Estimating and evaluating the statistics of gapped local-alignment scores

Area of Science:

• Bioinformatics
• Computational Biology
• Statistical Modeling

Background:

• Accurate estimation of alignment score distributions is crucial for biological sequence database searches.
• Existing methods like regression-based and lookup tables have limitations in precision.

Purpose of the Study:

• To develop and validate a novel maximum-likelihood-based algorithm for estimating score distributions.
• To compare the accuracy of this new algorithm against existing methods.
• To investigate the relationship between statistical accuracy and classification accuracy of p-values.

Main Methods:

• Developed a maximum-likelihood-based algorithm for score distribution estimation.
• Introduced a new method for measuring p-value accuracy.
• Explored mixture models and expectation maximization for score distribution modeling.
• Assessed classification accuracy using estimated p-values.

Main Results:

• The maximum-likelihood algorithm demonstrated superior accuracy over regression and lookup table methods.
• Ignoring low E-value scores during estimation did not significantly improve results over simpler methods.
• Paradoxically, less statistically accurate p-values sometimes yielded higher classification accuracy.

Conclusions:

• The maximum-likelihood approach provides a more statistically accurate method for estimating alignment score distributions.
• Statistical accuracy, not classification accuracy, should be the primary metric for evaluating similarity search methods.
• Further research into sophisticated modeling techniques did not yield significant improvements over simpler approaches.