MutaGAN: A sequence-to-sequence GAN framework to predict mutations of evolving protein populations

Area of Science:

• Computational Biology
• Virology
• Machine Learning

Background:

• Predicting pathogen evolution is crucial for disease control and treatment.
• Machine learning has not been extensively applied to predict viral evolutionary progeny.

Purpose of the Study:

• To develop a novel machine learning framework, MutaGAN, for accurate prediction of genetic mutations and viral evolution.
• To address the gap in using machine learning for predicting virus evolutionary trajectories.

Main Methods:

• Developed MutaGAN, a framework utilizing generative adversarial networks with sequence-to-sequence recurrent neural networks.
• Trained MutaGAN on a phylogenetic model of protein evolution with maximum likelihood tree estimation.
• Applied MutaGAN to influenza virus sequences from the NCBI Influenza Virus Resource.

Main Results:

• MutaGAN generated 'child' sequences from 'parent' sequences with a median Levenshtein distance of 4.00 amino acids.
• The framework successfully generated sequences containing known mutations for 72.8% of parent influenza sequences.
• Demonstrated MutaGAN's capability in predicting genetic mutations and evolutionary changes.

Conclusions:

• MutaGAN framework effectively predicts genetic mutations and evolutionary paths of viral populations.
• This tool has significant implications for pathogen forecasting and disease control strategies.
• MutaGAN shows broad utility for evolutionary prediction across various protein populations.