COVID-19 vaccine design using reverse and structural vaccinology, ontology-based literature mining and machine learning

Introduction

As one of the most successful medical inventions, vaccination has been used successfully against infectious diseases, such as smallpox and polio that have been completely or almost completely eradicated globally. The importance of vaccines is also manifested in the recent coronavirus disease 2019 (COVID-19) pandemic and the usage and rapid development of effective vaccines. There have been 121 COVID-19 vaccines that have entered or completed clinical trials as of 6 November 2021 [1]. Although the current licensed COVID-19 vaccines are effective, they could be subjected to further improvement in different ways, such as their effectiveness against potential new variants. On the other hand, there are still many other infectious diseases (tuberculosis, AIDS, malaria, etc.) for which we do not have effective vaccines. To face the ever-increasing threat of existing and new infectious diseases in the future, advanced methods for developing more effective and safe vaccines are in earnest demand.

Immunity is a complicated system, and different immune pathways and mechanisms need to operate in a coordinated manner to achieve protection from infectious diseases. Because of that, vaccines need to invoke different immune pathways/mechanisms. Therefore, rational vaccine development requires a comprehension of vaccine immune mechanisms. Immunity is the ability to distinguish ‘self’ and ‘non-self’ materials to eliminate the ‘non-self’ materials [2]. These ‘non-self’ materials are often called antigens, which are the parts of molecules (e.g. proteins) from the disease-causing microorganisms (also known as pathogens). Human immunity could be classified into innate and adaptive immunity. Innate immunity is a non-specific immune response that is antigen-independent and mostly mounts an immediate but short-living immune response against the ‘non-self’ objects, allowing time for adaptive immunity to occur. The innate immunity can also be enhanced after adequate priming or training, leading to ‘trained innate immunity’ [3]. Adaptive immunity is antigen-specific, which is the major immune mechanism stimulated by vaccines.

Vaccines work by inducing the body to gain adaptive immunity to a pathogen such that the body will produce a more robust immune response upon infection. Adaptive immunity includes humoral (antibody) response achieved by B-lymphocytes producing antibodies and cell-mediated immunity achieved by T-lymphocytes targeting specific cells. T cell responses can be primarily categorized into CD4 and CD8 T cell responses. CD4 (helper) T cells play a key role in releasing signals to aid both humoral and cell-mediated responses and the induction of long-term memory. CD8 (cytotoxic) T cells interact with cells presenting the antigen epitopes and induce the programmed cell death of the infected cells. In both humoral and cell-mediated immunity, immune cells rely on the recognition of B and T cell epitopes, which are the specific regions of an antigen recognized by the immune system. Hence, computational prediction of B and T cell epitopes is also an important step in rational vaccine design. However, our understanding of vaccine immune mechanisms is still limited. For example, we know neutralizing antibodies could be a correlate to protective immunity against COVID-19; however, we still do not have a clear understanding of the correlates of protection against severe symptoms. As a result, we often predict vaccine antigens without considering the specific immune mechanisms.

Rational vaccine design applies computational methods and prior knowledge to improve vaccine design. Different methods have been explored for effective vaccine design. Figure 1 provides a general overview of computational vaccine design approaches applied to COVID-19. These include reverse vaccinology (RV), structural vaccinology (SV), machine learning (ML) and ontology applications. Specifically, the pipeline includes, but not limited to, three stages: (i) gold standard collection and immune pattern/mechanism analysis, (ii) computational vaccine design for experiments and clinical trials and (iii) clinical trial and post-authorization vaccine usage surveillance. Different methods including RV, SV, ML and ontology-based methods can be applied to support these stages of study. For example, the identification of protective antigens (PAgs) from the literature could be facilitated by ontology-guided research, which is then utilized by RV to predict potential vaccine antigens and optimized using SV.

Herein, we discuss different vaccine design topics based on the classification of these three stages, with a focus on how these methods have been used to support rational COVID-19 vaccine design. Finally, we discuss potential avenues of future work that can be used to enhance RV.

