AntigenBoost: enhanced mRNA-based antigen expression through rational amino acid substitution

Abstract

Messenger RNA (mRNA) vaccines represent a groundbreaking advancement in immunology and public health, particularly highlighted by their role in combating the COVID-19 pandemic. Optimizing mRNA-based antigen expression is a crucial focus in this emerging industry. We have developed a bioinformatics tool named AntigenBoost to address the challenge posed by destabilizing dipeptides that hinder ribosomal translation. AntigenBoost identifies these dipeptides within specific antigens and provides a range of potential amino acid substitution strategies using a two-dimensional scoring system. Through a combination of bioinformatics analysis and experimental validation, we significantly enhanced the in vitro expression of mRNA-derived Respiratory Syncytial Virus fusion glycoprotein and Influenza A Hemagglutinin antigen. Notably, a single amino acid substitution improved the immune response in mice, underscoring the effectiveness of AntigenBoost in mRNA vaccine design.

Introduction

mRNA vaccines have emerged as a powerful alternative to conventional vaccines due to their high potency, safety, efficacy, rapid clinical development, and potential for low-cost manufacturing [1]. Unlike conventional vaccines, which often use inactivated viruses or protein subunits, mRNA vaccines leverage synthetic mRNA molecules that instruct cells to produce a protein mimicking a portion of the pathogen [2]. Optimizing mRNA sequences is fundamental for maximizing the therapeutic potential, safety, and efficiency of mRNA-based interventions, making them more effective and reliable for clinical use [3, 4].

Proper optimization of vaccine mRNA can reduce the dosage required for each injection leading to more efficient immunization programs [5]. The basic structure of mRNA consists of a protein-encoding open reading frame (ORF), 5′ and 3′ untranslated regions (UTRs), a 7-methylguanosine 5′ cap structure, and a 3′ poly(A) tail [6]. To enhance protein expression, various elements of the mRNA are optimized. Modifications to the UTRs, 5′ cap, and poly(A) tail are implemented to improve translational efficiency and increase mRNA stability [5, 7]. The ORF is optimized for codon adaptation to improve the translation elongation rate [8] and for secondary structure to achieve minimized energy conformation for mRNA stability [9]. Advanced bioinformatics tools, such as LinearDesign [10], mRNAid [11], and Codon Box [12], have been developed to mainly focus on optimizing the codon adaptation index (CAI) and the minimum free energy (MFE) of the mRNA structure [10, 13]. Additionally, mRNA-encoded antigens are structurally modified to adapt to a more stable conformation favoring an optimal immunogenicity [14–16].

Intrinsic ribosome destabilization (IRD) occurs in nascent polypeptides and hinders both prokaryotic [17] and eukaryotic [18] translation elongation. However, this phenomenon has often been overlooked during antigen design. In addition to the previously mentioned ORF optimization strategies, we believe that excluding IRD is also of great importance for efficient translation. In a previous study by Burke and his coworkers [19], the correlation between nascent dipeptide motifs and the mRNA translation level was investigated through a massively parallel assay involving combinations of 20*20 amino acids, providing systematic data for developing bioinformatics tools. It was found that combinations containing bulky and positively charged amino acids could destabilize mRNA during translation, especially when these dipeptide motifs are located on the β strand. Therefore, we hypothesize that identifying these dipeptide motifs responsible for low expression, termed low-expressed dipeptide (LEDipep) sites, and employing a conservative amino acid substitution strategy could enhance mRNA expression levels. Inspired by Burke’s results, we developed a tool named ‘AntigenBoost’ to (i) identify LEDipep sites on input sequence, (ii) evaluate all possible amino acid substitution strategies through a 2D scoring system, and (iii) validate potential mutational strategies via proper experiments (Fig. 1).

Figure 1.

Overview of the AntigenBoost workflow. The input is the target antigen sequence for optimization, and a reference protein structure helps to identify the β-strand segments of the target sequence. The reference should either be a homology structure from the PDB database or a predicted structure of the target sequence. Step 1. Use the reference structure to predict the β-strand segments of the target sequence through sequence alignment. Step 2. Search for destabilizing dipeptides in the target sequence. Step 3. Identify the LEDipep sites by combining the results from Step 1 and Step 2. Step 4. Provide mutational strategies for each LEDipep site by calculating the 〖ΔSeq〗_score for the potential improvement in mRNA expression and conservative score (BLOSUM62) for evolutionary acceptance. Step 5. Experimental validation for promising mutants.

