gcWGAN

gcWGAN generates de novo protein sequences conditioned on low-dimensional fold representations using a guided conditional Wasserstein Generative Adversarial Network (WGAN) to explore and expand protein sequence-structure relationships.

Key Features:

• Conditional Wasserstein GAN Framework: Uses a conditional WGAN with the Wasserstein distance in its loss function to provide stable gradients and improve design quality.
• Low-Dimensional Fold Space Representation: Constructs a low-dimensional, generalizable representation of fold space that serves as the conditional input to the generator.
• Ultrafast Sequence-to-Fold Predictor (Oracle): Integrates an ultrafast oracle that predicts sequence-to-fold mappings and provides feedback during training.
• Semisupervised Training Strategy: Trains on both labeled sequence-structure pairs and unlabeled sequence data to leverage diverse datasets.

Scientific Applications:

• Designing Novel Protein Folds: Generates protein sequences for arbitrary structural folds, including folds not present in the training set.
• Enhanced Design Success Rates: When evaluated over 100 novel folds, achieved higher success rates and covered approximately 3.5× more target folds than competing data-driven methods such as cVAE.
• Biologically Sound Designs: Produces designs validated as physically and biologically plausible using sequence- and structure-based predictors.
• Boosting Principle-Driven Methods: Provides design seeds and tailored design space to enhance principle-driven de novo methods such as RosettaDesign.

Methodology:

Trains a conditional WGAN on existing protein sequence–structure data conditioned on a low-dimensional fold representation, employs a semisupervised learning scheme using labeled and unlabeled sequences, and integrates an ultrafast sequence-to-fold predictor (oracle) as feedback within the loss function.