GXQ-Create: A Multimodal Virus Host Prediction Tool Integrating Genomic Features and Protein Language Models

GXQ-Create is an innovative multimodal virus host prediction tool that combines k-mer genomic features with the ESM-2 protein language model, using a late-fusion SVM architecture, and achieves a cross-validation accuracy of 96.4% in eukaryotic host prediction. This tool is of great value for preventing cross-species virus transmission and assessing the risk of emerging infectious diseases.

Understanding the interaction between viruses and hosts is at the core of virology and infectious disease research. Accurately predicting the potential host range is of great significance for preventing cross-species transmission. Traditional methods rely on genomic homology alignment, but their effectiveness is limited when dealing with new or rapidly evolving RNA viruses. In recent years, deep learning technology has provided possibilities for more intelligent prediction methods.

Dual-Modal Feature Extraction

1. k-mer Genomic Features: Extract k-mer frequency features from viral genomes, which reflect genomic composition and evolutionary patterns, with simple computation and strong robustness. 2. ESM-2 Protein Language Model: Introduce Meta AI's ESM-2 model, which obtains high-dimensional embeddings through pre-training on massive protein sequences, capturing information about protein evolution, structure, and function.

Late-Fusion Architecture

Adopt a late-fusion strategy: independently encode the two modal features → concatenate to form a joint representation → input to an SVM classifier for host prediction. This architecture avoids feature conflicts, and SVM has good generalization ability and interpretability.

Dataset and Evaluation

Trained and tested on eukaryotic hosts such as fungi, algae, protozoa, plants, and invertebrates, using cross-validation for evaluation, with an average accuracy of 96.4%.

Reasons for Performance Advantages

• Complementary modal information: k-mer (macro genomic patterns) and ESM-2 (micro protein functions) complement each other;
• Biological prior: ESM-2 pre-training contains evolutionary knowledge;
• Appropriate complexity: SVM is not prone to overfitting on small-sample biological data.

• Emerging Infectious Disease Monitoring: Quickly predict the potential hosts of unknown viruses, aiding risk assessment and prevention;
• Virus Evolution Research: Track host adaptation and evolution, understand cross-species transmission mechanisms;
• Agricultural Disease Prevention: Identify high-risk viruses, reduce agricultural losses;
• Biosafety Assessment: Evaluate changes in the host range of modified viruses, support biosafety reviews.

The code has been open-sourced on GitHub, implemented in Python, and depends on Biopython (sequence processing), PyTorch (ESM-2 operation), scikit-learn (SVM training), and NumPy/Pandas (data processing). Usage workflow: Prepare FASTA sequences → Extract k-mer and protein embeddings → SVM inference → View prediction results and confidence.

• Expand the host range to prokaryotes (bacteria, archaea);
• Integrate more modalities such as viral 3D structure and host receptor expression data;
• Explore end-to-end deep learning (graph neural networks/Transformers) to reduce manual feature engineering;
• Integrate into pathogen monitoring platforms to achieve real-time early warning.

GXQ-Create combines traditional bioinformatics methods with modern deep learning technology and performs excellently in virus host prediction tasks. Its open-source code provides references for computational biology, virology, and AI for Science research, helping to solve important virology problems.