Lupyan D, LeoMacias A, Ortiz AR'

1
Lupyan D, Leo-Macias A, Ortiz AR. A new
progressive-iterative algorithm for multiple
structure alignment. Bioinformatics, 2005, 21,
3255-3263.
Evening Journal Club 9/19/05 by Ilya Shindyalov
My comments are shown in such callouts.
MAMMOTH-mult server based on MStA algorithm
2

• Multiple structural alignment
• Structural bioinformatics
• Structural genomics
• Structure prediction
• Structure classification

Multiple structural alignment remains the problem
for which there is no good solution exists.
While the algorithm presented here claims
producing alignments of higher quality than
manual, this was only demonstrated for the whole
sample and for two easy (well conserved)
families immunoglobulins and globins. Thus
leaving the question open.
3
What MAMMOTH-mult delivers Biologically
meaningful trees and conservation at the SEQUENCE
and STRUCTURAL levels of FUNCTIONAL MOTIFS
4
MAMMOTH MAtching Molecular Models Obtained from
THeory- pairwise structure alignment algorithm
(Ortiz et al, 2002)
CE has 8, Dali has 5, MAMMOTH 7 !!!
Four steps (1) URMS (unit-vector root mean
square ) for hepta-peptides from Ca i to Ca
i1.
(2) Introduction of URMSR (Chew et al, 1999)
the expected minimum URMS distance between two
random sets on n unit vectors and score SAB
the score between two heptapeptides A and B.
This equation doesnt take into account protein
geometry.
5
MAMMOTH (cont.)
(3) DP (dynamic programming) with 4Å cutoff
similar to MaxSub (Siew et al, 2000).
DP should work only for similar structures only,
e.g. family level.
(4) Calculating P-value from Z-score.
where ? 0.5772, Euler-Mascheroni constant
(Gumbel, 1958)
6
MStA (Multiple Structure Alignment) algorithm
(1) All-against-all pairwise comparison with
MAMMOTH.
(2) Dendrogram by average linkage (Johnson,
Wichern, 1998) using the matrix from step 1.
If structures are rather equally-distant from
each other, the topology of dendrogram becomes
ambiguous.
(3) Following bootom-up over the tree, grouping
and realigning proteins using DP, MaxSub (Siew et
al, 2000), SIMPLEX optimization (Barton,
Sternberg, 1987).
7
MStA algorithm performance criteria
(1) core - core extension relative to the
shortest alignment strict core sequence - 100
conservation, structure 4Å used in
optimization loose core sequence - 66
conservation, structure 3Å used in comparisons
with other methods (both calculated for the
aligned positions only)
(2) RMScore mean RMS for the strict core
residues.
(3) norMD score quality of the alignment
(Thompson et al, 2001) norMD normalized Mean
Distance.
8
MStA algorithm training and testing sets
(1) HOMSTRAD set (Mizuguchi et al, 1998) 105
structural alignments of families manually
curated assumed a gold standard at the family
level here.
(2) CAMPASS set (Sowdhamini et al, 1998) 551
manual alignments of superfamilies manually
curated assumed a gold standard at the
superfamily level here.
(3) Multiprot set same data as in 2, realigned
with MultiProt (Shatsky et al, 2004).
(4) FSSP set (Holm, Sander, 1997) reprocessed
with MAMMOTH 1385 cliques were produced.
(5) Superfold set 2 superfolds immunoglobulins
(26) from (Bork et al, 1994) and globins from
SCOP. SCOP classification used as gold standard
here.
9
MAMMOTH-mult structural alignment for
immunoglobulins
10
MAMMOTH-mult structural alignment for
immunoglobulins (cont.)
11
MAMMOTH-mult structural alignment for globins
12
MAMMOTH-mult structural alignment for globins
(cont.)
13
(No Transcript)
14
(i) Results are quite impressive. (ii)
Independent benchmarking and comparison with
other methods is desired. (iii) Averages do not
tell about success with complex
families/superfamilies. (iv) Applicability for
above SCOP granularity superfamilies is
questionable.
Conclusions