ARTEMIS: A Novel Mass-Spec Platform for HLA-Restricted Self and Disease-Associated Peptide Discovery

Abstract

Conventional immunoprecipitation/mass spectroscopy identification of HLA-restricted peptides remains the purview of specializing laboratories, due to the complexity of the methodology, and requires computational post-analysis to assign peptides to individual alleles when using pan-HLA antibodies. We have addressed these limitations with ARTEMIS: a simple, robust, and flexible platform for peptide discovery across ligandomes, optionally including specific proteins-of-interest, that combines novel, secreted HLA-I discovery reagents spanning multiple alleles, optimized lentiviral transduction, and streamlined affinity-tag purification to improve upon conventional methods. This platform fills a middle ground between existing techniques: sensitive and adaptable, but easy and affordable enough to be widely employed by general laboratories. We used ARTEMIS to catalog allele-specific ligandomes from HEK293 cells for seven classical HLA alleles and compared results across replicates, against computational predictions, and against high-quality conventional datasets. We also applied ARTEMIS to identify potentially useful, novel HLA-restricted peptide targets from oncovirus oncoproteins and tumor-associated antigens.

Keywords: MHC class I; immunotherapy; ligandome analysis; mass spectrometry; peptide-HLA complex.

Figure 1

Using the SCD construct to recover and MS sequence HLA-restricted peptides. (A) Schematic representations highlight the differences between native, cell-surface pHLA and the engineered, secreted SCD construct. (B) Western blot analysis of SCD expression levels across alleles confirmed secretion of all tested SCDs, though at variable levels. Transduction, tissue culture, and Western conditions were closely matched across samples to allow meaningful comparisons of expression levels. The A3 SCD was tested with and without added cytokines (IFNγ, TNFα). Molecular weight markers (orange arrows) are indicated; the red arrow marks the expected PAGE mobility of an SCD with normal, expected N-glycosylation. (C) Chromatography retention times for nested peptides are graphed. The y-axis shows total number of unique peptide sequences in nested peptide sets in a co-elution time range; retention time intervals are indicated along the x-axis. (D) Accumulation rates of unique peptide sequences across five experimental replicates from the IP (green) and SCD (blue) datasets are compared (the pan-HLA IP dataset spans the three alleles in the haplotype of the cells analyzed, A3, B7, C7, but only A3 SCD results are shown). The y-axis shows total number of unique peptide sequences (also indicated in white in the bars) and progressively larger unions of five replicate datasets are indicated along the x-axis.

Figure 2

Intersections between SCD MS replicates. (A) Venn comparison of the unique peptides recovered by the A3 SCD transduced into cells cultured in the presence (blue circle) or absence (green circle) of cytokines to shift expression from conventional to immunoproteasomes. Numbers of peptides in the various intersection domains are indicated and the number of peptides that were culled because source proteins were listed in the MS CRAPome (29) are in parentheses, in gray. 89% of the cytokine-untreated peptides are observed in the cytokine-treated set. (B) Venn comparison of two A3 SCD experimental replicates are shown, labeled as in (A). Overlap is 83%. (C) Venn comparison of two A3 (shades of blue), B7 (shades of green), and C7 (shades of orange) SCD biological replicates are shown, labeled as in (A). Overlaps are 57% (A3), 49% (B7), and 53% (C7). Venn diagrams were generated with BioVenn (33).

Figure 3

Comparing SCD, IP, and sIP MS dataset intersections. (A) Venn comparisons of the intersections of the IP five experimental replicate union dataset (red) and unions of three single-replicate sIP (green) or SCD (blue) single-allele (A3, B7, C7) datasets. Overlap is 45% between the IP and SCD datasets, corresponding to observation of 63% of the IP peptides in the SCD dataset. The overlap is 25% between IP and sIP datasets. (B) Venn comparisons of the intersections of the IP five experimental replicate union dataset (red) and the union of the available multi-replicate SCD (blue) single-allele (A3, B7, C7) datasets. 80% of the IP peptides were observed in the SCD replicate-union dataset. Inset: Peptides in each of the three Venn domains were analyzed for binding with NetMHCpan 4.1; results are plotted as relative percentage “strong binder” (green), “weak binder” (yellow), or “non-binder” (orange), summed over lengths and the three alleles in the analysis. (C) Venn comparisons of the intersections of A3, B7, or C7 sIP/SCD single-run datasets. Overlaps are 28% (A3), 24% (B7), and 27% (C7). Venn diagrams were generated with BioVenn (33).