Gold standard protective vaccine antigens for rational vaccine design

The first step for rational vaccine design is to obtain high-quality data from known PAgs. Gold standard PAgs usually come from direct experimentally verified results. One of the most valuable experiments is the pathogen challenge assay, which evaluates how animals respond to a pathogen challenge after vaccination. Compared to non-vaccinated controls, the successfully vaccinated group should have statistically significant results against the pathogen challenge. Since humans are rarely challenged due to restricted guidelines, pathogen challenge assays are usually performed using laboratory animals. In addition to challenge assay, if a host immune response to a specific pathogen (e.g. neutralizing antibody against severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) [4]) correlates with the protection, the results from such immune response assay could also be used as experimental evidence.

Results from a large number of PAg studies have been collected in the Protegen database [5]. As of the end of 2021, Protegen contains 1547 PAgs encompassing 100 infectious diseases caused by bacteria, viruses and parasites and non-infectious diseases, including cancers and allergies. These PAgs were manually collected and curated from the literature with supporting experimental evidence (e.g. protection assay against a challenge or immune response assay correlates with protection). Analysis of Protegen data allowed us to identify patterns in PAgs that are associated with protection [6]. Another related database is the AntigenDB, an immuneinformatics database of pathogen antigens [7]. Compared with Protegen, AntigenDB focuses on epitope annotations but does not provide the evidence of protection for the antigens collected in the database.

Literature mining (LM) can be used to improve the search for PAgs and immune responses. For example, SciMiner is a tool for mining scientific literature using dictionary- and rule-based methods [8], which was applied to the study of vaccine-associated gene interaction networks with the support of biomedical ontology [9, 10]. In informatics, a formal ontology is a structured vocabulary of terms and relations among terms represented in a computer-interpretable format. The ontology-supported SciMiner demonstrated enhanced performance in identifying vaccine–gene and gene–gene linkages from publications [9–11]. Recently, LM was combined with mechanistic graphical modeling to consolidate existing knowledge around mRNA vaccines’ mode of action, the chain of immune events, to enhance the translatability of preclinical hypotheses [12].

The understanding of enriched patterns significantly supports vaccine design. Our previous Protegen data analysis found that specific patterns in criteria such as subcellular localization and adhesin probability have enriched patterns in PAgs compared to non-protective proteins [6]. For example, we found that 30% of PAgs and 0.4% of non-PAgs in Gram-negative bacteria are located in the outer member membrane of the bacterial cells (P-value <0.01), suggesting that the outer member membrane location is a good criterion for differentiating protective versus non-PAgs in Gram-negative bacteria. Similarly, we found that adhesins are more likely to be PAgs in bacteria, suggesting that adhesin probability is another good criterion for PAg prediction.

In addition to sequence analysis, it is also important to know when and where an antigen is expressed, which can be investigated using various omics data resources such as GEO [13] and ArrayExpress [14]. Enriched patterns of PAgs and their roles in stimulating protective immune responses can be identified from Omics gene expression data analysis. Many pathogen-specific Omics resources, such as the DualSeqDB, the host–pathogen dual RNA sequencing database for infection processes [15], and the Eukaryotic Pathogen, Vector and Host Informatics Resource (VEePathDB) [16], are also available to specifically study host–pathogen interactions. It is usually considered that pathogen proteins that are highly expressed in vivo are more likely to be PAgs than inactively expressed proteins. Such a rationale can help better identify protective vaccine antigens. Ontology has also been used to help analyze data for expression of the host response to vaccines [17–19] or pathogens [20–22]. This allows for the identification of patterns in host immune responses to different vaccine types [23] and potential adverse events (AEs) caused by vaccination.