Our initial efforts have focused on optimizing a vaccine for respiratory syncytial virus (RSV) infection. RSV is a common respiratory virus that causes global epidemic each year. RSV causes lower respiratory tract disease that may progress into life-threatening pneumonia and bronchiolitis [20]. To date, three RSV vaccines have been approved. GSK’s Arexvy is approved for individuals older than 50 years old [21, 22], while Pfizer’s Abrysvo is authorized for use in pregnant individuals and those over 60 [23–25]. Both vaccines consist of recombinant RSV fusion glycoprotein (F) with a stabilized prefusion (pre-F) conformation [26–28]. Additionally, Moderna’s mRESVIA, the first mRNA RSV vaccine also known as mRNA-1345, was recently approved to protect adults aged 60 and older from lower respiratory tract disease caused by RSV infection [29, 30].

We engineered an mRNA molecule NR135 [31] that expresses RSV pre-F. Using AntigenBoost, we identified several destabilizing dipeptides in NR135-encoded antigen and improved RSV-F expression through rationale amino acid substitution. Importantly, these substitutions also enhanced RSV-F levels in a different RSV-F mRNA NR091 [14, 32], leading to a stronger humoral immune response in mice. Additionally, we applied AntigenBoost-guided strategies to improve the expression of an Influenza A antigen. This work provides a new framework and a novel perspective for optimizing mRNA-encoded therapeutics.

Methods

LEDipep sites identification

The first step of AntigenBoost is to identify destabilizing dipeptides located on beta-strand (defined as LEDipep) sites for further optimization. In this step, the predicted secondary structure information for the target sequence is combined with the searched destabilizing dipeptide motifs to determine the LEDipep sites for mutational strategy design in the next step.

Secondary structure prediction

For RSV-F antigen, the PDB structure 8WSQ, showing 98% sequence identity with NR135, was selected as the reference structure. Sequence alignment with this reference allowed for the identification of the beta-strand locations in NR135.

For influenza haemagglutinin (HA) antigen, no available structure shared over 95% identity with the target HA sequence. Therefore, the predicted structure by AlphaFold2 was used to extract the locations of the beta-strands.

Destabilizing dipeptide search

DipepScore, quantifying the impact of specific dipeptides on the mRNA translation, is calculated using the formula:

Here, fc represents steady-state mRNA level measurements of 400 dipeptides (20 × 20 dipeptides, with eight replicates each) from Burke’s work [19], which reflects the effects of each codon pair on mRNA stability and translational efficiency. Here, we define destabilizing dipeptides with DipepScore <−1.8 including HR, II, KI, KV, KY, RH, VK, and YK.

Mutational strategy evaluation

The second step of AntigenBoost provides possible mutational strategies for the identified LEDipep sites with a 2D evaluation system, incorporating the calculation of two score functions for the potential improvement in mRNA expression and evolutionary acceptance.

Predicted mRNA expression scoring

Seq_Score is calculated using the formula:

Here, n represents the number of residues in the subpeptide, and DipepScore (𝑖) denotes the log fold change for the impact of each dipeptide pair.

Considering the translation process as a linear extension of the peptide chain, a point mutation strategy is developed by focusing on the dipeptide sites adjacent to the former and later residues. Rarely do standard LEDipep sites occur at the N-terminal or C-terminal ends, as they typically maintain secondary structures such as loops or helices.

The change in sequence score (ΔSeq_Score) is determined by the difference between the sequence score of the mutated LEDipep site and the original sequence score:

A positive ΔSeq_Score suggests a potential increase in mRNA expression level, indicating a favorable mutation. For comparative analysis, tetrapeptide sequences including a residue preceding and following the LEDipep site are extracted to guide the mutational strategy design.