Figure 4

Agreement between MS datasets and NetMHCpan predictions. (A) Peptides from the pan-HLA IP A3+B7+C7 dataset, or the single-run sIP and SCD A3, B7, and C7 datasets used in prior comparisons, were analyzed by NetMHCpan for binding to these three alleles. Results are plotted as “strong binder” (green), “weak binder” (yellow), or “non-binder” (orange) and are binned by peptide length. (B) Single-run SCD A3, B7, and C7 datasets were analyzed by NetMHCpan and plotted as in (A), comparing results from peptides identified at a 1% FDR cutoff or those added by expanding the cutoff to a 5% FDR. Results in (A, B) exceed 100% because NetMHCpan predicts that more peptides bind to multiple alleles than are observed in the MS datasets. In other words, many peptides were predicted to bind to more alleles in the set of three than were observed by MS, where most peptides were only identified binding to a single allele, contributing to overall binding scores of >100% when summed over all three alleles.

Figure 5

Comparisons of observed peptide length distributions recovered by MS. (A) The length distributions of peptides from the pan-HLA IP A3+B7+C7 dataset and the sIP and SCD union A3/B7/C7 datasets are displayed as pie charts. Peptide length distributions for single-run sIP and SCD analyses are compared, plotted as (B) absolute peptide numbers or (C) percentages. (D) Source protein subcellular localization profiles for the pan-A3/B7/C7 IP and union A3|B7|C7 sIP and SCD peptide datasets are shown as percentages. Compartments were labeled using Gene Ontology (GO) cellular compartment classifications for each peptide’s source protein: Nucleus [GO:0005634]; Mitochondrion [GO:0005739]; ER [GO:0005783]; Golgi [GO:0005794]; Cytosol [GO:0005829]; Membrane [GO:0016021]; or Extracellular [GO:0005576]/[GO:0005615] (36, 37).

Figure 6

Observed peptide length distributions recovered from SCDs by allele. The length distributions of peptides from merged SCD analyses across multiple replicates are displayed as pie charts. SCD allele and peptide length are shown as indicated. Numbers of replicates (“runs”), total peptides observed, and the number of unique peptides in the union datasets are indicated. Average peptide length across the merged datasets is shown in the in the center of each pie.

Figure 7

SCD-derived, allele- and length-specific peptide sequence logos. Allele-specific recognition motifs are presented as sequence logos, either as reference 9-mer logos from “naturally presented ligand” peptides available through the NetMHCpan motif viewer portal (column 1; http://www.cbs.dtu.dk/services/NetMHCpan/logos_ps.php) or as custom logos, generated as part of this work, from SCD-recovered peptides (columns 2 through 4). The x-axis reports position in the peptide, the y-axis reports information content of different residues at that position, in bits. Alleles are specified at left, and the total number of SCD-recovered peptides used to generate that logo is inset (columns 2 through 4). For SCD-recovered peptides, only logos generated from 8-mers, 9-mers, and 14-mers have been selected for display for simplicity. P1 positions that echo the P2 anchor position amino acid preference in 8-mer logos are indicated by red arrows and positions in 14-mer logos where glycine rises in abundance relative to 9-mers are indicated by yellow arrows.

Figure 8

Observed peptide overlaps in the A3/A11 HLA supertype. (A) The A3 biological replicate intersection is shown at top, echoing Figure 3C . Below are shown the 9-mer sequence logos for the residues in the three Venn domains, with numbers of peptides inset. The x-axis reports position in the peptide, the y-axis is scaled to the information content of different residues at that position. Overlap is 57%. (B) Venn analyses of SCD A3 (blue), A11 (purple), and A24 (red) results are shown with the numbers of peptides in each domain indicated. Below are shown the 9-mer sequence logos for the residues in the three SCD Venn domains, with numbers of peptides inset. The overlap between SCD A3 and A11 datasets is 27%. (C) Venn analyses of sIP A3 (blue), A11 (purple), and A24 (red) results are shown with the numbers of peptides in each domain indicated. [A24 results are shown as an orthogonal comparator.] Below are shown the 9-mer sequence logos for the residues in the three sIP Venn domains, with numbers of peptides inset. The overlap between sIP A3 and A11 datasets 16%. Venn diagrams were generated with BioVenn (33).

