In silico identification of vaccine targets for 2019-nCoV

Identification of 28 exact matches to SARS CoV immunogenic peptides by screening all epitopes deposited in IEDB

We collected all peptides in IEDB (3, as of 13-02-2020) reported positive in T cell assays and have human as the host organism. We then conducted a local sequence alignment of 10 2019-nCoV open reading frames (ORFs) against 35,225 IEDB peptides, and found 28 exact matches. Surprisingly, all identical hits (against peptides having sequence length greater than 3) were from SARS-CoV (Table 1, Data Table 1⁴). These peptides have been shown to bind various HLA alleles, although with higher tendency towards HLA-A:02:01, from both class I and class II, and can be target for CD8+ and CD4+ T cells respectively.

Identification of 22 2019-nCoV peptides with high degree of similarity to previously reported immunogenic viral peptides

In addition to 28 identical hits against SARS CoV, we observed a long tail in distribution of normalized alignment scores between 10 2019-nCoV ORFs and 35,225 IEDB peptides (Figure 1A, Methods). We therefore set out to further investigate potential vaccine targets among highly similar sequences.

Figure 1. 2019-nCoV peptides with high sequence similarity to immunogenic peptides in IEDB.

A. Comparison of normalized sequence alignment score for peptides with exact and non-exact matches. B. Number of target peptides grouped by their source organism.

The peptides having an exact sequence alignment with epitopes in IEDB had normalized alignment scores ranging from 4 to 6. Taking the normalized alignment score of exact matches as a reference, we extracted 2019-nCoV peptides having score greater or equal to 4. As illustrated in Figure 1A, we observed 45 and 11 peptides having normalized alignment score ≥ 4 and ≥ 5 respectively (Figure 1A inset). The target peptides were originated from 10 different sources (Figure 1B) where a total 36 peptides were derived from strains associated to SARS CoV. Of interest, we also observed 7 hits having high sequence similarity to targets from Homo sapiens.

In order to investigate the extent to which the difference between the source (2019-nCoV) and target (IEDB) peptides influences the immunogenicity of the source peptides we used a recently published immunogenicity model⁵ to predict and compare the immunogenicity between the source and target peptides (Data Table 2⁴).

We could see a similar (close to identical) immunogenicity scores for a number of IEDB and 2019-nCov peptides especially for those with high immunogenicity scores (Figure 2). While all 48 can be potential targets, of particular interest were those having higher immunogenicity score than IEDB peptides. Here, we list 22 out of 48 2019-nCoV peptides that scored higher compared to their targets that have been characterized to be immunogenic (Table 2). In this list 15 (68%) 2019-nCov peptides have a score higher than 0.5 whereas only 11(50%) of IEDB get a score immunogenicity score greater than 0.5.

Figure 2. Predicted immunogenicity for IEDB immunogenic vs. 2019-nCoV peptides.

2019-nCoV peptides having a high sequence similarity to immunogenic peptides and their targets were analysed for their immunogenicity potential by iPred algorithm.

It is worth noting that in general predicting immunogenicity of given a peptide is challenging and not a fully solved problem, and therefore current models for predicting immunogenicity are suboptimal. iPred is also not an exception. In fact, we could see that a substantial number of IEDB immunogenic peptides were scored < 0.5 (the threshold score used to classify immunogenic vs non-immunogenic). This led us to ask whether we can gather any other evidence of either immunogenicity or cross-reactivity.

De novo search of immunogenic peptides in 2019-nCoV proteome

As a complementary reciprocal approach, we conducted a de novo search of immunogenic peptides against the 2019-nCov proteome sequence. We scanned 9-mers from 2019-nCoV proteome with a window of 9 amino acids and step length of 1 amino acid (9613 in total). The immunogenicity of 9-mer peptides were predicted using iPred and MHC presentation scores were gauged using NetMHCpan 4.0⁶ for various HLA types. In this task, we focused on haplotypes common in Chinese and European populations, which include HLA-A*02:01, HLA-A*01:01, HLA-B*07:02, HLA-B*40:01 and HLA-C*07:02 alleles (Data Table 3⁴).

Based on MHC presentation and immunogenicity prediction, we detected 5 peptides predicted to bind 4 different HLA alleles of which 2 had strong immunogenicity scores (Figure 3). For those 65 strong binders to 3 different HLA types, 39 had immunogenicity scores ≥ 0.5 (Table 3). Collectively this analysis suggests a number of 9-mer immunogenic candidates for further experimental validation.

Figure 3. De novo search of 9-mer 2019-nCoV peptides with MHC presentation and immunogenicity potential.

The MHC binding was predicted for HLA-A*02:01, HLA-A*01:01, HLA-B*07:02, HLA-B*40:01 and HLA-C*07:02 alleles by NetMHCpan 4.0 and immunogenicity was predicted by iPred.

Immunogenicity of 2019-nCoV peptides to 1G4 CD8+ TCR molecule

While our de novo candidates are appealing shortlisted targets for experimental validation, it does not provide information about target T cell receptors (TCRs). We therefore set out to interrogate the possibility of cross reactivity with one well-studied TCR.