Conservative score

The optimization of LEDipep sites was guided by the BLOSUM62 substitution matrix [33] to facilitate a conservative mutational strategy. Prioritizing amino acid substitutions based on BLOSUM62 scores ensured minimal perturbation on protein structure and function. A higher BLOSUM62 score indicates a greater likelihood of observing the substitution in evolutionarily related sequences, suggesting a more conservative and functionally relevant mutation.

mRNA synthesis

All antigen sequences in this paper undergo codon optimization using Codon Box [12] to convert them into mRNA sequences for experimental validation. Circular plasmids pUC57 encoding the gene of target mRNA were synthesized in GenScript. Plasmids were linearized with BsaI (Novoprotein) treatment at 37°C overnight. Linearized plasmids were purified using Wizard® SV Gel and PCR Clean-Up System (Promega) following the manufacturer’s protocol. With linearized plasmids being used as templates, in vitro transcription was conducted at 37°C for 3 h using HiScribe® T7 High Yield RNA Synthesis Kit (NEB), where uridine triphosphate (UTP) was replaced by N1-Me-ψ-UTP (Hongene). The resulting products were treated with DNaseI (Novoprotein) at 37°C for 30 min. RNA was precipitated by 1.5-fold volume 7.5 M LiCl (Invitrogen) at −20 °C for 30 min, followed by 12 000 rpm centrifugation at 4°C for 15 min. The precipitates were washed with 70% ethanol for three times and redissolved in RNase-free water. The concentration of RNA was determined by a NanoDrop™ One Microvolume UV-Vis Spectrophotometer (ThermoFisher).

mRNA transfection

6 × 10⁵ HEK293T cells were seeded in 6-well plates. One hundred twenty-five nanograms of mRNA was transfected using Lipofectamine™ MessengerMAX™ Transfection Reagent (Invitrogen) following the manufacturer’s protocol. Cells were maintained in Dulbecco’s Modified Eagle Medium High Glucose (VivaCell) supplemented with 10% fetal bovine serum (VivaCell), 1% penicillin–streptomycin (Gibco), 1% nonessential amino acids (Gibco), and 1% sodium pyruvate (Gibco). Cells were cultured at 37°C with 5% CO₂. Cells were harvested and lysed using Radio-Immunoprecipitation Assay (RIPA) lysis buffer (Beyotime) supplemented with proteinase inhibitor (NCM) and BenzoNuclease® (Novoprotein) 24 h after transfection. Cell lysates were mixed with 5× Sodium Dodecyl Sulfate-Polyacrylamide Gel Electrophoresis (SDS-PAGE) Loading Buffer (NCM) and denatured at 100°C for 10 min.

Western blot analysis

Ten microliters of denatured cell lysates was loaded onto a 4%–12% SurePAGE precast mini polyacrylamide gel (GenScript). SDS-PAGE was run at 160 V for 40 min in MOPs buffer. Proteins were transferred onto a PVDF membrane (pre-activated by methanol) in NcmBlot Rapid Transfer Buffer (NCM) at 400 mA for 30 min. The membrane was blocked in 10% milk at room temperature for 1 h and incubated in primary antibody at room temperature for 2 h or at 4°C overnight. Primary antibodies were diluted in blocking buffer, including anti-RSV Fusion Antibody (1:200, Sino Biological), Influenza A Virus Hemagglutinin/HA Antibody (1:2000, Sino Biological), Glyceraldehyde 3-Phosphate Dehydrogenase (GAPDH) Monoclonal Antibody (1:10 000, Proteintech) and Alpha-Tubulin Mouse Monoclonal Antibody (1:10 000, Proteintech). The membrane was washed in Tris-Buffered Saline with Tween 20 (TBST) three times, 5 min each, and then incubated in secondary antibody at room temperature for 1 h. Secondary antibodies were diluted in blocking buffer, including Horseradish Peroxidase (HRP)-conjugated Affinipure Goat Anti-Rabbit IgG (H + L) and HRP-conjugated Affinipure Goat Anti-Mouse IgG (H + L) (15 000, Proteintech). The membrane was washed in TBST five times, 5 min each, and then developed in NcmECL SuperUltra reagents (NCM) for 1 min. Immunoblot was imaged in a 4600SF Chemiluminescent Imaging System (Tanon).

mRNA formulation and characterization

mRNA was encapsulated in lipid nanoparticles (LNPs) according to previous reports [34]. LNPs were prepared using a NanoAssemblr® microfluidic mixer (Precision Nanosystems). Different lipids were dissolved in ethanol at molar ratios of 47.4:10:40.8:1.8 (ALC0315: DSPC: steroid: ALC0159), while mRNA was diluted with citric acid buffer to a final concentration of 100 ng/μl. The two solutions were mixed at a volume ratio of 1:3 (lipid: aqueous) and a total flow rate of 5 ml/min. The resulting LNPs were dialyzed against 10 mM Tris buffer in Slide-A-lyzer® Dialysis Cassette G2 (Thermo) overnight. The mRNA encapsulation efficiency was determined using Quant-iT™ RiboGreen™ RNA Assay Kit according to the manufacturer’s instructions.