Figure 9

A3, A11, and A3 versus A11 KL sequence divergence. KL divergence between the amino acid frequency distributions at each position in the peptide alignments is plotted as sequence convergence (y-axis) versus percentage of peptides resampled (x-axis) for the SCD A3 9-mer dataset against itself (A), the SCD A11 9-mer dataset against itself (B), and the SCD A3 9-mer dataset against the SCD A11 9-mer dataset (C). Individual peptide position-by-position convergences are colored as indicated along the right. (D) A structural view highlights sequence differences between A3 and A11 affecting P2 amino acid preference. The molecular surface of A3 (2XPG.pdb; grey) is shown with the backbone of the bound peptide shown as a cartoon ribbon (purple) with the P1 and P2 side-chains shown in a licorice-stick representation.

Figure 10

Mapping ARTEMIS-identified peptides onto the LT sequence. (A) The sequence of the truncated form of MCV LT associated with cancer is shown mapped onto an oval. Residues in ARTEMIS-identified peptides are colored as indicated; phosphorylated threonine residues are marked with purple arrows. LT peptides predicted by NetMHCpan to bind to A2, A11, or A24 are shown around the outside of the oval, approximately positioned by their location in the sequence. Predicted strong binders are shown in bold, peptides predicted not to bind are shown in italics. Peptides identified by non-MS experimental methods are shown inside the sequence oval but outside the dotted line, colored by allele as indicated. All of these peptides are predicted to bind strongly to their cognate alleles. ARTEMIS-identified peptides are shown inside the dotted line, colored by cognate allele as indicated, with predicted binding/non-binding behavior shown in bold or italics. Phosphorylated residues in ARTEMIS-identified peptides are indicated by asterisks; threonines phosphorylated in LT in ARTEMIS-identified peptides are highlighted in yellow. Green dashed lines connect peptides identified by more than one method. (B) The crystal structure of A11 PDB accession code 3RL2 (51) is shown as a molecular surface, colored gray, with the backbone of the bound peptide shown in a cartoon representation, colored from blue to red, N- to C-terminus. The α-carbons of the first four residues in the peptide (P1 through P4) are marked with spheres. The eye symbol indicates the viewpoint shown in (C). (C) A view down the peptide binding groove of A11 from the perspective indicated in (B). Phosphothreonine residues have been modeled into the P2 and P4 positions, and a phosphoserine has been modeled into the P3 position. Phosphate groups have been shown in a ball-and-stick representation, colored by atom type.

Figure 11

Mapping ARTEMIS-identified peptides onto the MSLN sequence. The sequence of MSLN is shown mapped onto an oval. The leader peptide is colored gray, the furin cleavage site is colored red, the 4x potential N-glycan sites are colored orange, and the GPI signal sequence is colored purple. The red dashed line separates the MSLN proprotein sequence into MPF and MSLN moieties. ARTEMIS-identified peptides are shown inside the oval, approximately positioned by their location in the sequence, colored by cognate allele as indicated, with predicted binding/non-binding behavior by NetMHCpan shown in bold or italics as indicated. 57% of the MPF/MSLN sequence is presented as A2, A11, or A24 peptides identifiable via ARTEMIS. The MPF moiety of the MSLN fusion precursor proteins is highly expressed as a secreted protein via transduction in HEK293 cells, ten-fold or more higher in culture supernatants than SCDs, so may contaminate the isolated SCDs. However, observed coverage was comparable over the two moieties, including peptides from the leader sequence (which are not derived from a secreted contaminant but from endogenously expressed MPF/MSLN), suggesting that contamination was not a huge issue. 115 A2, A11, and A24 peptides were predicted to bind, using NetMHCpan, that were not observed by ARTEMIS. These peptides have not been shown to avoid excessive clutter.