Improved prediction of virus-human protein-protein interactions by incorporating network topology and viral molecular mimicry

This study introduces vhPPIpred, a machine learning method that improves virus-human protein-protein interaction prediction by incorporating network topology and viral molecular mimicry, supported by a rigorously curated benchmark dataset that demonstrates superior performance and efficiency over existing tools.

The Gist

Imagine your body is a massive, bustling city. Inside this city, there are millions of workers (human proteins) who constantly shake hands, exchange packages, and work together to keep the city running. This network of handshakes is called the Protein-Protein Interaction (PPI) network.

Now, imagine a virus as a master thief trying to break into this city. To succeed, the thief doesn't just smash the doors; they try to blend in. They might wear a disguise that looks like a legitimate worker, or they might try to shake hands with the city's most important officials to gain access.

This paper introduces a new, super-smart detective tool called vhPPIpred designed to predict exactly how these viral thieves will try to shake hands with human workers.

Here is the breakdown of how they built this tool and why it's a game-changer, explained simply:

1. The Problem: The "Fake" Test

Previously, scientists tried to build computer programs to predict these viral handshakes. But they had a major flaw: they were cheating on their own tests.

Imagine a teacher giving a math test to students, but the test questions are identical to the homework the students just did. Of course, the students will get 100%! But that doesn't mean they actually understand math; they just memorized the answers.

Many previous computer programs were trained on data that overlapped with their test data. They looked "smart" because they were just remembering old cases, not learning new patterns. Also, they didn't have a good list of "fake" interactions (times when a virus didn't shake hands with a human) to learn from.

2. The Solution: A Rigorous New Exam

The authors of this paper decided to build a brand new, fair exam.

• The Benchmark Dataset: They carefully curated a list of known viral handshakes (positive examples) and a list of interactions that definitely didn't happen (negative examples).
• The "No Cheating" Rule: They made sure the viruses and humans in the "training" group were completely different from those in the "testing" group. It's like teaching a student with a textbook on cats and then testing them on dogs. If the student still gets it right, they actually understand the concept of "animals," not just the specific pictures they memorized.

3. The Detective Tool: vhPPIpred

Once they had a fair exam, they built a new AI detective called vhPPIpred. Instead of just looking at the "ID cards" (the genetic sequence) of the virus and the human, this detective looks at four clever clues:

1. The "Face" (Sequence Embedding): It uses a super-advanced AI (ProtT5) to read the protein's "face" and understand its deep structure, not just its letters.
2. The "Family History" (Evolutionary Info): It looks at how the protein has changed over millions of years to see what it's good at.
3. The "Popularity Contest" (Network Topology): This is a brilliant insight. The detective knows that viral thieves prefer to shake hands with the most popular human workers (those with the most connections). If a human protein is a "celebrity" in the city network, the virus is more likely to target them.
4. The "Disguise" (Molecular Mimicry): Viruses are masters of disguise. They often mimic the shape of human proteins to trick the system. The tool checks: "Does this viral protein look like a human protein that already shakes hands with the target?" If yes, it's a high-risk interaction.

4. The Results: Beating the Competition

The authors put their new detective against five other famous detectives (previous methods).

• The Scoreboard: On the new, fair exam, vhPPIpred won hands down. It was much better at spotting the real threats and ignoring the false alarms.
• The Speed: It was also faster and used less computer memory than the deep-learning-heavy competitors, making it practical for real-world use.

5. Real-World Superpowers

Why does this matter? The paper shows two amazing things this tool can do:

• Finding the Front Door (Receptors): Viruses need a specific "front door" (receptor) to enter a cell. This tool successfully predicted which human proteins act as these doors for various viruses, helping scientists understand how infections start.
• Predicting the "Badness" (Virulence): Can we tell if a new virus will be deadly just by looking at its predicted handshakes? Yes! The authors showed that by analyzing the pattern of handshakes a virus makes, they could predict how dangerous (virulent) the virus would be. It was more accurate than looking at the virus's DNA alone.

The Big Picture

Think of vhPPIpred as a high-tech security system for our cellular city. By understanding the "social network" of our cells and how viruses try to hack that network, we can:

• Find new drugs to block the handshake.
• Predict how dangerous a new virus might be before it spreads.
• Understand the rules of the game that viruses play to infect us.

This study didn't just build a better tool; it built a better rulebook for testing these tools, ensuring that future discoveries in fighting viruses are based on solid, honest science.

Technical Summary

1. Problem Statement

Protein-protein interactions (PPIs) between viruses and humans are fundamental to viral infection mechanisms, pathogenesis, and the development of antiviral therapies. While experimental methods (e.g., Yeast Two-Hybrid, AP-MS) exist, they are costly, time-consuming, and often yield false positives. Computational prediction methods have emerged as a solution but face two critical limitations:

• Lack of Rigorous Benchmarking: Existing methods often suffer from data leakage due to overlapping proteins between training and test sets, leading to overestimated performance. Furthermore, the construction of reliable negative datasets (non-interacting pairs) is challenging, as random sampling often introduces false negatives.
• Insufficient Feature Representation: Current machine learning approaches primarily rely on sequence embeddings and evolutionary profiles, often neglecting critical biological context such as the topology of the host interaction network and the specific mechanism of "viral molecular mimicry" (where viruses mimic host ligands to hijack host proteins).