Immune response detection in vivo

Female BALB/c (an inbred laboratory strain) mice (6–8 weeks, 18–20 g) were divided into three groups, eight mice each. One microgram of LNP-formulated mRNA or Phosphate-Buffered Saline (PBS) was injected intramuscularly on Day 0 and Day 21, respectively. The mice were sacrificed on Day 35, and blood was collected. Binding antibody titers were determined by enzyme-linked immunosorbent assay. One hundred microliters of 1 μg/mL Human RSV (A2) Fusion glycoprotein/RSV-F Protein (Sino Biological) was coated in each well of a 96-well plate at 4°C overnight. The plate was blocked in 200 μl of 5% BSA in TBST at 37°C for 1 h and then washed three times by TBST. Mice serum was incubated at 37°C for 1 h, starting at 1 in 100 dilution and then four-fold serial dilutions to prepare eight dilutions in total. After washes, 100 μl of Goat Anti-Mouse IgG (H + L)-HRP (1:6000 diluted, SouthernBiotech) was added and incubated at 37°C for 1 h. The plate was then thoroughly washed by TBST for three times. One hundred microliters of 3,3′,5,5′-Tetramethylbenzidine (TMB) Chromogen Solution (Invitrogen) was added in each well and incubated at room temperature for 15 min, followed by the addition of 100 μl of stop solution (NCM). OD450 (Optical Density at 450 nm) was measured in a microplate reader (BioTek). If the optical density (OD) was greater than 0.1and larger than 2.1-fold of negative control serum (P/N > 2.1, P is OD450 of the experimental serum at certain dilutions; N is OD450 of the negative control serum at the same dilution), the sample was considered as positive. The antibody titer was determined as the highest dilution of the positive serum. The limit of this assay was 100–1 638 400.

Results

AntigenBoost predicted LEDipeps on RSV-F

According to Burke’s work [19], destabilizing dipeptides forming extended β strands hinder ribosomal progression and downregulate mRNA levels in vivo. Here, we developed algorithms to identify destabilizing dipeptides, including HR, II, KI, KV, KY, RH, VK, and YK. Those located or partially located on beta-strands were designated as LEDipeps (Fig. 1).

NR135 incorporated designs to enhance the immunogenicity and stability of RSV-F protein after four cycles of screening [31]. The major designs are DS2-Cav1 and SC. DS2-Cav1 increases RSV-F trimerization and physical stability through double disulfide bond engineering and cavity filling. SC further stabilizes the trimeric prefusion conformation by abolishing furin cleavage sites and fusing F1/F2 with a Glycine-Serine (GS) linker. Additionally, the cytoplasmic tail was removed while the transmembrane domain was maintained to anchor the trimeric protein on cell membrane (Fig. 2A). Using homology reference structure (PDB ID: 8WSQ), several LEDipeps were identified in NR135 by AntigenBoost, including I²⁵²I²⁵³, K³²⁰V³²¹, K³⁵⁵I³⁵⁶, I³⁹²I³⁹³, and V⁴³⁰K⁴³¹. To preserve antigen immunogenicity, LEDipep sites within the key antigenic sites (Ø, I, II, III, IV, and V) on pre-F trimer [35, 36] were excluded. Two target LEDipep sites, I²⁵²I²⁵³ and K³²⁰V³²¹, were selected for further optimization (Fig. 2A). Structural analysis confirmed that both sites were parts of beta-sheets (Fig. 2B).

Figure 2.

