase overall protein expression.15
In vitro susceptibility to decapping of transcripts capped with analogs YK-CAP-01–07 was investigated by incubation of 5′-capped short RNAs and human Dcp2 (hDcp2) at 37 °C for 45 minutes, followed by separation of capped and decapped RNAs by polyacrylamide gel electrophoresis (PAGE). The gels were stained with SYBR Green II, visualized and the intensity of the bands corresponding to capped and decapped RNAs was quantified using a Typhoon FLA 7000 instrument. The results, shown in Fig. 3c and Fig. S17 (ESI†), indicated that the YK-CAP analogs obviously improve the stability of 5′-capped mRNAs, with percentages of decapped RNA to total RNA with YK-CAP-01–06 of 11.8%, 10.3%, 11.5%, 10.2%, 11.5% and 9.8%, respectively, compared to 48.6% for the reference analog CleanCap® Reagent AG. In particular, RNA capped with YK-CAP-06 has the largest resistance against hDcp2, which suggested that bulky substituents at the C4′ position in the ribose ring can improve mRNA stability against decapping enzymes.
Translational efficiency of 5′ capped mRNAs with YK-CAPs
To investigate translational efficiency in vitro and in vivo, 5′ capped mRNAs with YK-CAPs analogs and CleanCap® Reagent AG were encapsulated in lipid nanoparticles (LNPs), which were formulated with ionizable cationic lipid YK-009, 1,2-distearoyl-sn-glycero-3-phosphocholine (DSPC), cholesterol and 1,2-dimyristoyl-rac-glycero-3-methoxypolyethylene glycol-2000 (DMG-PEG) at a molar ratio of 50
:
10
:
38.5
:
1.5 using microfluidic mixing, according to our previous publications and ratio optimization study.16 The resulted LNPs were characterized as excellent physicochemical properties, with a hydrodynamic diameter of 60 to 90 nm, polydispersity index (PDI) of less than 0.1, and encapsulation efficiency of higher than 90% (Fig. 4a–c).
 |
| Fig. 4
In vitro and in vivo delivery of LNP-mRNA-capped with cap analogues. (a) Particle size of capped mRNA LNPs. (b) PDI of capped mRNA LNPs. (c) Encapsulation efficiency of capped mRNA LNPs. (d) Relative luciferase expression of HEK-293T, C2C12, A204 and DC2.4 cells after incubation with LNPs-Fluc mRNA for 24 h. (e) Protein expression of HEK-293T cells following 24 h of incubation with RSV mRNA-capped with cap analogues. (f) Protein expression of HEK-293T cells following 24 h of incubation with VZV mRNA-capped with cap analogues. (g) Quantified total luminescence of the muscle sites of female BALB/c mice injected via an intramuscular route at the indicated time points. (h) Bioluminescence image of female BALB/c mice injected via an intramuscular route at the indicated time points. Statistical significance was calculated using one-way ANOVA (analysis of variance) with Dunnett's multiple comparisons test for 4a–4f and a two-way ANOVA with Tukey's multiple comparisons test for 4g, (ns: no significant difference, *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001). Data represent the mean ± SD. | |
Then, we prepared a series of mRNA-LNPs encoding different proteins and investigated the translational activity in various cultured cell lines to study the biological activity of mRNAs with YK-CAPs in vitro. Fluc-encoding mRNAs capped with YK-CAP-01–07 and CleanCap® Reagent AG were transfected into human kidney embryonic cells (HEK293T), human striated muscle sarcoma cells (A204), mouse myoblasts cells (C2C12) and mouse dendritic cells (DC2.4), and the luminescence dependent on the Fluc protein production was determined (Fig. 4d). It was found that the translational properties of mRNA varied among different cell lines. In HEK293T cells, the translation efficiencies of YK-CAP-02 and YK-CAP-04 capped mRNAs were 1.5- and 1.7-fold higher than that of CleanCap® Reagent AG-capped RNA, whereas the translation efficiencies of YK-CAP-07-capped mRNAs was 70% of that of CleanCap® Reagent AG-capped RNA. CleanCap® Reagent AG (3′OMe) was also evaluated as the control to transfect into HEK293T cells (Fig. S18, ESI†). CleanCap® Reagent AG (3′OMe) has similar properties compared with CleanCap® Reagent AG. As a result, we used CleanCap® Reagent AG as the sole control group in the subsequent experiments. A similar trend could be observed in A204 cells. On the other hand, in C2C12 and DC2.4 cells, the mRNA carrying YK-CAP-03 analogs demonstrated were 1.6- and 2.5-fold higher than that of commercially available CleanCap® Reagent AG-capped RNA, respectively, whereas the mRNA carrying YK-CAP-02, YK-CAP-05 and YK-CAP-06 demonstrated similar transfection efficiency to that of CleanCap® Reagent AG-capped RNA. Next, we investigated whether mRNAs capped with the YK-CAP-01–06 would yield more protein than transcript capped mRNAs with CleanCap® Reagent AG in specific more complex mRNA vaccines for infectious diseases, i.e. respiratory syncytial virus (RSV) and varicella-zoster virus (VZV). As expected, RSV preF-mRNA capped with YK-02–04 yielded 4.3 × 103 ng, 5.0 × 103 ng and 4.4 × 103 ng protein produced by each 3 × 105 cells, respectively, compared to 3.5 × 103 ng for reference analogs CleanCap® Reagent AG (Fig. 4e). In the case of VZV gE-mRNA, we also observed 1.4- to 1.6-fold higher expression of the protein from mRNAs capped with YK-CAP-02–04 (Fig. 4f).