Compared to genome sequence analysis, omics data analysis is more complex due to the requirement of considering various conditions under which Omics assays are performed. Ontology provides a feasible way to deal with the complexity of the diverse conditions tested. For example, three independent high-impact studies [24–26] of transcriptomic gene expression profiles in Yellow Fever vaccine 17D immunized human subjects led to overlapping but quite different results of enriched genes [27, 28]. The Vaccine Investigation Ontology (VIO) [28], together with the Vaccine Ontology [18, 29], has been used to model and standardize various vaccine study conditions. These ontologies were designed to describe different variables for vaccine design. VIO allowed us to identify the underlying causes of the overlapping but different transcriptomic gene expression profiles found in the three studies. Such a strategy can also be used to study the immune responses stimulated by other vaccine types. The COVID-19 pandemic prompted the Coronavirus Infectious Disease Ontology (CIDO) [30, 31] development to integrate and provide knowledge representation for all aspects of coronavirus diseases. The VIO and CIDO can be integrated to better support omics data annotations and analyses, leading to enhanced future vaccine design.

RV and its application to COVID-19 vaccine design

The conventional vaccine discovery method is performed via in vitro screening in the laboratory. However, this method is resource-consuming and cannot be applied to all pathogens since not all pathogens can be cultured and not all proteins can be purified for testing. With the advent of high-throughput sequencing technologies, a very large number of microbial genome sequences have become publicly available. This fostered an innovative vaccine design approach, termed RV, which starts with a computational prediction of vaccine candidates by genome-based bioinformatics analysis [32]. The first RV study identified vaccine candidates against meningococcal disease from the whole genome sequences of the disease-causing bacteria [33]. From the genome sequence of Neisseria meningitidis serogroup B strain MC58, 600 genes were predicted to encode surface-exposed or exported proteins, and 350 from the list were successfully cloned and expressed in Escherichia coli, purified and used to immunize mice. Within 18 months, 25 of these were shown to induce bactericidal antibodies, which correlate with vaccine efficacy in humans [33]. The four most immunogenic and conserved antigens are included in the Bexsero® vaccine (formerly known as 4CMenB) [34], which was licensed within a relatively short period in early 2013 in the EU and 2015 in the United States [35].

The first RV study’s great success led to the development of many RV prediction programs [36]. The existing open-source RV prediction programs could be characterized based on their algorithmic approaches or input feature types (Figure 2). The first category includes rule-based filtering RV programs, including NERVE [37], Vaxign [38, 39], Jenner-predict server [40] and VacSol [41]. In particular, Vaxign has been applied to predict vaccine candidates for more than 20 pathogens including Acinetobacter baumannii [42, 43], African swine fever virus [44], Brucella melitensis [45], Clostridioides difficile [46, 47], Helicobacter pylori [48], Mycobacterium spp. [49] and S. pneumoniae [50]. Vaxign has also been used in a Reverse Microbiomics strategy to predict autoantigens and virulence factors in dysbiotic microbiomes [51]. Both NERVE and Vaxign are rule-based filtering RV programs that predict and use features such as subcellular localization, adhesin probability, transmembrane helices and immune epitopes. NERVE is a command-line program, while Vaxign is the first web-based vaccine design program. From the comparison of different RV methods, Vaxign exhibits the best specificity while sacrificing sensitivity [39]. Another potential subclass of rule-based RV is used by the program ReVac, developed for bacterial pathogens [52]. Alternative to filtering, it implements a ranking system for every gene as a potential vaccine candidate based on control antigens acquired from databases such as Protegen and AntigenDB. It surveys each vaccine candidates’ positive and negative features at the nucleotide and amino acid level, utilizing several prediction tools and ranking schemes, which attempt to counterbalance any individual tools’ potential for inaccurate feature predictions. As it considers a pathogen’s pangenome, ReVac has identified species-conserved known and novel vaccine candidates for Moraxella catarrhalis [52], non-typeable Haemophilus influenzae [52] and S. pneumoniae [53].

