Generative Protein Design: From Deep Learning Algorithms to Translational Applications.

Area of Science:

• Computational Biology
• Biochemistry
• Machine Learning

Background:

• Protein design traditionally relied on energy-function optimization.
• Deep learning has shifted the paradigm towards probabilistic generative modeling.
• Advancements in protein representation are key to this transition.

Purpose of the Study:

• To review the algorithmic basis of deep learning in protein design.
• To classify current generative protein design methodologies.
• To summarize evaluation principles and applications of generative protein design.

Main Methods:

• Review of protein representations: sequence-centered, graph-based, and SE(3)-equivariant manifolds.
• Classification of design approaches: sequence-structure decoupled, hybrid, and co-design.
• Discussion of specific techniques within each paradigm (e.g., hallucination, backbone generation, joint generative formulations).

Main Results:

• Deep learning enables sophisticated protein design through advanced representations.
• Three main design paradigms (decoupled, hybrid, co-design) offer diverse strategies.
• Established evaluation principles ensure physical validity and functional relevance.

Conclusions:

• Generative protein design is evolving from mere structure generation to programmable engineering of biological functions.
• The field is rapidly advancing due to innovations in deep learning and protein representations.
• Future directions point towards designing complex biological functions with high precision.