Protein structure alignment significance is often exaggerated

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning Applications

Background:

• Machine learning models generate vast numbers of predicted protein structures.
• Existing structure search algorithms struggle with computational efficiency and accurate statistical significance estimation at scale.
• Convergent evolution leads to similar structural motifs in unrelated proteins, inflating false positive rates in alignments.

Purpose of the Study:

• To develop computationally efficient algorithms for searching large protein structure databases.
• To provide robust statistical significance estimates for protein structure alignments.
• To address the overestimation of significance by current methods due to convergent evolution.

Main Methods:

• Investigated the tendency of unrelated proteins to evolve similar motifs.
• Analyzed the performance of popular structure search and alignment algorithms.
• Developed and validated a novel method for estimating statistical significance (E-values) in large-scale structure searches.

Main Results:

• Unrelated proteins exhibit convergent evolution of structural motifs, leading to false positives.
• Previous methods overestimated statistical significance by up to six orders of magnitude.
• The novel E-value estimation method is accurate, scalable with database size, and robust to fold diversity.

Conclusions:

• Accurate statistical significance estimation is crucial for large-scale protein structure analysis.
• The new method provides reliable significance values, overcoming limitations of prior approaches.
• An online service implementing the novel method is available at https://reseek.online.