Machine Learning-Driven Antigen Selection Reveals Conserved T-Cell Targets for Broad Coronavirus Vaccination

This study demonstrates that machine learning-guided selection of conserved non-spike T-cell targets enables the development of multiepitope mRNA vaccines capable of eliciting robust, broad-spectrum T-cell immunity against diverse coronaviruses, offering a scalable alternative to spike-focused approaches.

The Gist

The Big Picture: Why We Need a New Kind of Coronavirus Shield

Imagine the coronavirus (and its cousins) as a master of disguise. Currently, most vaccines act like a "Wanted Poster" that shows a picture of the criminal's face (the Spike protein). This works great when the criminal wears that specific face. But, the virus is a shape-shifter; it changes its face (mutations) to escape the police. When the face changes, the old "Wanted Poster" becomes useless.

This paper proposes a different strategy: Instead of chasing the face, let's catch the criminal by their fingerprints.

The "fingerprints" in this story are the virus's internal parts (non-spike proteins). These parts are like the engine of a car or the wiring inside a house. They are essential for the virus to survive, so it can't change them easily without breaking itself. These parts are conserved—they look almost the same across different versions of the virus and even across different species of coronaviruses.

The Problem with Old Maps

Scientists have been trying to find these "fingerprints" for a while, but it's like trying to find a needle in a haystack by guessing. There are millions of possible "needles" (peptides) in the virus's genetic code. Checking them one by one in a lab is slow, expensive, and often misses the best targets.

The Solution: The "AI Detective"

The researchers in this paper built a super-smart AI Detective (called the NEC Immune Profiler).

1. The Search: The AI scanned the entire genetic library of coronaviruses (like reading every book in a massive library).
2. The Filter: It looked for two things:
   • Conservation: Is this part of the virus the same in SARS-CoV-1, SARS-CoV-2, MERS, and bat viruses? (The "Fingerprint" must be unique and unchanging).
   • Visibility: If we put this piece in a human body, will the immune system's "police officers" (T-cells) actually see it and recognize it as a threat?
3. The Shortlist: The AI picked the best candidates—parts of the virus that are both unchanging and highly visible to the immune system.

The Experiment: Testing the Theory

The team didn't just trust the computer; they put their theory to the test in three ways:

1. The "Gym Workout" (Human Lab Test)
They took blood samples from healthy people who had been vaccinated or infected before. They gave these blood cells a "workout" by exposing them to the AI-selected virus pieces.

• The Result: The immune cells woke up! They recognized these "fingerprints" immediately. The more "universal" the piece was (found in many different virus types), the stronger the immune response. It was like showing a suspect's fingerprint to a detective, and the detective immediately said, "I know this guy!"

2. The "Factory Test" (Protein Production)
They built a new type of vaccine using mRNA (the same technology used in Pfizer and Moderna shots). Instead of coding for the Spike protein, these mRNA instructions coded for a long string of the AI-selected "fingerprints."

• The Result: When they put this mRNA into cells, the cells successfully built the protein string and displayed the "fingerprints" on their surface, exactly as the AI predicted. The factory worked.

3. The "Field Test" (Mouse Vaccination)
Finally, they vaccinated mice with these new mRNA vaccines.

• The Result: The mice developed a strong army of T-cells. These T-cells were ready to fight. In fact, the immune response was just as strong (and in some cases stronger) than the response to the standard Spike-based vaccines.

The Analogy: The "Universal Key"

Think of the current Spike-based vaccines as keys that only open one specific door. If the lock changes (a new variant), the key doesn't fit.

This new approach creates a Master Key (or a "Skeleton Key"). Because the internal parts of the virus (the "fingerprints") are so similar across almost all coronaviruses, this Master Key can unlock the immune system's defense against:

• The current virus.
• Future variants of the current virus.
• Completely different coronaviruses that might jump from animals to humans in the future (zoonotic spillover).

Why This Matters

• Future-Proofing: We are preparing for the next pandemic before it happens. If a new "Bat Virus" jumps to humans, this vaccine might already work because it targets the parts of the virus that haven't changed.
• Harder to Escape: It is much harder for a virus to mutate its internal "engine" than its outer "face." If it changes the engine, it stops working.
• Broader Protection: It doesn't just stop you from getting sick; it trains your immune system to recognize the virus family as a whole, potentially preventing severe disease even if you get infected.

The Bottom Line

This study proves that we can use AI to find the "Achilles' heels" of coronaviruses—parts that the virus cannot change. By building vaccines that target these unchanging parts, we can create a shield that is much harder for the virus to break, offering protection that lasts longer and covers a wider range of threats than our current vaccines.

Technical Summary

1. Problem Statement

Current coronavirus vaccines primarily rely on the Spike (S) protein to induce neutralizing antibodies. While effective against matched strains, this approach is vulnerable to antigenic drift and lineage-specific escape mutations, necessitating frequent updates. Furthermore, Spike-focused immunity offers limited protection against future zoonotic spillovers from diverse betacoronavirus (β-CoV) lineages.
There is a critical need for variant-tolerant vaccine strategies that target conserved internal viral proteins (non-structural regions). These regions are under strong functional constraint, evolve more slowly, and are less exposed to immune selection pressure. However, systematically identifying conserved T-cell epitopes across the entire β-CoV proteome and validating their immunogenicity in a unified workflow (combining computational prediction, human ex vivo screening, and in vivo validation) has remained limited.

