Navigating the amino acid sequence space between functional proteins using a deep learning framework.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning in Biology

Background:

• Understanding protein sequence-function relationships is crucial for evolution, disease research, and protein design.
• The complexity of protein sequence space makes it challenging to map sequences to functions.
• Generative models offer a powerful approach to learn and replicate data specificity, aiding in the deciphering of complex biological systems.

Purpose of the Study:

• To develop and apply an unsupervised generative model, an Adversarial Auto-Encoder (AAE), for generating novel protein sequences.
• To explore the capabilities of AAEs in capturing sequence patterns associated with protein functions and relationships between sequence positions.
• To investigate methods for generating intermediate protein sequences and transferring functional properties between protein families.

Main Methods:

• An Adversarial Auto-Encoder (AAE), an unsupervised generative model, was employed.
• The AAE was tested on three protein families: sulfatase, HUP, and TPP.
• Two novel sampling strategies, latent space interpolation and latent space arithmetic, were developed and analyzed.

Main Results:

• Clustering of encoded sequences in the AAE's latent space showed high functional homogeneity.
• Latent space interpolation generated meaningful biological sequences in previously uncharted evolutionary areas.
• Latent space arithmetic successfully transferred functional sequence properties between different protein sub-families, as evidenced by 3D structure models.

Conclusions:

• Deep learning frameworks, like AAEs, can effectively model biological complexity.
• This study provides new tools for exploring amino acid sequence and functional spaces.
• The generative capabilities of AAEs open new avenues for protein design and understanding molecular evolution.