Immunopeptidomics-guided identification of functional neoantigens in non-small cell lung cancer

This study demonstrates that an immunopeptidomics-guided approach, which integrates whole exome sequencing, transcriptomics, and mass spectrometry to filter neoantigens based on observed HLA peptide presentation, significantly improves the success rate of identifying functional neoantigens for personalized cancer vaccines in non-small cell lung cancer patients.

The Gist

Imagine your body is a bustling city, and your immune system is the police force. Usually, this police force patrols the streets looking for criminals (cancer cells) and arrests them. But cancer cells are clever masterminds; they wear disguises to look like normal citizens, so the police ignore them.

The Goal: Finding the "Disguise"
Scientists want to create a personalized "Wanted Poster" for each cancer patient. These posters highlight a unique mutation (a "neoantigen") that the cancer cell is wearing. If the police (T-cells) see this specific disguise, they can attack the cancer.

The problem? There are millions of possible disguises a cancer cell could wear, but the police only recognize a tiny handful. Traditional methods are like throwing darts blindfolded at a massive board of potential disguises. They usually miss the mark, succeeding only about 6% of the time.

The New Strategy: The "Criminal Record" Approach
This paper describes a smarter way to find the right disguise. Instead of guessing, the researchers looked at the "criminal records" of the cancer cells in 24 lung cancer patients.

Here is how they did it, using a simple analogy:

1. The "Wanted" List (Genomics)

First, they took a snapshot of the cancer's DNA (the "Wanted" list). This gave them a list of thousands of potential mutations. However, just because a mutation exists in the DNA doesn't mean the cancer cell actually displays it on its surface for the police to see.

2. The "Surveillance Camera" (Immunopeptidomics)

This is the game-changer. The researchers used a high-tech microscope (Mass Spectrometry) to look directly at the surface of the cancer cells. Think of this as checking the surveillance footage to see exactly what the criminals are actually wearing right now.

They found that in most cases, the cancer cells were wearing very few of the "disguises" predicted by the DNA list. In fact, for one patient, they found one specific mutation on the surface that no one had predicted before.

3. The "Local Police" Insight (The Cohort Study)

The researchers realized that every patient's "police force" (their HLA system) has different preferences. Some police officers only arrest people wearing red hats; others only look for blue shoes.

• Old Way: Guessing what the police might like based on general rules.
• New Way: Looking at the specific "arrest records" of the patient's own immune system. They asked: "What kind of peptides (disguises) does this specific patient's immune system usually catch?"

4. The Result: A Much Better Success Rate

By combining the DNA list with the "surveillance footage" (what is actually being displayed) and the "arrest records" (what the patient's immune system likes), they created a highly targeted filter.

• The Old Method: Success rate of ~6%.
• The New Method: They tested this on 6 patients and found strong immune responses in 5 of them (83% success rate).

They didn't just find the one mutation they saw directly; they used the pattern of what was visible to predict other hidden mutations that the immune system would recognize.

The "Bayesian" Detective Work

To make sure their method wasn't just luck, they used a statistical tool called "Bayesian modeling." Imagine a super-detective who looks at all the cases in the city and says:

• "In this neighborhood (Lung Cancer Type A), the police usually catch criminals wearing long coats."
• "But in this specific house (Patient A147), the police seem to prefer short jackets."

This model helped them understand that every patient has a unique "fingerprint" of what their immune system is good at spotting.

Why This Matters

Lung cancer is a tough opponent, and current treatments often fail. This study shows that by listening to the specific "language" of a patient's immune system and looking at what the cancer is actually showing, we can design vaccines that are much more likely to work.

In short: Instead of guessing which key fits the lock, the researchers looked at the lock itself to see which keys had already been used, then made a perfect copy of that key. This turned a 6% chance of success into an 83% chance.

Technical Summary

1. Problem Statement

Non-small cell lung cancer (NSCLC) remains a leading cause of cancer mortality, with limited long-term survival even with modern checkpoint inhibitor therapies. Personalized neoantigen vaccines represent a promising precision medicine strategy, yet their clinical translation is hindered by the difficulty of identifying therapeutically relevant neoantigens.

• The Bottleneck: Current in silico prediction methods rely on genomic and transcriptomic data to predict which mutations will result in HLA-presented peptides. However, these methods suffer from low success rates (approximately 6% of candidates elicit T-cell responses) because they cannot account for the complex biological filters of antigen processing and presentation.
• The Gap: Direct observation of neoantigens via mass spectrometry (immunopeptidomics) is rare (estimated at ~0.5% of missense mutations) due to sensitivity limits and the vast complexity of the immunopeptidome. Consequently, researchers lack a robust framework to filter thousands of predicted candidates down to the few that are biologically presented and immunogenic.

2. Methodology

The authors developed a proteogenomics workflow integrating multi-omics data with Bayesian statistical modeling to improve neoantigen selection for 24 NSCLC patients (15 Adenocarcinoma [LUAD], 9 Squamous Cell Carcinoma [LUSC]).