Indeed, most RV applications published to date were focused on one or a few genomes of a given organism of interest and did not account for the generally high amount of genomic diversity observed among isolates of a species, i.e. the species’ pangenome [54, 55]. Knowledge of the population structure of the organism, the proportion of isolates carrying an antigen of interest and the degree of variation of that antigen within those populations are paramount to implement in future RV methods to improve vaccine development. PanRV is a RV tool for identifying potential vaccine candidates in microbial pangenomes [56].

In recent years, advances in ML and the accumulation of vaccine implementation data [5, 57] enabled the development of next-generation ML-based RV tools. These ML-RV tools include VaxiJen [58] and Vaxign-ML [59]. Vaxign-ML was developed and systematically evaluated using nested 5-fold cross-validation and leave-one-pathogen-out validation and had the highest predictive performance compared to other existing RV tools (including rule-based and ML-based methods) using a benchmarking dataset. The Vaxign2 vaccine design framework integrated both Vaxign and Vaxign-ML to facilitate RV-based vaccine antigen identification [59]. Two other studies [60, 61] also performed ML analysis for vaccine design; however, there is no source code or web tool available from these studies.

RV prediction of protective vaccine antigens is often followed by epitope prediction, which is a major topic of immune-informatics [62–64]. There are two major classes of epitopes that are usually analyzed: B cell and T cell epitopes [65, 66]. Epitope-based predictions focus on scanning possible epitopes across antigens. Currently, we can achieve high accuracy (over 90%) in the prediction of MHC Class I and II T cell epitopes, relatively low accuracy in continuous (2D) B-cell epitopes and much lower accuracy in discontinuous (3D) B-cell epitopes [62]. The Immune Epitope Database Analysis Resource (IEDB-AR) stores many manually curated epitopes [67, 68]. The IEDB population coverage provides an estimation of how broad a given epitope(s) would cover the target population, allowing for vaccines to account for higher coverage for a given country [69]. Among many epitope predicting tools available [62], Epivax’s iVax toolkit is a commercial software tool kit designed to handle multiple sequences for their analysis [70] and was utilized for a COVID-19 vaccine design [71]. These IEDB tools have been used to predict candidate targets for immune responses to SARS-CoV-2 [72] and to study the impact of different SARS-CoV-2 variants on the total CD4+ and CD8+ T cell immune responses [73].

Table 1 lists 14 major RV studies applied to SARS-CoV-2 vaccine design. These studies utilized a mix of ML or filter-based methods applied to proteins or epitopes, illustrating the diversity of the RV approaches used to tackle COVID-19. Studies that were published after late 2021 typically incorporated sequence data from different SARS-CoV-2 variants, while earlier ones utilized only the original Wuhan reference strain. These early or late studies also tended to utilize the same set of webtools or approaches for prediction and validation, such as netMHCpan [74], NetMHCIIpan [75], JCAT [76] and Vaxijen [86]. Many studies also assessed their predicted epitopes and vaccine constructs’ immunogenicity by docking with toll-like receptors or human leukocyte antigen. While all studies predicted the SARS-CoV-2 S protein or epitopes from SARS-CoV-2 S as part of their vaccine design, epitope-based studies tended to include non-S protein epitopes.

The epitope prediction studies typically suggested a combination of epitopes from multiple SARS-CoV-2 proteins for a vaccine construct. These include the identification of non-structural protein as having potential epitopes for vaccination [77, 78]. Non-structure proteins (nsps) are derived from the SARS-CoV-2 orf1ab polyprotein that plays an important role in viral replication [79]. Nsp3 was suggested to be included as part of a cocktail vaccine with the S protein [78]. Safavi et al. constructed a multi-epitope vaccine candidate out of six different nsps and the S-ACE2 binding domain for their vaccine construct [80]. SARS-CoV-2 orf10 [77] was another potential candidate chosen due to dissimilarity to other proteins. These epitope prediction studies utilized the methods that were previously described and as such are well tested and supported by existing research tools and databases. Moreover, prior knowledge and experiments on SARS and Middle East respiratory syndrome guided antigen selection and vaccine design.

