Sequence comparison has become a major tool for biological research in the post-genomic era, forming the basis for functional annotation, classification, and analysis of evolutionary relationships. At the residue level, however, the relation between sequence, structure and function can often be obscure, and examples abound of proteins with a clear functional and homologous relationship but sharing negligible similarity at the sequence level.

Structural alignment therefore is the method of choice for reliable homology assessment and derived features like functional classification and phylogeny. This importance is reflected in the number of tools available for structural alignment, such as DALI [¹], SSAP [²], STRUCTAL [³], MAMMOTH [⁴], CE [⁵] and COMPARER [⁶] (for recent reviews on the topic, see Kolodny et al. [⁷] and Mayr et al. [⁸]). Databases for functional classification such as CATH [⁹], FSSP [¹⁰] and PASS2 [¹¹] each derive directly from the use of one or more of these methods, whereas for SCOP expert input in the structural classification is deemed critical [¹²]. Structural alignments are also routinely used for benchmarking sequence alignment methods. A number of databases have been developed for this purpose, among which BAliBASE [¹³], HOMSTRAD [¹⁴] and SABmark [¹⁵] are widely used. These databases often rely on expert knowledge and include a notion of 'core blocks', i.e. where alignment ambiguity does not occur and hence can be trusted. The general problem of uncertainty in sequence alignment has recently been discussed by Wong et al. [¹⁶]. Due to the complexity of interpreting non-trivial alignment regions, these are often omitted in large-scale evolutionary analyses, even though there is ample evidence for their fundamental importance [¹⁶,¹⁷]. An approach to pinpointing alignment ambiguity is the generation of ensembles of suboptimal alignments [¹⁸], but computational demands remain prohibitive for genome wide studies.

Recent structural alignment methods have started to place emphasis on dealing with structural flexibility, such as FATCAT [¹⁹], MultiProt [²⁰], MATT [²¹] and RAPIDO [²²]. This may increase the consistency of alignments produced by each of these methods, but does not address the intrinsic ambiguity arising from structural divergence. The fundamental issue is whether a one-to-one equivalence exists between residues from different proteins that could be expressed as one definite alignment between sequences [¹⁸]. This is illustrated in Figure 1, where we show that a single insertion can lead to ambiguity in the functional correspondence between most residues in the loop.

To further elucidate the effect of structural diversity on structural alignment, we prepared two distinct comprehensive sets of alternative structures for proteins from the HOMSTRAD database of homologous protein families. The first set comprises proteins for which alternative crystal structures are available. The other set is derived from molecular dynamics simulations to explore a more extensive spectrum of possible structures. An overview of our analysis procedure is outlined in Figure 2.

Our main results show that in many cases structural variation strongly affects structural alignments, even for highly similar sequences. Moreover, the derived alignment appears to be highly sensitive to even small conformational changes of the proteins. The uncertainty in pairing up structural equivalent residues makes it difficult to determine which alignment alternative would describe most closely the functional relationship between the proteins. To address this issue, we show how alternative alignment visualizations may be used to exploit the information contained within variable alignment regions.

Structural variation, as presented here by alternative crystal structures and molecular dynamics simulations, has a profound effect on structural alignment. The sensitivity to structural variation is a bottleneck for the effective application of structural alignment approaches. This undermines the current basis of all sequence alignment methodologies and is an underestimated problem for the homology assessment used in structural and functional classification. The GroES 'mobile loop' example demonstrates how functionally essential protein regions can coincide with variable structural alignment segments. Our database should therefore be useful for alignment verification and delineation of functionally important protein regions.

The HOMSTRAD database of homologous structure alignments [¹⁴] was used as a source to select homologous proteins with known structure. HOMSTRAD families containing two homologous proteins (A and B in Figure 2) were selected. The corresponding structures were retrieved from the PDB [²⁸] and taken as reference. For each reference structure, after equilibration, molecular dynamics simulations were performed for up to 10 ns, and snapshot structures were stored every 1 ns. Standard solvated conditions in the Gromos 43a1 forcefield [²⁹] and the Gromacs simulation package [³⁰] were used (details summarized in additional file 2). In addition, for each reference structure, we retrieved all alternative PDB structures with 100% sequence identity. In the subsequent analysis only the residues corresponding to the HOMSTRAD sequences were used.

From each pair of reference HOMSTRAD structures, we constructed reference alignments with the widely used structural alignment tool DALI [¹]. We also used DALI to create pairwise alignments between each reference structure and the alternatives of the other reference structure (PDB and snapshots). The sequence differences between the alignments were calculated using Sum-of-Pairs (SP) scoring implemented in the BAliBASE alignment comparison tool [¹³]. SP scores range from 0 (non-identical) to 1 (identical sequence alignments). Finally we calculated the root mean square deviation (RMSD) between the C? atoms of the alternative structures and their reference structure using the McLachlan algorithm [³¹] as implemented in the program ProFit version 2.5.3 [³²].

Our final database consists of 496 proteins (divided over 341 families) for which 3309 snapshot structures could be made and 565 proteins (divided over 395 families) for which we found in total 2998 alternative crystal structures with redundant sequences. A full list of all aligned structures and relevant details is provided in additional file 3.

All authors designed the research, analyzed the results and wrote the paper. WP and KAF performed the research. All authors read and approved the final manuscript.