Furthermore, we conducted the evaluation of in vivo expression using Fluc-mRNA capped with YK-CAP-01–06 and CleanCap® Reagent AG. The LNPs formulated with 5′ capped Fluc-mRNA were administered into mice via intravenous injection, and at 6-, 12-, 24-, 48-, 96- and 168-hours post-injection, the luciferase activity was determined (Fig. 4g and h). The highest activities of differently capped mRNAs were observed at 6 h or 12 h, followed by a gradual decline, but the expression levels remained detectable up to 7 days. A notable difference between YK-CAP-02–04 and CleanCap® Reagent AG-capped mRNAs (1.9–2.8-fold higher expression at 6 h time point and 2.2–2.9-fold at 12 h) was also shown. This suggests that these novel mRNA Cap1 analogs possess greater capacity to execute higher translational activity.
Superior therapeutic and preventive activity of 5′ capped mRNAs with YK-CAPs
Recent advances in mRNA vaccine technology have revolutionized the development of prophylactic and therapeutic vaccines. A critical factor influencing their efficacy and safety is the 5′ cap structure, which determines the stability, translational efficiency, and immunogenicity of the mRNA. Cap analogs that balance translation efficiency and low immunogenicity are crucial to avoid interferon-driven antigen silencing—a phenomenon observed with earlier uncapped mRNAs for tumor associated antigen vaccines.17 In the RSV vaccine landscape, Moderna's mRNA-1345 demonstrated ∼68% efficacy against RSV-associated acute respiratory disease in older adults, highlighting the importance of optimized capped mRNA in achieving robust neutralizing antibody responses.18 For VZV, where glycoprotein E (gE)-targeting mRNA vaccines can elicit immune responses that are markedly superior to those induced by live-attenuated vaccines,19 cap modifications could enhance stability and improve translation efficiency. Preliminary data indicate that certain synthetic caps (e.g., YK-CAP-02, YK-CAP-03 and YK-CAP-04) increase protein yield in striated muscle sarcoma cells and dendritic cells (Fig. 4d), potentially boosting vaccine potency in immunocompromised populations.
To further validate if mRNAs capped with YK-CAPs can exert therapeutic and preventive effects, three different mRNAs that are relevant for therapeutic or preventive applications (OVA-, RSV preF- and VZV gE-mRNA) were used in in vivo experiments in mice. YK-CAP-02, YK-CAP-03 and YK-CAP-04 were chosen as the research caps, which performed excellently based on the previous evaluations, and commercially available CleanCap® Reagent AG as the reference CAP. First, the antigen-specific CD8+ T cell response in the spleens of C57BL/6J mice was evaluated. Immunizing the C57BL/6J mice with three doses of OVA mRNA-LNP (10 μg mRNA per mouse), encoding the SIINFEKL peptide, led to significantly higher proliferation of OVA specific cytotoxic T lymphocytes (CTLs) in the spleen compared with the blank control group (DPBS), which suggested that OVA-mRNA capped with YK-CAP-02, YK-CAP-03, YK-CAP-04 can induce effective immune responses, facilitating a tumor immunotherapy effect (Fig. S19, ESI†). Then, the antitumor effect was investigated using the OVA mRNA encoding SIINFEKL peptide, a cancer-associated antigen, in an OVA-expressing MC38 murine model for colorectal carcinoma. Mice were SC inoculated with MC38-OVA cells on the right flank (day-3). Once the tumor volume reached approximately 50 mm3, the mice were immunized with 10 μg of OVA mRNA capped with YK-CAP-02, YK-CAP-03, YK-CAP-04 and CleanCap® Reagent AG, via IM injection at intervals on days 0, 3, 7 and 14. Statistically significant (vs. controls) inhibition of tumor growth was observed in mice that received OVA-mRNA, capped with YK-CAP-02, YK-CAP-03, YK-CAP-04 and CleanCap® Reagent AG. And all of the mice treated with OVA-mRNA capped with YK-CAP-02, YK-CAP-03 and YK-CAP-04 at different time points exhibited smaller tumor volumes than those with CleanCap® Reagent AG, although there was no statistical significance (Fig. 5b). Simultaneously, the survival of mice vaccinated with varying capped OVA-mRNAs was significantly prolonged compared to the control group until the humane endpoint was reached (a tumor volume of 2000 mm3) (Fig. 5c). Moreover, compared to mice vaccinated with CleanCap® Reagent AG capped OVA-mRNA, those vaccinated with OVA-mRNA vaccines capped by YK-CAP-02, YK-CAP-03, and YK-CAP-04 exhibited slightly extended survival times.
