Reply to Tan et al.: Differences between real and simulated proteins in multiple sequence alignments.

Tan et al. (1) comment on our earlier paper regarding the accuracy of multiple sequence alignments (MSAs) using different guide tree topologies (2). We stress that the scope of our result was confined to alignments of very large numbers of protein sequences with known structures, where accuracy was measured against structure-based alignments. We point out that this result could not be translated to a strictly phylogenetic view. Tan et al. (1) demonstrate that, using a phylogenetic perspective, one can get the opposite result to ours. Given how they configure their test system, Tan et al.'s result is to be expected and easy to explain. If one simulates MSAs with many indels at random locations and then tests correspondence between alignments, including gaps in the test, then guide tree topology must have a huge effect. This is more or less inevitable. Our benchmark test sets do not have gaps at random locations. Gaps are mostly confined to loops between the main secondary structure regions. During evolution indels may occur in secondary structure elements, but rarely. Occurring indels may be cancelled out by compensating events that restore length and periodicity of the element. In contrast, gaps in loops are common and tolerated during evolution. This extreme imbalance in indel frequency has been well known for decades (3). The parameterization for the ALF simulation program comes partly from ref. 4, which describes such an imbalance with indels predominantly at exposed positions in structures. ALF can be used to simulate alignments with indel probabilities across sites from a distribution. Tan et al. (1) chose a uniform distribution. One has to ask what kind of sequences these simulated ones might be most similar to in reality. What kinds of biological sequences allow indels equally easily at any position? Such sequences may exist in intergenic regions but will be difficult to align after even moderate sequence divergence. Equal probabilities of indels at all sites suggest sequences not under any selective or structural constraint. All our sequences are proteins with 3D structural information and constrained structure. Our main tests used a combination of PFAM sequences and Homstrad structure-based alignments. We also used Balibase but only to make a minor point. With the large tests the effect we described was mainly clear for more than 1,000 sequences; that is the upper limit of the tests in Tan et al. (1). On a much smaller scale, we can see …