The Transitive/Tree-based Consistency Score (TCS) web server allows biologists to estimate the local reliability of protein multiple sequence alignments (MSAs) for structural and evolutionary modeling purposes using the TCS metrics (1). The server makes it possible to colorize a pre-computed MSA according to its reliability and to filter out or down-weight unreliable positions for further evolutionary and structural analyses. A pre-computed MSA can be estimated from any available third party method. The web server is fully connected with the PhyML (2) tree reconstruction web server. In theory, it can also be used for DNA and RNA analyses though the method has only been validated on proteins so far.

The TCS server addresses an important need in the biological community. While the development of novel MSA methods has received a large amount of attention over the last two decades, the problem of quantifying the relative reliability of such alignments has been a less popular investigation topic. This is certainly surprising, given the approximate nature of all available algorithms and their frequently reported incapacity to deliver structurally correct models, even when using the best available methods (3). In practice, most publications describing novel MSAs come along with benchmarks that report the average accuracy of given software on specific benchmarks—usually made of structure-based MSAs. Given the variability in performance of even the best packages, these average figures are of little practical use for biologists. They merely allow bet hedging when picking up a given method, but they give no guarantee of estimating the best possible model on a given data set (4). This falls short of biologists’ expectations, which mostly involve assembling the best possible MSA or identifying the trust worthiest columns when doing homology or phylogenetic modeling.

The issue of identifying the most evolutionary or structural correct positions in an MSA has been addressed from two complementary perspectives. When doing phylogenetic reconstruction, the most widely used methods for MSA post-processing, Gblocks (5) and trimAl (6), filter out positions on the basis of their conservation and indel indexes. This filtering is done under the assumption that such positions are more likely to be inaccurately aligned. These methods have initially been validated on simulated data sets for their capacity to improve phylogenetic reconstruction although effectiveness of this approach was recently questioned in two independent reports (1,7). The most recent generation of methods (HoT (8), GUIDANCE (9), PSAR (10)) estimate reliability indexes by considering the stability of the MSA under some sort of re-alignment perturbation (sequence flipping, guide tree variation or sequence removal). TCS belongs to this group of methods and makes it possible to estimate the reliability of any position in an MSA by considering the stability of its alignment when independently re-aligning all possible pairs of sequences within the considered MSA.

In a recent publication (1), the TCS was extensively benchmarked for its capacity to identify positions in which the aligned residues are most likely to be structurally homologous. The results indicate that given the most widely used MSA methods, TCS is significantly more discriminative than HoT and GUIDANCE while being significantly faster than GUIDANCE. Regardless of the MSA methods, TCS can identify ~40% of the correctly aligned residues with a confidence of 90% or higher. Using appropriate simulated and empirical benchmarks, we were also able to show that the proper use of the scoring positions—through up-weighting or filtering—makes it possible to build significantly more accurate phylogenetic trees, with most alignment methods reaching comparable accuracy levels when doing TCS-based post-processing. Altogether, these results indicate that the TCS reliability metrics can be used to improve both the structural and the evolutionary modeling capacity of most multiple sequence aligners. Its systematic usage should help decrease the uncertainty resulting from the use of different alignment methods and therefore increase the robustness and reproducibility of biological results established on the basis of MSA modeling.

Given a pre-computed MSA, the TCS algorithm estimates the reliability of every pair of aligned residues (PairTCS), every individual residue (ResidueTCS), every MSA column (ColumnTCS), every MSA sequence (SequenceTCS) and the whole alignment (AlignmentTCS). The procedure implemented in the web server can be summarized as follows.

The first computation step involves extracting the sequences two by two in order to build a T-Coffee library. A T-Coffee library is a list of residue pairs weighted according to their probability to be part of the final MSA. These pairs do not have to be consistent with each other and a valid library can contain incompatible pairs. In theory users can compile these libraries using any suitable combination of pairwise and MSA methods. By default, the library is computed by applying a pair hidden Markov model adapted from ProbCons (11) onto every pair of sequences and by keeping all residue pairs having a posterior alignment probability higher than 0.9. For every pair of sequences, this roughly amounts to select the optimal alignment along with a few high scoring sub-optimal alignments. Other library schemes are supported including Fast-Coffee that involves running three fast aligners onto the sequences (MAFFT (12), MUSCLE (13) and Kalign (14)) and populating the library with all pairs of residues aligned in the three resulting MSAs. The library computation is the most computationally intense step and therefore the most limiting. The running time of default library requires O(N²L²), where N is the number of sequences and L is the length of the MSA.

