Protein Language Models Expose Viral Immune Mimicry.

Area of Science:

• Computational virology
• Immunoinformatics
• Machine learning in biology

Background:

• Viruses employ molecular mimicry to evade host immune responses, complicating immune recognition and therapeutic interventions.
• Distinguishing viral from host proteins is crucial for understanding viral pathogenesis and developing effective countermeasures.
• Existing methods face challenges in accurately identifying viral proteins, particularly those involved in immune evasion.

Purpose of the Study:

• To develop and validate a machine learning model using pretrained protein language models (PLMs) to differentiate viral from human proteins.
• To identify and characterize viral proteins that are frequently misclassified by the model, indicating potential immune evasion strategies.
• To gain mechanistic insights into viral immune escape by integrating PLMs with explainable AI techniques.

Main Methods:

• Leveraging pretrained protein language models (PLMs) for sequence-based classification of viral and human proteins.
• Integrating PLMs with explainable AI models to interpret the features driving protein classification and misclassification.
• Evaluating model performance using Receiver Operating Characteristic Area Under the Curve (ROC-AUC) and analyzing misclassified sequences.

Main Results:

• The developed model achieved state-of-the-art performance with a ROC-AUC of 99.7%.
• Misclassified viral proteins (3.9% of sequences) were characterized by low immunogenicity and belonged to human-specific viral families associated with chronic infections.
• Analysis revealed overlapping biophysical signals confound both immune systems and machine learning models in identifying immune-evading viral proteins.

Conclusions:

• Coupling PLMs with explainable AI provides a powerful approach for computational virology and understanding viral immune escape mechanisms.
• The findings highlight the challenges posed by viral molecular mimicry and low-immunogenicity proteins for both natural immunity and artificial intelligence.
• This research offers implications for the rational design of vaccines and strategies to combat persistent and pathogenic viral infections.