SSM vs. others: 1GPU:A

 
Materials from this page cannot be reproduced without permission from the authors.
Comparisons made on November 2002, using current versions of VAST, CE, DALI, DEJAVU and SSM v1.22 from 20/11/2002.
 

CONTENTS
  1. VAST
  2. CE (Combinatorial Extension)
  3. DALI
  4. DEJAVU
  5. Conclusion
 
1GPU:A (678 residues)
TRANSKETOLASE COMPLEX WITH REACTION INTERMEDIATE
28 longest helices and 16 strands were used for SSE matching.
 

1.  V A S T    (server)

Figure 1GPU:A-1 shows the Ca-alignment lengths obtained from SSM and VAST for different structural neighbours (as chosen by VAST). As seen from the picture, only 42 PDB entries were reported by VAST as statistically significant structural neighbours. First 14 PDB entries in Figure 1GPU:A-1 represent highly similar structural neighbours, so that all residues of input are aligned to them by both SSM and VAST. Other entries represent more remote structures, with less than 50% of 1GPU:A residues aligned to them. The Figure demonstrates that SSM and VAST produce very similar results. They fully agree in the classification of highly similar and remote structural neighbours. For the latter ones, SSM offers shorter Ca-alignments on comparison with VAST.


  Figure 1GPU:A-1.
Length of Ca-alignment as a function of PDB entry, obtained by SSM (black line) and VAST (red line). Details of the calculations are given here.
 

As may be derived from Figure 1GPU:A-2, all longer alignments from VAST have higher RMSDs. Both SSM and VAST produce alignments with RMSDs in a fairly reasonable range (less than 5 Å).


  Figure 1GPU:A-2.
RMSD of Ca-alignment corresponding to data in Figure 1GPU:A-1. Details of the calculations are given here.
 

The match index of SSM and VAST alignments, shown in Figure 1GPU:A-3, makes it clear that despite the differences in alignment lengths and RMSDs (cf. Figures 1GPU:A-1 and 1GPU:A-2), the principal quality of SSM and VAST alignments is virtually the same. As seen from Figure 1GPU:A-3, SSM and VAST match indexes replicate each other down to small local details, with SSM match index being slightly higher on average. Given the agreement in match indexes and a moderate difference in alignment lengths and RMSDs for remote structures, one may conclude that SSM and VAST differ in balancing the compromise between alignment length and RMSD (SSM tends to shorter alignments at lower RMSDs at same principal quality of alignment).


  Figure 1GPU:A-3.
Match Index corresponding to data shown in Figure 1GPU:A-1. Details of the calculations are given here.
 

Figure 1GPU:A-4 shows that SSM and VAST generally agree in P-values of the alignments. The most difference in P-values is observed in the regions of highly similar and most remote structures. As seen from the Figure, SSM makes more differentiation of statistical significance between highly similar structures, and disagrees with VAST in the evaluation of several low-similarity matches.


  Figure 1GPU:A-4.
P-values corresponding to matches shown in Figure 1GPU:A-1. Details of the calculations are given here.
 

Z-scores of Ca-alignments, given by SSM, are almost an exact replica of those obtained from VAST, but 2.5 times lower (cf. Figure 1GPU:A-5). Apart from that difference, VAST and SSM Z-scores show a very nice agreement.


  Figure 1GPU:A-5.
Z-scores corresponding to matches shown in Figure 1GPU:A-1. Details of the calculations are given here.
 

 

 

2.  C E (Combinatorial Extension)    (server)

On comparison with VAST, CE gives much more hits (cf. Figures 1GPU:A-6 and 1GPU:A-1). It does not recognize the input and gives 1TRK:A as the closest structural prototype. It finds 12 closest structural neighbours to 1GPU:A (all residues of the latter are aligned), all the rest represents very remote structures, with less than 25% input's residues aligned for most of them. As seen from the picture, SSM shows reasonable agreement with CE on the length of Ca-alignments. The difference between SSM and CE alignments is found within 15% in most cases, however for individual entries it may reach 50%. It is possible to say that on average, CE and SSM produce similar-length alignments meaning that there is no significant systematic deviation.


  Figure 1GPU:A-6.
Length of Ca-alignment as a function of PDB entry, obtained by SSM (black line) and CE (red line). Details of the calculations are given here.
 