Uniquely, the protein prediction studies exhibited greater variety in the techniques used to analyze the SARS-CoV-2 gene sequences. Besides the use of the tools described earlier, regular expression and distance algorithms were also utilized for RV data analysis. Regular expressions were used to identify potential epitopes that are not found in human hosts to minimize the chance of the immune system recognizing the vaccine candidate epitope in humans [81]. Another study utilized a distance algorithm to identify common protein interaction motifs between SARS-CoV-2 and humans [82]. A third instead identified nullomers [83], amino acid sequences that were as far apart from any known human acid amino sequence as possible, to identify potential vaccine candidates. However, the results of these unique approaches were not well cited in other papers and have not been advanced for vaccine design. Therefore, it is unclear if these methods are better than the current epitope-based standard.

SV and its application for COVID-19 vaccine design

SV is a rational method to address vaccine design issues regarding suboptimal stability, safety, immunogenicity or generation of broad protection against all isolates of a pathogen [84, 85]. The first proof-of-concept study enhanced the immunogenicity of the fusion (F) glycoprotein of respiratory syncytial virus (RSV) by fixing the conformation-dependent neutralization-sensitive epitopes [86]. The F glycoprotein contributes to the fusion of the RSV and host cell membranes and is a primary target for vaccine development. For decades, researchers have been using the post-fusion F glycoprotein as a vaccine candidate, but it did not provide any protection in challenge studies. An investigation into the conformational rearrangement of this protein between its metastable pre-fusion and stable post-fusion identified a change in its epitope content [87]. The less stable form of the pre-fusion F glycoprotein has more and better epitopes. Therefore, a vaccine candidate that fixed the F glycoprotein in its pre-fusion conformation induced a more potent neutralizing antibody response. The discovery of this structure-based approach has revolutionized vaccine development (Figure 3).

Following the RSV vaccine success, SV has been applied to vaccine design for other pathogens, including SARS-CoV-2 and other coronaviruses (Table 2). Similar to the RSV, the spike (S) glycoprotein of the SARS-CoV-2 plays a crucial role in mediating virus entry and has been the primary target of many vaccines currently in clinical trials or authorized for emergency use [1]. Since the cryo-EM structure of the S protein [88] and the neutralizing antibodies that bind to the S protein [89, 90] were determined, SV approaches have been applied to optimize the S protein structure as a vaccine candidate. For example, Henderson et al. controlled the S protein’s receptor-binding domain (RBD) between the ‘up’ and ‘down’ configurations to induce immunogenicity [91]. On the other hand, structural modifications were also performed on the native S protein to stabilize the S protein in its pre-fusion form [92], a strategy similar to the RSV vaccine development [93]. This led to the use of a double proline added to stabilize the S protein and lock its conformation after assembly.

The pre-fusion S protein design was used for the mRNA vaccines developed by Pfizer [94] and Moderna [95] and for other vaccine candidates. For example, the Johnson and Johnson adenoviral vaccine incorporates double proline substitutions on the furin-like domain that yielded more neutralizing antibodies [92]. In addition, other studies for SV focused on incorporating features of the SARS-CoV-2, such as the presence of a furin cleavage site between two of its subunits [96]. The Biovacc-19 vaccine study proposed targeting non-human-like epitopes to avoid immune system detection [97]. Dae et al. described the suitability of the S protein’s RBD for protein subunit vaccines [98]. RBD can be a monomer or bind to itself as a dimer and be stabilized with a tandem repeat single chain.

