Vaxign2: the second generation of the first Web-based vaccine design program using reverse vaccinology and machine learning

Vaxign filtering-based protective antigen prediction

Vaxign is the first web-based vaccine design program using RV. The first generation of Vaxign applies a filtering-based method to select vaccine antigen candidates based on the user's prior knowledge of the target pathogen's pathogenesis. A typical workflow involves the following components: (i) subcellular localization computed by PSORTb program (19); (ii) transmembrane domains computed using TMHMM 2.0 with default settings (20); (iii) adhesin probability calculated using SPAAN program with default settings (21); (iv) similarity to host proteins (human/mouse/pig) using BLAST and NCBI protein database (22).

Vaxign-ML machine learning-based protective antigen prediction

With the advance of machine learning and accumulation of manually collected protective antigens in Protegen (23), Vaxign-ML was created and significantly improved vaccine antigen prediction (24). In brief, Vaxign-ML combined the protein sequences’ biological and physicochemical properties as the input features to train five different machine learning models. The input protein sequences were extracted from the Protegen database, a continuous effort over the past ten years collecting and annotating experimentally verified protective antigens (23). All machine learning models were evaluated and selected based on nested five-fold cross-validation and leave-one-pathogen-out validation. The best performing model, extreme gradient boosting, was used to build the Vaxign-ML program.

Vaxitop epitope prediction and IEDB verified epitope mapping

However, the Vaxign and Vaxign-ML predicted PAgs could be further investigated for their immunogenic potential as vaccine candidates before experimental verification. Vaxign2 provides the immunogenicity assessment by the post-prediction analysis workflow. The principal mechanism of vaccines is the adaptive immune response: humoral (antibody) and cell-mediated responses. The protection offered by these immune responses is primarily mediated by B cells and T cells. In particular, T cell response can be mainly categorized into CD4 (helper) and CD8 (cytotoxic) T cell responses, which are induced by epitopes bound to major histocompatibility complex (MHC)-II, and MHC-I molecules, respectively. Therefore, it is essential to evaluate the predicted PAgs based on their B cell and T cell epitopes.

Vaxign2 supports MHC-I and MHC-II T cell epitope predictions for input proteins via Vaxitop. In brief, all the epitopes’ Position-Specific Scoring Matrix (PSSM) for different MHC-I or MHC-II alleles are generated by MEME (25) based on known epitope data from the IEDB immune epitope database (26). Then the input proteins are scanned for epitopes by the PSSMs. The P-value for the predicted epitope binding to PSSMs is calculated by the MAST sequence homology search algorithm (25). Besides epitope prediction, Vaxign2 also supports the mapping of IEDB experimentally verified T cell and B cell epitopes to the input proteins (26).

Population coverage prediction

As mentioned in the previous section, epitopes bound to the MHC-I or MHC-II molecules are presented to T cells to induce an immune response. However, human MHC molecules are highly polymorphic, and the expression of different MHC molecules is significantly impacted by human genetic variation. Thus, it is essential to determine if the predicted PAg contains a set of epitopes capable of binding to different MHC molecules and offers a broad coverage to the world population. Based on the result from Vaxitop MHC-I and MHC-II T cell epitope prediction, Vaxign2 can also calculate the population coverage of the input proteins using the IEDB Population Coverage Tool (27). The predicted population coverage of the different countries is also visualized and highlighted in the world map.

Protein function and orthologs prediction

The sequences of all PAgs are scanned for functional domains, including Clusters of Orthologs (COG) and Gene Ontology (GO) terms, as well as possible orthologous proteins using HMMER2 (http://hmmer.org/) with the hidden Markov models downloaded from the EggNog database (28).

Use Case 1: dynamic analysis of SARS-CoV-2 S protein evaluation

The SARS-CoV-2 S protein is a commonly used vaccine antigen in current COVID-19 vaccine development. Figure 2 showed how Vaxign2 was used to dynamically assess the S protein as a vaccine target by Vaxign/Vaxign-ML, and to evaluate the immunogenicity and biological functions of S protein in post-prediction analyses. The SARS-CoV-2 S protein's NCBI protein ID (YP_009724390.1) was input to the Vaxign2 dynamic analysis (Figure 2A). Vaxign2 computed Vaxign/Vaxign-ML results, including the Vaxign-ML score and adhesin probability. Vaxign-ML predicted S protein to be a good vaccine antigen with a score of 97.6 (Figure 2B). Vaxign calculated S protein's adhesion probability of 0.635; with the cutoff of 0.51, this protein was protected to be an adhesin contributing to viral entry into the host cell. The Vaxign/Vaxign-ML results both suggested S protein as a promising vaccine antigen target.

