Application of protein structure alignments to iterated hidden Markov model protocols for structure prediction

Abstract

Background: One of the most powerful methods for the prediction of protein structure from sequence information alone is the iterative construction of profile-type models. Because profiles are built from sequence alignments, the sequences included in the alignment and the method used to align them will be important to the sensitivity of the resulting profile. The inclusion of highly diverse sequences will presumably produce a more powerful profile, but distantly related sequences can be difficult to align accurately using only sequence information. Therefore, it would be expected that the use of protein structure alignments to improve the selection and alignment of diverse sequence homologs might yield improved profiles. However, the actual utility of such an approach has remained unclear.

Results: We explored several iterative protocols for the generation of profile hidden Markov models. These protocols were tailored to allow the inclusion of protein structure alignments in the process, and were used for large-scale creation and benchmarking of structure alignment-enhanced models. We found that models using structure alignments did not provide an overall improvement over sequence-only models for superfamily-level structure predictions. However, the results also revealed that the structure alignment-enhanced models were complimentary to the sequence-only models, particularly at the edge of the "twilight zone". When the two sets of models were combined, they provided improved results over sequence-only models alone. In addition, we found that the beneficial effects of the structure alignment-enhanced models could not be realized if the structure-based alignments were replaced with sequence-based alignments. Our experiments with different iterative protocols for sequence-only models also suggested that simple protocol modifications were unable to yield equivalent improvements to those provided by the structure alignment-enhanced models. Finally, we found that models using structure alignments provided fold-level structure assignments that were superior to those produced by sequence-only models.

Conclusion: When attempting to predict the structure of remote homologs, we advocate a combined approach in which both traditional models and models incorporating structure alignments are used.

Figure 1

Relative performance of single-master HMMs, SLAHMMs, and the combined models with differing iteration parameter sets (PS), presented as a coverage vs. theoretical errors per query (EPQ) plot. The different parameter sets are defined in Table 2 and explained in the text. Values for correct assignments are truncated at 600 in order to emphasize differences between the various methods (no method had an error below 600 correct assignments).

Figure 2

Relative performance of different types of HMMs in assignment of structure to sequence probes, presented as a coverage vs. error plot. SLAHMM-CW refers to models built in the same way as SLAHMMs, but using only sequence information to align the SCOP domains rather than a structural alignment (see text). Iterative parameters used for construction of all models were from PS1 (Table 2). Values for correct assignments are truncated at 600 in order to emphasize differences between the various methods (no method other than SLAHMM-CW had errors below 600 correct assignments).

Figure 3

Venn diagram describing coverage overlap of the three primary model sets from PS1, when using a strict cutoff of 80 incorrect assignments (theoretical EPQ ~0.05). The numbers shown in parentheses near each model type designation refer to the total number of correct matches made by that model type prior to the cutoff point. Identical matches of the same probes by some or all of the three different methods are provided by the numbers in the set diagram. The completely unique matches by single-master HMMs and SLAHMMs are color coded to match the circle for that model type. "All Models" denotes the assignments made by the combined database of SLAHMMs and single-master HMMs used together in a single search.

Figure 4

Relative performance of different types of HMMs in assignment of fold-level structure to sequence probes, presented as a coverage vs. error plot. Details of model types are provided in the text. Iterative parameters used for construction of all models were from PS1 (Table 2).

Figure 5

Comparison of HMMs built using an older protein sequence database for iterative construction ("old db") with those built using a current sequence database ("new db"), presented as a coverage vs. error plot. Results are colored similarly for corresponding model types, with the results based on the older database in a lighter color. A different version of the HMMER software was also used for the two result sets; details of model types and construction are provided in the text. Iterative parameters used for construction of all models were from PS1 (Table 2).