Transcript Document

Macromolecular Structure Database
Structural Database Infrastructure Services for Europe
www.ebi.ac.uk/msd
EMBL-EBI
The MSD databases
The MSD actually consists of two separate
databases:
 the archive database is highly normalized, with
thousands of relationships linking some 400
tables; the deposition database is the
definitive archive for all structural data at MSD
 the search database is a much simpler,
denormalized database, with data items
duplicated and aggregated into 40 much wider
tables, making it more amenable to searching
and retrieval of data : the MSDSD
EMBL-EBI
What is the MSDSD database
 A relational database primarily developed in Oracle
that stores the data derived from the PDB together
with reference and other derived information
 Simple to understand for the novice biologist and
fast in performance for the database non-expert
 Originates from the internal MSD archive database
that ensures accuracy and data integrity
 In MSDSD naming and other summary information
is repeated from every level of the hierarchy to the
next one in order to be closer to the familiar PDB
data
EMBL-EBI
Main MSDSD features
 The symmetry has been expanded and the
information of the quaternary biological
assemblies is directly available
 External Information like binding sites and secondary
structure has been derived on the assembly level
 The original PDB asymmetric unit is also available
 Includes and provides clear database relations
with the ligands “data mart” and other reference
information
 Includes information and cross-references to
external databases (NCBI taxonomy, UniProt,
SCOP etc)
EMBL-EBI
DATA Analysis
Data mining is a term that is applied very loosely
within bioinformatics to describe any type of data
analysis. Almost without exception the analysis of
molecular biology is hypothesis based where the
search for information has a target that is defined by
the knowledge of the biological context of the data.
EMBL-EBI
Data Mining
“Analysis of data in a database using tools which
look for trends or anomalies without knowledge of
the meaning of the data.”
“True data mining software does not just change
the presentation, but discovers previously unknown
relationships among the data.
(Webopedia and other technical dictionaries)”
was first “invented” by IBM
EMBL-EBI
Traditional analysis is via “verification-driven
analysis”
Requires hypothesis of the desired
information (target)
Requires correct interpretation of proposed
query
Discovery-driven data mining
Finds data with common characteristics
Results are ideal solutions to discovery
Finds results without previous hypothesis
EMBL-EBI
So what is Hypothesis driven
data analysis ?
 Define a target = hypothesis
 Search for target
 There are/are-not “hits”
 Verify/negate hypothesis
 Distribution is centred on target
“catalytic triad” :
Atomic coordinates :
Mathematical graph :
HIS,ASP,SER :
text string matching
coordinate superposition
graph matching
data hierarchy knowledge
EMBL-EBI
For example, it is possible to find the presence of
catalytic triads within the PDB by selecting an example
structure and then using a matching technique such as
coordinate superposition or graph analysis to screen this
against all the coordinate data within the PDB. This will
identify the presence of similar residue configurations to
the search target and result in a distribution of hits
centered on the original search model.
HOWEVER we can only find similar objects distributed
about this target.
EMBL-EBI
Discovery-driven data mining
 Finds data with common characteristics
 Results are ideal solutions to discovery
 Finds results without previous hypothesis
 Target is mathematical – so has no scientific
dependency
EMBL-EBI
Mining techniques
 Creation of predictive models : future data
expectation
 Link analysis : connections between data
objects
 Database segmentation : classification
 Deviation detection : finding outliers.
EMBL-EBI
So what is this data mining ?
 Given multiple sets of primary data)
 Characters, numbers, Function(numbers),….
 Find anomalies
 To many : numerical occurrence
 Data variation : Derivatives
 Singularities
 Correlations and clusters
Finds new things !
But not what it means !
 Within primary data
 with other data (dependent variables)
EMBL-EBI
Discovery driven data mining of the PDB
 Analysis of 3-dimensional coordinates
 Defined common patterns of atomic interactions
locally
 DB segmentation - active sites & common packing features
 Link analysis - Similarity between different functional
group
 Defined globally
 DB segmentation - common patterns of super-secondary
str’
 Link analysis - common folds in diverse protein families
 Outlier detection - unique folds
 Nucleic Acid sequences
 Define information content using information
technology
EMBL-EBI
Issues
 Systematic “error” propagates as solution
 300 lysozyme structures return as a strong solution
 Results cannot be found below the noise level
 Need to characterise the noise level
 Need to improve signal/noise ratio (S/N) to see information
 Target is not biologically defined
 It does not give you the biological answer
 Results should reproduce known biology
 Can give you new results not previously observed
EMBL-EBI
Data selection
 Cannot leave in 300 lysozyme structures !