CE and SSM show a general agreement in RMSDs of Ca-alignments as well (cf. Figure 1GPU:A-7). Both servers produce alignments with RMSD in a reasonable range (5 Å and less), except for one remote structure aligned by SSM with a higher root mean square deviation. Figure 1GPU:A-7 makes and impression that on average, RMSD from SSM are by 0.5 Å lower than those from CE.


  Figure 1GPU:A-7.
RMSD of Ca-alignment corresponding to data in Figure 1GPU:A-6. Details of the calculations are given here.
 

The match indexes, calculated from SSM and CE results, indicates a slightly better quality of SSM alignments for individual structures (cf. Figure 1GPU:A-8). A detail analysis of the Figure reveals that SSM and CE alignments have indeed a very close match indexes, which is an indication of a good agreement in principal quality of alignment. The differences in alignment lengths and RMSDs for most structures, visible in Figures 1GPU:A-6 and 1GPU:A-7 should therefore be attributed to the difference in balancing the compromise between them, employed by the servers. Usually, CE favours longer alignments at higher RMSDs, however in the particular example of 1GPU:A one cannot probably say so. Just as RMSDs and alignment lengths (cf. Figures 1GPU:A-6 and 1GPU:A-7), match index shows a clear distinction between highly similar and remote structural neighbours.


  Figure 1GPU:A-8.
Match Index corresponding to data shown in Figure 1GPU:A-6. Details of the calculations are given here.
 

With the exception for the highly similar structures, Z-scores from SSM agree reasonably well with those from CE (cf. Figure 1GPU:A-9). In comparison with CE, we usually apply a factor of 2 to CE Z-scores. As seen from the Figure, in the particular example of 1GPU:A this worsenes the comparison. For highly similar structures, SSM gives a considerably higher Z-score, making a much more pronounced distinction between similar and dissimilar structures.


  Figure 1GPU:A-9.
Z-scores corresponding to matches shown in Figure 1GPU:A-6. Details of the calculations are given here.
 

 

 

3.  D A L I    (server)

DALI output represents a clear intermediate between results obtained from VAST and CE (cf. Figures 1GPU:A-1, 1GPU:A-6 and 1GPU:A-10). DALI recognizes the input structure and its 13 closest structural neighbours, aligning all input's residues to them. Comparison with SSM for remote structures gives approximately 15% deviation in most instances, On comparison with VAST and CE, although in some cases the difference may reach 50%. Clearly enough, on average SSM offers a somewhat longer, than DALI's, alignments.


  Figure 1GPU:A-10.
Length of Ca-alignment as a function of PDB entry, obtained by SSM (black line) and DALI (red line). Details of the calculations are given here.
 

Figure 1GPU:A-11 shows that SSM produces slightly longer alignments at comparable RMSDs for most structures. On top of that, RMSDs from DALI show spikes, the highest of which reaches 19.9 Å, which is definitely beyond a reasonable range. With the exception of these spikes, SSM and DALI keep RMSDs in approximately the same corridor of values. Comparison of Figures 1GPU:A-10 and 1GPU:A-11 suggests that SSM makes alignments of a better quality (longer alignments at similar RMSDs).


  Figure 1GPU:A-11.
RMSD of Ca-alignment corresponding to data in Figure 1GPU:A-10. Details of the calculations are given here.
 

Indeed, the match index, calculated from SSM results, is on average somewhat higher than that obtained from DALI's output (cf. Figure 1GPU:A-12). Higher match index generally indicates a higher quality of 3D alignment (longer alignments at same RMSD), therefore one can conclude that for most SSM alignments, presented in the Figure, have a better quality on comparison with those from DALI. This difference is not typical for the comparison with DALI; on contrary, in most cases the principal quality of SSM and DALI alignments, as meaured by match index, is found nearly identical. Confusingly, pronounced spikes in DALI RMSDs, which are clearly seen in Figure 1GPU:A-11, do not have a visible reflection in the match index. A detail analysis of Figure 1GPU:A-12 shows, however, that RMSD spikes drop the match index to nearly zero (correspond Figures 1GPU:A-11 and 1GPU:A-12 in the locations of the spikes), and although the absolute difference between SSM and DALI match indexes in those points is not big, their relative difference is quite significant.


  Figure 1GPU:A-12.
Match Index corresponding to data shown in Figure 1GPU:A-10. Details of the calculations are given here.
 

Z-scores from SSM and DALI show a general similarity (cf. Figure 1GPU:A-13). As usually, DALI gives higher Z-scores for highly similar structures, and lower Z-scores for the remote ones. Often, DALI Z-scores compare better with minus logarithm of SSM's P-values (black line in Figure 1GPU:A-13); that does not apply to the present case of 1GPU:A.


