Conditional generative modeling for de novo protein design with hierarchical functions

Area of Science:

• Computational biology
• Machine learning
• Protein engineering

Background:

• Protein design is crucial for medicine and biotechnology but computationally intensive.
• Existing generative models address specific protein design sub-problems, not general-purpose design.
• Machine learning offers powerful tools for complex problems like protein generation.

Purpose of the Study:

• To develop a general-purpose protein design method conditioned on functional labels.
• To establish an evaluation framework for generative protein models.
• To explore novel protein functions through conditional generation.

Main Methods:

• Developed ProteoGAN, a conditional generative adversarial network.
• Devised an evaluation scheme with biological and statistical metrics.
• Analyzed model performance against baselines and hyperparameters.

Main Results:

• ProteoGAN outperforms classic and recent deep-learning methods for protein sequence generation.
• The model demonstrates effective conditioning on Gene Ontology functional labels.
• Analysis provides insights into model behavior and hyperparameter influence.

Conclusions:

• ProteoGAN represents a significant advancement in general-purpose, functionally-conditioned protein design.
• The developed evaluation metrics offer a standard for assessing generative protein models.
• Future work can explore generating proteins with novel functions by combining labels.