2. Methodology

The study employed an integrated workflow combining artificial intelligence (AI) with multi-modal experimental validation:

• AI-Driven Antigen Prioritization:

  • Data Curation: Compiled and curated a non-redundant reference proteome of ~4,000 β-CoV sequences from BV-BRC, NCBI, and UniProt.
  • Prediction Model: Utilized the NEC Immune Profiler (NIP), an AI platform modeling antigen processing, proteolytic cleavage, and HLA class I binding.
  • Scoring: Generated Antigen Presentation (AP) scores for all 9-mer and 10-mer peptides across 156 common human HLA class I alleles.
  • Selection Strategy: Identified "immunogenic hotspots" (regions with high AP scores) and clustered them by sequence similarity. Within these clusters, regions with high evolutionary conservation across distinct β-CoV taxa were selected.
  • Optimization: A multi-objective algorithm selected epitope sets balancing high AP scores, broad taxonomic conservation, and wide HLA coverage.

• Construct Design:

  • Designed three mRNA constructs (NEC-T4, NEC-T5, NEC-T6) encoding concatenated T-cell epitopes predominantly from non-structural proteins.
  • Architecture: Included an N-terminal signal sequence and a DC-LAMP targeting domain to direct antigens to the endosomal/lysosomal compartment, enhancing MHC presentation.
  • Formulation: mRNA was synthesized using CleanCap technology and formulated into Lipid Nanoparticles (LNPs).

• Experimental Validation:

  • In Vitro (Human):
    • Ex Vivo Vaccination: PBMCs from healthy donors (pre-vaccinated and infected with SARS-CoV-2) were stimulated with peptide pools or mRNA-loaded autologous APCs.
    • Expansion: Cells underwent a 35-day expansion protocol with IL-2/IL-7 and restimulations.
    • Readouts: Flow cytometry for Activation-Induced Markers (AIM: CD137, CD40L) and cytokines (IFN-γ, TNF) to assess reactivity and polyfunctionality.
    • Immunopeptidomics: Transfected Expi293F cells (expressing HLA-A02:01, A03:01, B*07:02) were analyzed via LC-MS/MS to confirm intracellular expression, processing, and HLA class I presentation of the encoded peptides.
  • In Vivo (Murine):
    • Model: HLA-A2.1 transgenic mice (CB6F1) immunized with mRNA-LNPs (NEC-T4, NEC-T5) vs. controls (Comirnaty, Spike-LNP, GFP-LNP).
    • Readout: Triple-color FluoroSpot assay (IFN-γ, IL-2, TNF) on splenocytes to measure antigen-specific T-cell responses.

3. Key Contributions

• Validation of Conservation as a Predictor: Demonstrated a strong positive correlation between evolutionary conservation across β-CoV taxa and functional T-cell immunogenicity in humans.
• Integrated Workflow: Established a scalable pipeline linking AI-based antigen prioritization directly to functional human immune screening and in vivo vaccine testing.
• Non-Structural Targeting: Successfully identified and validated highly immunogenic T-cell targets within non-structural viral proteins (e.g., Orf proteins), which are typically less variable than Spike.
• mRNA Platform Feasibility: Proved that multiepitope mRNA constructs encoding conserved regions can be efficiently expressed, processed, and presented to induce robust polyfunctional T-cell responses.

4. Key Results

• Conservation vs. Immunogenicity:
  • Peptides conserved across >5 distinct CoV taxa elicited significantly stronger T-cell responses (Reactvity Scores) than those restricted to fewer taxa (Median relative reactivity: 2.6 vs. 0.0001 for CTLs; p = 0.0071).
  • 72% of broadly conserved peptides elicited CTL responses compared to 30% of less conserved peptides.
  • 81% of the most immunogenic broad-conservation sequences mapped to non-structural regions.
• Ex Vivo Human Response:
  • The ex vivo vaccination protocol successfully expanded antigen-specific CD4+ and CD8+ T cells from all tested donors.
  • Polyfunctionality: Stimulated T cells exhibited high levels of polyfunctionality (co-expression of CD137, CD40L, IFN-γ, TNF).
  • mRNA Constructs: NEC-T4 and NEC-T5 induced dose-dependent T-cell expansion. While CTL responses were variable, Th responses were robust and highly polyfunctional.
• In Vivo Murine Response:
  • Immunization with NEC-T4 and NEC-T5 induced robust, antigen-specific T-cell responses in HLA-A2.1 mice.
  • Magnitude: Responses (measured in Spot Forming Units) were comparable to or exceeded those of the licensed Comirnaty vaccine and full-length Spike-LNP controls.
  • Quality: Induced strong triple-positive (IFN-γ+IL-2+TNF+) T-cell populations, indicating high-quality, durable immunity.
• Immunopeptidomics: Confirmed that the mRNA constructs were translated, processed, and presented on HLA class I molecules, validating the computational predictions.

5. Significance and Implications

• Pandemic Preparedness: This study provides a blueprint for developing "variant-agnostic" vaccines. By targeting conserved internal proteins, future vaccines could offer broader protection against emerging β-CoV lineages and potential zoonotic spillovers, reducing the need for constant reformulation.
• Complementary Strategy: While Spike-based vaccines are crucial for preventing severe disease via neutralizing antibodies, T-cell-focused vaccines targeting conserved regions provide a complementary layer of defense, potentially preventing infection and transmission even when antibody levels wane or variants escape.
• Scalable Framework: The integration of machine learning with functional validation offers a rapid, rational design approach applicable not only to coronaviruses but to other rapidly evolving viral families.
• Clinical Potential: The demonstration that conserved epitopes can drive robust, polyfunctional T-cell immunity in both human ex vivo systems and transgenic mice supports the progression of these candidates toward clinical efficacy and protection studies.

Limitations Noted: The study focused on acute immunogenicity rather than long-term durability or protective efficacy against viral challenge. The in vivo model was limited to HLA-A2.1 transgenic mice, which does not capture the full diversity of human HLA backgrounds. Future work requires heterologous challenge studies and evaluation in broader human cohorts.