AUTHORS
Luis O Correa-Medero, PhD, Marta Wilbrink2, Jessika Baral3, Emma Duggan4, Kasidet Manakongtreecheep, PhD5, Natalie Clark, PhD6, Steven A Carr, PhD7, Derin B. Keskin8, Catherine J. Wu, MD9, Jenn G Abelin, PhD10, Nir Hacohen, PhD11
AFFILIATIONS
1 Broad Institute of MIT & Harvard, Cambridge, Massachusetts, United States
2 Broad Institute of MIT & Harvard, Cambridge, Massachusetts, United States
3 Broad Institute of MIT & Harvard, Cambridge, Massachusetts, United States
4 Broad Institute of MIT & Harvard, Cambridge, Massachusetts, United States
5 Broad Institute of MIT & Harvard, Cambridge, Massachusetts, United States
6 Broad Institute of MIT & Harvard, Cambridge, Massachusetts, United States
7 Broad Institute of MIT & Harvard, Cambridge, Massachusetts, United States
8 Dana-Farber Cancer Institute, Boston, Massachusetts, United States
9 Dana-Farber Cancer Institute, Boston, Massachusetts, United States
10 Broad Institute of MIT & Harvard, Cambridge, Massachusetts, United States
11 Broad Institute of MIT & Harvard, Cambridge, Massachusetts, United States
ABSTRACT
New genomic and analytic platforms are increasingly revealing novel classes of antigens that promise to expand the immunologically targetable landscape of hematologic malignancies, which have been conventionally characterized as lower mutation burden and hence have only a limited repertoire of canonical neoantigens. Such antigens include those arising from the 'dark proteome', including those derived from noncanonical open reading frames (ncORFs) and endogenous retroviruses (ERVs), and these offer a compelling alternative pool of broadly shared, cancer-specific targets.
The potentially vast number of candidate ncORF antigens creates a prioritization bottleneck. Current affinity-based predictors fail to implicitly encode peptide-human leukocyte antigen (pHLA) stability, a superior and well-acknowledged determinant of CD8? T cell immunogenicity. We hypothesized that a scalable, user-defined stability assay combined with a pan-allelic predictive model could systematically define the structural determinants of stability, robustly predict in vivo immunogenicity, and effectively prioritize dark proteome targets for antigen-targeting therapies for blood malignancies. We thus developed HLAglyph, a highly scalable platform for generating user-defined pHLA stability data. HLAglyph couples oligonucleotide-encoded peptide libraries expressed in E. coli (Pepyrus) with single-chain HLA-beta-2-microglobulin constructs and chaperone enzymes to enable cytoplasmic refolding (HLAplex). Following purification, pHLA complexes were subjected to thermal proteome profiling (TPP) across a 30°C to 75°C temperature gradient and quantified using TMT-enabled Mass Spectrometry to derive high-confidence melting temperatures (Tm). To enable in silico modeling, we developed HLAres, an ESM-2–based pan-allelic stability predictor finetuned from a binding-prediction backbone via regression against HLAglyph Tm data. Immunogenicity was evaluated both computationally via curated CEDAR datasets and functionally in vivo via IFN-γ ELISpot in HLA-A*11:01 transgenic mice.
Profiling of sequence-divergent allotypes (A*11:01, A*02:01, B*07:02) revealed that HLAglyph generates robust, highly reproducible Tm data (standard deviation < 3°C across replicates), recovering a substantially broader sequence space than endogenous cell-line TPP. Binding affinity correlated only weakly with measured Tm, demonstrating that stability constitutes non-redundant biophysical information. Our predictive model, HLAres, accurately learned these determinants, achieving Spearman correlations > 0.8 and root-mean-square errors (RMSE) < 2°C on held-out peptides, while successfully generalizing to unseen alleles.
Crucially, both measured and predicted stability successfully stratified CD8? T cell immunogenicity. In transgenic mouse models immunized with an affinity-matched panel of neoantigen-like peptides, T cell responses strictly tracked with pHLA stability. Furthermore, among intermediate-affinity binders (%rank 0.5–2), a range where affinity-based ranking fails to differentiate targets, stability effectively separates immunogenic and non-immunogenic peptides.
Our experimental and machine learning based approaches overcome the limitations of current antigen prioritization pipelines by enabling the scalable measurement and accurate prediction of pHLA stability. By resolving immunogenicity in sequence spaces where binding affinity is uninformative, this framework provides a highly discriminating tool to prioritize vast pools of shared ncORF and ERV targets, accelerating the development of potent, off-the-shelf cancer antigen targeting therapies, such as personalized vaccines for hematologic malignancies.
DISCLOSURES
Luis O Correa-Medero, PhD: No relevant conflicts of interest to declare.