 |
| Fig. 5 Therapeutic and preventive activity of 5′ capped mRNAs with YK-CAPs. (a) Schematic of the tumor vaccination regimen. Mice were inoculated the tumor cells on day-3 and received intramuscular injections on days 0, 3, 7 and 14. (b) Tumor volumes of the MC38-OVA tumor model in the immunotherapeutic study. (c) Survival curves of the MC38-OVA tumor model in the immunotherapeutic study. (d) Schematic of the RSV or VZV vaccination regimen. Each mouse received intramuscular injections on days 0, 28 and 56. Serum was collected on days 14, 42 and 70. (e) RSV PreF IgG titerin serum after 14 d post-injection of RSV mRNA-capped with cap analogues via intramuscular injection. (f) VZV gE IgG titer in serum after 14 d post-injection of VZV mRNA-capped with cap analogues via intramuscular injection. Statistical significance was calculated using a two-way ANOVA with Tukey's multiple comparisons test (ns: no significant difference, *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001). Data represent the mean ± SD. | |
Encouraged by the results showing that mRNA capped with YK-CAP-02, YK-CAP-03 and YK-CAP-04 encoding more RSV preF and VZV gE proteins (Fig. 4e and f), we investigated the preventive activity of mRNA encoding RSV preF and VZV gE proteins in two different mRNA vaccines for infectious diseases. Mice were treated with RSV preF- and VZV gE-mRNA-LNP vaccines (2 μg per mouse) via intramuscular injection on days 0, 28 and 56, respectively. Then, blood from the orbital sinus of the mice was collected on days 14, 42 and 70 followed by centrifugation to separate the serum, and then the levels of humoral immune IgG antibodies were detected. The results are shown in Fig. 5e and f. For both RSV preF- and VZV gE-mRNA vaccines, mRNAs with all selected analogs exhibited IgG antibody titers much more than those of CleanCap® Reagent AG, indicating that these analogs can elicit more robust immune response, significantly enhancing antibody levels in mice. Therefore, the novel cap analogs play a crucial role in improving the stability and efficacy of mRNA vaccines. During the above experiments, no gross adverse effects of mRNA (capped with either selected YK-CAPs or CleanCap® Reagent AG) administration were noticed in mice, such as weight loss, lethargy, anorexia, diarrhea, ruffled hair coat or neurological impairment, which could preliminarily reveal the safety of these YK-CAP analogs.
Safety of 5′capped mRNAs with YK-CAPs in vivo
The safety of these 5′capped mRNAs with YK-CAP-02–04 encapsulated in LNPs were evaluated to ensure their suitability for clinical use. First, routine blood examination was conducted to evaluate the potential liver and kidney toxicity. No significant changes were observed in the serum levels of aspartate transaminase (AST), alanine transaminase (ALT), uric acid (UREA) or creatinine (CREA) compared to the control group and the group with CleanCap® Reagent AG, indicating the absence of liver and kidney toxicity (Fig. 6a). Then, the major organs, including the heart, liver, spleen, lungs, and kidneys, were collected from mice for histopathological examination, which showed that no obvious pathological changes were observed in any of the organs (Fig. 6b).
 |
| Fig. 6
In vivo safety evaluation of 5′capped mRNAs with YK-CAPs. (a) Serum levels of aspartate transaminase (AST), alanine transaminase (ALT), uric acid (UREA), and creatinine (CREA) in BALB/c mice vaccinated with LNPs encapsulated with 5′capped OVA mRNAs with YK-CAP-02–04 and CleanCap® Reagent AG, measured on day 12 after three vaccinations on days 0, 3, and 7 (n = 3). The blank refers to DPBS-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 5′capped OVA-mRNA-LNPs and DPBS (n = 3). (c) Serum cytokines at 6 h after one vaccination (n = 3). Statistics were assessed by a one-way ANOVA with Dunnett's multiple comparisons test. ns means no significant difference. Data represent the mean ± SD. | |
The hematological changes of these 5′capped mRNAs with YK-CAP-02–04 encapsulated in LNPs administered via intramuscular injection were also evaluated in Sprague-Dawley rats. No significant changes occurred in the cell counts of lymphocytes after immunization, while transient slight increases in the cell counts of white blood cells, monocytes, neutrophils, eosinophils, and basophils were observed in the LNPs encapsulated with 5′capped OVA mRNAs with YK-CAP-02–04 and CleanCap® Reagent AG groups shortly after immunization, and then normalized to levels comparable to the DPBS group (Fig. S20, ESI†).
The 5′caps can be sensed by endosomal pathogen recognition receptors (PRRs) (toll-like receptor (TLR) 3/7/8) and several cytosolic receptors (e.g. retinoic acid-inducible gene-I (RIGI)-like receptors), which subsequently mediates pro-inflammatory cytokine release.20 The levels of key cytokines, including IL-1β, TNF-α, IL-6, and IFN-γ, were assessed. Compared to the control group, cytokine levels generally increased (Fig. 5b). However, compared to the group with CleanCap® Reagent AG, there were no significant changes in the levels of IL-1β, TNF-α, and IL-6 and IFN-γ after treated with LNP formulations encapsulating YK-CAP-02–04 capping mRNAs (Fig. 6c). These results support the safety of the 5′capped mRNAs with novel YK-CAPs in mice.
