KMAD

1 Introduction

More than 30% of all human proteins contain unfolded regions (Pentony and Jones, 2010). This stands in marked contrast to how little we know about them (Van der Lee et al., 2014). Intrinsically disordered proteins (IDPs) lack a stable tertiary structure, and thus lack a hydrophobic core. They thus also lack an active site. IDPs often interact with other proteins by means of short linear motifs (SLiMs), which are very abundant in disordered regions. This is a fundamental component of cell signalling (Gibson, 2009). These IDP characteristics have many consequences for aligning their sequences and thus for their study in general. Algorithms underlying existing MSA software are directly or indirectly based on knowledge obtained from studying 3D protein structures. Aligning a hydrophobic residue in the protein core with a hydrophilic one is penalized heavily when aligning ordered proteins, but this is much less the case when aligning IDPs simply because IDPs do not have a hydrophobic core. In addition, folded proteins contain large protein–protein interaction surfaces, ion binding sites, active sites or other 3D motifs that MSA software capitalizes on, while IDPs generally consist of short motifs surrounded by stretches of highly variable length and composition. The main function determinants of IDPs are SLiMs, and posttranslational modifications (PTMs).

For the specific purpose of aligning IDPs, we introduce Knowledge-based multiple sequence alignment (MSA) for IDPs, or KMAD, that incorporates SLiM, domain and PTM annotations in the alignment procedure. Obviously, the inclusion of this knowledge will cause these motifs to line up in the final MSA without certainty that this reflects biological reality. This way, however, KMAD generates hypotheses that can be validated experimentally to progress, for example, protein engineering, drug design or the analysis of genetic disorders. In these research fields, scientists most often are interested in one single protein and want to gather information for this one protein. In the particular case of IDPs such information normally relates to SLiMS and PTMs and how conserved these are in an MSA. In the HSSP project (Sander and Schneider, 1991), the concept of the insertion-free MSA was introduced specifically for this purpose (see Fig. 1). Although KMAD can produce ‘normal’ MSAs (i.e. with insertions and deletions possible in all sequences), we tend to use it insertion-free when studying IDP MSAs.

Open multimedia modal

MSA validation is a hotly debated issue that seems mostly unsolved (Iantorno et al., 2014). Even worse, for the same set of sequences different MSAs can be produced that all seem correct (Iantorno et al., 2014). Alignments extracted from structure superpositions are generally considered the gold standard, but many examples exist in which a structure superposition-derived MSA does not reflect well the sequence family’s phylogeny (Iantorno et al., 2014). The MSA benchmark suites BAliBASE (Thompson et al., 2005^; Perrodou et al., 2008) and PREFAB (Edgar, 2004) are mostly structure superposition based, albeit that superpositions used in these suites do not always agree (Edgar, 2010) with external sources such as CATH (Sillitoe et al., 2015) or SCOP (Andreeva et al., 2008). It is not clear how these discrepancies arose but it is known that the same structures sometimes can be superposed differently depending on the algorithm used and the parameter choices (Irving et al., 2001^; Konagurthu et al., 2006). Edgar (2010) recently studied quality measures for protein alignment benchmarks and concluded that ‘protein alignment assessment is more challenging than generally realized’; he also concluded that BAliBASE block identifications do not correlate well with conserved protein secondary structures. We used BAliBASE and PREFAB to validate KMAD’s MSAs and the MSAs produced by a series of well-known alignment programs, and conclude that the differences in alignment validation score between the six methods are smaller than the ‘noise’ in BAliBASE (www.cmbi.ru.nl/kmad/balibase/). It should be noted that KMAD was designed to produce insertion-free alignments (see Fig. 1) that best allow for transfer of information from the whole alignment to the one sequence of interest while BAliBASE and PREFAB contain ‘complete’ alignments that better reflect the underlying phylogeny.

2 Methods

Sickmeier et al., 2007Larkin et al., 2007Sievers et al., 2011Katoh et al., 2002Notredame et al., 2000Edgar, 2004The Uniprot Consortium, 2014Altschul et al., 1990EBateman et al., 2004Blom et al., 1999Dinkel et al., 2013Ashburner et al., 2000

The KMAD server can annotate IDPs in the MSA. IDPs are either obtained from the D2P2 database of disorder predictions (Oates et al., 2013), or by running the freely available disorder predictors: GlobPlot (Linding et al., 2003), DISOPRED (Ward et al., 2004), SPINE-D (Zhang et al., 2012), PSIPRED (Jones, 1999), PreDisorder (Cheng et al., 2005) and IUPred (Dosztányi et al., 2005).

KMAD was designed for the optimal alignment of IDPs. Proteins with a stable tertiary structure (non-IDPs) often also contain PTMs, SLiMs and domains. KMAD can be used for the alignment of non-IDPs too, but better results will be obtained if non-IDPs are first aligned with software optimized for that task (i.e. any software other than KMAD) followed by fine-tuning the resulting alignment with KMAD. We call this the refinement option of KMAD.