A. Data Generation & Integration:

• Genomics/Transcriptomics: Whole Exome Sequencing (WES) and RNA-seq were performed on tumor and matched normal tissues to identify somatic missense variants and gene expression levels.
• Immunopeptidomics: Mass spectrometry (MS) was used to directly identify HLA Class I and Class II peptides eluted from tumor tissues.
• Workflow:
  1. Prediction: Used pVACseq to generate a list of putative neoantigens based on WES/RNA-seq data and patient HLA allotypes.
  2. Filtering (The Core Innovation): Instead of relying solely on binding affinity scores, the authors filtered predictions based on empirical evidence:
     • Did the source protein yield peptides in the patient's specific immunopeptidome?
     • Did the peptide length match the patient's HLA allotype preferences?
     • Was there auxiliary biological support (e.g., protein localization in specific cellular compartments, COSMIC mutation data, Human Protein Atlas relevance)?
  3. Selection: Candidates were manually curated to select 9–14 peptides per patient for functional testing.

B. Functional Validation:

• Patient-matched Peripheral Blood Mononuclear Cells (PBMCs) were stimulated with synthetic mutant and wild-type peptides.
• Immune responses were measured using IFN-γ ELISPOT assays to detect T-cell reactivity.

C. Statistical Modeling:

• The authors employed Bayesian multilevel regression models to retrospectively analyze the immunopeptidome data.
• The model quantified the relative contributions of donor-specific factors (individual HLA presentation propensity) versus protein-specific factors (abundance, cellular compartment) in determining peptide presentation.
• This modeling validated the rules-based selection strategy and provided a statistical framework for understanding presentation probabilities.

3. Key Results

A. Cohort Characterization:

• The cohort exhibited high Tumor Mutational Burden (TMB), with a predominance of C>A transversions (smoking signature).
• Peptidome Similarity: The study found that tumor immunopeptidomes largely sample the same source proteins as healthy lung tissue (92% overlap for HLA-I), suggesting that neoantigen discovery must rely on specific mutations rather than unique protein expression.
• Compartmentalization: HLA-I peptides were preferentially derived from nuclear/cytosolic proteins, while HLA-II peptides came from membrane/extracellular proteins. This correlation was used to filter candidates (e.g., prioritizing nuclear proteins for HLA-I predictions).

B. Neoantigen Identification Success:

• Direct Observation: Only one neoantigen (ALYREF/ALYREF-D10Y) was directly observed via MS in the entire cohort of 24 patients, confirming the rarity of direct detection.
• Functional Validation: The proteogenomics-guided approach was tested on 6 patients (3 LUAD, 3 LUSC).
  • Success Rate: 5 out of 6 patients (83%) showed strong functional T-cell responses to at least one neoantigen.
  • Candidate Efficiency: 9 out of 70 tested peptides (13%) elicited strong responses. This is a two-fold improvement over the standard ~6% success rate of purely in silico prediction methods.
  • Specific Findings: Responses were observed for both HLA-I and HLA-II candidates in LUAD, but only HLA-II in LUSC. One patient (A116) showed no response, which the Bayesian model retrospectively predicted as having a low general propensity for peptide presentation.

C. Bayesian Modeling Insights:

• Donor Specificity: Donor-specific effects were the dominant variance component in peptide presentation, particularly for LUSC (8x and 22x greater than protein effects for HLA-I and II, respectively).
• Protein Precedence: If a donor presented any peptide from a specific protein, the probability of presenting a neoantigen from that same protein increased by 23–41%. Conversely, if a protein was widely presented across the cohort, the probability of an additional peptide being presented decreased, highlighting the need for patient-specific filtering.

4. Key Contributions

1. Proof-of-Concept for Immunopeptidomics-Guided Selection: Demonstrated that using patient-specific immunopeptidome data as a filter significantly outperforms standard genomic prediction algorithms for neoantigen selection.
2. Rules-Based Filtering Framework: Established a practical workflow that prioritizes candidates based on:
   • Evidence of source protein processing (presence in the peptidome).
   • Cellular compartment alignment (nuclear vs. extracellular).
   • HLA allotype length preferences.
3. Statistical Validation: Provided a Bayesian framework that quantifies the "presentation propensity" of individual donors, explaining why some patients respond to vaccines while others do not, and validating the empirical rules used for selection.
4. High Success Rate: Achieved an 83% functional response rate in a small clinical cohort, a significant leap from historical benchmarks.

5. Significance

• Clinical Impact: This study addresses the critical "needle in a haystack" problem in personalized cancer vaccines. By shifting from purely predictive algorithms to a proteogenomic validation strategy, the authors have created a more reliable pipeline for identifying neoantigens that actually engage the immune system.
• Precision Medicine: The findings emphasize that neoantigen selection cannot be a "one-size-fits-all" approach. It must account for the unique donor-specific presentation biology (HLA heterozygosity, processing efficiency) and the specific source protein context.
• Future Directions: The study suggests that future vaccine development should incorporate patient-specific immunopeptidomics (even if direct neoantigen detection is rare) to filter candidate lists. It also highlights the potential of Bayesian modeling to refine machine learning algorithms for neoantigen prediction by incorporating donor-specific variance components.

In summary, this paper provides a robust, data-driven methodology that bridges the gap between genomic mutation prediction and biological reality, significantly enhancing the efficacy of neoantigen identification for NSCLC immunotherapy.