Calibrating E-values for hidden Markov models using reverse-sequence null models.

Abstract

MOTIVATION: Hidden Markov models (HMMs) calculate the probability that a sequence was generated by a given model. Log-odds scoring provides a context for evaluating this probability, by considering it in relation to a null hypothesis. We have found that using a reverse-sequence null model effectively removes biases owing to sequence length and composition and reduces the number of false positives in a database search. Any scoring system is an arbitrary measure of the quality of database matches. Significance estimates of scores are essential, because they eliminate model- and method-dependent scaling factors, and because they quantify the importance of each match. Accurate computation of the significance of reverse-sequence null model scores presents a problem, because the scores do not fit the extreme-value (Gumbel) distribution commonly used to estimate HMM scores' significance. RESULTS: To get a better estimate of the significance of reverse-sequence null model scores, we derive a theoretical distribution based on the assumption of a Gumbel distribution for raw HMM scores and compare estimates based on this and other distribution families. We derive estimation methods for the parameters of the distributions based on maximum likelihood and on moment matching (least-squares fit for Student's t-distribution). We evaluate the modeled distributions of scores, based on how well they fit the tail of the observed distribution for data not used in the fitting and on the effects of the improved E-values on our HMM-based fold-recognition methods. The theoretical distribution provides some improvement in fitting the tail and in providing fewer false positives in the fold-recognition test. An ad hoc distribution based on assuming a stretched exponential tail does an even better job. The use of Student's t to model the distribution fits well in the middle of the distribution, but provides too heavy a tail. The moment-matching methods fit the tails better than maximum-likelihood methods. AVAILABILITY: Information on obtaining the SAM program suite (free for academic use), as well as a server interface, is available at http://www.soe.ucsc.edu/research/compbio/sam.html and the open-source random sequence generator with varying compositional biases is available at http://www.soe.ucsc.edu/research/compbio/gen_sequence