An Integrated Computational Antigen Discovery Pipeline with Hierarchical Filtering for Emerging Viral Variants

This paper presents an integrated computational pipeline utilizing hierarchical filtering and machine learning to rapidly identify and optimize conserved antigen candidates for emerging viral threats, as demonstrated by its successful application to Rift Valley fever, Mayaro, and SARS-CoV-2 viruses.

The Gist

Imagine the world of viruses as a massive, chaotic library where every book is a slightly different version of a dangerous story. Some stories are old classics (like the original SARS-CoV-2), while others are constantly being rewritten with new chapters and typos (the new variants).

Your immune system is like a security guard trying to stop these stories from entering the building. To do its job, the guard needs to recognize specific "keywords" or "clues" hidden inside the virus books. These clues are called epitopes. If the guard can spot the right keyword, it can sound the alarm and destroy the virus.

The problem is that there are millions of possible keywords in a single virus book, and the virus keeps changing its spelling. Finding the right keyword to build a vaccine is like trying to find a single specific needle in a haystack that is on fire and constantly rearranging itself.

This paper introduces a super-smart, automated detective team (a computational pipeline) designed to find those needles faster than any human could. Here is how they do it, broken down into simple steps:

1. The Great Data Dump (Data Collection)

First, the team gathers every single copy of the virus story they can find from the library (GenBank). They don't just look at one version; they look at hundreds of variations to understand how the virus changes its spelling. This ensures they aren't just designing a vaccine for one specific virus, but for the whole family of viruses.

2. The "Needle" Hunt (Epitope Prediction)

The team uses a fleet of different AI tools (like BepiPred and AlphaFold) to scan the virus books.

• The Analogy: Imagine you have 10 different detectives looking at the same crime scene. One detective uses a magnifying glass, another uses a metal detector, and a third uses a thermal camera.
• The Strategy: Instead of trusting just one detective, the pipeline uses a consensus rule. If only one detective says, "This is the clue!" the team ignores it. But if three detectives all point to the same spot, the team marks it as a "High Confidence Candidate." This filters out the false alarms.

3. The "Do Not Touch" Zones (Filtering)

Even with a shortlist, there are still too many candidates. The team applies strict filters:

• The "Buried Treasure" Filter: Some clues are hidden deep inside the virus, where the immune system guard can't reach them. The team uses a tool called "Solvent Accessibility" to throw away any clues that are buried. They only keep the ones exposed on the surface, like a flag waving in the wind.
• The "Sugar Coating" Filter: Viruses often cover themselves in sticky sugar molecules (glycans) to hide. The team identifies these sugar spots and removes them from the list because the immune system can't grab onto them easily.

4. The "Mutation Gym" (Optimization)

This is where the pipeline gets really clever. Sometimes, the best clue isn't perfect. The team uses a powerful AI model (called ESM) to act like a gym trainer for the virus.

• The Analogy: Imagine you have a key that almost fits a lock, but it's a little stiff. The AI tries thousands of tiny tweaks to the key's shape to see if it fits better.
• The Goal: They tweak the viral clues to make them:
  1. Stronger: So the immune system recognizes them instantly.
  2. Safer: So they don't accidentally trigger an allergy or poison the body.
  3. Stable: So they don't fall apart before they can do their job.

5. The Final Showdown (Testing on Real Viruses)

The team tested this detective system on three different viruses:

• SARS-CoV-2 (The Coronavirus): They successfully found the "keywords" that the most powerful neutralizing antibodies use to stop the virus. They even found clues that work across all the different variants (Alpha, Delta, Omicron, etc.), proving their method finds the "universal" parts of the virus that don't change.
• RVFV & MAYV (The Unknowns): They applied the same logic to Rift Valley Fever and Mayaro viruses. Even without as much prior data, the pipeline narrowed down millions of possibilities to a tiny, manageable list of high-quality candidates that look very promising for future vaccines.

Why This Matters

Traditionally, finding these vaccine ingredients is like searching for a needle in a haystack by hand. It takes years and costs millions of dollars.

This paper presents a robotic vacuum cleaner that can suck up the whole haystack, sort the hay from the needles, and hand you the best needles in minutes. By using this "hierarchical filtering" (filtering in layers, getting stricter each time), they can rapidly design vaccines that are ready for the next pandemic before it even starts.

In short: They built a digital factory that takes a virus, strips away the noise, finds the best targets, and tweaks them to be the perfect vaccine ingredients, all before a human scientist even has their morning coffee.

Technical Summary

1. Problem Statement

Emerging and evolving viral pathogens (e.g., SARS-CoV-2, Rift Valley fever virus, Mayaro virus) pose significant global health challenges. Traditional vaccine and therapeutic development is often slow, expensive, and labor-intensive. While computational methods have advanced in protein structure prediction and machine learning (ML), existing approaches for antigen discovery suffer from several limitations:

• Narrow Scope: Many tools focus on isolated tasks (e.g., single epitope prediction) rather than end-to-end pipelines.
• Manual Intervention: Current multi-epitope vaccine designs often require heavy manual curation and reliance on specific docking tools.
• Lack of Variant Awareness: Many pipelines fail to account for viral mutations and sequence variability, reducing their efficacy against emerging strains.
• Computational Cost: Structure-based docking and iterative manual filtering are computationally taxing and slow.

There is a critical need for an automated, scalable, and mutation-aware computational pipeline that can rapidly narrow the antigen search space to identify conserved, neutralizing, and safe epitope candidates.