  Figure 1GPU:A-13.
Z-scores corresponding to matches shown in Figure 1GPU:A-10. Details of the calculations are given here.
 

 

 

4.  D E J A V U    (server)

DEJAVU does not recognize the input structure and does not find any of its closest structural neighbours. It gives 1ZPD and 1BFD as the closest prototypes of the input. As seen from Figure 1GPU:A-14, all but two of DEJAVU matches have less than 15% of the input residues aligned. Typically for the comparison with DEJAVU, SSM produces noticeably longer Ca-alignments; most of SSM alignments are 40% longer than those obtained from DEJAVU.


  Figure 1GPU:A-14.
Length of Ca-alignment as a function of PDB entry, obtained by SSM (black line) and DEJAVU (red line). Details of the calculations are given here.
 

Figure 1GPU:A-15 shows that longer SSM alignments come at the expense of higher RMSDs. DEJAVU produces considerably shorter RMSDs, as compared to SSM for virtually all structures. While DEJAVU evidently keeps RMSD in the range below 2.25 Å, SSM allows for RMSD up to 5 Å and even more in a few instances.


  Figure 1GPU:A-15.
RMSD of Ca-alignment corresponding to data in Figure 1GPU:A-14. Details of the calculations are given here.
 

In spite of considerable differences in alignment lengths and RMSDs, seen in Figures 1GPU:A-16 and 1GPU:A-16, the match indexes, calculated from SSM and DEJAVU results, are in fact close to each other. A detail analysis of data presented in Figure 1GPU:A-16 shows that SSM and DEJAVU curves correlate very well, although on average SSM's match index is slightly higher. Comparison of Figures 1GPU:A-14, 1GPU:A-15 and 1GPU:A-16 makes it clear that at similar quality of 3D alignments, SSM and DEJAVU differ significantly in balancing the alignment length and RMSD.


  Figure 1GPU:A-16.
Match Index corresponding to data shown in Figure 1GPU:A-14. Details of the calculations are given here.
 

P-values from DEJAVU are noticeably lower than those given by SSM (cf. Figure 1GPU:A-17). DEJAVU assignes zero P-values to the matches with the most similar structures from its findings, which we find somewhat oddish. Zero P-value means that there is no chance to get match of a better quality, that is, by definition, matching a structure to itself. Therefore assigning zero P-values to matches with evidently remote structural neighbours looks confusing.


  Figure 1GPU:A-17.
P-values corresponding to matches shown in Figure 1GPU:A-14. Details of the calculations are given here.
 

Z-score is, generally, a reflection of P-value. According to that, DEJAVU assigns very high Z-scores to the matches with zero P-values (compare Figures 1GPU:A-17 and 1GPU:A-18). As seen from Figure 1GPU:A-18, DEJAVU tends to higher Z-scores on comparison with SSM. Although a very general trend in SSM and DEJAVU Z-scoring may be identified as similar (decreasing Z-score with increasing dissimlarity), the overall agreement can hardly be rated as a very good one.


  Figure 1GPU:A-18.
Z-scores corresponding to matches shown in Figure 1GPU:A-14. Details of the calculations are given here.
 

 

 

5.  Conclusion

The principal quality of 3D Ca-alignments, as measured by match index, is in a good agreement between results produced by SSM and those obtained from other servers. SSM makes alignments of a somewhat better, on average, quality as measured by match index. Different servers differ in solving the compromise between alignment length in RMSD. In that respect, SSM agrees reasonably well with VAST, CE and DALI, but shows a considerable difference from DEJAVU.

SSM alignments are relatively close to those obtained from VAST, CE and DALI, meaning that the average deviation of the alignment lengths is sufficiently low. DEJAVU fails to identify the input structure and all closest structural neighbours. The difference in alignment lengths between SSM and DEJAVU is significant. The corresponding RMSDs from VAST, CE and DALI are close to those offered by SSM. In several instances, DALI produces very high RMSDs, which may be an indication of a problem with its superposition algorithm. RMSDs from DEJAVU are considerably lower than those from SSM, however corresponding to them match indexes agree reasonably well with those calculated from SSM results. SSM agrees well with P-values and Z-scores given by VAST. Comparison with CE, DALI and DEJAVU show only qualitative agreement in general trends in Z-scoring.