In addition, the second strategy for SV is to introduce new epitopes and correspondingly enhance the immunogenicity of the vaccine candidate. For example, a single epitope mutation in the H7N9 hemagglutinin vaccine enhanced it from eliciting non-neutralizing antibody immune responses [99, 100] to a 4-fold increase in neutralizing antibodies without changing its conformation [101]. Ong et al. [102] utilized a structure-based evolutionary approach to generate possible modified S protein candidates. This was done through a new algorithm called EvoDesign that generated mutations of the S protein with added CD4 T cell epitopes while keeping the same B cell epitope profile [102]. This approach could be utilized to address possible SARS-CoV-2 mutants. However, this method is limited to computational prediction and has yet to be verified experimentally. A common problem with all these in silico tools and methods is the lack of experimental verification, which complicates the assessment of advantages and disadvantages among different methods.

COVID-19 has already acquired new mutations that increase transmission and reduce the effectiveness of current vaccines. The D614G mutation of the S protein has been shown to increase infectivity, possibly due to increased thermal stability of the S protein molecule [103]. This also leads to concerns about immune escape from the new variants [104, 105]. Immune escape can be dealt with by creating new vaccines that incorporate conserved epitopes across multiple sequences or by forecasting potential mutations. Several recent studies attempt to incorporate multiple COVID-19 strains [106, 107] or Betacoronavirus species [108] using antigens already identified by RV. The alternative approach to deal with immune escape is to instead identify a set of epitopes that are resistant to virus mutation [109]. There is also work in developing algorithms and methods to predict potential S protein mutations that can cause immune escape. So far, these methods include mapping the location of antibody-resistant mutations [110] or utilizing natural language processing (NLP) to identify semantic changes that represent viral escape [111].

It is noted that different RV and SV strategies are often not directly comparable. As introduced above, some of the RV strategies may target on the protein level and some on the epitope level. It could also be a combination of focusing on a specific protein (such as the Spike protein of SARS-CoV-2) or all the proteins of a pathogen. Different SV strategies have also been developed, including the conformational changes on specific proteins [86], the addition of new epitopes [102] or maintenance of the RBD and then N-terminal domain [112, 113]. It is also possible to combine different strategies to achieve complementary results.

Vaccine efficacy and adverse event analysis to support vaccine design

Post-authorization and post-licensure vaccine success are informative to vaccine efficacy and vaccine design. In the fight against the COVID-19 pandemic, the marketed COVID-19 vaccines were all initially authorized for emergency usage, followed by later stage formal licensure approval. Different COVID-19 vaccines offer differing levels of neutralizing antibodies and protection, which is important to guide further vaccine design [114, 115].

Another aspect of vaccine usage studies is related to vaccine AEs (VAEs), which are undesirable outcomes associated with vaccinations. Knowledge of VAE results can inform the efficacy and safety of existing vaccines, further guiding future vaccine design [17, 116–119]. For example, a new vaccine-induced immune thrombotic thrombocytopenia (VITT) was associated with the Ad26.COV2. S (i.e. Janssen vaccine) and ChAdOx1 (i.e. Oxford-AstraZeneca vaccine) nCoV-19 vaccines that use replication-incompetent adenoviral vectors [120]. This observation caused a pause in the authoritative use of the vaccine in the United States from March to April 2021 while the FDA investigated this association (https://www.fda.gov/news-events/press-announcements/fda-and-cdc-lift-recommended-pause-johnson-johnson-janssen-covid-19-vaccine-use-following-thorough). This scenario confirms the importance of examining the vaccine safety and its integration for better vaccine design. A recent bioinformatics study did not find any evidence of significant homology of SARS-CoV-2 spike protein sequences to myocarditis-associated antigens [121]. However, the COVID-19 mRNA vaccines, including Pfizer-BioNTech BNT162b2 and Moderna mRNA-1273 vaccines, have been reported to be associated with myocarditis [122]. Although rare, severe AEs like VITT and myocarditis have raised concerns in term of the vaccine safety in specific populations [122]. Therefore, it is critical to analyze VAE data to identify potential vaccine safety issues and understand possible underlying molecular mechanisms.