 Select by sequence similarity at 70% exact
alignment
 Different “phase space” to select data
 Remove structures with resolution < 2.5A
 Remove NMR (different statistics)
 Remove pre-1982 etc.
 Geometrical analysis criteria to check for
outliers
EMBL-EBI
Off the shelf products
 Main problem – they “all” do column
correlation – but this requires row analysis
 Ie you can find whether x coordinates are more
correlated to y coordinates than z coordinates
 Slow
 I tried the above on 1e3 of data and it took hours;
not much chance on 1.6e9 data then.
 Money often
EMBL-EBI
Local atomic interactions
 Data
 3D coordinates
 Atom types
 Residue types
 Convert coordinates to
distances - easier to
compare, no need to
superpose coordinates.
 Create 3D Hash table of
triplets of distances
between “points”
EMBL-EBI
Local atomic interactions
 Merge triplets
 Any pair of N-fold
interactions are a (N+1)
interaction if they have
(N-1) equivalence.
 Just keep going until
no more (N+1)
interaction are found.
 Time = 8 seconds
 (Digital alpha ES40)
EMBL-EBI
Local atom interactions
 Define key atoms/groupsof-atoms as run time
parameters.
 Solves problem of residue
symmetry
 Approximation for speed
 This is a hypothesis
 External definition of
residue equivalence (PHE 
TYR) for released data.
 Improves Signal/Noise ratio.
 This is a hypothesis
EMBL-EBI
Is this data mining ?
 Basic 3D correlation of distances is
 Program can be run without any prior definitions.
 Addition of key atoms and residue equivalence
introduces biology and chemistry
 introduces hypothesis regarding what is important.
 Without adding this information you get very little out.
 Improvement to the method should spot this
without being told !
EMBL-EBI
Re-implemented
 This idea has been re-implemented
 Core analysis on distance only
 Statistical analysis of residue equivalence is
carried out – will find residue equivalence
 Bit slower now – 2 minutes
 To use MSD assembly data
 Must be able to normalise by chain similarity
to remove common features due to structure.
 Can use MSD similarity tables for this.
EMBL-EBI
Refining the answers
 This analysis produces approximate
geometrical results
 For each “solution”, a second full All vs.
All LSQ overlay is performed
 handles symmetry in D,E,R,P,T
 handles different residue overlays
 Clusters results using average linkage
 Writes average + superposed coordinates
+ ligands.
EMBL-EBI
Catalytic quartet
EMBL-EBI
Electrostatic interaction
Ligands are
found close by
rather than
associated with
the residues
EMBL-EBI
N-linked glycosolation binding site +?
 Spot the non-sugar
 This glycosolation
site is the same as
active site found in
“1a53” – indol-3glycerolphosphate
synthase
EMBL-EBI
Summary
 Creates 1000’s of results
 Returns many metal and catalytic sites
 50% have at least 2 of 3 residues as
sequence neighbours
 30% have associated ligands
http://www.ebi.ac.uk/msd-srv/MSDtemplate/
See
T.J.Oldfield (2003) PROTEINS: Structure, Function, and Genetics 49, 510-528.
T.J.Oldfield (2002) Acta Cryst. D57, 1421-1427
EMBL-EBI
Data mining – not idiot proof
 Date of birth and age will give 100 %
correlation
 Authors for structure submission will be
correlated to authors on primary citation.
 “Lysozyme” is the most common fold
pattern
 36 spelling’s of E.Coli will mask results.
 Requires representative sets
 Statistically valid ones too !
 Signal/Noise ratio is a problem : hit the
noise and the calculation grows rapidly
EMBL-EBI
Other methods
 Representative sets and clustering
 Another talk
 Data mining fold
 Information technological analysis of genomes