Experimental
Datasets
This study utilized a dataset comprising 123 mRNA cap analogs sourced from patent data.21–23 Each analogue was characterized by its expression profile and the corresponding SMILES-encoded structural sequence. To classify the expression levels, a threshold of 1 was established. Analogs with expression values exceeding 1 were categorized as “high” (label of “1”), while those with values at or below 1 were deemed “low” (label of “0”). The generative model for mRNA cap analog design was trained exclusively on the “high” expression dataset. In contrast, the discriminative model, tasked with predicting expression levels, was trained on the entire dataset encompassing both “high” and “low” expression analogs.
Sequence tokenization
The SMILES sequences representing the structural information of the mRNA cap analogs were tokenized using the byte pair encoding (BPE) algorithm.24 This data preprocessing technique involves iteratively merging the most frequent pairs of sub-tokens within the dataset. Initially, each character in the SMILES string is considered an individual token. Subsequently, the BPE algorithm repeatedly merges the most frequent pairs of sub-tokens, effectively creating a vocabulary of sub-tokens with varying lengths. This process continued until a total of 1335 merges were performed, resulting in a final vocabulary size of 261 unique sub-tokens. Finally, the tokenized SMILES sequences were converted into numerical representations based on their corresponding token indices in the BPE vocabulary. These numerical encodings served as the input for training the AI models.
Model architecture
The mRNA cap analog generative model was built upon the GPT architecture. First, we employed an embedding layer to embed SMILES strings. The embedding space was set to a size of 260, with a feature dimension of 256. Subsequently, we utilized transformer encoder layers to extract features from the embedded representations. Each transformer encoder layer consists of a pre-layer normalization layer, a multi-head attention layer, a post-layer normalization layer, and a feed-forward neural network layer. In total, we used three transformer encoder layers, and the feature dimension was 256. Finally, we de-coded the latent features obtained from the transformer encoder to generate SMILES strings (Fig. 1a).
Each layer in this model comprises a sequence of components: layer normalization followed by residual connections, a masked multi-head attention layer, another set of layer normalization and residual connections, and finally, a feed-forward neural network (MLP). The mRNA cap analogs discriminative model architecture commences with an input embedding layer that processes the tokenized SMILES sequences. Subsequently, two transformer encoder layers are employed, each comprising layer normalization, residual connections, masked multi-head attention with three heads, and a multi-layer perceptron (MLP). Following these encoder layers, a final MLP is utilized for binary classification, predicting whether the expression level of the mRNA cap analogs is high or low (Fig. 1b).
Model implementation
The preprocessed data were divided into a training set (60%), a validation set (20%), and a test set (20%). All training processes were executed on a server equipped with four NVIDIA A100 80GB SXM GPUs, utilizing the PyTorch framework (version 2.1.2)25 with Python 3.11.7 and the Nadam optimizer.26 The experimental environment was Alibaba Cloud Linux 3 (OpenAnolis Edition 3.2104 U10). During training, a batch size of 32 and a learning rate of 0.001 were employed, with 500 training epochs conducted. The detailed information of hyper-parameters is shown in Table S1 (ESI†).
To evaluate the mRNA cap analogs generative model, an RNN (recurrent neural network) model was trained on the same data. The perplexity of both the GPT-based model and the RNN model was calculated, and the results were compared to assess the performance of the GPT-based generative model.
Methods
The crude products were purified with a BUCHI Pure C830 chromatography purification system in the modified mononucleoside synthesis process for YK-CAP-01–07, while the AKTA Avant150 protein purification system and Agilent 1260 Infinity II HPLC system were applied in the cap analog synthesis process. The LCMS and 1H-NMR were measured on a Vanquish-ISQ EM LC/MS System and Quantum-Iplus AS400/ZA0028 NMR spectrometer, respectively.
Animal studies
All the animal experiments were kept to the National Regulation of China for Care and Use of Laboratory Animals (approval number: HLZG-DWLL-2024-0506-02). All female BALB/c mice, female C57BL/6 mice and female Sprague-Dawley rats were purchased from Beijing Huafukang Biotech Co., Ltd (Beijing, China). The mice were housed in a ventilated, temperature-controlled, and illumination-controlled animal facility without specific pathogens (20 ± 2 °C; 50% ± 10%; illumination, 8:00–20:00; dark, 20:00–8:00). For antitumor evaluation, euthanasia was performed on surviving animals once the tumor volume exceeded 2000 mm3.
Synthesis of cap analogs
Procedures for the synthesis and characterization of the cap analogs are given in the ESI.†
Preparation of linearized plasmids for DNA templates
All antigen gene fragments were artificially synthesized and cloned into the pVAX vector (ThermoFisher) by GenScript Biotech Corporation, resulting in the successful construction of circular plasmids. The linearized template preparation involved linearization of the plasmid DNA using the BsaI-HFv2 restriction enzyme (New England Biolabs). The digestion system consisted of 1 μg of circular plasmids, 5 μL of 10× rCutSmart Buffer, 1 μL of BsaI-HFv2, and approximately 44 μL of DNase/RNase-free water. The digestion reaction was carried out at 37 °C for 15–30 minutes. After the digestion reaction, the linearized product was purified using a DNA Purification Kit (Tiangen Biotech), and the concentration of the product was measured using the microvolume UV-vis spectrophotometer (Denovix).
