Fold2Seq

Fold2Seq generates protein sequences conditioned on target 3D topological folds using a transformer-based generative framework to support protein engineering and protein design.

Key Features:

• Transformer-Based Architecture: Learns sequence embeddings and captures long-range dependencies in protein sequences via a transformer model.
• Fold Embedding via Secondary Structural Elements: Represents target folds using densities of secondary structural elements within 3D voxels to encode spatial topology.
• Performance and Reliability: Demonstrates improved or comparable performance versus data-driven deep generative models and physics-based approaches like RosettaDesign on test sets with single, high-resolution, complete structure inputs, evaluated for speed, coverage, and reliability.
• Robustness to Imperfect Inputs: Offers advantages over structure-based deep models and RosettaDesign when handling low-quality, incomplete, or ambiguous input structures.

Scientific Applications:

• Protein engineering: Design novel protein sequences that conform to specified 3D topological folds.
• Drug discovery: Generate candidate protein sequences with target structural characteristics relevant to therapeutic design.
• Synthetic biology: Create diverse protein sequences with desired folds for engineered biological systems.

Methodology:

Jointly learns sequence embeddings with a transformer model while embedding target folds using 3D voxel-based densities of secondary structural elements.