ViralMap: Predicting Features in Viral Proteins from Primary Sequence

ViralMap is a multi-label deep learning model that leverages ESM-2 representations to predict diverse structural and functional features directly from viral protein sequences, providing a scalable tool to accelerate antigen engineering and vaccine design for emerging pathogens.

The Gist

Imagine you've just discovered a new, mysterious alien virus. You have its genetic code (the "blueprint"), but you have no idea what the virus's weapons look like, how they work, or how to build a shield (a vaccine) against them.

In the past, figuring this out was like trying to understand a complex machine by taking it apart piece by piece in a dark room. It took months or years of lab work.

Enter ViralMap. Think of ViralMap as a super-smart, instant translator that can look at a virus's raw genetic blueprint and immediately tell you exactly how its weapons are built, where they are located, and how they function.

Here is a simple breakdown of how it works and why it matters:

1. The Problem: The "Instruction Manual" is Missing

Viruses are like sneaky thieves. To stop them, we need to build vaccines that target their specific "weapons" (proteins). But to design a vaccine, scientists need a detailed map of these weapons.

• The Old Way: Scientists used to run a virus's protein sequence through a dozen different, clunky computer programs. It was like trying to assemble a piece of furniture by using a hammer, a screwdriver, a wrench, and a tape measure, all from different brands that didn't fit together. It was slow, confusing, and often missed details.
• The Viral Virus Problem: Most computer programs were trained on human or animal proteins. Viruses are weird, fast-evolving, and very different from us. So, those old tools often got confused when looking at viral proteins.

2. The Solution: The "Swiss Army Knife" Translator

The authors built ViralMap, a single, all-in-one tool designed specifically for viruses.

• The Brain: ViralMap uses a massive AI brain (called ESM-2) that has "read" millions of protein sequences. It's like a librarian who has read every book in the universe and can instantly recognize patterns.
• The Job: You feed it a raw string of letters (the virus's protein sequence). In return, it spits out a detailed "feature list" for that protein.

3. What Does It Actually Find? (The 10 Features)

ViralMap doesn't just say "this is a protein." It acts like a high-tech detective, highlighting 10 specific things on the protein that are crucial for making a vaccine:

• The Anchors (Topology): It tells you which parts of the protein are stuck in the virus's skin, which stick out into the air (to be attacked by antibodies), and which are hidden inside. Analogy: It's like a map showing which parts of a submarine are underwater and which are above the surface.
• The Glue (Disulfide Bonds): It finds the chemical "staples" that hold the protein's shape together. If you break these, the weapon falls apart.
• The Scissors (Cleavage Sites): It finds where the virus cuts its own proteins to activate them. Analogy: It's like finding the perforated line on a ticket that needs to be torn to work.
• The Camouflage (Glycosylation): It spots the sugar coats the virus wears to hide from our immune system. Analogy: It's like a spy wearing a disguise; ViralMap points out exactly where the disguise is.
• The Flexible Parts (Disordered Regions): It finds the floppy, wiggly parts of the protein that might be hard to target.
• The Fusion Engines (Coiled Coils): It finds the spring-loaded mechanisms viruses use to punch into our cells.

4. Why Is This a Big Deal?

The paper mentions the "100 Days Mission." This is a global goal to be able to create a vaccine for a brand-new, unknown disease (like "Disease X") within 100 days of it being discovered.

• Speed: ViralMap can take a sequence from a new virus and generate this entire "feature map" in seconds.
• Accuracy: When the authors tested it on famous viruses like SARS-CoV-2 (the virus that causes COVID-19) and HIV, ViralMap got the details right, even for parts of the virus it had never seen before.
• Simplicity: Instead of juggling ten different software programs, scientists now have one button to press.

The Bottom Line

Imagine you are a mechanic. Before, if a new, strange car rolled into your shop, you had to guess how the engine worked by looking at it. Now, you have a scanner that instantly tells you: "Here is the fuel line, here is the spark plug, here is the part that is broken, and here is how to fix it."

ViralMap is that scanner for viruses. It turns a confusing string of genetic code into a clear, actionable blueprint, helping scientists design better vaccines faster to stop the next pandemic before it spreads.

Technical Summary

1. Problem Statement

The development of effective viral vaccines requires precise engineering of viral proteins (antigens) to enhance immunogenicity, stability, and expression. This engineering process relies on comprehensive sequence annotation profiles that identify critical features such as:

• Topology: Transmembrane domains, signal peptides, and cytoplasmic/extracellular localization.
• Post-Translational Modifications (PTMs): N-glycosylation sites, furin cleavage sites, and disulfide bonds.
• Structural Features: Coiled coils and intrinsically disordered regions (IDRs).

Current Limitations:

• Fragmentation: Existing tools are typically single-task models (e.g., one for transmembrane domains, another for glycosylation). Combining them into a pipeline is technically difficult due to incompatible dependencies and licensing.
• Generalizability: Most general protein annotation tools are trained on non-viral datasets. Viral proteins evolve rapidly and diverge significantly from host proteins, leading to poor performance when these tools are applied to emerging pathogens.
• Speed vs. Accuracy: The "CEPI 100 Days Mission" aims to create vaccines for "Disease X" within 100 days. Current workflows cannot rapidly convert raw viral genomic sequences into the actionable annotation profiles needed for antigen design.

