Transcript The Domain Structure of Proteins

The Domain Structure of
Proteins: Prediction and
Organization.
Golan Yona
Dept. of Computer Science
Cornell University
(joint work with Niranjan Nagarajan)
Golan Yona, Cornell University
PDB: 1a8y 367aa long MKIIRIETSRIAVPLTKPFKTALRTVYTAESVIVRITYDSGAVGWGEAPPTLVITGDSM…………
Golan Yona, Cornell University
The domain structure of a protein





A domain is considered the fundamental
unit of protein structure, folding,
function, evolution and design.
Compact
Stable
Folds independently?
Has a specific function
Golan Yona, Cornell University
A protein is a combination of
domains
Protein1
Protein2
Protein3
Golan Yona, Cornell University
Any signals that might indicate
domain boundaries?
A very weak signal if any in the
sequence
 Usually domain delineation is done
based on structure
 Best methods available – manual!
 But structural information is sparse..

Golan Yona, Cornell University
Definitions and assumptions
Domain: continuous sequence that
corresponds to an elemental building
block of protein folds.
 A subsequence that is likely to be stable
as an independent folding unit.
 Was formed as an independent unit,
and later was combined with others –
more complex functions.
 There are traces of the autonomous
units..

Golan Yona, Cornell University
First step..
Gather data – database search
 Histogram of matches is informative but
noisy

sequence

Mutations, insertions, deletions,
conflicting evidence
Golan Yona, Cornell University
Previous methods





Methods based on the use of similarity searches and
knowledge of sequence termini to delineate domain
boundaries using heuristics/rules (MKDOM, Domainer,
DIVCLUS, DOMO).
Methods that rely on expert knowledge of protein
families to construct models like HMMs to identify other
members of the family (Pfam, TigrFam, SMART).
Methods that try to infer domain boundaries by using
sequence information to predict tertiary structure first
(SnapDragon. Rigden’s covariance analysis)
Methods that use multiple alignments to predict domain
boundaries (PASS, Domination).
Others..(e.g. CSA and DGS = guess based on size)
Golan Yona, Cornell University
How do you evaluate the different
methods?
No universal measures
 A variety of qualitative and quantitative
evaluation criteria, external resources
and manual analysis are used to verify
domain boundaries

Golan Yona, Cornell University
Method outline



Source/test data – SCOP
Processed data - alignments
Learning system:
– Domain-information-content scores
– NN
– Probabilistic model

Evaluation

“A Multi-Expert System for the Automatic
Detection of Protein Domains from Sequence
Information” Niranjan Nagaragan and Golan Yona, in
the proceedings of RECOMB2003
Golan Yona, Cornell University
Overview
Intron Boundaries
Seed Sequence
DNA DATA
blast search
Sequence Participation
Multiple Alignment
Secondary Structure
Entropy
Neural Network
Correlation
Contact Profile
Physio-Chemical Properties
Final Predictions
Golan Yona, Cornell University
The source/test data set

PDB structures with their partitions into
domains as defined in SCOP:
– 1ctf: domain1 1-76 domain2 77-123

Remove sequences shorter than 40 aa
and almost identical entries
Golan Yona, Cornell University
Alignments







Search each query against a database of ~1 million
non-redundant sequences
Remove fragments first
Two phase alignment procedure
– First phase: blast
– Second phase: multiple iteration psi-blast
Select one representative from each group of similar
proteins
Remove proteins that are less than 90% covered
(missing information)
Number of domains ranging from 1-7
Final set: 605 multi-domain proteins and 576 single
domain proteins (1/4)
Golan Yona, Cornell University
The domain-information-content of
an alignment column
Measures that (are believed) to reflect
structural properties of proteins
 A total of 20 measures

–
–
–
–
–
Conservation measures
Consistency and correlation measures
Measures of structural flexibility
Residue type based measures
Predicted secondary structure information
– Intron-exon data
Golan Yona, Cornell University
Conservation measures

Entropy: some positions are more conserved
than others

Class entropy: some positions have preference
towards a class of amino-acids (similar physiochemical properties)

Evolutionary pressure (span): sum of pairwise
similarities
Motivation: consider the mutual similarity of amino acids
Golan Yona, Cornell University
Consistency and correlation
measures


All domain appearances should maintain its integrity
Consistency: difference in sequence counts

Asymmetric correlation: consistency of individual
sequences.

Symmetric correlation: reinforcement by missing
sequences

Measures are averaged over a window
Golan Yona, Cornell University
Consistency and correlation
measures – cont.

Sequence termination: strong but
elusive
– Fragments
– Premature halt in alignment
– Loosely aligned

Product of left and right termination
scores: given c sequences that
terminate at a position, with evalues
e1,e2,e3,…ec
Golan Yona, Cornell University
Golan Yona, Cornell University
Measures of structural flexibility

Indel entropy: variability indicates
structural flexibility (likely to occur near
domain boundaries)

Correlated mutations: indicative of
contacts
Contact profiles
Golan Yona, Cornell University
Contact profile
Golan Yona, Cornell University
Residue type based measures
hydrophobic vs. hydrophilic
 cystines and prolines
 Classes of amino acids

Predicted secondary structures
Helices and strands are rigid
 Loops are more abundant near domain
boundaries

Golan Yona, Cornell University
Intron-exon data

Exon boundaries are expected to
coincide with domain boundaries
1
2
1
2
1
3
3
2
Protein1
Protein2
Protein3
Golan Yona, Cornell University
Score refinement and normalization
Smoothing using a window w
(optimized)
 Unification to a single scale – zscore
over all positions