AntigenBoost identified LEDipeps on RSV-F. (A) Linear diagram of the antigen design for RSV-F protein variants NR091 and NR135. NR091 is derived from DS-Cav1 and includes furin cleavage sites (residues 109 and 137), resulting in the cleavage of F0 into F1 and F2 in vivo. NR135 replaces these furin cleavage sites and links F1 and F2 with a GS linker. The cytoplasmic tail is also removed in NR135. Both NR091 and NR135 preserve the transmembrane domain. LEDipep sites identified by AntigenBoost are highlighted in the inset. SP, signal peptide; p27, the p27 peptide removed by furin cleavage in NR091; TM, transmembrane domain; CT, cytoplasmic tail. (B) Identification of LEDipep sites on the structure of RSV-F. The reference RSV-F structure (PDB ID, 8WSQ) is shown in surface representation, sharing a sequence identity of 98% with NR135. One monomer of internal protein structure is shown in cartoon. LEDipep sites, I252I253 and K320V321, are shown in sticks.

Amino acid substitution increased RSV-F expression in vitro

Here we applied a 2D scoring system to evaluate all the 19 possible amino acid substitution strategies for each amino acid. In this system, the ΔSeq_Score represents the predicted change of mRNA expression level after substitution and the conservative score indicates the conservation after mutation (Fig. 3A). For site I²⁵²I²⁵³, we found several promising substitution strategies in the top-right corner. I252L was selected due to its isomer relationship and its high potential to increase expression, with a conservative score of 2 and ΔSeq_Score of 1.15. The LEDipep site K³²⁰V³²¹ presented a completely different pattern in the 2D evaluation. Highly conservative substitutions tend to offer a limited improvement in expression levels, necessitating a balance between the two scores. We selected three strategies, namely, K320Q, K320M, and K320F, along the diagonal edge with ΔSeq_Scores of 0.47, 1.22, and 1.55 respectively, representing different levels of conservation: 1, −1, and −3 (Supplementary Table 1). These strategies were chosen for experimental validation.

Figure 3.

Amino acid substitution increased RSV-F expression in vitro. (A) RSV-F amino acid substitution strategies from AntigenBoost with a 2D scoring system. Each amino acid substitution strategy is represented in a scatter plot according to its 〖ΔSeq〗_score against conservative score. Left: I252I253. Right: K320V321. The shaded area includes promising strategies given by AntigenBoost with 〖ΔSeq〗 _Score ≥ 0.4 and Conservative Score ≥ −1. Selected mutants for experimental validation are depicted by triangles, while other strategies are shown as dots. (B) Conservative substitutions of I252 and K320 on NR135 largely improved RSV-F expression in HEK293T cells. (C) Conservative substitutions of I290 and K368 on NR091 improved RSV-F expression in HEK293T cells. Cells were transfected with an equal amount of control (ctrl, unmuted) or mutated mRNA. Mock cells were not transfected. Cells were lysed 24 h after transfection, and total protein was analyzed by western blot. RSV-F and reference protein α-tubulin were detected by specific antibodies.

We synthesized mRNA with I252L, K320F, K320M, and K320Q single amino acid substitution within the ORF of NR135. Additionally, we investigated combinations of I252L with each of the three K³²⁰ mutations. An equal amount of these mRNA was transfected into HEK293T cells. Cells were lysed 24 h after transfection, and RSV-F expression was detected by western blot analysis. I252L, K320M, and K320Q substitutions upregulated the RSV-F level by ~1.5- to 2-fold (Fig. 3B). The combination of I252L with either K320M or K320Q also increased RSV-F expression but did not show obvious synergistic effects. Interestingly, the K320F substitution decreased RSV-F level, while the I252L/K320F combination rescued it. These results confirmed the negative effects of the LEDipep identified in RSV-F. Moreover, the conservative score is of great importance to amino acid substitution. The less conservative K320F mutation is likely to affect the conformation of the antigen, leading to a decreased RSV-F level that could be detected by a specific antibody.

To further confirm the significance of these amino acid substitutions, we applied I252L and K320Q in a different RSV-F molecule NR091, corresponding to I290L and K368Q respectively. NR091 encodes DS-Cav1 RSV-F where the furin cleavage sites and C-terminus were maintained, resulting in uncleaved F0 or cleaved F1 and F2 products (Fig. 2A). According to the western blot results, both I/L and K/Q mutations could upregulate RSV-F levels, especially F1 (Fig. 3C). Therefore, our amino acid substitution strategies were effective across different antigen designs.