The S protein was then evaluated for its immunogenicity and functional profile by Vaxign2 post-prediction analyses. Vaxitop predicted 94 MHC-I (Supplementary Table S3) and 54 MHC-II (Supplementary Table S4) unique promiscuous epitopes for S protein (P-value ≤ 0.01) (Figure 2C). The MHC-I & -II reference alleles represent the majority of human MHC alleles in the world population (33,34), and epitope promiscuity is defined to bind four or more MHC-I or MHC-II alleles in the reference set (35). Vaxign2 also found 12 and 45 verified epitopes for T and B cells, respectively (Figure 2D, Supplementary Tables S5 and S6). Furthermore, S protein was predicted to have high population coverage in most countries (Figure 2E). Note that some countries with low or no predicted population coverage might be due to the lack of reported allele frequencies in the Allele Frequency Net Database (36) and did not reflect the actual population coverage. Vaxign2 also computed the Gene Ontology (GO) terms for S protein and identified virulence-related terms (Figure 2F), such as viral entry into host cell (GO:0046718), host cell surface receptor binding (GO:0046789), and receptor-mediated virion attachment to host cell (GO:0046813) (Supplementary Table S7). Finally, a total of 51 S protein orthologs were identified (Figure 2G, Supplementary Table S8) in Orthocoronavirinae, which is a subfamily related to human coronaviruses. In summary, the Vaxign2 post-prediction analyses suggested S protein had good epitope profiles and contributed to an important role in viral infection. Such analyses provided by Vaxign2 provided valuable biological rationales on the selection of S protein as a vaccine candidate. Indeed, S protein has been the primary target of many COVID-19 vaccines such as Pfizer and Moderna (16,17).

Use Case 2: pre-computed queries for coronaviruses vaccine selection

The complete proteome of SARS-CoV-2 was uploaded to the Vaxign2 pre-computed queries and was compared to seven other coronaviruses (Figure 3). The causative agents for the Middle East respiratory syndrome (MERS) and Severe acute respiratory syndrome (SARS) are MERS-CoV and SARS-CoV, respectively. SARS-CoV, SARS-CoV-2, and MERS-CoV are all beta-coronaviruses, which are very virulent and cause severe respiratory syndromes. On the other hand, human coronavirus OC43 (HCoV-OC43) and HKU1 (HCoV-HKU1) belong to the beta-coronavirus, while human coronavirus 229E (HCoV-229E) and NL63 (HCoV-NL63) are alpha-coronaviruses. These four strains only cause mild cold symptoms in humans. In addition to the human coronaviruses mentioned above, a murine coronavirus MHV-1 was also included in the comparison to SARS-CoV-2. The hypothesis is that some coronavirus virulence factors only exist in the severe form of SARS-CoV/SARS-CoV-2/MERS-CoV but not in the other mild or non-human coronaviruses. The pre-computed coronavirus results in Vaxign2 could be queried (Figure 3A) to address this hypothesis. Specifically, our Vaxign2 query found seven proteins that were conserved in the three virulent human coronaviruses (SARS-CoV, SARS-CoV-2 and MERS-CoV), but not in the other five mild or non-human coronaviruses. These seven proteins included Non-structural protein 7–10 (nsp7–10), Uridylate-specific endoribonuclease (nendoU), 2′-O-methyltransferase (2′-O-MT), and nucleocapsid phosphoprotein (N) (Figure 3B). Among the seven conserved proteins, three proteins (nsp8–10) were predicted as adhesion proteins by Vaxign, but only nsp8 protein was predicted to be PAg by Vaxign-ML. Therefore, nsp8 was selected for further analysis (Figure 3C). In particular, the genome group phylogeny analysis (Figure 3D) showed that nsp8 was predicted to be more closely related to the SARS-CoV than MERS-CoV and the other four mild human coronaviruses (Figure 3D). It could be a feasible strategy to create a COVID-19 cocktail vaccine, as described in our COVID-19 vaccine prediction study (37), that combines multiple proteins to target different aspects of host immunity for better protection.