DeepIDDP

DeepIDDP predicts inter-domain distance maps to improve assembly and full-chain structure prediction of multi-domain proteins.

Key Features:

• Neural network with attention mechanisms: Uses a neural network architecture incorporating attention mechanisms to capture complex inter-domain interactions.
• Novel inter-domain features: Introduces two novel inter-domain features specifically tailored to enhance prediction of distances between domains.
• Data enhancement (DPMSA): Employs the DPMSA data enhancement strategy to mitigate the absence of co-evolutionary information in target proteins.
• Integration with SADA (SADA-DeepIDDP): Integrates predicted inter-domain distances into the SADA domain assembly method, producing the hybrid SADA-DeepIDDP model.

Scientific Applications:

• Benchmarking vs trRosettaX/trRosetta: On multi-domain protein benchmarks, SADA-DeepIDDP improved inter-domain distance prediction accuracy by 11.3% versus trRosettaX and 21.6% versus trRosetta.
• Domain assembly improvement: The hybrid domain assembly model increased domain assembly accuracy by 2.5% relative to SADA.
• Reassembly of AlphaFold models: Reassembling human multi-domain models from the AlphaFold database with TM-scores ≤ 0.80 produced an average TM-score improvement of 11.8%.
• Research utility: Applicable to structural biology and computational biochemistry for studying complex protein architectures, probing protein function, and prioritizing therapeutic targets.

Methodology:

Neural network with attention mechanisms; two novel inter-domain features; DPMSA data enhancement strategy; integration of predicted distances into SADA to form SADA-DeepIDDP.