Searching for protein variants with desired properties using deep generative models.

Area of Science:

• Biotechnology
• Computational Biology
• Protein Engineering

Background:

• Protein engineering seeks to enhance protein functionality for various applications.
• Deep learning models capture protein sequence features but struggle with long-range dependencies.
• Existing generative models require improvement in representing relationships between amino acid sites in longer sequences.

Purpose of the Study:

• To develop a deep learning model that improves representation learning for longer protein sequences.
• To enhance the similarity between generated protein sequences and original sequences.
• To leverage the positional relationship of protein sequences in latent space for variant discovery.

Main Methods:

• Proposed a Temporal Variational Autoencoder (T-VAE) model comprising an encoder and a decoder.
• Utilized dilated causal convolution in the encoder to expand receptive fields for better long-sequence encoding.
• The decoder generates variants that closely resemble the original protein sequences.

Main Results:

• T-VAE achieved a higher person correlation coefficient and lower mean absolute deviation in predicting protein fitness compared to other models.
• Demonstrated superior representation learning for longer protein sequences.
• Achieved a 12.9% improvement in sequence identity between generated and input data compared to the baseline model.

Conclusions:

• T-VAE exhibits enhanced capabilities in representation learning for longer protein sequences.
• The model shows significant advantages in generating protein variants with high sequence identity.
• T-VAE offers a promising approach for discovering novel protein variants with improved functional properties.