ProDESIGN-LE

ProDESIGN-LE designs protein sequences by predicting amino acid types for positions in a target backbone using a transformer trained on concise local-environment representations.

Key Features:

• Local-environment representation: Uses a concise representation of each residue's local environment to capture interactions within its immediate structural surroundings.
• Transformer model: Trains a transformer to learn correlations between local environments and amino acid types for residue prediction.
• Sequence assignment: Assigns residue types to positions in a target backbone based on model predictions to ensure structural compatibility.
• Performance: Designs sequences with an average runtime of approximately 20 seconds per protein.
• Structural fidelity: Produces designed sequences whose predicted structures have an average TM-score exceeding 0.80 to their target structures.

Scientific Applications:

• Rational protein engineering: Designs sequences for diverse proteins, including naturally occurring and hallucinated (hypothetical) proteins, to explore or optimize protein structure and function.

Methodology:

The method uses a concise representation of each residue's local environment, trains a transformer model to learn relationships between local environments and amino acid types, and uses the trained model to assign residue types to a target backbone.