Amino acid substitution induced stronger humoral immune responses in vivo

To test if amino acid substitution can improve mRNA vaccine efficacy in vivo, we immunized BALB/c mice with LNP-formulated RSV-F mRNA and boosted them at Day 21. Serum was collected at Day 35 (Fig. 4A). NR091 and the I290L mutant were selected since the Isoleucine-Leucine (IL) mutation seemed to induce a higher level of expression than the K368Q mutant (Fig. 3C). The I290L mutant induced about a two-fold higher level of IgG that recognizes the RSV-F antigen (Fig. 4B). Neither vaccine affected the body weight growth of the mice being investigated (Fig. 4C). Therefore, LEDipep optimization through minor mutation could elicit stronger protective titers, which is important for mRNA vaccine design.

Figure 4.

Amino acid substitution induced stronger humoral immune responses in vivo. (A) Immunization plan. BALB/c mice were primed by 1 μg LNP-encapsulated NR091 (ctrl) or I290L mutant RSV-F mRNA on Day 0 and boosted with the same vaccine on Day 21. PBS was injected as a negative control. Serum was collected on Day 35. (B) I290L mutated RSV-F vaccine induced higher levels of specific IgG. Serum IgG titers from individual mice are represented by dots, with the geometric mean and 95% confidence intervals for each group. The limit of detection was shown by a dashed line. Statistical difference compared to the control group was assessed using GraphPad Prism 8 t-test. ^*P < .05. (C) The tested RSV-F vaccines did not affect mice growth. Weekly body weights were recorded, and the mean growth values for each group were shown with their standard errors (SEM).

AntigenBoost improved influenza HA mRNA expression in vitro

We ran AntigenBoost on HA from Influenza Antigen A H3N2 /Darwin/9/2021 strain (EPI2495551), to test if this algorithm can be wildly applied in mRNA vaccine design. Using a predicted structure by AlphaFold2 [37], we identified five LEDipeps including V⁴²K⁴³, V¹⁴⁶K¹⁴⁷, K²⁷⁵I²⁷⁶, V²⁵³K²⁵⁴, and K⁴⁸⁴I⁴⁸⁵. We omitted V¹⁴⁶K¹⁴⁷, K²⁷⁵I²⁷⁶, and K⁴⁸⁴I⁴⁸⁵ within epitopes [38] or near the head domain of HA [39], which is critical for receptor binding and targeted by strain-specific neutralizing antibodies. This ensures that the antigen remains recognizable to the immune system. V⁴²K⁴³ and V²⁵³K²⁵⁴ were continued with amino acid substitution, both of which were on beta-sheets as well as within the relatively conserved stem domain (Fig. 5A).

Figure 5.

AntigenBoost improved influenza HA mRNA expression in vitro. (A) Identification of LEDipep sites on the predicted structure of HA. The predicted monomeric structure of HA by AlphaFold2 is shown in cartoon, overlaid with trimeric surface representation (PDB ID: 6PDX). Functional regions comprise the head domain, stem domain, and TM domain. LEDipep sites on the stem domain, I252I253 and K320V321, are shown in sticks. (B) HA amino acid substitution strategies from AntigenBoost with a 2D scoring system. Each amino acid substitution strategy is represented in a scatter plot according to its 〖ΔSeq〗_Score against Conservative Score. Left: K275I276. Right: K484I485. The shaded area includes promising strategies given by AntigenBoost with 〖ΔSeq〗_Score ≥ 0.4 and Conservative Score ≥ −1. Selected mutants for experimental validation are depicted by triangles, while other strategies are shown as dots. (C, D) The effects of single/double amino acid substitutions on HA expression in vitro. HEK293T cells were transfected with an equal amount of WT or mutated mRNA. Mock cells were not transfected. Cells were lysed 24 h after transfection, and total protein was analyzed by western blot. Influenza A HA and reference protein α-tubulin were detected by specific antibodies.