The major types of VAE data include VAE case reports, electronic health records and the literature. The VAE Case Reporting System (VAERS; https://vaers.hhs.gov/) is a US CDC/FDA-led online system that allows the public, health professionals and vaccine manufacturers to report AEs following vaccination [123–125]. We have recently developed Cov19VaxKB, a web-based integrative COVID-19 vaccine knowledgebase that not only includes COVID-19 vaccine knowledge but also provides web interfaces for querying and analyzing COVID-19 vaccine AEs downloaded from the VAERS database [1]. VAERS AE data are standardized using the medical dictionary for regulatory activities (MedDRA) [126]. To further support advanced AE classification, the Ontology of Adverse Event (OAE) was developed [127–129]. OAE was used to analyze VAE patterns in VAERS and help identify patterns of gene expression caused by VAEs [117, 127, 130–133]. Furthermore, OAE can be used as a VAE modeling system to model the relations among AE severity and their occurrence across certain demographics, such as in males or females [118, 134]. Different statistical methods have also been developed to associate specific AEs with different vaccines under specific conditions [135]. Statistical and ontological methods have been developed to analyze VAERS AE case report data associated with monovalent and combination vaccines against hepatitis A and B diseases [136]. While the VAE analyses reported AEs significantly enriched in comparison with other vaccines, it also supported the assertion that COVID-19 vaccines are safe [1].

Using literature data, the AEs after intended or accidental usage of Brucella vaccines were also collected and analyzed, leading to better understanding of VAEs from animal vaccines against zoonotic brucellosis [137]. Many NLP tools such as SciMiner [8] and cTAKES [138] can be used to support AE-related EHR and literature data mining. Ontology development significantly supports AE LM [23, 139–142]. Ontology-based LM of vaccine–gene or pathogen–disease associations has been used to detect the associations between pathogens and phenotypes, supporting infectious disease research [29, 143, 144]. ML tools can then be implemented to analyze mined AE data [145–148].

Future directions

Most existing RV methods use either biological properties or physicochemical properties of proteins in their training datasets. Significant correlations have been reported for properties such as subcellular localization, adhesin probability and peptide signaling [6]. Meanwhile, physicochemical properties, such as hydrogen bond association in spike protein subunits, were significantly associated with S protein mutations in SARS-CoV-2 variants of concerns [170]. Rational vaccine design can be inferred by further investigating protective immune mechanisms. For example, new criteria for vaccine design might be identified by comparing the relations between virulence factors (VFs) and PAgs. VFs are molecules that allow pathogens to overcome host defense mechanisms and cause disease. A total of 5304 VFs supported by experimental evidence (e.g. loss or reduction of pathogenicity in the host after the VF gene mutation) are curated and stored in the Victors database [171]. Many of these VFs are also used as PAgs for vaccine development due to their important roles in pathogenicity and protective antigenicity. An early study shows that VFs and PAgs have overlapping but differential profiles [171]. However, more work is needed to deeply compare the pools of VFs and PAgs to identify more fundamental mechanisms and relations among these two types, which will also support rational vaccine design. Similarly, the mechanisms underlying the effects of different SARS-CoV-2 mutations, beyond those for S proteins, are not well characterized.

SV research currently shows infrequent usage of ML algorithms for SV enhancement. Except for the recent use of agent-based modeling for in silico verification of vaccine candidates, ML methods were ancillary [172]. One potential expansion for SV design would be in the use of evolutionary algorithms. Evolutionary algorithms apply multiple rounds of assessment and replacement with modification of inputs to optimize the results to a given function over multiple iterations based on their fitness. However, for SV, there still is a lack of good scoring functions to ensure that the evolutionary algorithms will generate optimized vaccine candidates, instead of maximizing some arbitrary mechanic. While the SV study by Ong et al. [102] provides promising in silico prediction results, only one generation of structural candidates was performed. Their study could be augmented by developing and using a good fitness function, which can then be used to support multiple generations of computation and vaccine candidate prediction.

