1. No category

The Protein Interaction Prediction
Engine (PIPE) Project
Frank Dehne
School of Computer Science
Carleton University
Ashkan Golshani
Institute for Biochemistry
Carleton University
PIPE Project
Project started in 2003
Multi-disciplinary team:
Computer Science
– F.Dehne
Biochemistry
– A.Golshani
– M.Dumontier
– J.Greenblatt
Biomed. Eng.
– J. Green
Graduate Students:
S.Pitre, C.North,
A.Amos-Binks, A.
Schoenrock, M.Alamgir,
Bahram Samanfar,
Mohsen Hooshyar, ...
Equipment:
256 core PC Cluster
1168 core Sun T2 “Victoria
Falls” Cluster
32,000 core Blue Gene/Q
Basics of Proteins
Why do we study protein-protein interactions (PPIs)?
1.Proteins often perform their function(s) by interacting
with other proteins
For example: RNA pol II
1. PPIs form a major group of interaction networks within a cell
i.e. help define the biology of a cell as a system.
3rd yr
Biology undergrads
Biology Admin staff
Biology Profs
Biology TA
Conclusion:
He is most likely a biology student.
Nuclear
Export
Variant
Histone
Ubuiquination
TREX
CPF
Mediator
COMPASS
ER to Golgi
Transport
Elongator
RNAP II
Biochemical Methods
Tandem affinity purification (TAP)

Yeast two-hybrid screen

Co-immunoprecipitation

Bimolecular fluorescence complementation

Affinity electrophoresis

Pull-down assays

Label transfer

Phage display

In-vivo crosslinking

Chemical cross-linking

Strepprotein interaction experiment

Frank
Dehne ● www.dehne.net
Quantitative
immunoprecipitation combined with

Tandem Affinity Purification
(TAP)
Do YGL227W and YMR135C interact?
Frank Dehne ● www.dehne.net
TAP tag
Y2H
Computation-based methods may offer improved alternatives.
Can we detect PPI based
on primary sequence
only?
- No pre-defined information
- Can be applied to all proteins
- Can be applied to all genomes, even newly
sequenced ones
Working Hypothesis
Certain regions of interactions can be very small
(20 - 40 amino acids)

Interaction “Codes”

Known Protein Interactions
species
# proteins
# protein
pairs
# known
interactions
# unknown
interactions
S. cerevisiae
6,300
19,867,056
15,151
???
C. elegans
23,684
280,454,086
6,607
???
H. sapiens
22,513
253,406,328
41,678
???
S. cerevisiae
C. elegans
H. sapiens
Basic PIPE Algorithm
String comparison
Match =
(Sum of pairwise PAM
values > Threshold)
PIPE Output
PIPE Output
PIPE: Detecting Novel Protein-Protein Interactions
S. cerevisiae
Banting and Best Institute of
Medical Research, Toronto
Protein complex: YGL227W, YMR135C, YIL017C, YDL176W, YIL097W, YDR255C, YBR105C
PIPE: Elucidating the Architecture of Protein Complexes
S. cerevisiae
PIPE: Reducing False Positives
Eliminate “popular
motifs” via median
filter
PIPE's Prediction Accuracy
Many Other Methods...
Types of Protein Interaction Prediction Methods:

Phylogenetic profiling

Identification of homologous interacting pairs

Identification of structural patterns (Van der
Waals)

Bayesian network modelling

3D template-based protein complex modelling

Supervised learning (SVM)
Park's Comparison Experiment
Park (BMC Bioinformatics, 2009, 10:419) compared the four best methods

[M1] Martin et.al. (Bioinformatics 2005,21(2):218-226): protein pair is encoded
by a product of signatures which is then classified by a support vector
classifier

[M2] PIPE

[M3] Shen et.al. (Proc Natl Acad Sci USA 2007, 104(11):4337-4341): each
protein sequence is encoded by a feature vector that represents the
frequencies of 3 amino acid-long subsequences, and feature vectors are
concatenated for a pair of proteins and classified by SVM.

[M4] Guo et.al. (Nucl Acids Res 2008,36(9):3025-3030): each protein
sequence is encoded by a feature vector that represents auto-correlation
values of 7 different physicochemical scales, and feature vectors are
concatenated for a pair of proteins and classified by SVM.

Consensus Method: “Vote” among M1-M4.
Park's Comparison Experiment
From: Park, BMC
Bioinformatics, 2009,
10:419
Global Scan of Entire Protein Interaction Network
species
# proteins
# protein
pairs
# known
interactions
S. cerevisiae
6,300
19,867,056
15,151
C. elegans
23,684
280,454,08
6
6,607
H. sapiens
22,513
253,406,32
8
41,678
Open Problem...
Challenges:



Large number of protein pairs (requires high speed, SVM not possible)
Small number of true positives (very sparse, ~ 0.1 % density)
Requires very high specificity ~99.95 % (i.e. less than 0.05% false
positive rate) – Otherwise: #false positives > #true positives
PIPE's Prediction Accuracy
PIPE Sequential Performance Improvements
Character based amino acid representation was
converted into binary encodings. Removed need
for character-to-index lookup in PAM120.