Scoring results from AntigenBoost of V⁴²K⁴³ and K⁴⁸⁴I⁴⁸⁵ shared similar patterns with K³²⁰V³²¹ during RSV-F optimization, indicating several promising mutation strategies for HA: K43M, V42M, K43E, and K43Q for site V⁴²K⁴³ and K484M, K484E, and K484Q for site K⁴⁸⁴I⁴⁸⁵ (Fig. 5B). Following the successful precedent in the RSV-F protein (K320Q), we selected K43Q and K484Q with of 0.49 and 0.84, respectively, for experimental validation (Supplementary Table 2). Interestingly, neither K43Q nor K484Q alone showed obvious effects compared to the wildtype (WT) mRNA (Fig. 5C). A synergistic effect was observed with an ~1.6-fold increase of HA expression after K43/484Q double mutations in HEK293T cells (Fig. 5D). Thus, AntigenBoost is applicable in custom mRNA design to improve antigen expression levels. Besides, V⁴²K⁴³ and K⁴⁸⁴I⁴⁸⁵ on HA are found conservative across different H3N2 strains recommended by World Health Organization as vaccine composition in the past 5 years (Supplementary Fig. 1), suggesting that the K43/484Q double mutational strategy could be generalized to other strains in HA mRNA vaccine design.

Discussion

The immune response to an mRNA-based vaccine depends on its immunogenicity and constant expression, which can be influenced by translational efficiency and mRNA stability. These factors have been optimized through UTR design [7, 40], codon optimization [10, 13], and nucleic acid modifications [41]. In addition to these mRNA design improvements, amino acid mutations are employed to stabilize the translational product in a favored conformation. Techniques such as proline substitution, disulfide crosslinking, and cavity filling have been widely applied in the antigen development of SARS-CoV-2 [16, 42, 43], human immunodeficiency virus [44], influenza [45, 46], and RSV [14, 15, 31]. Despite these strategies, there remains potential to improve translation through subtle amino acid adjustments.

Research has shown that nascent peptides affect mRNA stability through amino acid mutations, particularly those with bulky or positively charged amino acids. Positively charged amino acids are thought to slow elongation, possibly due to electrostatic interactions between the nascent peptide and the ribosomal exit tunnel [47]. Studies have found that nascent peptides with multiple arginine and lysine form arresting peptide intermediates, which disappear when these positively charged amino acids are replaced with glutamine [47]. Positively charged residues, especially lysine and arginine, significantly impact ribosomal velocity more than factors like codon usage and transcript secondary structure [48]. Even a single positive charge can slow the ribosome’s progression [48]. Po et al. found that side-chain size inversely correlates with elongation rate, possibly due to van der Waals interactions between the introduced amino acid and the tunnel walls [49]. Burke’s research revealed that a minimal combination of certain bulky and positively charged nascent peptides forming extended β-strand structures is sufficient to induce ribosome slowdown, thereby negatively regulating gene expression [19]. This finding aligns with our selected destabilizing dipeptides (HR, KI, KV, KY, RH, VK, and YK), except for II, which consists of two identical bulky amino acids without positive charge.

In this study, by substituting II with LI, we successfully improved protein expression by ~1.5–2-fold (Fig. 3B), despite leucine being nonpolar and sharing an identity bulkiness score with isoleucine [50]. This result is consistent with Burke’s linear statistical model, which demonstrates a moderate relevance of mRNA levels to the isoelectric point and bulkiness of amino acids (adjusted R square = 0.25) [19]. The interaction between nascent peptides and ribosome elongation is complex and requires further research to identify additional factors, such as protein folding and their contributions to mRNA levels [51]. Therefore, we use experimentally determined steady-state mRNA levels to build our algorithm, providing a robust prediction for the effects of dipeptides on mRNA expression by incorporating currently unknown factors.

We incorporated structural analysis to determine if destabilizing peptides are located on beta-strand as an essential criterion for determining LEDipep sites. S4PRED [52], the secondary structure prediction software used in Burker’s work [19] with an accuracy of 75.3% [53], was insufficient for residue-level precision in this study. To address this issue, we proposed two solutions.

For antigens with available homology structures of high sequence identity (>95%), our algorithm can predict the secondary structure of the target protein through pair-wise alignment with the homologous sequence of experimentally determined structure. This method is based on the principle that proteins with high sequence identity typically have high structural and functional similarity, making them reliable templates [54]. Furthermore, Wilson et al. found that protein domains with over 95% sequence identity exhibit minimal structural deviation, with an average root mean square deviation (RMSD) of ~0.3 Å [55]. For antigens without available experimental structures, advanced software such as AlphaFold2 [37] and RosettaFold [56] can be used to predict the protein conformation. Among the two approaches for secondary structure prediction, experimental structures are prioritized due to their accuracy in representing protein conformations. However, when high-quality homology structures are unavailable or partially missing, using predicted structures from AlphaFold2 is a viable alternative due to its high accuracy in predicting α-helical, β-hairpin, and disulfide-rich peptides, despite the high computational demands [57]. In this study, both approaches provided reliable secondary structure predictions, leading to the successful identification and optimization of LEDipeps (Figs 2B and 5A). Notably, the recently released AlphaFold3 [58], which offers superior structure predictions compared to previous versions, is highly recommended for future applications of AntigenBoost. Future research will evaluate the precision of using predicted and experimental structures in LEDipep identification to achieve a consensus prediction.

