FUpred

FUpred predicts protein domain boundaries from sequence by integrating contact maps generated by deep residual neural networks with coevolutionary precision matrices to optimize intra-domain contacts and minimize inter-domain contacts.

Key Features:

• Contact map generation: Generates contact maps using deep residual neural networks.
• Coevolutionary integration: Incorporates coevolutionary precision matrices alongside contact maps.
• Boundary optimization: Identifies domain boundary locations by optimizing the number of intra-domain contacts while minimizing inter-domain contacts.
• Discontinuous and multi-domain detection: Predicts discontinuous domains and multi-domain compositions from sequence alone.
• Benchmarking: Evaluated on 2,549 proteins and achieved a Matthew's correlation coefficient (MCC) of 0.799 for single- and multi-domain classification.
• Comparative performance: Reported to outperform the best existing machine learning or threading-based methods by 19.1% (or 5.3%); for discontinuous domains achieved a domain boundary detection score of 0.788 and a normalized domain overlapping score of 0.521, improvements of 17.3% and 23.8% versus the best control method.

Scientific Applications:

• Domain composition inference: Determining domain compositions from protein sequences for single-, multi-, and discontinuous-domain proteins.
• Structural biology: Providing accurate domain boundaries to support structural analysis and interpretation.
• Protein structure and function analysis: Facilitating analyses that link domain architecture to protein function and modeling.

Methodology:

Integrates contact maps produced by deep residual neural networks with coevolutionary precision matrices and locates domain boundaries by optimizing intra-domain contacts while minimizing inter-domain contacts.