ViraHinter: a dual-modal artificial intelligence framework for predicting virus-host interactions

ViraHinter is a novel dual-modal deep learning framework that integrates structural generation and sequence representation to accurately predict virus-host interactions, outperforming existing models in identifying novel host factors and enabling broad-spectrum antiviral discovery across diverse viral families.

The Gist

Imagine a virus as a master thief trying to break into a massive, high-tech bank (the human body). To succeed, the thief doesn't just smash the door; they need to find the specific keys, security codes, and internal maps that allow them to sneak past guards, open vaults, and steal the money (our cells' resources) to make copies of themselves.

For a long time, scientists have been trying to map out exactly which "keys" (human proteins) each "thief" (virus) uses. But this is incredibly hard work. It's like trying to find a specific needle in a haystack by looking at it under a microscope one by one. It takes forever, and many needles remain hidden.

Enter ViraHinter, a new artificial intelligence tool that acts like a super-smart detective who can predict exactly which keys a thief will use, even if they've never seen that specific thief before.

Here is how ViraHinter works, broken down into simple concepts:

1. The Two-Brain System (Dual-Modal AI)

Most old AI tools for this job were like detectives who only read the thief's resume (the genetic sequence). They could guess what the thief might do based on their past history, but they missed the physical reality of the crime.

ViraHinter is different because it has two brains working together:

• Brain A (The Sequence Reader): It reads the genetic "resume" of the virus and the human, just like the old tools.
• Brain B (The 3D Architect): It builds a 3D model of what the virus and human proteins look like when they try to shake hands. It checks if their shapes actually fit together like a lock and key.

By combining the "resume" with the "3D blueprint," ViraHinter can spot connections that other tools miss. It's like knowing a thief wears a red hat (sequence) and knowing they are tall enough to reach the high window (structure).

2. The "Crystal Ball" for New Threats

The real magic of ViraHinter is its ability to generalize. Imagine you've studied how a pickpocket steals from a tourist in Paris. Usually, you can't guess how they will steal from a tourist in Tokyo because the situations are different.

But ViraHinter learns the underlying rules of pickpocketing. So, if a brand-new virus appears (like a new strain of flu or a novel coronavirus), ViraHinter can look at it and say, "Even though I've never seen this specific virus, its shape and genetic code suggest it will likely use these specific human proteins to get in."

3. Finding the "Universal Keys"

The researchers used ViraHinter to study two major groups of viral thieves: Coronaviruses (like SARS-CoV-2) and Influenza (the flu).

They discovered something fascinating:

• The "Pan-Viral" Keys: They found that different types of coronaviruses all use the same few human "keys" to break in. It's like realizing that every burglar in the city uses the same master key for the back door.
• The "Flu" Network: Similarly, they found 33 human proteins that are essential for all types of flu viruses (H1N1, H3N2, H9N2), even though the flu viruses look very different on the outside.

4. Why This Matters: The "Broad-Spectrum" Shield

Why is finding these "universal keys" so important?

Currently, when a new flu strain or a new coronavirus appears, we have to start from scratch to find a cure. It's like making a new lock for every new thief.

But if we know that all these viruses rely on a specific human protein (like a protein called RAB11A for flu or RAB8A for coronaviruses), we can design a drug that jams that specific human protein.

• The Analogy: Instead of trying to catch every different thief, you just change the lock on the back door so no one can open it.
• The Result: This could lead to "broad-spectrum" antivirals—medicines that work against many different viruses at once, including ones we haven't even discovered yet.

The Bottom Line

ViraHinter is a powerful new tool that stops us from having to guess and check every single virus interaction manually. It uses a mix of genetic reading and 3D modeling to predict how viruses hijack our bodies.

By identifying the "weak spots" in our own biology that viruses love to exploit, it gives scientists a roadmap to build better, broader defenses against future pandemics. It turns the search for cures from a game of "find the needle in the haystack" into a game of "here is the exact location of the needle."

Technical Summary

1. Problem Statement

Protein-protein interactions (PPIs) between viruses and their hosts are fundamental to viral infection, replication, and pathogenesis. Despite the existence of high-throughput experimental mapping (e.g., AP-MS, proximity labeling), the virus-host interactome remains largely uncharacterized due to:

• Experimental Limitations: High cost, labor intensity, and difficulty in capturing transient or low-affinity interactions.
• Data Scarcity: Limited coverage for emerging pathogens and non-model organisms.
• Sequence Divergence: Viral proteins evolve rapidly with low sequence homology across families, making homology-based transfer of knowledge (e.g., from Influenza to Coronaviruses) ineffective.
• Computational Gaps: Existing deep learning methods often rely solely on sequence features (missing spatial constraints) or are computationally prohibitive for proteome-wide screening (e.g., AlphaFold 3, RoseTTAFold2). There is a need for a framework that generalizes across viral families and predicts interactions for unseen viruses.

2. Methodology: The ViraHinter Framework

A. Data Curation: High-Fidelity Atlas
The authors constructed a unified virus-host PPI atlas by integrating data from four major repositories: IntAct, BioGRID, VirHostNet, and VirusMentha.

• Filtering: Only physical associations were retained; colocalization and genetic interactions were excluded.
• Confidence Stratification: Instead of a single threshold, source-specific rules were applied:
  • IntAct: MI score > 0.6 (High), 0.4–0.6 (Medium).
  • BioGRID: Low-throughput physical (High); validated high-throughput (Medium).
  • VirHostNet: Binary assays (High); complex assays (Medium).
  • VirusMentha: Multi-validated records only (High).
• Dataset Size: The final benchmark set contained 10,451 high-confidence PPIs (700 viral, 4,202 human proteins), while the full resource contained 43,942 PPIs.
• Negative Sampling: A "hard-negative" set was generated using MMseqs2 clustering to exclude proteins with >60% sequence identity to known binders, ensuring the model learns true interaction rules rather than sequence similarity.