“Sliding window” process was improved to use
incremental updates.

Pre-computed all possible protein fragment
comparisons and stored all matches of similar
fragments in a hash table.

Large Scale Parallelization: MP-PIPE
Architecture:

Cluster of multi-core processors

One MP-PIPE worker per proc.

Each worker with multiple threads
H.sapiens protein pairs
Summary of PIPE Results
PIPE's superior performance and prediction accuracy enabled
the first ever complete scan of entire protein interaction networks
species
# proteins
# protein
pairs
# known
interactions
# novel
PIPE pred.
*
S.
cerevisiae
6,300
19,867,056
15,151
14,438
1 hour
C. elegans
23,684
280,454,08
6
6,607
32,548
1 week
H.sapiens
22,513
253,406,32
8
41,678
130,470
3 months
* False positive rate: 0.0001


256 core PC Cluster
1168 core Sun T2 “Victoria
Falls” Cluster
Running time
Listens from overall PPI profile
Previously reported PPIs
Novel PPIs (current study)
Profile of PPI pairs on the basis of cellular process
Babu et al Nature 489: 585-589
H.sapiens dsDNA break
repair
l



Blue: Proteins known to be
involved in dsDNA break
repair
Green: Known interaction
Red: Novel interactions
discovered by PIPE
Yellow: Novel proteins likely
involved in dsDNA break
repair
Frank Dehne ● www.dehne.net
Online Resources
InSiPS: In-Silico Protein
Synthesizer
Participants:
l
l




Frank Dehne, Computer Science (PI)
Andrew Schoenrock, Computer Science (grad.stud.)
Sylavin Pitre, Computer Science (postdoc)
Ashkan Golshani, Biochemistry (Co-PI)
Dan Burnside, Biochemistry (grad.stud.)
Houman Moteshareie, Biochemistry (grad.stud.)
Frank Dehne ● www.dehne.net
In Silico Protein Synthesizer
(InSiPS)
The In Silico Protein Synthesizer (InSiPS) is a
computational tool that can synthesize proteins with
specific protein-protein interaction prediction profiles.
Given a set of target proteins and a set of non-target
proteins, InSiPS can generate a protein sequence that is
predicted to interact with the target proteins and
predicted not to interact with the non-targets.
Frank Dehne ● www.dehne.net
Frank Dehne ● www.dehne.net
Fitness Function
Frank Dehne ● www.dehne.net
Performance on BGQ
Frank Dehne ● www.dehne.net
Performance on BGQ
Frank Dehne ● www.dehne.net
Target Proteins (Yeast)
Frank Dehne ● www.dehne.net
Parameter Settings
Frank Dehne ● www.dehne.net
“Good” Cases
Frank Dehne ● www.dehne.net
“Bad” Cases
Frank Dehne ● www.dehne.net
Experimental Verification
Task:
Design a protein that attaches to a yeast protein involved in
DNA repair, thereby blocking its function.
Target Yeast protein:
YAL017W (PSK1) – DNA repair
InSiPS generated protein:
“Anti-PSK1”:
HHHHHHSDNEHLHKCQRLKTRWKMARQFSDPQHNMYWIINWAQAM
NIHADQNQEEEEELHDASVNNAEQYMAQCAPEEACQYPVRRSYGLH
ATNCIERRKCCMIMYQHPTCRQWEAKNTCAISRAGKGVYWKGIIFMRA
WKHWCTRRLVQ
fitness: 0.465163
target score: 0.71832232
max non-target score: 0.35243136 (YLL039C)
avg non-target score: 0.0720702297
Frank Dehne ● www.dehne.net
Experimental Verification
UV Light
PSK1
DNA
X
mRNA
mRN
A
Protein
mRN
Protein
A
mRNA
mRN
A
Protein
mRN
Protein
A
Deletion
DNA
DNA
Frank Dehne ● www.dehne.net
X
Anti-PSK1
mRNA
mRN
A
Protein
mRN
Protein
A
Experimental Verification
WT +
Anti-PSK1
expressed
PSK1
knockout
Decreasing cell density
WT
WT
(empty
vector)
Expression of Anti-Psk1 causes sensitivity to UV light.
Equal numbers of cells serially diluted and exposed to 30s of UV light
Frank Dehne ● www.dehne.net
Future Projects
Drug Design:
Design synthetic proteins that attach to critical
proteins in viruses, thereby inhibiting the virus.
Possible Targets:
- E-coli
-…
- HIV
HIV
(Capsid protein)
Frank Dehne ● www.dehne.net