In vitro transcription of mRNA
Fluc mRNA, OVA mRNA, preF mRNA and gE mRNA capped with different cap analogs were produced by in vitro transcription. The NTPs, the CleanCap® Reagent AG and CleanCap® Reagent AG (3′OMe), nuclease inhibitor, inorganic pyrophosphatase and T7 RNA polymerase were purchased from Hongene Biotech Co. Ltd. For each mRNA synthesis, DNase/RNase-free water was first added in a 200 μL PCR tube, followed by 2 μL 10× buffer, 2 μL for each 100 mM NTPs (ATP, CTP, GTP, 1-methyl-pseudouridine-5′-triphosphate), and 2 μL 100 mM cap analogs. After each addition, the mixture was thoroughly mixed and then gently centrifuged to collect the contents. Next, a 20 U RNase inhibitor, 0.05 U inorganic pyrophosphatase, 50 U T7 RNA polymerase, and 1 μg linearized DNA template were added and the volume was adjusted to 20 μL with DNase/RNase-free water. After 2 hours incubation at 37 °C, DNase I (1 U) was introduced in the system and continued to incubate at 37 °C for another 30 minutes to remove the DNA template. Then, RNA purification was undertaken using an AKTA avant25 (Cytiva) with 5 mL POROS GoPure Oligo dT25 columns (ThermoFisher). The purified mRNA is resuspended in DNase/RNase-free water, and the IVT yield was quantified using a microvolume UV-vis spectrophotometer (Denovix) and the integrity and purity were evaluated using a capillary electrophoresis system (Agilent 5200 Fragment Analyzer).
Determination of the capping efficiency of mRNA
To determine the capping efficiency, magnetic bead purification was utilized. Prior to measurement, the mRNA was first hybridized with biotinylated DNA–RNA hybridization probes. Next, 120 μL of the mRNA sample was incubated with 100 μL of pre-treated streptavidin-encapsulated magnetic beads at room temperature for 30 minutes, with gentle mixing throughout the incubation. Subsequently, 20 μL of RNase H (5 U μL−1) was added, and the mixture was incubated at 37 °C for 3 hours, with mixing every half hour. After incubation, the magnetic beads were washed, and then 100 μL of 75% methanol heated to 80 °C was added to the beads. The mixture was heated on a hot plate to 80 °C for 3 minutes, then placed on a magnetic rack to collect the supernatant. The sample was dried using a SpeedVac centrifuge at room temperature for 45 minutes until the volume was reduced to 10 μL. The sample was then re-suspended in 50 μL of 100 μM EDTA/1% MeOH and was ready for LC-MS analysis. Since capped and non-capped bases have a significant difference in the molecular weight, the capping efficiency of mRNA transcription initiated by different capping analogs can be determined by the molecular mass difference.
RNA decapping assay using hDcp2
Short RNAs (27-nt) were synthesized with modified cap analogs first as described previously.8 Then, the capped RNAs (30 ng each) were subjected to digestion with 10 nM human Dcp2 (Biofeng) in 1× MDE (50 mM Tris–HCl pH 7.5, 50 mM NH4Cl, 0.1% poloxamer 188, 1 mM DTT, and 5 mM MgCl2). Reactions were performed at 37 °C for 45 min. After terminating, the hDcp2-treated RNAs were resolved electrophoretically on denaturing the 15% acrylamide/7 M urea 1× TBE gel stained with SYBR Green II (Lonza) and visualized using a Typhoon FLA 7000 (GE Healthcare).
Preparation of LNP-mRNAs
Lipid nanoparticles were prepared by mixing the lipids and mRNA. Overall, ionizable cationic lipid YK-009, 1,2-distearoyl-sn-glycero-3-phosphocholine (DSPC, IVT), cholesterol (IVT) and 1,2-dimyristoyl-rac-glycero-3-methoxypolyethylene glycol-2000 (DMG-PEG 2000, Sinopeg) at a molar ratio of 50
:
10
:
38.5
:
1.5 were fully dissolved in ethanol to obtain a lipid–ethanol solution. The varied capped Fluc-mRNA, OVA-mRNA, RSV preF-mRNA or VZV gE-mRNA were diluted with a 20 mM citrate buffer (pH = 4–5). Next, the lipid and mRNA solution (with a volume ratio of 1
:
3) were mixed via microfluidics at a total flow rate of 12 mL min−1. Then, the formulations were diluted against the 10-fold volume of PBS (phosphate-buffered saline) buffer (pH = 7.4, Life Science), followed by the 300 kDa PES (polyether sulfone) ultracentrifugal filters to remove the ethanol. Finally, the LNP encapsulated Fluc-mRNA, OVA-mRNA, RSV preF-mRNA or VZV gE-mRNA were obtained after diluting the mixtures with PBS to a certain volume and filtering through a 0.22 μm filter.
Physiochemical characterization of LNP-mRNAs
The particle size and polydispersity index (PDI) of the formulations were ascertained using dynamic light scattering with a Malvern Zetasizer Ultra. To begin, a 25 μL aliquot of each formulation was diluted to a final volume of 125 μL with normal saline and subsequently introduced into the sample cell. Each sample was subjected to triplicate measurements to ensure the accuracy and reproducibility of the results. The encapsulation efficiency was determined using a Quant-iT™ RiboGreen RNA Quantitation Kit (ThermoFisher).