B. Model Architecture: Dual-Modal Fusion
ViraHinter is a dual-modal deep learning framework that fuses structure-informed representations with sequence-derived embeddings.

1. Structure Branch:
   • Initialized from IntFold (a general protein structure prediction model).
   • Uses a 48-block Pairformer stack to generate single and pair representations.
   • Employs an iterative diffusion module to generate all-atom complex structures.
   • Key Innovation: Intra-chain information is masked in the pair representation to force the model to focus on intermolecular compatibility (interface signals) rather than monomeric folding.
2. Sequence Branch:
   • Extracts embeddings using ESM-2 (150M) protein language models for both viral and human partners.
3. Fusion & Prediction:
   • Features from both branches are fused. Structure-derived features are transferred via a stop-gradient connection to prevent backpropagation from destabilizing the structure module.
   • The fused features pass through additional Pairformer blocks and a gated linear layer to output the interaction probability.
   • The model simultaneously predicts interaction likelihood and generates the complex structure.

C. Training Strategy

• Two-Stage Training: Adaptation on a broad medium-confidence atlas followed by fine-tuning on the high-confidence subset.
• Leakage Control: Training and test sets were split by viral sequence identity (max 60% identity between train and test viral proteins) to ensure the model generalizes to unseen viral families.

3. Key Results

A. Benchmark Performance
ViraHinter was evaluated against state-of-the-art baselines: AlphaFold 3 (AF3), RoseTTAFold2-PPI (RF2-PPI), and RoseTTAFold2-Lite.

• Severe Class Imbalance (1:1000 ratio): ViraHinter achieved an AUPR of 0.44, significantly outperforming RF2-PPI (0.28), AF3 (0.23), and RF2-Lite (0.10).
• Interface Size: ViraHinter maintained superiority across weak (≤10 residues), medium, and strong interfaces. For weak interfaces, ViraHinter scored 0.11 vs. AF3's 0.00.
• Generalization (Viral Held-Out): When tested on viral proteins with ≤60% identity to training data, ViraHinter achieved an AUPR of 0.50, a 4.5-fold improvement over AF3 (0.11). This demonstrates robustness against rapidly evolving viral sequences.
• Cross-Viral Performance: Consistently outperformed baselines across diverse viral groups including SARS-CoV, Influenza A, HIV, HPV, and Herpesviruses.

B. Validation on Coronaviruses

• Evidence-Aware Ranking: On an independent set of 958 SARS-CoV-1/2 and MERS-CoV interactions, ViraHinter successfully prioritized pairs with independent experimental validation (Low-throughput evidence) over those with no evidence.
• Pan-Coronavirus Discovery: Identified 7 shared host factors (e.g., RAB8A, RAB5C, CUL2, PABPC1) across SARS-CoV-1, SARS-CoV-2, and MERS-CoV.
• Structural Conservation: Structural modeling of the NSP7-RAB8A interface revealed that despite sequence divergence in NSP7 across the three viruses, they converge on an identical functional cleft on RAB8A, suggesting a conserved, broad-spectrum therapeutic target.

C. Validation on Influenza A Viruses

• Subtype Analysis: Applied to H1N1, H3N2, and H9N2 (including a novel H3N2 strain isolated in the authors' lab).
• Shared Core: Identified 33 common host factors targeted by all three subtypes.
• Novel Findings: 12 of these factors (e.g., SFN, RAB11A, CDK1) lacked physical interaction records in existing databases but were linked to immune regulation in literature.
• Structural Conservation: Modeling of HA-RAB11A interactions showed that despite antigenic drift in HA, the binding mode to RAB11A is structurally conserved across subtypes, highlighting RAB11A as an evolutionary bottleneck.

4. Key Contributions

1. Dual-Modal Architecture: Successfully integrates structural geometry (via diffusion/Pairformer) with evolutionary sequence context (via ESM embeddings), overcoming the limitations of sequence-only or structure-only models.
2. Generalization to Unseen Viruses: Demonstrated the ability to predict interactions for viral families with low homology to training data, a critical capability for emerging pathogens.
3. Discovery of Conserved Mechanisms: Uncovered "evolutionary bottlenecks"—host factors and binding interfaces that are structurally conserved across divergent viral strains (e.g., RAB8A for Coronaviruses, RAB11A for Influenza).
4. Evidence-Aware Prioritization: The model effectively distinguishes high-confidence, experimentally verified interactions from noise, aiding in the selection of targets for wet-lab validation.

5. Significance

• Broad-Spectrum Antivirals: By identifying host factors essential for multiple viral subtypes or families, ViraHinter provides a roadmap for developing broad-spectrum antivirals that are less susceptible to viral mutational escape.
• Accelerated Drug Discovery: The framework enables systematic screening of host factors for all known human-infecting viruses, rapidly generating hypotheses for therapeutic targets.
• Pandemic Preparedness: The ability to model interactions for novel or zoonotic viruses with limited sequence data makes ViraHinter a vital tool for rapid response to emerging infectious diseases.
• Methodological Advancement: Sets a new benchmark for PPI prediction, proving that integrating structural generation with sequence representation yields superior performance in complex, imbalanced biological datasets.