MSA quality was evaluated in two ways. First, MSAs made with Clustal Omega and KMAD were visually inspected in light of the known biology. Second, a quantitative analysis was performed with the protein linear motif benchmark for MSA tools from the BAliBASE suite, and with the PREFAB suite.

3 Implementation

Open multimedia modal

Sander and Schneider, 1991Vriend, 1990Joosten, 2007Fig. 1Venselaar et al., 2010Gotoh (1982)Needleman and Wunsch, 1970Midic et al., 2009Figure 2

R_fR_f

4 Results

KMAD overcomes several problems encountered with other methods when aligning IDPs. Two examples are shown in Figures 3 and 4. The first example (Fig. 3) illustrates the advantage of KMAD’s use of SLiMs. Clustal Omega spreads out the LIG_BIR_II_1 motif. In the KMAD alignment, the motif is lined up nicely in all sequences but 2. Even if this predicted SLiM should not be aligned as in the B panel of Figure 3, the bioscientist will still benefit from awareness of the presence of this conserved motif in the N-terminal segment.

Open multimedia modal

Open multimedia modal

The second example (Fig. 4) illustrates the use of PTMs. The phosphorylated serines at positions 74 and 75 in the first sequence are shifted to the left relative to the columns of phosphorylated serines at positions 75 and 76 in the other four sequences. A gap at position 73 in the query sequence would be a better solution in this case, but because we want to keep the first sequence indel-free, an insertion is introduced in all other sequences. The phosphorylated serine at position 94 in the first sequence seems better aligned with the serines at position 92 in sequences 2 and 3, and indeed, KMAD finds this solution.

Eventually, glycosylated asparagines at position 102 in sequences 4 and 5 should most probably be aligned to asparagines at position 105 from sequences 1–3. All problems mentioned above are solved by KMAD with no harm to the rest of the alignment. The problem that the predicted threonine phosphorylation sites around position 106 cannot be lined up is beyond KMAD’s reach.

Figures 3 and 4 provide two examples in which knowledge of the presence of sequence motifs is used to better align those motifs. This method is of course highly cyclic and we therefore compared the performance of KMAD with five other MSA methods using the BAliBASE and PREFAB suites. Most proteins in these benchmarks have a stable 3D structure, while KMAD was designed for the alignment of IDPs. We therefore used KMAD in refinement mode starting with Clustal Omega alignments. To avoid circularity when testing KMAD on the BAliBASE motif reference set we performed a leave-one-out validation, i.e. in each alignment the motif annotated by BAliBASE was excluded from the KMAD annotations.

The score differences in Table 1 are all within the margin of error, given that noise in the benchmarks (www.cmbi.ru.nl/kmad/balibase/) contributes a few percent to the scores.

Open multimedia modal

We chose four sequence families for testing the CPU performance of the six methods. The sequences in these families on average are about 150, 250, 550 and 750 amino acids long, respectively. From each family subsets of 5, 50, 500, 5 K, 20 K and 80 K were selected and these 24 groups of sequences were aligned with each of the six methods using each time one core on the same computer that had more than adequate RAM, so that the wall-time of each calculation provided a good measure for the CPU performance of the method. The results are visualized in Figure 5. In this plot, some points are missing because we stopped each alignment when after 1 week no result had been returned yet. Clearly, KMAD outperforms the other methods for large alignments.

Open multimedia modal

5 Discussion

In the field of sequence alignment research it is common practice to compare the sequence alignments obtained with MSA software with those that are obtained from structure superpositions (Nguyen and Pan, 2013). IDPs do not possess a static 3D structure so that this method is not applicable here. We have discussed several examples in which KMAD produces IDP alignments that intuitively feel correct, and several more examples are worked out in detail at the associated website. It should be noted, however, that the features used for alignment quality determination are the same as those used for producing the MSA. This is not a very elegant method, but given the nature of IDPs probably the best that can be done. KMAD certainly will emphasize functionally important IDP residues and regions, and thus will provide a basis for subsequent experiments needed to shed light on the sequence structure function relation of this intriguing branch of the protein kingdom.

6 Usage

KMAD is available as a standalone version and as a web-server. The standalone version consists of three parts to (1) obtain information about the features (IDPs, PTMs, SLiMs and domains) and map them onto the sequences; (2) KMAD itself to use all information to align the sequences; and (3) the output scripts. These programs are available with installer, documentation, etc., from the projects website. KMAD accepts a query sequence and runs its whole pipeline automatically to predict disordered regions, align sequences or annotate alignments. The standalone version and the web-server both allow user to change parameters, including user-defined features. A REST API is available, so that users can access KMAD programmatically without the need to install the standalone version.