Decreasing the number of false positives in sequence classification.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Probabilistic models in sequence analysis often assign probabilities to most sequences.
• Bayesian binary classification with null models is common in tools like HMMER, SAM, and INFERNAL.
• Null models, such as uniform or genomic distributions, are used for comparison.

Purpose of the Study:

• To evaluate the impact of different null models on sequence classification results.
• To minimize false predictions and reduce biological validation costs.
• To compare the performance of various null models in sequence analysis.

Main Methods:

• Assessed null models including uniform, genomic, family-specific, and target distributions.
• Utilized Bayesian binary classification for sequence analysis.
• Conducted tests on randomly generated DNA and protein sequences, and a benchmark with P. falciparum.

Main Results:

• The target null model consistently yielded the lowest number of false positives with random sequences.
• The uniform model showed a GC bias, potentially increasing false positives for sequences with extreme compositional bias.
• Benchmark tests confirmed the superiority of the target model in specific cases.

Conclusions:

• Both uniform and target null models are suitable for classification.
• The target model offers higher specificity and is more dependable for biological validation, particularly with compositionally biased sequences.
• The GC bias of the uniform model is a critical factor not previously highlighted.