Prediction of virus-host associations using protein language models and multiple instance learning

Area of Science:

• Virology
• Bioinformatics
• Machine Learning

Background:

• Predicting virus-host associations is crucial for identifying viral hosts, especially in microbiomes where most virus hosts remain unknown.
• Understanding these associations aids in discovering novel viruses that may infect humans and animals.

Purpose of the Study:

• To introduce EvoMIL, a novel deep learning method for predicting virus-host associations using only viral sequences.
• To identify critical viral proteins involved in host specificity determination.

Main Methods:

• EvoMIL integrates a pre-trained protein language model (ESM) with attention-based multiple instance learning for protein-level predictions.
• The method utilizes protein embeddings, outperforming traditional sequence composition features like amino acids and DNA k-mers.

Main Results:

• EvoMIL demonstrated significant median F1 score improvements: 10.8%, 16.2%, 4.9% for prokaryotic hosts and 1.7%, 6.6%, 11.5% for eukaryotic hosts.
• Binary classifiers achieved high AUC values (>0.95 for prokaryotic, 0.8-0.9 for eukaryotic hosts).
• Identified key viral proteins contributing to host specificity.

Conclusions:

• Protein embeddings provide robust signals for predicting virus-host associations.
• EvoMIL offers an effective approach for identifying virus hosts and key proteins, advancing our understanding of virus-host interactions.