Luminescence detection of Fluc-mRNA
Prior to transfection, in vitro transcription was performed using different cap analogs with the firefly luciferase-encoding sequence and then the Fluc mRNA-LNPs were prepared following the method mentioned previously. Next, the resulting formulations were transfected into HEK293T, A204, C2C12, and DC2.4 cell lines, respectively. Cells were seeded at a density of 10
000 per well in a 96-well plate. When the cell confluence reached approximately 80%, 50 ng of mRNA was transfected into each well. After transfection, the cells were incubated in a 37 °C, 5% CO2 incubator for 24 hours. The growth medium was then removed from the wells, and the cells were washed with PBS. After centrifugation to remove PBS, 50 μL of 1× passive lysis buffer was added, and the cells along with all liquids were transferred to the microcentrifuge tubes, followed by centrifugation to obtain the sample. 20 μL of the sample were mixed with 100 μL of the Dual-Lumi™ II Firefly Luciferase Detection Reagent (Beyotime) at room temperature and the fluorescence expression intensity of each well was detected with a microplate reader (BioTek Synergy H1) at 450 nm.
Protein expression studies of RSV preF- and VZV gE-mRNA in cells
The resulting RSV preF-mRNA or VZV gE-mRNA LNP formulations (0.5 μg mRNA) were transfected into the HEK293T cells that were seeded at a density of 2.5 × 105 per well in a 12-well plate and incubated in a 37 °C, 5% CO2 incubator for 16 hours. The growth medium was then removed from the wells, and the cells were washed with PBS. After centrifugation to remove PBS, 50 μL of RIPA lysis buffer with a 0.5 μL protease inhibitor was added and placed at 0 °C for 30 min. To obtain the protein samples, the cells along with all liquids were transferred to the microcentrifuge tubes and centrifuged at 12
000 rpm for 10 min to collect the supernatant. Finally, the concentration of the protein samples was determined with a BCA assay kit and the protein expression efficiency of varied capped RSV preF- and VZV gE-mRNA vaccines was detected by ELISA using a Human RSV-preF ELISA Kit (Sino Biological, KIT11049PR) and VZV-gE ELISA Kit (Acro Biosystem, RAS-A185), respectively.
Delivery performance of LNP-mRNA in vivo
The female BALB/c mice (4–6 weeks) were employed to evaluate the in vivo delivery performance of the LNP-mRNA with different cap analogs. Briefly, each of the capped FLuc-mRNA LNPs (containing 10 μg mRNA) was injected into the mice via an intramuscular (IM) route. At 6 h, 12 h, 24 h, 48 h, 96 h, and 128 h post-injections, the mice were injected intraperitoneally with 3 mg D-luciferin potassium. After 5 min, the bioluminescence signals were detected with an IVIS Spectrum instrument (AniView 600).
OVA-specific T-cell analysis
Female C57BL/6J mice, aged 6–8 weeks, were vaccinated via IM injection on days 0, 3, and 7, each injection containing 10 μg of 5′capped OVA-mRNA encapsulated in LNPs. After 5 days post-administration, the mice were sacrificed and the spleen was collected. Splenocytes of the immunized mice were blocked with anti-CD16/32 at 4 °C for 10 min and stained with fluorophore-labeled antibodies of Tselect H-2Kb OVA Tetramer-SIINFEKL for 20 min. Then, the cells were washed with 1× FACS buffer three times and stained with fluorophore-labeled anti-CD8 antibodies for another 20 min. The stained cells were washed with 1× FACS buffer and analyzed by flow cytometry.
Anti-tumor immunotherapy in vivo
Before commencing the study, female C57BL/6J mice, aged 6–8 weeks, were allowed to acclimate for at least three days. They were provided sufficient food and sterile water, maintained under a controlled environment with a 12-hour light/dark cycle, temperature maintained at 22 °C ± 2 °C, and relative humidity at 55% ± 15%. MC38-OVA cells were cultured at 37 °C in a 5% CO2 incubator in a complete growth medium, according to the instruction manual. For harvesting, cells were treated with 0.25% trypsin-EDTA and then resuspended in Dulbecco's phosphate-buffered saline (DPBS, Life Science). Then, MC38-OVA cells were subcutaneously (SC) injected into the right flanks of female C57BL/6J mice at a dosage of 2 × 105 cells per 100 μL per mouse to establish a subcutaneous MC38-OVA tumor model. Vaccination was typically initiated when the tumor volume reached approximately 50 mm3. The mice were vaccinated via IM injection on days 0, 3, 7, and 14 post-cell implantations (each injection containing 10 μg of OVA-mRNA). A control group was also established, consisting of mice injected with an equal volume of PBS solution. Each group contained eight mice. Tumor diameters were measured every three days starting from day 0 post-tumor inoculation.
