Transcript The presentation part II

Structure-Sequence alignment
“Structure is better preserved than sequence”
Query sequence
MTYKLILNGKTKGETTTEAVDAATAEKVFQYANDN
GVDGEWTYTE
Non-redundant templates of structures:
Me!
Me!
Me!
Me!
How can we match a sequence and a structure?
MTYKLILNGKTKGETTTEAVDAATAEKVFQYANDNGVDGEWTYTE
MVNGLILNGKTK------------------------AEKVFQYANDNGVDGEWTYTE
Sequence: Similar Sequences take this
structure (but remember – sequence is
less preserved than structure…)
trp (W):
probably not
here!
Solvation: which AAs are buried?
Pair-Interaction:
How well do AAs get along
(Positive hate positive? Maybe
not…?)
more:
• 2nd structures prediction.
• 2nd structures constraints (β-strands
forming β -sheets…)
• etc.
“An Efficient and Reliable Protein Fold
Recognition Method for Genomic
Sequences”
David T. Jones (1999)
“What a good presentation!”
B. Raveh (2003)
GenTHREADER overview:
Query sequence
Templates
MTYKLILNGKTKGETTTEAVDAAT
AEKVFQYANDNGVDGEWTYTE
For each template (in the Brookhaven PDB):
• Construct a profile sequence
• Align with query sequence
• Calculate structural parameters (“to be continued…”)
• send parameters to a well-trained NEURON NETWORK (like
PSIPred…)
• OUTPUT: match confidence & alignment
STAGE 1: Building a profile for each
template
1. Start with sequence of template peptide:
“MTPAVTTYKLVINGKTLKGETTTKAVDAETAEKAFK
QYANDNGVDGVWTYDDATKTFTVTC”
2. Run BLASTP on OWL non-redundant
protein sequence data bank, with sequence as
input.
3. Take all sequences with E-Value < 0.01.
4. Align using MULTAL – multiple sequence
alignment method.
5. Construct a sequence profile based on
BLOSUM 50 matrix.
STAGE 2: Align sequence with a
profile
MTYKLILNGKTKGETTTEAVDAATAEKVFQYANDNGVDGEWTYTE
SCORE = ?
Length of query sequence = ?
Length of template profile = ?
Length of alignment itself = ?
STAGE 3: calculate (some) structural
parameters
In stage 2, the sequence was aligned to a profile of the structure.
The aligned sequence is now imposed on the 3D structure of the
template, and used for ENERGY POTENTIALS calculation.
STAGE 3: structural parameters (cont.):
E-Pair (pair interaction potential)
• an energy potential for the probability of the interactions
observed in this structure.
• Distance and sequence separation between certain atoms of two
different amino-acids are measured (Cβ – Cβ , Cβ - N, Cβ – O, etc.)
• Statistics of known structures were gathered and weighted.
• The observed interactions are compared to the statistics
• An energy potential is calculated
aa 39
• In essence: the smaller E-Pair, the better.
aa 157
STAGE 3: structural parameters (cont.):
E-Solv (solvation potential)
• Degree of burial (DOB) for an amino acid: “the number of other
Cβ atoms located within 10Å of the residue’s Cβ atom”
• In general, hydrophobic amino acids like to be buried, safely
away from water.
• Hydrophilic acids might like the
outside world better.
• Each amino acid DOB is calculated.
• It’s compared to statistical occurrence.
• ΔEsolv(AA,r) = -RT ln( f(AA,r) / f(r) )
Cβ
Cβ
Cβ
Cβ
Cβ
10Å
Cβ
STAGE 4:
send it all to the (trained) Neuron Network
Ouput is a score between 0-1 – translated to confidence
level (Low, Medium, High & Certain)
See this page on the web
Who trains the Neural network?
• Representatives were taken for different fold types in CATH
(“T-Level”).
• CAT numbers were used for comparing pairs.
• 9169 chain pairs
• 383 pairs shared a common domain fold (= should give a
positive answer)
• The network was trained with these pairs.
Neural network – black box?
Confidence assignment
CERTAIN
LOW
MEDIUM
HIGH
GenTHREADER – what to do with it?
Results on a ‘classic’ test set of 68 proteins:
• High true-positive rate:
73.5% correctly recognized, 48.5% with CERTAIN.
• Extremely reliable:
Every “CERTAIN” prediction was correct.
• Fast automatic method.
• For 22 of 68 proteins, alignment is over 50% accurate.
• Let’s go analyze the Mycoplasma Genitalium with it!
Whole Genome Analysis with GenTHREADER
Mycoplasme Genitalium genome analysis – ONE DAY ONLY!
ORF MG276 of mycoplasma gen.:
spotting a remote homologue
• MG276 is an “Adenine Phospho-ribosyl-transferase”
(but this information is not given to GenTHREADER)
• 1HGX is a template of other Phospho-ribosyl-transferase.
• It has only 10% sequence identity with our MG276!
• It was found by GenTHREADER as a certain match
• E-Pair saved the situation!
• But how do we know it’s true?
1HGX
template
Ligand binding site of 1HGX template
Substrate
ORF MG276 of mycoplasma gen.:
supporting evidence for 1HGX as a template
1. Substrate
binding sites
preserved
2. Secondary
structure
prediction of
MG276 is
similar
3. We cheated
all along…
ORF MG353 of mycoplasma gen.:
an ORF with no known function
• MG353 – no homologues found in databases
• 1HUE is a template of an “Histone-like” protein
• Very low sequence similarity with our MG353.
• It was found by GenTHREADER as a certain match
• Striking similarity in DNA Binding region
despite overall low sequence similarity
GenTHREADER improvements:
(McGuffin, Jones - may 2003)
• PSI-BLAST, PSI-PRED (2nd stuructures), some more…
• Some Results:
AB-INITIO FOLDING - ROSETTA
(Simons et al 1997, 1999, Bystroff & Baker 1998, Bonneau et al 2001)
Prediction of a protein fold from scratch?
Method I: physically simulate protein folding
Problem: CPU time
Practical for short peptides
APKFFRGGNWKMNGKRSLG
ELIHTLGDAKLSADTEVVCGI
APSITEKVVFQETKAIADNKD
WSKVEVHESRIYGGSVTNCK
ELASQHDVDGFLVGGASLKP
VDGFLHALAEGLGVDINAKH
Method II: check probability for all possible conformations
Problem: infinite search space
Solution: use mother nature – decrease search space
Decreasing the search space using
elements from short peptides:
• Take fragments of short peptides (3 residues – 9 residues
long).
• Join them together
• Keep the 2nd structures constant.
• “Play” with the angles of loop residues.
• RESULT: 200,000 decoy structures
In addition - I-Sites prediction
13 local-structure 3D motifs with sequence
profiles:
•Strong independence of motifs (fold-initiation sites?)
•complements secondary structure
Find the correct fold for a given
sequence (back to threading…)
P(structure) – sequence independant
P(sequence | structure):
•2nd structure packing
•Solvation
•Strand hydrogen bonding
•2nd structure – amino acid (proline
in helix, etc.)
•Strand assembly in sheets
•Structure compactness
•Frequency of I-Sites 3D motifs
•Etc.
•Pair Interaction
•I–Sites prediction for this
sequence(3D motifs) – did not
contribute to performance
•Etc.
RESULTS in CASP 4 – Baker’s a winner…
native
structures
vs.
predicted
models