Utilizing Computational Machine Learning Tools to Understand Immunogenic Breadth in the Context of a CD8 T-Cell Mediated HIV Response

Abstract

Predictive models are becoming more and more commonplace as tools for candidate antigen discovery to meet the challenges of enabling epitope mapping of cohorts with diverse HLA properties. Here we build on the concept of using two key parameters, diversity metric of the HLA profile of individuals within a population and consideration of sequence diversity in the context of an individual's CD8 T-cell immune repertoire to assess the HIV proteome for defined regions of immunogenicity. Using this approach, analysis of HLA adaptation and functional immunogenicity data enabled the identification of regions within the proteome that offer significant conservation, HLA recognition within a population, low prevalence of HLA adaptation and demonstrated immunogenicity. We believe this unique and novel approach to vaccine design as a supplement to vitro functional assays, offers a bespoke pipeline for expedited and rational CD8 T-cell vaccine design for HIV and potentially other pathogens with the potential for both global and local coverage.

Keywords: CD8 T-cells; HIV; T-cell epitopes; machine learning; vaccines.

Figure 1

Frequency of each HLA Class I allele (HLA-A, HLA-B, and HLA-C) represented within IAVI Protocol C Alleles. Red boxes demarcate the allele frequencies contained within 13 pre-selected volunteers (Table 2) with percentage coverage listed above each stacked histogram plot. Seventeen Individual alleles contribute to HLA-A analysis, 21 Individual alleles contribute to HLA-B analysis, and 13 Individual alleles contribute to HLA-C analysis.

Figure 2

Two-dimensional representation of HLA diversity using Principal Component Analysis (PCoA). A HIV-1 Gag binding profile was predicted for every HLA allele using NetMHCpan and a set of transmitted founder sequences. The binding profile of each volunteer (red dot) was defined by taking the union of predicted binding for each of their HLA alleles. PCoA was performed using the pairwise similarity matrix of all volunteers, revealing distinct clusters of individuals. A subgroup of 13 volunteers were chosen to provide optimal coverage of the HLA binding profiles (blue dots).

Figure 3

Coverage per predicted peptide calculated against a defined set of HLA alleles. Size of segments on X axis from left to right represents cumulative, combined HLA allele frequencies that are iteratively removed from the analysis, starting with least frequent alleles. Blue line—modeling using predicted 300 peptides. Red line—modeling using predicted 250 peptides. Green line—modeling using predicted 200 peptides. Purple line—modeling using predicted 150 peptides.

Figure 4

Affinity plots for all predicted peptides with conservation of ≥2.2% (n = 14,953). (A)—Predicted peptide affinity (Rank Binding) vs. primary associated HLA. (B)—Predicted peptide affinity (Rank Binding) vs. peptide frequency within transmitted founder proteome. (C)—Predicted peptide frequency vs. primary associated HLA.

Figure 5

Cumulative coverage distribution plots of full length transmitted founder gag sequences using a 3-select coverage model and a 1% Binding Threshold, 3-Select best (red), and 3-Select random (blue). P-values calculated using Kolmogorov-Smirnov test.

Figure 6

IFNγ ELISpot responses observed in HIV+ Volunteers. (A)—Number of total ELISpot responses observed in volunteers whose transmitted founder proteome sequence was included within the in-silico prediction (Seq In: N = 10) and volunteers whose transmitted founder proteome sequence was not included within the in-silico prediction (Seq Out: N = 13). Shapiro-Wilk values Seq In: W = 0.7887, p = 0.0008. Seq Out: W = 0.8976, p = 0.0315. Mann-Whitney test, p = 0.6215. (B)—Correlation of total number of ELISpot responses in volunteers whose transmitted founder proteome sequence was included within the in-silico prediction against the order of priority the sequence was predicted to occur (Spearman Correlation; r = 0.1356, p = 0.2209). (C)—Correlation of total number of ELISpot responses in volunteers whose transmitted founder proteome sequence was included within in silico prediction against the % coverage each epitope represented (Spearman Correlation; r = 0.2695, p = 0.0357).