Detection of humoral immune IgG antibody levels induced by varied capped RSV preF- and VZV gE-mRNA vaccines
The female BALB/c mice that were 6–8 weeks were randomly assigned with 6 mice for each group. Then, they were injected varied capped RSV preF- or VZV gE-mRNA vaccines (2 μg per mouse) via IM on days 0, 28 and 56, respectively. Then, blood from the orbital sinus of the mice was collected on days 14, 42 and 70, and allowed to keep at room temperature for 2 hours, and then placed in a 4 °C refrigerator for 4 hours. A yellow supernatant, which was the serum, was obtained by centrifuging the blood at 4000 rpm for 10 minutes. Next, the serum samples were diluted and added into a 96-well plate that was pre-coated with an antigen. After incubated at room temperature for 120 minutes, the mouse IgG antibody was conjugated by horseradish peroxidase (HRP) and continuously incubated for 60 minutes in the dark at room temperature. Finally, the concentration of the IgG antibody was calculated by comparing the absorbance values with the standard curve once the 3,3′,5,5′-tetramethylbenzidine (TMB, ThermoFisher) was added and the absorbance at 450 nm was measured with a microplate reader.
Toxicity evaluation
To evaluate the safety of these 5′capped OVA mRNAs encapsulated in LNPs, female C57BL/6J mice (6–8 weeks) were intramuscularly injected with varying capped mRNA LNPs (10 μg of mRNA), with DPBS used as a control. Serum samples were collected 6 h post-injection and the cytokine levels of IL-1β, TNF-α, IL-6, and IFN-γ were determined with an ELISA plate reader, according to the manufacturer's instructions (Elabscience). Sera were obtained on day 12 after three injections on days 0, 3, and 7 to assess liver and kidney function (Automated Biochemical Analyzer). The major organs, including the heart, liver, spleen, lungs, and kidneys, were collected on day 12 after three injections on days 0, 3, and 7 to prepare the histopathological sections. Female Sprague-Dawley rats (6–8 weeks) were vaccinated with LNPs encapsulated with 5′ capped OVA mRNAs with YK-CAP-02–04 and CleanCap® Reagent AG (50 μg of mRNA) on day 0, with DPBS used as a control. The counts of various white blood cells (white blood cells, neutrophils, basophils, monocytes, eosinophils and lymphocytes) were determined with an automatic blood analyzer on days -2, 5 and 14.
Statistical analysis
Statistical analyses were performed using GraphPad Prism 8.0.2 (GraphPad Software). All the data are presented as the mean ± SEM. Statistical differences were analyzed by the one-way ANOVA with Dunnett's multiple comparisons test or two-way ANOVA with Tukey's multiple comparisons test. All tests were considered statistically significant when the p value was < 0.05.
Conclusions
Cap analogs that mimic the cap structure at the 5′ end of eukaryotic mRNAs have a significant impact on mRNA stability, translational efficiency, and immunogenicity. Cap1 analogs make the mRNA co-transcriptional capping process simple and high-yielding, with the capping efficiency reaching over 95%. Consequently, a novel approach for designing mRNA Cap1 analogs with optimized biological activity through a GPT-based generative model is presented in this work. YK-CAP-01–07 were predicted to exhibit high expression levels by the screening model and were subsequently synthesized for further experimental validation. The capping efficiency and mRNA yield of YK-CAP-01–06 during in vitro transcription had no statistical difference with those of the commercially available CleanCap® Reagent AG, and they also exhibited better stability against the decapping enzyme hDcp2. Notably, mRNAs capped with YK-CAP-02, YK-CAP-03, and YK-CAP-04 demonstrated superiority over those with CleanCap® Reagent AG in translational efficiency in various cell lines and in vivo imaging experiments, confirming their ability to enhance mRNA stability and expression efficiency. Finally, three different therapeutically or preventively relevant mRNA were used in in vivo experiments in mice to further validate mRNAs capped with these three selected YK-CAPs exerting therapeutic and preventive effects. In conclusion, the novel cap analogs with modified ribose play a crucial role in improving the stability and efficacy of mRNA vaccines and there is enormous potential for application in the development of tumor immunotherapy or preventive mRNA vaccines.
Author contributions
G. S. and F. Y. provided project administration and supervision; H. Z. performed conceptualization, investigation, data curation, formal analysis, software and writing – original draft; J. M. performed the formal analysis, writing – original draft and writing – review and editing; T. M., Y. M., L. J., L. L., Y. L., and K. D. performed methodology; M. Z. and D. H. performed validation.
Data availability
The data supporting this article have been included as part of the ESI.†
Conflicts of interest
All authors were employed by the Beijing Youcare Kechuang Pharmaceutical Technology Co. Ltd. G. S., H. Z., Y. M., L. J., Y. L., K. D., and D. H. are inventors on a pending patent related to this work filed by the Beijing Youcare Kechuang Pharmaceutical Technology Co. Ltd.
Acknowledgements
We thank Dr Dongyang Liu and Dr Chan Yang for their helpful advice and discussion.
Notes and references
- M. Shanmugasundaram, A. Senthilvelan and A. R. Kore, Chem. Rec., 2022, 22, e202200005 CrossRef CAS.
- M. Gu and C. D. Lima, Curr. Opin. Struct. Biol., 2005, 15, 99–106 CrossRef CAS.
- F. Borbolis and P. Syntichaki, FEBS J., 2002, 289, 1457–1475 CrossRef PubMed.
- M. Warminski, P. J. Sikorski, J. Kowalska and J. Jemielity, Top. Curr. Chem., 2017, 375, 1–29 CrossRef CAS.