The library is then used to estimate the TCS score of every pair of aligned residues in the target MSA using the following procedure (see (1) for details). Given two aligned residues R^x_i (ith residue of sequence x) and R^y_j (jth residue of sequence y) forming the (R^x_i,R^y_j) pair within the target MSA, the TCS score is estimated by identifying within the library all compatible pairs, R^x_iR^z_k and R^z_kR^y_j, containing a common intermediate residue, R^z_k, from a third sequence z. Two compatible pairs define a triplet whose score is assigned the lowest score of the two pairs. All triplets’ scores are then added up and the resulting sum is normalized by the sum of the scores of all the pairs involving either R^x_i or R^y_j ; thus yielding a value between 0 and 1 named PairTCS of (R^x_i,R^y_j).

The other metrics are directly derived for the PairTCS. The ResidueTCS score is assigned to every residue and obtained by averaging the PairTCS score over all pairs involving the considered residue and any another residue appearing in the same column of the target MSA. The ColumnTCS score is estimated by averaging the PairTCS score over all pairs of aligned residues in each target MSA column. The SequenceTCS and the AlignmentTCS are obtained by averaging all the ResidueTCS in the sequence and over the entire alignment, respectively. The ResidueTCS, ColumnTCS, SequenceTCS and AlignmentTCS are available in score_ascii output format of the web server. The PairTCS is not a default output (sp_ascii) and is only available using the command line version of T-Coffee (http://tcoffee.readthedocs.org/en/latest/TCS.html).

Once the metrics have been computed, they are used for filtering out the less reliable columns (or residues), coloring every position in the target MSA and generating bootstrap replicates (for tree reconstruction) using the ColumnTCS as a weight when randomly selecting columns for inclusion in a bootstrap replicate with the same column size of the target MSA (1). The filtered MSAs and the bootstrap replicates can be either downloaded for further usage or sent to the PhyML web server. PhyML can accept the replicates input as multiple data sets (option ‘Number of data sets’ in PhyML server) and it will generate an individual tree for each replicate instead of the consensus-bootstrap tree.

Input

MSAs (CLUSTAL, FASTA, MSF, Phylip or Nexus format) can either be cut/pasted or uploaded. For unaligned sequences, users can run any T-Coffee method first and then click a blue button ‘Core/TCS’ in the ‘Send results’ section, which will automatically initialize pre-filled TCS service with T-Coffee MSA as input. By default the server is set to compute a Mproba_pair T-Coffee pairwise library and to generate the filtered alignments along with its other outputs by removing all columns having a scaled ColumnTCS score (range 0–9) less than 4. Users can change this behavior by triggering the ‘Show more options’ hyperlink that makes it possible to change the library computation methods, the filtering threshold and/or the filtering on residues rather than columns. The advance mode also makes it possible to retain empty columns after filtering (Figure ​(Figure1c).1c). Email address is optional but advisable when submitting computationally intensive jobs. The ‘Submit’ button triggers the job processing.

Open in a separate window

Figure 1.

(a) Graphic output (score_html). The SCORE is the AlignmentTCS normalized to a maximum of 1000. The sequence list below indicates the relative SequenceTCS of the considered sequences with values normalized to a maximum of 100. In the alignment below, residues are color coded from blue (ResidueTCS = 0) to dark pink (ResidueTCS >= 9). The line below the alignment indicates the consensus score of every column (ColumnTCS). (b)score_ascii output of the same alignment. Residues are replaced by their ResidueTCS normalized between 0 and 9. (c) Filtered output in CLUSTAL format. Columns having ColumnTCS less than 4 have been removed (default). The advanced option, Remove empty columns, is disable such that empty columns are kept for illustration purposes.

We describe the TCS server, a web-based tool able to identify the most correct regions within any pre-computed MSA, regardless of the chosen aligner. To the best of our knowledge, TCS is the first reliability index that has been validated for its relevance at improving both homology modeling and phylogenetic reconstruction. The TCS server also makes it possible to depart from using the standard similarity based filtering schemes of Gblocks or trimAl thus allowing more information to be retained when estimating trees. Overall, this server is able to reduce alignment error (uncertainty) on downstream biological analyses. Future improvements will extend TCS capacity to handle RNA alignment.