De Novo Protein Design for Novel Folds Using Guided Conditional Wasserstein Generative Adversarial Networks.

Area of Science:

• Computational Biology
• Protein Engineering
• Machine Learning

Background:

• Limited diversity in known protein structural folds despite vast sequence data.
• Challenges in predicting sequence-structure relationships for novel folds.
• Need for advanced models to design proteins with arbitrary structural architectures.

Purpose of the Study:

• To develop deep generative models for designing protein sequences for novel structural folds.
• To reveal underlying sequence-structure relationships.
• To explore uncharted protein sequence space.

Main Methods:

• Developed semisupervised guided, conditional, Wasserstein Generative Adversarial Networks (gcWGAN).
• Constructed a low-dimensional representation of fold space for conditional input.
• Incorporated an ultrafast sequence-to-fold predictor (oracle) to guide WGAN training.
• Utilized a semisupervised strategy leveraging sequence data with and without paired structures.

Main Results:

• gcWGAN generated more successful designs and covered 3.5 times more target folds than cVAE on novel folds.
• Designs were assessed as physically and biologically sound by predictors.
• Achieved comparable or better fold accuracy with greater sequence diversity and novelty compared to cVAE.
• The model enhanced de novo protein design methods like RosettaDesign.

Conclusions:

• gcWGAN effectively designs proteins for novel structural folds by learning generalizable principles.
• The model explores new sequence space, expanding possibilities in protein engineering.
• Developed tools (data, code, models) are publicly available for research.