The Janus-faced E-values of HMMER2: extreme value distribution or logistic function?

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Protein domain annotation using tools like HMMER is crucial for understanding protein function.
• HMMER2 remains relevant for ongoing research and database evaluation, despite the release of HMMER3.
• E-value computation in HMMER2 involves a logistic function or an extreme value distribution (EVD), with the lower value reported.

Purpose of the Study:

• To investigate the impact of HMMER2's E-value calculation method on protein domain annotation.
• To identify potential conflicts arising from the switch between logistic function and EVD in E-value computation.
• To evaluate the reliability of automated annotation decisions influenced by these E-value calculations.

Main Methods:

• Analysis of E-value computation in HMMER2 across thousands of domain models using Pfam release 23.
• Comparison of E-values derived from the logistic function versus EVD at critical score regions.
• Examination of annotation discrepancies in global and fragment search modes.

Main Results:

• A switch from EVD to the logistic function occurs for a significant percentage of domain models as scores increase (99% in global mode, 75% in fragment mode).
• A critical score region with conflicting E-values exists when the breakpoint E-value exceeds a user-defined threshold (e.g., 0.1).
• This conflict impacts automated annotation for 185 domain models in hmmpfam and 1,748 in hmmsearch global modes, affecting 24.4% of hits.

Conclusions:

• The automatic switch to a logistic function in HMMER2 can lead to conflicting E-values and potentially false protein domain annotations.
• The logistic function may not always be an appropriate alternative to EVD for E-value calculation in HMMER2.
• Annotation decisions guided by HMMER2's E-values, especially in automated pipelines, require careful consideration due to these statistical model switching issues.