How reliably can we predict the reliability of protein structure predictions?

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Biology

Background:

• Comparative methods are standard for in silico protein structure prediction, relying on multiple sequence alignments.
• Improving multiple alignment quality is crucial, as errors lead to inaccurate predictions.
• Weak evolutionary signals can hinder accurate alignment, challenging structure prediction.

Purpose of the Study:

• To evaluate the utility of alignment posterior probabilities in assessing the reliability of protein structure predictions.
• To investigate the correlation between alignment probabilities and secondary structure prediction accuracy.
• To explore the relationship between alignment probabilities and 3D structural information.

Main Methods:

• Utilized stochastic sequence alignment methods to generate posterior distributions of multiple alignments.
• Calculated alignment posterior probabilities for each column within the distribution.
• Predicted secondary structures using four distinct protocols on the HOMSTRAD database.

Main Results:

• Alignment posterior probabilities correlate with secondary structure prediction reliability, with varying strengths across protocols.
• Secondary structure prediction reliability shows the strongest correlation with posterior probabilities derived from the full alignment distribution.
• Alignment posterior probabilities also correlate with 3D distances between C-alpha atoms in superimposed structures.

Conclusions:

• Alignment posterior probabilities offer a priori error detection for comparative protein models at the sequence alignment level.
• These probabilities serve as a valuable metric for assessing the confidence in predicted protein structures.