2. Methodology

The authors propose a three-stage integrated computational pipeline designed to process viral sequences, predict epitopes, and screen candidates through hierarchical filtering.

Stage 1: Data Collection and Preprocessing

• Input: Viral strain sequences and literature-derived structural data (PDB).
• Strategy: The pipeline gathers comprehensive sequence information to account for genetic variability within viral families. It integrates sequence searches with literature reviews to ensure robustness against sequence variations.

Stage 2: Epitope Analysis and Consensus Filtering

• Prediction Tools:
  • Linear Epitopes: Predicted using sequence-based tools (BepiPred3, BepiBlast, BepiPred2).
  • Non-linear Epitopes: Protein structures are generated using AlphaFold2 (AF2) and AlphaFold3 (AF3) if experimental structures are unavailable. These structures are analyzed using DiscoTope2 and DiscoTope3.
• Consensus Strategy: To reduce false positives, the pipeline employs a consensus approach. Epitopes are retained only if supported by multiple tools:
  • Comb: Supported by at least one tool.
  • Cons2: Supported by at least two tools.
  • Cons3: Supported by at least three tools.
• Physicochemical Filtering:
  • Solvent Accessibility (SAS): Tools FreeSASA and NetSurfP-3 are used to filter for exposed residues (threshold: 20 Ų). Buried residues are discarded.
  • Glycosylation: Sites identified by NetNGlyc-1.0 are excluded to avoid steric hindrance from glycan molecules.
  • Literature Integration: Known epitopes from literature are merged into the candidate pool.

Stage 3: Mutation Screening and Property Optimization

• Conservancy Analysis: The IEDB conservancy tool evaluates epitope stability across variant sequences, calculating identity percentages and match rates.
• Mutation Design (ESM): The Evolutionary Scale Model (ESM), a masked language model, is used to predict the probability ratio of mutated residues versus wild-type residues.
  • Objective: Introduce mutations to enhance antigenicity (using VaxiJen 3.0) while maintaining safety (non-toxic via ToxinPred 3.0, non-allergenic via AlgPred 2) and proper folding (via ESM).
  • Thresholding: Mutations are screened based on ESM score percentiles (85th, 90th, 95th, 99th) to select high-quality candidates.

3. Key Contributions

• Integrated Pipeline: The first work to combine diverse computational tools (structure prediction, ML-based epitope prediction, and evolutionary models) into a unified, automated antigen discovery workflow.
• Hierarchical Filtering: Introduction of a multi-step consensus and physicochemical filtering strategy (SAS, glycosylation, consensus levels) that drastically reduces the search space without losing high-confidence candidates.
• Mutation-Aware Design: Utilization of ESM to guide mutation strategies that specifically aim to boost immunogenicity and safety profiles, addressing the gap in variant-specific antigen design.
• Open-Source Framework: The pipeline is designed to be extensible, allowing for the continuous addition of new tools and data repositories for different viral families.

4. Results

The pipeline was validated on three distinct viral families: SARS-CoV-2 (Coronavirus), Rift Valley Fever Virus (RVFV) (Phlebovirus), and Mayaro Virus (MAYV) (Alphavirus).

• SARS-CoV-2 Validation:

  • Search Space Reduction: The initial pool of 1,169 predicted residues was reduced to 127 (Cons3 + SAS + Glyco filtering). Further restriction to the Receptor-Binding Domain (RBD) and conservancy analysis narrowed this to 41 residues.
  • Ground Truth Validation: The pipeline successfully identified 16 validated epitopes from literature, with 11 located in conserved regions.
  • Neutralizing Antibody Overlap: The pipeline identified epitopes overlapping with the binding interface of the broadly neutralizing antibody BA7535 (PDB: 8H7Z), which is effective against multiple variants (Alpha, Delta, Omicron, etc.). 6 of 10 residues in the BA7535 interface were found in the candidate pool, with 4 in conserved regions.
  • Screening Efficiency: At the 90th percentile ESM threshold, the number of candidates dropped from 201 to just 10, demonstrating high filtering efficiency.

• RVFV and MAYV Case Studies:

  • Sequence Diversity: DiMA analysis identified conserved regions (e.g., Gn-region for RVFV, E2-glycoprotein for MAYV) to focus the search.
  • Filtering Impact: The hierarchical filtering reduced candidate pools significantly:
    • MAYV: From 233 initial residues to 117 (Strict suggestion).
    • RVFV: From 804 initial residues to 398 (Strict suggestion).
  • Mutation Insights: For RVFV, a specific epitope at residue index 563 was identified. It neighbors a mutation (index 566) found in the MP-12 vaccine strain. The ESM score for this mutation was in the top 4, validating the use of ESM scores to guide mutation design.

5. Significance and Future Outlook

• Accelerated Development: The pipeline offers a powerful framework to expedite the identification of antigen candidates, potentially shortening the timeline for vaccine and therapeutic development against emerging pathogens.
• Broad Efficacy: By focusing on conserved regions and utilizing mutation screening, the pipeline aims to design antigens with broad efficacy against viral variants, a critical need for future pandemics.
• Limitations: The current study relies on third-party tools that may have limitations in accuracy. The final candidates have not yet been experimentally validated in wet-lab assays (e.g., serum assays).
• Future Directions: The authors suggest integrating ground-truth experimental data to refine scoring metrics, employing multi-objective optimization, and developing "agentic" workflows to dynamically tailor the pipeline to user needs.

In conclusion, this work presents a robust, data-driven foundation for open-source antigen design, successfully demonstrating its ability to filter vast sequence spaces into a manageable set of high-potential, conserved, and safe epitope candidates.