Golan Yona, Cornell University
Maximizing the information content
of scores
Opt for the most distinct distributions of
domain positions vs. boundary positions
 Affected by the parameters (w
smoothing factor) and x (boundary
window size)
 Use the Jensen-Shannon divergence
measure

Golan Yona, Cornell University
Examples
Golan Yona, Cornell University
Even measures with identical
distributions may be informative in a
mutli-variate model
 To simplify model only the top 12 are
selected

Golan Yona, Cornell University
The learning system
A neural network is trained to model
effectively the complex decision
boundary surface
 Predicts correctly 94% of domain
positions and 88% of the transitions in
the test set
 Also tried mapping from multiple
positions (local input neighborhood) to
single/multiple output

Golan Yona, Cornell University
Overview
Intron Boundaries
Seed Sequence
DNA DATA
blast search
Sequence Participation
Multiple Alignment
Secondary Structure
Entropy
Neural Network
Correlation
Contact Profile
Physio-Chemical Properties
Final Predictions
Golan Yona, Cornell University
Hypothesis evaluation

Simple model: refine predictions
– Significant fraction of the positions in a
window centered at x should be predicted
as transitions
– Order transitions by their quality (depth of
the minima) and reject all transitions that
are within 30 residues from already
predicted transitions
Golan Yona, Cornell University
The domain generator model
Multiple hypotheses – find the “best
one”
 Assume a model: random generator
that moves repeatedly between a
domain state and a linker state and
emits one domain or transition at a
time according to different source
probability distributions.
 Total probability is the product

Golan Yona, Cornell University
Formally..
S = D1 D2
Dn
 We are given a sequence S (multiple
alignment) of length L and a possible
partition into n domains D=D1,D2,..Dn of
lengths l1,l2,..,ln (NN output)
 Find the partition that will maximize the
posterior probability P(D/S)

Maximize the product of the likelihood
and the prior
Golan Yona, Cornell University
Calculating the prior P(D)

For an arbitrary protein of length L what is the
probability to observe D

Approximate using a simplified model: given
the length of the protein, the generator selects
the number of domains first and then selects
the length of one domain at a time,
considering the domains that were already
generated.
Golan Yona, Cornell University
The prior probabilities
Approximate P0(li/L) by P0(li) normalized
to the relevant range.
 P0(li/L) is derived based on
experimental data

Golan Yona, Cornell University
The prior probabilities (cont.)

Calculate Prob(n/L) = Prob(n,L)/P(L)

1

2
Golan Yona, Cornell University
The likelihood



Use probabilities of observed scores
considering the two different sources
The model D partitions the sequence S into n
domains and n-1 transitions: D1,T1,D2,T2,…,Tn1,Dn that correspond to the subsequences
s1,t1,s2,t2,..,tn-1,sn
Assume domains are independent of each
other (additional test can be used)
Golan Yona, Cornell University
…likelihood

Each term P(si/Di) and P(tj/Tj) is a
product over the probabilities of the
individual positions, each one is
estimated by the joint probability
distribution of the 12 features

How to estimate this probability?
(independence assumption does not
hold)
Golan Yona, Cornell University
Golan Yona, Cornell University
Likelihood of individual position

Given k random variables X1,X2,..,Xk their joint
prob. Distribution

Use first order dependencies
For each pair, calculate the distance between
the joint prob. Distribution and the product of
the marginal distributions

Golan Yona, Cornell University

Sort all pairs based on their dependency, and
pick the most dependent one (denoted by Y1,
Y2) and start the expansion

Select the next one based on the strongest
dependency with variables that are already in
the expansion
Golan Yona, Cornell University


Denote by Z=PILLAR(Y) the random variable
that Y is most dependent on
Of all possible dependencies involving Y3 pick
P(Y3/Z) and add it to the expansion
Proceed until you exhaust all variables
 Maximize support, minimize error
 The expansion is different for domain and
transition regions

Golan Yona, Cornell University
Finally..

Enumerate all possible hypotheses,
calculate the posterior probability for
each one, and output the one that
maximizes the prob.
Golan Yona, Cornell University
Summary of results




Distance accuracy: average distance of the predicted transitions
from their associated SCOP transition points.
Distance sensitivity: average distance of SCOP transitions from
their associated predicted transition points.
Selectivity: percentage of correct predictions (within 10 residues
from SCOP transitions)
Coverage: percentage of correctly identified SCOP transitions (within
10 residues from predicted transitions)
Golan Yona, Cornell University
Examples
PDB ID: 2gep
 Domain Definition:
8-72, 73-272, 273-352, 353-497
 Predicted Domains:
1-75, 76-270, 271-352, 353-497
 PFam Definition:
1-67,
273-345, 356-425
Golan Yona, Cornell University
Examples
PDB ID: 1b6s chain D
 Domain Definition:
1-78, 79-276, 277-355
 Predicted Domains: 1-73, 74-271, 272-355
 PFam Definition:
30-167
Golan Yona, Cornell University
Examples
PDB ID: 1acc
 Domain Definition:
14-735
 Predicted Domains:
1-158, 159-583, 584-735
 PFam Definition:
103-544
Golan Yona, Cornell University
Conclusions







A method for predicting the domain structure of a
protein from sequence information alone
Protein/DNA data, multiple features, optimization based
on information theory principles, learning system and
final prediction using the domain-generator model (with
confidence values).
Exhaustive hypothesis evaluation
Fully automatic and fast
Perform very well even compared to the best manual
and semi-manual methods out there (also on CATH
data)
Dare to say …can be used to verify domain assignments
based on structural data
Improvements: other learning systems, more features
Golan Yona, Cornell University
Acknowledgments

Niranjan Nagarajan
SCOP
 CATH
 PSI-BLAST
 Pfam
 InterPro


NSF
Golan Yona, Cornell University