DEMO2

DEMO2 assembles full-length structural models of multi-domain proteins by predicting inter-domain spatial restraints with deep residual convolutional networks and integrating those restraints with analogous multi-domain template alignments from the PDB to guide L-BFGS-based model construction.

Key Features:

• Domain-level input: Starts from individual domain models as the basis for full-length assembly.
• Inter-domain restraint prediction: Predicts inter-domain spatial restraints using deep residual convolutional networks.
• Template integration: Incorporates analogous multi-domain template alignments sourced from the Protein Data Bank (PDB).
• Hybrid energy function: Uses a hybrid energy function that combines predicted deep-learning restraints with template-based restraints.
• Model construction: Constructs full-length structural models via L-BFGS simulations guided by the hybrid energy function.
• Model ensembles: Generates up to five distinct full-length structure models per protein.
• Output details: Outputs deep-learning inter-domain restraints and top-ranked multi-domain structure templates.
• Benchmark performance: Demonstrated superior performance in large-scale benchmarks and the blind CASP14 experiment.

Scientific Applications:

• Drug discovery: Provides full-length multi-domain protein models to inform target characterization and structure-based studies.
• Functional annotation: Enables inference of protein function and domain interplay from assembled full-length structures.
• Evolutionary studies: Supports analysis of domain architecture and evolutionary relationships across multi-domain proteins.

Methodology:

Starts from individual domain models, predicts inter-domain spatial restraints using deep residual convolutional networks, and constructs full-length models via L-BFGS simulations using a hybrid energy function that combines the predicted deep-learning restraints with analogous multi-domain template alignments from the PDB.