ProteinSolver

ProteinSolver designs protein sequences that fold into specified three-dimensional geometries by applying deep graph neural networks to the protein sequence design problem.

Key Features:

• Deep Graph Neural Networks: Employs a graph neural network architecture to model protein sequence design as a constraint satisfaction problem (CSP).
• Extensive Training Dataset: Trained on a dataset of over 70 million real protein sequences corresponding to more than 80,000 known structures.
• Rapid Sequence Design: Generates candidate protein sequences rapidly that conform to specified target geometries.
• In Silico Benchmarking: Evaluates designed sequences using energy-based scores, molecular dynamics simulations, and structure prediction methods.
• Experimental Validation: Demonstrated proof-of-principle by generating sequences matching the serum albumin structure, with top designs synthesized and assessed by circular dichroism.

Scientific Applications:

• Protein engineering: Produces novel sequences for engineering proteins with defined structural frameworks.
• Structural biology: Tests hypotheses about sequence–structure relationships and folding outcomes.
• Synthetic biology: Supplies designed protein sequences for constructing synthetic biomolecular systems.
• Drug discovery: Supports design of protein scaffolds and potential binders relevant to therapeutic development.
• Enzyme design: Enables design of enzymes with tailored structures for specific catalytic functions.
• Biomaterials: Facilitates creation of protein-based biomaterials with specified geometries.

Methodology:

Frames protein sequence design as a constraint satisfaction problem and applies deep graph neural networks; trained on >70 million sequences corresponding to >80,000 structures; evaluates designs using energy-based scores, molecular dynamics simulations, and structure prediction methods.