General Intelligence Framework to Predict Virus Adaptation Based on a Genome Language Model.

Area of Science:

• Virology
• Computational Biology
• Genomics

Background:

• Viral pandemics often originate from animal viruses adapting to humans.
• Predicting viral adaptation from genetic data is difficult due to incomplete sequences and limited training labels.
• Existing models struggle with insufficient data and blind sequence prediction.

Purpose of the Study:

• To develop a semi-supervised General Intelligence framework to predict Virus Adaptation (GIVAL).
• To create a novel language model (vBERT) for embedding viral protein sequences.
• To overcome limitations in predicting viral adaptation with inadequate data.

Main Methods:

• Developed GIVAL, a semi-supervised framework utilizing a language model (vBERT) for viral protein sequence embedding.
• Pretrained vBERT using hidden Markov model-contextualized tokens.
• Evaluated vBERT against other models (DNABERT-2, proteinBERT, ESM-2, Transformer, Word2Vec) for distinguishing viral proteins.
• Assessed GIVAL's accuracy and fault tolerance with insufficient training labels.

Main Results:

• vBERT outperformed existing models in classifying viral proteins based on labels like serotypes and mutations.
• GIVAL demonstrated higher accuracy and better fault tolerance in virus adaptation prediction with limited or noisy labels.
• GIVAL successfully predicted increased human adaptation in equine-origin IAVs and bovine H5N1 IAVs.
• GIVAL identified an adaptation shift in MERS-CoV-like virus variants towards SARS-CoV-2 receptors.
• GIVAL quantified incremental adaptation in monkeypox virus variants, correlating with human case increases.

Conclusions:

• GIVAL provides a generally intelligent framework for genotype-based virus adaptation prediction.
• The framework effectively handles insufficient labels and blind sequence inputs.
• GIVAL has potential applications in predicting other genotype-to-phenotype relationships in viruses.