FUpred: detecting protein domains through deep-learning-based contact map prediction

Area of Science:

• Computational biology
• Structural bioinformatics
• Protein domain analysis

Background:

• Protein domains are independently folding and functional units crucial for protein structure and function.
• Accurate domain boundary assignment is essential but challenging, especially for proteins with discontinuous domains.
• Existing algorithms lack efficiency and accuracy in predicting domain boundaries from sequence alone.

Purpose of the Study:

• To develop an efficient and accurate algorithm for predicting protein domain boundaries from sequence.
• To specifically address the challenge of predicting domains in proteins with discontinuous domain architectures.
• To improve the understanding of protein structure and function through precise domain identification.

Main Methods:

• Developed FUpred, a novel algorithm utilizing deep residual neural networks and coevolutionary precision matrices to generate contact maps.
• Employed a strategy to maximize intra-domain contacts and minimize inter-domain contacts within predicted contact maps.
• Validated FUpred on a large dataset of 2549 proteins, including those with discontinuous domains.

Main Results:

• FUpred achieved a Matthew's correlation coefficient of 0.799 for single- and multi-domain classifications, outperforming existing methods by 19.1% (machine learning) and 5.3% (threading).
• For discontinuous domains, FUpred demonstrated superior performance with domain boundary detection (0.788) and normalized domain overlapping scores (0.521), exceeding control methods by 17.3% and 23.8%, respectively.
• The algorithm successfully identifies domain composition from sequence data, offering a significant advancement for complex protein structures.

Conclusions:

• FUpred provides a highly accurate method for predicting protein domain boundaries directly from amino acid sequence.
• The algorithm offers a breakthrough in analyzing proteins with discontinuous domains, a previously challenging area.
• This work establishes a new pathway for precise protein domain composition analysis, particularly from sequence data alone.