T cell cross-reactivity has been instrumental for the T cell immunity against both tumor antigens and external pathogens. In that regard, a number of T cells have been extensively characterized including 1G4 CD8+ TCR, which is known to recognize the ‘SLLMWITQC’ peptide presented by HLA-A*02:01. We therefore set out to leverage the data from a recently published study⁷ and exploit the possibility of cross reactivity of this TCR to any 2019-nCoV peptide.

Here, we scanned all 9-mers from the 2019-nCoV proteome (9613 peptides) with Binding, Activating and Killing Position Weight Matrices (PWM, see the method section) and associated each peptide with the geometric mean of these three assays as a measure of immunogenicity (Data Table 4⁴). The distributions of binding, activation and killing scores along with their multiplicative score and geometric mean are illustrated in Figure 4. Based on geometric mean, we observed 20 2019-nCoV peptides with a score > 0.8 and 516 peptides > 0.7. The 9-mer peptides with geometric mean > 0.7 and positive HLA-A*02:01 binding prediction by NetMHCpan 4.0 are listed in Table 4.

Figure 4. Distribution of 1G4 TCR positional weight matrix scores for 2019-nCoV peptides.

The positional weight matrices were obtained from 7 and 9613 9-mers generated from 10 2019-nCoV ORFs were computed for their TCR recognition potential.

We further analysed the MHC binding propensities and gathered peptides not only predicted positive by NetMHCpan but also to have leucine (L) and valine (V) in anchor positions 2 (P2) and 9 (P9) respectively. Previous studies have shown that for MHC-I HLA-A02:01 specific peptides, 9-mer peptides with leucine at P2 and valine at P9 are preferably presented on the surface of HLA-A02:01⁸. Looking at the LV peptide, we identified 44 2019-nCoV peptides of which 2 peptides had immunogenicity score > 0.7 and 12 peptides > 0.6 (Table 5). Thus, here we provide the list of peptides that are potential targets for 1G4 TCR recognition for subjects with HLA-A02:01 haplotype.

Data acquisition

2019-nCoV open reading frame sequences were downloaded from NCBI (MN908947.3). All sequences subjected for analysis are deposited in GitHub repository.

Data analysis

All subsequent analyses have been conducted in R 3.6.1.

Sequence similarity comparison

The sequence similarity between 2019-nCoV open reading frames and previously characterized immunogenic peptides in IEDB was analysed by local alignment using R ‘pairwiseAlignment’ function from Biostrings v2.40.2 package. The local alignment utilized BLOSUM62 matrix, gapOpening of 5 and gapExtension of 5. The alignment score was normalized by length of target peptides.

In extracting peptides with the exact sequence alignment with epitope sequences in IEDB, only peptides with more than 3 amino acids in length were shortlisted.

Immunogenicity prediction

We have used iPred⁵ to predict immunogenicity of each given peptide. Briefly, iPred employs peptides’ length and physicochemical properties of amino acids modelled by sums of ten Kidera factors and associates a score to each peptide reflecting its likelihood of recognition by a T cell.

Predicting presentation by MHCs

In order to predict peptide binding to MHC we used NetMHCpan V4⁶. This version of NetMHCpan that comes with a number of improvements, incorporate both eluted ligand and peptide binding affinity data into a neural network model to predict MHC presentation of each given peptide.

Predicting cross reactivity to 1G4 TCR

To gauge the level of 1G4 TCR cross-reactivity to list of 2019-nCoV virus, we have leveraged the data from a recently published study⁷. 1G4 or NY-ESO-1-specific TCR is a very well-studied and clinically efficacious TCR which recognize the peptide ‘SLLMWITQC’ presented by HLA-A*02:01. Karapetyan et al. have recently provided data from three experimental assays reflecting Binding, Activating and Killing upon each mutation at each position of all possible 9-mers using these three datasets. In a similar way to the original paper, we trained three Position Weight Matrices named B, A and K respectively from Binding, Activating and Killing assay. We defined the cross-reactivity score of a given 9-mer sequence as the geometric mean of B, A and K.

We then scanned 2019-nCoV virus protein sequence with each of B, A and K PWMs and associated each of 9613 9-mers with a cross reactivity score. At the same we utilized NetMHCpan and associated each 9-mer with its presentation score. Our final list of cross-reactive candidate peptides were those with a cross-reactivity sore >= 0.8 and reported as strong binders from NetMHCpan and have ‘L’ and ‘V’ amino acids at anchor positions. The custom R codes are accessible from GitHub repository (see software availability⁴).

Source data

2019-nCoV open reading frame sequences were downloaded from NCBI (MN908947.3).

Underlying data

Zenodo: In silico identification of vaccine targets for 2019-nCoV (Data tables). http://doi.org/10.5281/zenodo.3676886¹⁰

This project contains the following underlying data:

Data are available under the terms of the Creative Commons Attribution 4.0 International license (CC-BY 4.0).