Transcript Detecting the Domain Structure of Proteins from Sequence Information

Detecting the Domain
Structure of Proteins from
Sequence Information
Niranjan Nagarajan and Golan Yona
Department of Computer Science
Cornell University
What’s and Why’s

Why?




Function Prediction
Improved Alignments and more accurate
Evolutionary Studies
Protein Design
What?


Delineating Sequence Contiguous Domains
Work exclusively on Sequence Information
Past Work



The Pfam Protein Families Database, Bateman et al
(2002) Nucleic Acids Research 30:276-280
ProDom and ProDom-CG: tools for protein domain
analysis and whole genome comparisons, Corpet et al
(2000) Nucleic Acids Research 28:267-269
Automated Protein database classification: I.
Integration of compositional similarity search, local
similarity search and multiple sequence alignment. II.
Delineation of domain boundaries from sequence
similarities, Jerome et al (1998) Bioinformatics
14:164-187
Overview of the Process
Seed Sequence
Sequence Participation
blast search
Multiple Alignment
Secondary Structure
Entropy
Neural Network
Correlation
Contact Profile
Physio-Chemical Properites
Final Predictions
Motivation




Simple and Extensible
Tests an array of novel sources of
information
Automated method based on statistical
analysis of the scores
Domain transition signals are learned
rather than programmed in
Score Design




Efficiently Computable
Yields single value per profile column
Robustness to Alignment inaccuracies
Useful in distinguishing in-domain from
out-domain columns in isolation or in
combination with other scores
Correlation

Measures the conservation of the
alignment in a region
High Correlation
Low Correlation
Entropy

Estimates the diversity of the amino-acid
distribution for a column
Low Entropy
High Entropy
Sequence Participation

Identifies and quantifies the significance
of regions where there is a major
change in sequence participation
Secondary Structure

Uses psipred secondary structure
predictions for the seed sequence
Contact Profile

Contacts are predicted based on correlated mutation
values that are significantly larger than random
values
Physio-Chemical Properties


We tested properties like Hydrophobicity,
Molecular Weight, and Charge and various
classifications of the amino acids for their
information content
Scores were calculated by:
 Using the classification to assign values in
the range [0, 1] to every residue
 Taking the average of the values for a
profile column
Generating the Data Set


Seed Sequences: 4810 non-redundant (95% identity) PDB
sequences that are at least 40 amino acids long (PDB data as of
may 2002)
Alignments:





The seeds were blasted against a composite non-redundant
database with 693,912 non-fragmented entries
The resulting hits were compiled in a database
The seeds were queried using PSI-BLAST (until convergence)
against these smaller databases to generate the alignment
Domain Definitions: Definitions in SCOP 1.57 were used (seeds
with inconsistent definitions or less than 90% coverage were
removed)
The final set, after filtering to ensure to ensure a balance in the
number of single (576) and multi-domain (605) proteins,
contained 1181 seed proteins and their alignments
Massaging and Optimizing the Scores



Scores were smoothed over various smoothing
windows to test the importance of evening out local
fluctuations
Scores were normalized to ensure that values from
different proteins were comparable
The size of the smoothing window was optimized
using the Jensen-Shannon Divergence between the
distributions for in-domain and out-domain columns
Designing and Training the
Neural Network


Matlab’s Neural Network Toolbox was used to design and train
networks
Network Properties:
 Feed-Forward Back Propagation network with Tangent
Sigmoid activation function
 Current best network takes in 11 inputs and has two hidden
layers with 10 and 5 neurons respectively
 Neural network trained on a set of 484 proteins with a
validation set of 237 proteins and test set of 460 proteins
 Best network has accuracy of 91% for in-domain and 70%
for out-domain columns in test set
From Neural Network to
Cutpoint Predictions


A column is predicted as a cutpoint if a significant fraction of
columns in a window centered at it are predicted as being outdomain
For regions with multiple cutpoints near one another, minimas of
the smoothed prediction curve are used to decide the most
suitable cutpoint
Comparative Results
Average accuracy in residues
Our Method
43 (48)
Pfam
38 (43)
ProDom
10 (10)
SMART
14 (14)
Tigr
5 (3)


Average sensitivity in residues Percentage Accuracy Percentage Coverage
32 (36)
47 (21)
49 (22)
14 (22)
47 (25)
78 (42)
89 (89)
35 (24)
9 (7)
73 (74)
38 (27)
26 (19)
97 (94)
35 (18)
2 (1)
Accuracy evaluates predictions with respect to the
true definitions
Sensitivity evaluates true definitions with respect to
the definitions
Examples





Seed Number: 9847
PDB ID: 1b6s chain D
Domain Definition:1-78, 79-276, 277-355
Predicted Cutpoints: 73, 271
PFam Definition: 30-167
More Examples





Seed Number: 11791
PDB ID: 1acc
Domain Definition: 14-735
Predicted Cutpoints: 158, 583
PFam Definition: 103-544
Highlights



Correctly predicts domain definitions for 237
(52%) of the proteins in the test set thus
comparing favorably with PFam (258 and
56%)
The procedure is simple and fast and
comparable in accuracy and coverage to
PFam
General purpose method for delineating
domain boundaries that relies solely on
sequence information