Transcript Slides

Swiss-Prot 20 Years
Prediction of Subcellular
Localization of Proteins
~ Past, Present, and Future ~
Human Genome Center, Inst. Med. Sci.,
University of Tokyo
Kenta Nakai
20 Years Ago..
• I became a graduate student
in Prof. Minoru Kanehisa’s lab
• I wanted to write a program
that interprets the information
encoded in DNA sequences
• But biology is full of
exceptions
Diagnosis System of Bacterial
Infections (MYCIN 1974)





Enter Information about the patient. (Name, Age, Sex, and Race)
Are there any positive cultures obtained from SALLY?
…
Has SALLY recently had symptoms of persistent headache or other
abnormal neurologic symptoms (dizziness, lethargy, etc.)?
…
INFECTION-1 is MENINGITIS
+ <ITEM-1> MYCOBACTERIUM-TB [from clinical evidence only]
+…
[REC-1] My preferred therapy recommendation is as follows:
1) ETHAMBUTAL
Dose: 1.289 (13.0 100mg-tablets) q24h PO for 60 days [calculated
on basis of 25 mg/kg] then 770 mg (7.5 100mg-tablets) q24h PO ..
Knowledge Base for Automatic
Reasoning
• Knowledge is represented as a collection of “if-then”
rules, which are chained to make the system solve a
realistic problem
Rule 123
If: the gram stain of the organism is negative
and: the aerobicity of the organism is anaerobic
and: the morphology of the organism is rod
then: the genus of the organism is bacteroides
with a certainty factor of 0.6
Working Memory
Name: Sally
Age: 42 years
Sex: Female
Race: …
Expert Systems
Knowledge Base
Inference Engine
Sample Problem
Prediction of
Subcellular
Localization
Typical Sorting Signals
Signal Function
Example
Import into nucleus
-P-P-K-K-K-R-K-V-
Export from nucleus
-L-A-L-K-L-A-G-L-D-I-
Import into mitochondria
<-MLSLRQSIRFFKPATRTLCSSRYLL-
Import into plastid
<-MVAMAMASLQSSMSSLSLSSNS
FLGQPLSPITLSPFLQG-
Import into peroxisomes
-S-K-L->
Import into ER
<-MMSFVSLLLVGILFWAT
EAEQLTKCEVFN-
Return to ER
-K-D-E-L->
Amino Acid Composition
• Another good clue for
prediction
• Suited for machine
learning
Outer membrane proteins and
periplasmic proteins of Gramnegative bacteria
PSORT (I)
• Nakai & Kanehisa, 1991, 1992
• Expert system using about 100 “If-then” rules
signal peptide
signal cleavage site
(Specific Signals)
TMS in Mature Part
MTS
Topology
KDEL
GY
KK
GPI
ERM PM LSM
TMS
NLS
Topology
TMS
Apolar
SKL
Topology
motif
ERL LSL OT ERM PM MT MTMT MT NC PX ERM PM GG CP
OM IM IT MX
Papers and the web server
• Nakai & Kanehisa, Proteins 1991
– cited 295 times
• Nakai & Kanehisa, Genomics 1992
– cited 961 times
– 34 in 2006
• Web server since 1993
Limitations of PSORT
• Relatively low accuracy possibly because of the
complexity of the sorting mechanisms
• It is difficult to optimize the certainty parameters
assigned for each rule
• It is tedious to update the knowledge base with the
growth of the training data
PSORT II
• Nakai & Horton, 1997,
1999 (cited 638 times)
k=3
• Machine learning
• kNN (k-nearest neighbor)
method
Q
iPSORT: Bannai et al. 2002
Rule 1
A protein has an SP if the sum of hydropathy index values within [6,25]
exceeds 18.3
Rule 2
A protein has either an mTP or a cTP if it contains less than 3 D/Es
within [1,30] and if it contains a motif similar to 11212111, where
2=(I,R),3=(D,E,H,K,N),1=otherwise
Rule 3
A protein has an mTP if it satisfies Rule 2, if the sum of isoelectric
point values within [1,15] exceeds 93, and if it contains a motif similar
to 12211221, where 2=(K,R),3=(I,P),1=otherwise
PSORTb and PSORT.ORG
• Gardy et al. 2003, 2004
– Contribution from a
Canadian group (Brinkman
lab)
• Update for bacterial proteins
WoLF-PSORT
• Horton et al. 2006
• Latest PSORT update for
eukaryotic proteins
• WoLF: Women only Love
Fools!?
Current Dilemma
• More data are necessary to improve the training
process
• The practical value of prediction methods becomes
less with the growth of experimental data
• Moreover, the more we investigate, the more the
number of exceptions grows
It’s a General Problem
• Gene Finding
• Prediction of Protein Structure
• …
• Knowing the answer of a problem before we become
to know how to solve it
Similarity search against the data of typical model
organisms will become enough in many cases
New Generation Predictors
• Should be useful to engineer proteins for their
targeting sites
• Should complement errors of proteome analyses (i.e.,
isoforms with differential localization)
• Comprehensively example-based rather than
statistical feature-based (such as amino acid
composition)
Biology is like Linguistics
• Both are naturally born and full of exceptions
• There may not exist “general principles”
Future of Sequence Analysis
• It will become “DNA linguistics”
• Large dictionaries (databases) will contain both
general cases and exceptions
• Such databases may be a sort of knowledge base
that can be used to simulate the subcellular
processes
Past, Present, and Future
• Past
– Expert system-based predictions
• Present
– Machine learning-based predictions
• Future
– Combination of both?
– Revival of knowledge bases to simulate cellular processes?
Acknowledgments
• Minoru Kanehisa
• Paul Horton
• Hideo Bannai, Satoru Miyano
• Jennifer Gardy, Fiona Brinkman
• And all the other people who contributed to the
PSORT project!