2. Methodology

The authors developed vhPPIpred, a machine learning-based framework designed to address the aforementioned limitations through rigorous dataset construction and multi-source feature integration.

A. Benchmark Dataset Construction

To ensure unbiased evaluation, the authors constructed a high-quality benchmark dataset with strict separation criteria:

• Data Sources: Positive PPIs were curated from eight databases (BioGRID, IntAct, VirusMentha, etc.), resulting in 16,314 high-confidence physical interactions.
• Clustering & Splitting: Viral and human proteins were clustered using MMseqs2 at 40% sequence identity. The dataset was divided into six groups based on viral protein clusters.
• Independence: Groups were constructed to ensure non-overlapping PPIs and minimum sequence similarity between training and test sets for both viral and human proteins.
• Negative Sampling: Negative samples were generated using mammalian virus proteins (known not to infect humans) and human proteins that do not interact with viruses, ensuring biological plausibility for non-interaction. A 1:10 positive-to-negative ratio was maintained.

B. Feature Engineering

vhPPIpred integrates four distinct feature types to capture intrinsic protein properties and interaction-specific patterns:

1. Sequence Embedding: 1,024-dimensional vectors generated by ProtT5-XL-U50 (a pre-trained protein language model).
2. Evolutionary Information: 20-dimensional vectors derived from Position-Specific Scoring Matrices (PSSM) via PSI-BLAST.
3. Network Topology: The degree of the human protein in the Human Protein-Protein Interaction Network (HPPIN), based on the hypothesis that highly connected hubs are more likely targets.
4. Viral Molecular Mimicry: A novel feature quantifying the similarity between a viral protein and the "neighbor" proteins (those interacting with the target human protein) in the HPPIN, combined with the interaction scores of those neighbors. This captures the mechanism where viruses mimic host ligands.

Note: PCA was applied to reduce the dimensionality of ProtT5 and PSSM embeddings before model input.

C. Machine Learning Model

• Algorithm Selection: Six algorithms (Random Forest, AdaBoost, XGBoost, SVM, KNN, Naïve Bayes) were evaluated. XGBoost was selected as the base algorithm due to superior performance (AUROC 0.92, AUPRC 0.67).
• Optimization: Hyperparameters were tuned via grid search and cross-validation.

3. Key Contributions

1. Standardized Benchmark Dataset: The creation of a rigorously curated dataset with non-overlapping clusters and biologically informed negative sampling, serving as a new standard for evaluating virus-human PPI predictors.
2. Novel Feature Integration: The first method to explicitly incorporate viral molecular mimicry and host network topology into PPI prediction, moving beyond pure sequence-based approaches.
3. Open-Source Tool: The release of vhPPIpred as a publicly available tool (GitHub) for the research community.

4. Results

Performance on Benchmark Dataset

• Comparison: vhPPIpred outperformed five state-of-the-art methods (HVPPI, LSTM-PHV, Cross-Attention_PHV, MultiTask-Transfer, TransPPI).
• Metrics: It achieved the highest Precision (0.952), AUROC (0.921), and AUPRC (0.680).
• Re-evaluation: When competing methods were retrained on the new benchmark (removing data leakage), their performance dropped significantly (e.g., Cross-Attention_PHV AUPRC dropped from 0.109 to 0.088), confirming that previous results were overestimated.

Performance on Independent Datasets

vhPPIpred was tested on three independent datasets (Yang's HHV-5 dataset, Zhou's SARS-CoV-2 dataset, and DeNovo dataset):

• Yang's Dataset: Achieved the highest AUROC (0.801) and F1-score (0.329).
• Zhou's Dataset: Despite the difficulty of predicting novel virus interactions, vhPPIpred achieved the highest accuracy (0.173) compared to near-zero performance by other methods.
• DeNovo Dataset: Achieved a superior accuracy of 0.705, significantly outperforming competitors (0.158–0.229).

Computational Efficiency

• Time & Memory: vhPPIpred demonstrated high computational efficiency, maintaining stable runtime and memory usage even as sample sizes increased to 100,000, outperforming deep learning-heavy methods like TransPPI and Cross-Attention_PHV which showed sharp resource spikes.

Downstream Applications

1. Viral Receptor Identification: In a test against known Receptor Binding Protein (RBP)-receptor pairs, vhPPIpred identified 7 true pairs in the Top 10 and 18 in the Top 50, significantly outperforming other methods.
2. Virulence Prediction: By representing predicted PPIs as graphs and using Graph Convolutional Networks (GCN) to extract features, the model predicted viral virulence (High vs. Low) with an AUROC of 0.848. This outperformed methods based solely on viral genome (AUROC 0.790) or proteome (AUROC 0.830) sequences, validating that PPIs are strong determinants of viral phenotypes.

5. Significance

• Scientific Impact: The study resolves the issue of data leakage in PPI prediction benchmarks, providing a more realistic assessment of model capabilities.
• Biological Insight: It validates the hypothesis that viral molecular mimicry and host network topology are critical, previously underutilized factors in virus-host recognition.
• Practical Application: The tool offers a rapid, cost-effective approach for:
  • Identifying unknown viral receptors.
  • Inferring the virulence of emerging viruses (crucial for early warning systems).
  • Guiding antiviral drug discovery by pinpointing key host targets.

In conclusion, vhPPIpred represents a significant advancement in computational virology by combining rigorous data curation with biologically informed feature engineering to deliver a robust, efficient, and highly accurate prediction tool.