We developed a 2D scoring system consisting of ΔSeq_Score and conservative score to evaluate amino acid substitution strategies. ΔSeq_Score, the primary metric, assesses potential improvement by measuring how well substitutions mitigate issues caused by intrinsically disordered dipeptides. When selecting an optimal mutant, a high ΔSeq_Score is prioritized, as it indicates the effectiveness of the strategy in enhancing protein expression, while an acceptable conservative score helps avoid harmful mutations that inhibit protein folding and stability. This system enabled successful enhancements in RSV-F and HA expression through I/L, K/M, or K/Q substitutions. Notably, for the LEDipep sites (KV, VK, and KI), K/Q and K/E substitution strategies yield similar scores. K/Q substitutions are preferred for preserving polarity and hydrogen bonds with surrounding residues, whereas K/E substitutions introduce a disruptive charge reversal. The increased expression from lysine mutations might be attributed to the elimination of a ubiquitination site. However, an authentic ubiquitination site requires degron recognition and a structurally disordered segment beyond a neighboring ubiquitinated lysine [59]. Since our LEDipeps didn’t meet these criteria, it is unlikely that this lysine significantly contributes to ubiquitin-mediated protein degradation.

Notably, the HEK293T cell line was used in this study for experimental validation due to its several advantages. The HEK293T cell line is known for its high transfection efficiency, high yields of transfected genes, rapid growth rate, and human glycosylation pattern. These benefits make the HEK293T cell line widely used for studies involving mRNA transient transfection [7, 61]. However, as the goal of this study is to develop an algorithm to address common issues in the field of mRNA therapy, we will validate the AntigenBoost in various cell lines and primary cells to highlight the significance of our results.

There are a few caveats in our model. Currently, AntigenBoost is not able to predict the optimal strategy, and laboratory validation is essential to confirm effective mutants to improve antigen expression. Based on successful optimization for the two antigens, we conclude that the ΔSeq_Score should exceed 0.4 and the conservation score should be at least equal to −1. As shown in the shaded area in our scoring system, these standards offered a range of amino acid substitution strategies, narrowing down amino acid substitution strategies to a manageable number for experimental validation (Figs 3A and 5B). Incorporating parameters like ΔΔG for mutational protein stability [60] could refine the strategy further. Moreover, the current algorithm cannot predict synergistic effects across multiple LEDipep sites, which were observed in HA but not in RSV-F. This complexity likely arises from differences in antigen structures, especially since the two example antigens here are both metastable.

In summary, we developed AntigenBoost, a robust bioinformatics tool that identifies and optimizes destabilizing peptides, as a complementary method for improving mRNA vaccine design. With the growing field of mRNA therapy, our work is valuable not only for vaccine development but also for other mRNA-based therapeutics.

Key Points

• AntigenBoost is a bioinformatics tool that efficiently identifies and optimizes destabilizing nascent dipeptides in a specific antigen.

• AntigenBoost provides a 2D mutational scoring system to stabilize these dipeptides through rationale amino acid substitutions, therefore enhancing antigen expression.

• With the help of AntigenBoost, we successfully improved the expression of mRNA-encoded antigens of RSV and Influenza A.

Funding

The work presented here was funded by NanoRibo (Shanghai) Biotechnology.

Data availability

The datasets and source code are publicly available at https://github.com/YumGao/AntigenBoost

Author contributions

Y. Gao: Investigation, Methodology, Software, Visualization, Writing – original draft; S. Zhu: Investigation, Methodology, Validation, Writing – original draft; H. Li: Methodology; X. Hao: Methodology; W. Chen: Methodology; D. Pan: Project administration; Z. Qian: Conceptualization, Writing – review & editing.