2. Methodology

ViralMap is a multi-label annotation model designed specifically for eukaryotic viral proteins. It predicts ten distinct annotation classes directly from primary amino acid sequences.

A. Data Curation

• Source: Downloaded ~4 million eukaryotic viral proteins from UniProt (excluding phages).
• Filtering: Applied strict filters for annotation quality (score ≥4.0), sequence length (100–1,024 residues), and removal of fragments/polyproteins.
• Clustering & Selection: Used MMseqs2 to cluster proteins at 60% identity. A custom annotation-aware selection algorithm was applied to choose representative proteins. This algorithm prioritized:
  • High-quality evidence (experimental > manual > automatic).
  • Coverage across all ten annotation classes.
  • Preservation of sequence diversity within clusters.
• Final Dataset: 8,238 proteins (3.13 million residues).
• Splitting: Data was split into 5 folds using a viral-family-aware strategy to ensure strict homology separation between training, validation, and test sets, preventing data leakage.

B. Model Architecture

ViralMap utilizes a two-stage architecture:

1. Base Model (Feature Extraction):
   • Built upon the ESM-2 protein language model (33 layers, 650M parameters).
   • The final four transformer layers are fine-tuned; earlier layers are frozen to preserve general representations.
   • Outputs 1,280-dimensional embeddings for each residue.
   • A 2-layer fully connected neural network (with dropout and ReLU) maps embeddings to 10 independent probability scores per residue (one for each annotation class).
2. Post-Processing Module:
   • Individual Residue Classes: For N-glycosylation, furin cleavage, disulfide bonds, and chain boundaries, an F2-optimized threshold (prioritizing recall) is applied to convert probabilities to binary predictions.
   • Region-Based Classes: For topology (signal peptides, transmembrane, cytoplasmic, extracellular) and structural features (coiled coils, IDRs), probabilities are decoded using Hidden Markov Models (HMMs) with the Viterbi algorithm.
     • A 4-state HMM enforces mutual exclusivity for topology.
     • 2-state HMMs are used for signal peptides, coiled coils, and disordered regions.
     • Transition matrices and start priors are learned from the training data.

C. Training Strategy

• Loss Function: Binary cross-entropy with two weighting schemes to address severe class imbalance:
  • Intra-class imbalance: Upweighting positive residues.
  • Inter-class imbalance: Using the effective number of samples framework.
• Optimization: AdamW optimizer, mixed precision (fp16), and early stopping based on macro-averaged PR-AUC.

3. Key Contributions

• Unified Framework: ViralMap is the first model to simultaneously predict ten diverse annotation classes (topology, PTMs, and structural features) in a single pass, replacing fragmented multi-tool pipelines.
• Viral Specialization: Unlike general tools, ViralMap is explicitly trained on curated eukaryotic viral data, addressing the unique evolutionary divergence of viral proteins.
• Sequence-Only Input: The model requires only primary amino acid sequences, enabling immediate functional profiling of novel viruses without the need for structural data or homology searches.
• Scalable Pipeline: Designed for rapid deployment in pandemic response scenarios (e.g., CEPI 100 Days Mission).

4. Results

The model was evaluated on held-out test folds against established benchmark tools (DeepTMHMM, NetNGlyc, ProP, DeepCoil, AIUPred).

• Overall Performance: ViralMap achieved a PR-AUC ≥ 0.75 for 7 out of 10 classes.
• Benchmark Comparisons:
  • N-Glycosylation: ViralMap significantly outperformed NetNGlyc in both precision (0.648 vs. 0.270) and recall (0.913 vs. 0.827), demonstrating it learns contextual features beyond simple motif detection.
  • Furin Cleavage: Achieved high recall (0.780) compared to ProP (0.440), despite the sparsity of annotated cleavage sites.
  • Topology: Performance was competitive with DeepTMHMM. ViralMap showed higher recall for transmembrane domains but slightly lower precision.
  • Intrinsically Disordered Regions (IDRs): Achieved the highest PR-AUC (0.942) with a much better precision/recall balance than AIUPred (Precision: 0.832 vs. 0.388).
  • Coiled Coils: While globally difficult to predict, ViralMap successfully localized functional coiled-coil motifs in viral fusion machinery.
• Case Studies:
  • SARS-CoV-2 Spike: ViralMap correctly identified the signal peptide, transmembrane domain, furin cleavage sites (S1/S2 and S2'), coiled-coil regions (HR1/HR2), and all 22 reference N-glycosylation sites. It successfully generalized to a novel family (Coronaviridae) not seen during training.
  • HIV-1 Env (gp160): Accurately predicted topology, the furin cleavage site separating gp120/gp41, the dense glycan shield (29 sites), and disulfide bonds.

5. Significance

ViralMap addresses a critical bottleneck in pandemic preparedness: the conversion of raw viral genomic data into actionable biological insights. By providing a unified, sequence-based framework that is specialized for viral proteins, it enables researchers to:

1. Rapidly characterize emerging pathogens (e.g., "Disease X").
2. Identify targets for antigen engineering (e.g., stabilizing domains, removing glycosylation shields).
3. Accelerate the timeline for vaccine design, potentially saving millions of lives by meeting the 100-day vaccine development goal.

The tool is implemented in Python, available as a command-line interface, and will be accessible via a web interface at the Disease X Knowledge Base.