- J. Pelletier, M. Schmeing and N. Sonenberg, WIREs RNA, 2021, e1636 CrossRef CAS.
- M. Warminski, E. Trepkowska, M. Smietanski, P. J. Sikorski, M. R. Baranowski, M. Bednarczyk, H. Kedzierska, B. Majewski, A. Mamot and D. Papiernik,
et al.
, J. Am. Chem. Soc., 2024, 146, 8149–8163 CrossRef CAS PubMed.
- K. Drazkowska, R. Tomecki, M. Warminski, N. Baran, D. Cysewski, A. Depaix, R. Kasprzyk, J. Kowalska, J. Jemielity and P. J. Sikorski, Nucleic Acids Res., 2022, 50, 9051–9071 CrossRef CAS.
- P. J. Sikorski, M. Warminski, D. Kubacka, T. Ratajczak, D. Nowis, J. Kowalska and J. Jemielity, Nucleic Acids Res., 2020, 48, 1607–1626 CrossRef CAS PubMed.
- J. M. Henderson, A. Ujita, E. Hill, S. Yousif-Rosales, C. Smith, N. Ko, T. McReynolds, C. R. Cabral, J. R. Escamilla-Powers and M. E. Houston, Curr. Protoc., 2021, 1, e336 CrossRef PubMed.
-
R. I. Hogrefe, A. Lebedev, A. P. McCaffrey and D. Shin, WO2017053297 A1, TriLink BioTechnologies, Inc., 2017.
- A. Senthilvelan, T. Vonderfecht, M. Shanmugasundaram, I. Pal, J. Potter and A. R. Kore, Org. Lett., 2021, 23, 4133–4136 CrossRef CAS PubMed.
- A. Senthilvelan, T. Vonderfecht, M. Shanmugasundaram, J. Potter and A. R. Kore, Bioorg. Med. Chem., 2023, 77, 117128 CrossRef CAS PubMed.
- M. Kozarski, K. Drazkowska, M. Bednarczyk, M. Warminski, J. Jemielity and J. Kowalska, RSC Adv., 2023, 13, 12809–12824 RSC.
- A. Radford, J. Wu, R. Child, D. Luan, D. Amodei and I. Sutskever, OpenAI Blog, 2019, 1, 9 Search PubMed.
- A. M. Rydzik, M. Warminski, P. J. Sikorski, M. R. Baranowski, S. Walczak, J. Kowalska, J. Zuberek, M. Lukaszewicz, E. Nowak and T. D. W. Claridge,
et al.
, Nucleic Acids Res., 2017, 45, 8661–8675 CrossRef CAS.
- 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, 2202590 CrossRef CAS PubMed.
- D. Kairuz, N. Samudh, A. Ely, P. Arbuthnot and K. Bloom, Front. Immunol., 2022, 13, 1018961 CrossRef CAS PubMed.
- E. Wilson, J. Goswami, A. H. Baqui, P. A. Doreski, G. Perez-Marc, K. Zaman, J. Monroy, C. J. A. Duncan, M. Ujiie and M. Rämet,
et al.
, N. Engl. J. Med., 2023, 389, 2233–2244 CrossRef CAS PubMed.
- M. A. Monslow, S. Elbashir, N. L. Sullivan, D. S. Thiriot, P. Ahl, J. Smith, E. Miller, J. Cook, S. Cosmi and E. Thoryk,
et al.
, Vaccine, 2020, 38, 5793–5802 CrossRef CAS.
- S. Meulewaeter, Y. Zhang, A. Wadhwa, K. Fox, I. Lentacker, K. W. Harder, P. R. Cullis, S. C. de Smedt, M. H. Y. Cheng and R. Verbeke, J. Mol. Biol., 2024, 436, 168385 CrossRef CAS.
-
J. Zhang, L. Zhang, Y. Zhou, J. Chen, J. Zhang, S. Li, Y. Liu, C. Guo, W. Tang, F. Wu, et al. WO2024051696 A1, Guangzhou Henovcom Bioscience, 2024.
-
J. Zhang, L. Zhang, Y. Zhou, J. Chen, J. Zhang, S. Li, Y. Liu, C. Guo, W. Tang, F. Wu, et al. AU 2022430003 B9, Guangzhou Henovcom Bioscience, 2022.
-
C. Xu, M. Houston, A. Lebedev, J. M. Henderson, A. Ujita, I. Koukhareva, C. Zhao and I. V. Ilichev, WO2023147352 A1, TriLink BioTechnologies, Inc., 2023.
-
R. Sennrich, B. Haddow and A. Birch, arXiv, 2015, preprint: not peer-reviewed, arXiv:1508.07909 DOI:10.48550/arXiv.1508.07909.
-
A. Paszke, S. Gross, S. Chintala, G. Chanan, E. Yang, Z. DeVito, Z. Lin, A. Desmaison, L. Antiga and A. Lerer, Automatic Differentiation in Pytorch, 2017 Search PubMed.
-
T. Dozat, Incorporating Nesterov momentum into Adam, 2016 Search PubMed.
|
This journal is © The Royal Society of Chemistry 2025 |
Click here to see how this site uses Cookies. View our privacy policy here.