Immunopeptidomic Data Integration to Artificial Neural Networks Enhances Protein-Drug Immunogenicity Prediction

Abstract

Recombinant DNA technology has, in the last decades, contributed to a vast expansion of the use of protein drugs as pharmaceutical agents. However, such biological drugs can lead to the formation of anti-drug antibodies (ADAs) that may result in adverse effects, including allergic reactions and compromised therapeutic efficacy. Production of ADAs is most often associated with activation of CD4 T cell responses resulting from proteolysis of the biotherapeutic and loading of drug-specific peptides into major histocompatibility complex (MHC) class II on professional antigen-presenting cells. Recently, readouts from MHC-associated peptide proteomics (MAPPs) assays have been shown to correlate with the presence of CD4 T cell epitopes. However, the limited sensitivity of MAPPs challenges its use as an immunogenicity biomarker. In this work, MAPPs data was used to construct an artificial neural network (ANN) model for MHC class II antigen presentation. Using Infliximab and Rituximab as showcase stories, the model demonstrated an unprecedented performance for predicting MAPPs and CD4 T cell epitopes in the context of protein-drug immunogenicity, complementing results from MAPPs assays and outperforming conventional prediction models trained on binding affinity data.

Keywords: MHC-II prediction; artificial neural-networks; bioinformatics; immunopeptidomics; machine-learning; protein-drug immunogenicity.

Figure 1

Pipeline of protein-drug (Infliximab) immunopeptidome profiling. Infliximab-pulsed DCs were lysed and HLA-DR-peptide complexes were purified with a pan-specific antibody (L243). Next, LC-MS/MS was performed, identifying 15,240 unique ligands. MAPPs self-proteins were used to train the artificial neural network model, NNAlign_MAC. Infliximab MAPPs peptides were pooled from different donors and used to compare to the predicted MHC-II hot-spots regions. Finally, T cell experiments were used to validate regions and select protein-drugs residues prone to introduce modifications in order to avoid immunogenicity.

Figure 2

HLA-DR peptide distribution, binding motifs and amino acid frequencies. (A) MAPPs peptide frequency from all donors combined associated to each HLA-DR gene (HLA-DRB1, DRB3, DRB4, and DRB5) after NNAlign_MAC deconvolution. Percentages (and absolute numbers) are shown for the peptides assigned to each allelic variant. (B) Motif deconvolution obtained by NNAlign_MAC per donor. NNAlign_MAC allele logos were built with all peptides from each MS data set assigned for that particular allele. The number after the allele name reflects the number of peptides found in that dataset for the given allele (Example: DRB3^*03:01-7 peptides). NetMHCIIpan motifs were built from top 1% scoring prediction of 100,000 random peptides evaluated using the list of alleles expressed in each donor sample. Motif logos were build using Seq2Logo with default settings. (C,D) Amino acid frequency comparison of in-house MAPPs and peptides from binding affinity (BA) assays (C) and mass spectrometry (MS) eluted ligands (D) collected from IEDB. For each comparison, 500 peptides per allele were selected at random per each allele (DRB1^*01:01, DRB1^*04:01, DRB1^*07:01, DRB1^*11:01, and DRB1^*15:01) and pooled together before the amino acid frequency was calculated.

Figure 3

NNAlign_MAC improves Infliximab MAPPs predictions. (A) Infliximab profile predictions were generated with NNAlign_MAC (red), MixMHC2pred (purple), and NetMHCIIpan (blue), and benchmarked against Infliximab experimental MAPPs (green). Promiscuity profiles were generated for each method, selecting the protein-drug predicted peptides below a defined %Rank threshold, and stacking the peptides over the protein sequence (see section Materials and Methods). The correlation of the different profiles to MAPPs data (Spearman correlation coefficient, SCC) is shown in matching colors for each prediction method. Different %Rank values were selected for each method according to its best predictive power (NNAlign_MAC = 1, MixMHCIIpred = 0.5, and NetMHCIIpan = 2) (Supplementary Figures 1–3). Complementarity determining regions (CDRs), were calculated with the DomainGapAlign tool of IMGT.org (CDR1-IMGT:27-38; CDR2-IMGT: 56-65; CDR3-IMGT:105-117), both for infliximab heavy and light chain variable domains (blue rectangles). (B) Scatter plots of the predicted profiles in (A) for NNAlign_MAC, MixMHC2pred, and NetMHCIIpan vs. MAPPs. Both SCC and PCC are shown for Infliximab heavy (Infliximab_HC, red) and light chain (Infliximab_LC, blue). The discrete patterns in the x-axis of the plots are explained by the maximum number of alleles predicted for each method (MixMHC2pred is only available for a limited set of alleles).

Figure 4

NNAlign_MAC Infliximab and Rituximab MAPPs prediction. (A) Infliximab MAPPs peptides were collected from Hamze et al. (15) and compared to in-house MAPPs profiles. Note that the collected dataset only contained peptides mapped to the variable regions of Infliximab light and heavy chains. NNAlign_MAC prediction promiscuity profiles and SCC correlation against the two datasets are shown in matching colors (red). (B) Scatter plot of the two MAPPs profiles [In-house vs. Hamze et al. (15)] for the heavy (red) and light chains (blue) of Infliximab protein-drug. (C) NNAlign_MAC correlation to the combined dataset [In-house + Hamze et al. (15)]. (D) Scatter plot of the NNAlign_MAC prediction vs the combined Infliximab MAPPs profile. (E) NNAlign_MAC correlation to Rituximab MAPPs data collected from Hamze et al. (15). (F) Scatter plot of the NNAlign_MAC prediction vs Rituximab MAPPs profile from the same publication.

Figure 5

T cell evaluation of MAPPs and NNAlign_MAC identified hot-spot regions. (A) Schematic with of the location of 6 and 9 ELISpot tested peptides for the light and heavy chain of Infliximab, respectively. The peptides were identified with a color code covering three regions, (1) regions predicted by NNAlign_MAC where MAPPs peptides were found (Magenta: LC_1, LC_5, LC_6, HC_2, HC_3, HC_4, HC_6, HC_8, HC_9); (2) regions predicted by NNAlign_MAC, with no MAPPs peptides (Cyan: LC_2, LC_4); and regions were the methods identified none or very few ligands (Yellow: LC_3, HC_1, HC_5, HC_7). (B) IFN-γ ELISpot test for Infliximab peptides selected in (A). Each boxplot was constructed from the two-individual donor-response measurement replicas to each peptide assessed. Units in IFN-γ production are expressed as counts ΔSFU per million (subtracting the average background for each donor assessment). The fraction number over each peptide line corresponds to the number of donors with a significant ELISpot response (4 times over the average background for the two independent measurements).

Figure 6

NNAlign_MAC is able to predict Infliximab-and Rituximab CD4 T cell epitopes. CD4 T cell epitope sequences identified by Hamze et al. (15) mapped to (A) Infliximab and (C) Rituximab variable regions of the light chain and heavy chain (orange dotted lines). NNAlign_MAC predicted profiles (Materials and Methods, profile generation) and SCC correlation to MAPPs are displayed in red. Scatter plot of the NNAlign_MAC prediction profiles vs MAPPs for (B) Infliximab and (D) Rituximab T cell responses from Hamze et al. (15).