Due to the emergence of new variants of SARS-CoV-2, many authorized vaccines are less efficacious than initially designed. The need for updated vaccines has prompted the development of another round of vaccine development to account for emerging variants [80, 106, 169]. While the current vaccines are still efficacious enough against variants of concern, this may not remain true for future emerging variants. The existing vaccine design tools surveyed in this article can easily be reapplied to new variants to update or create alternative vaccine design. However, there is still a time delay to develop and test the updated vaccines.

The rapid evolution of new variants requires a rational strategy to balance vaccine coverage for the short and long term. This development is aided by the large volume of COVID-19-related data allowing for the analysis of mutations that are evolutionarily favored [173]. By using the vast number of mutated SARS-CoV-2 sequences, it is possible to identify mutational patterns that result in changes in immunogenicity, antibody reaction and vaccine efficacy. We may further design more powerful vaccines, e.g. by using conserved sequences or incorporating critical mutations, against the infection of various viral variants. The IEDB has incorporated COVID-19 variants of concern, allowing for further vaccine design across variants [73]. While recent studies have proposed methods to predict relevant mutations for new variants [174], these are still underdeveloped.

There are still other major concerns in COVID-19 vaccine design and usage. SARS-CoV-2 has been shown to jump across multiple mammal species, such as mice [175], ferrets [176] and white deer [177]. This promiscuity in reservoirs allows for viral populations to evolve into different strains or variants, potentially returning as variants of concerns for humans [178]. Moreover, as different vaccines can differ in their immunogenic profiles [179], it is possible for different vaccines to differ in effect against different SARS-COV-2 variants. While evidence shows that the currently approved vaccines are still effective against different viral variants, albeit reduced, how effective they are is inconsistent across multiple studies [180, 181]. Furthermore, insufficient vaccine coverage in the general population allows for COVID-19 to continue to spread and mutate into new variants. The United States, as of 7 April 2022, only has 66.46% of its population fully vaccinated based on the COVID-19 Map dashboard hosted at the Johns Hopkins Coronavirus Resource Center [182]. While AE surveillance can be used to assuage the public and identify potentially unsafe vaccines, a vaccine can only help when it is used. These concerns make it important but challenging to design more effective and safer COVID-19 vaccines against various variants.

Much research is still needed in different directions. High-quality benchmarking of PAgs, such as the Protegen data collection, still need to be enhanced to support better vaccine design tool development. The effects of VAEs can be better analyzed and incorporated for further vaccine design. The function of mutations can be analyzed to discover patterns that change vaccine efficacy or immunogenicity and allow for the development of vaccines that can account for new, emerging strains. Additional programs, including new webtools for vaccine design, can be developed. There is currently no web tool for designing SV-optimized PAgs [98, 183]. The incorporation of more advanced ML tools to increase the accuracy of in silico predictions is also needed to improve RV and SV development. Successful modeling of human immune responses through the use of ML would be able to provide quick and safe preliminary results to identify optimized antigens. Another potential utility for ML algorithms is to identify which vaccine would have the greatest effect for specific demographics as part of personalized medicine. While these ML tools require good data sets that provide robust and accurate results, standardized ontologies can be further developed and applied to support AI-ready data and metadata standardization and more advanced vaccine design.

Data availability

Not applicable.

Authors’ contributions

Y.H. initiated the study. A.H. generated data. E.O., J.H., A.D. and H.T. edited the manuscript. All coauthors discussed and co-prepared the manuscript.

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Author Biographies

Anthony Huffman is a PhD student at the University of Michigan Medical School, Ann Arbor, MI, USA. He is focused on ontology standardization and analysis of gene expression related to vaccines.

Funding

National Institute of Allergy and Infectious Diseases (NIAID) grants (1UH2AI132931, 1R01AI081062); a COVID-19 research grant (U072807) from Michigan Medicine–Peking University Health Sciences Center Joint Institute for Clinical and Translational Research (to Y.H.).