Conditional Random Fields for the Prediction of Signal Peptide Cleavage Sites

1
Conditional Random Fields for the Prediction of
Signal Peptide Cleavage Sites

• M.W. Mak
• The Hong Kong Polytechnic University

S.Y. Kung Princeton University
2
Contents

• Introduction
• Proteins and Their Subcellular Locations
• Importance of Protein Cleavage-Site Prediction
• Information in Amino Acid Sequences
• Existing Approaches to Cleavage Site Prediction
• Conditional Random Field (CRF)
• CRF for Cleavage Site Prediction
• Experiments and Results
• Effectiveness of Different Feature Functions
• Effect of Varying Window Size
• Fusion with SignalP

3
Proteins and Their Destination

• A protein consists of a sequence of amino acids.
• Newly synthesized proteins need to pass across
  intra-cellular membrane to their destination.

http//redpoll.pharmacy.ualberta.ca
4
Signal Peptide

• A short segment of 20 to 100 amino acids (known
  as signal peptides) contains information about
  the destination (address) of the protein.
• The signal peptide is cleaved off from the
  resulting mature protein when it passes across
  the membrane.

http//nobelprize.org
Mature protein
Source S. R. Goodman, Medical Cell Biology,
Elsevier, 2008.
Signal Peptide
Cleavage Site
5
Importance of Cleavage Site Prediction

• Defects in the protein sorting process can cause
  serious diseases, e.g., kidney stone

Source http//nobelprize.org/nobel_prizes/medicin
e/laureates/1999/illpres/diseases.html
6
Importance of Cleavage Site Prediction

• Many proteins (e.g. insulin) are produced in
  living cells. To cause the proteins to be
  secreted out of the cell, they are provided with
  a signal peptide.

Bioreactor
Source http//nobelprize.org/nobel_prizes/medicin
e/laureates/1999/illpres/diseases.html
7
Information in Sequences

• Signal peptides contain some regular patterns.
• Although the patterns exhibit substantial
  variation, they can be detected by machine
  learning tools.

Rich in hydrophobic AA
Cleavage Site
8
Existing Methods

• Weight matrices (PrediSi)
• Neural Networks (SignalP 1.1)
• HMMs (SignalP 3.0)

9
Weight Matrices
15 Positions
20 AA
t -1 t t1
M A R S S L F T F L C L A V F I N G C L S Q I E Q
Q
Score at position t 1608678771310686
067178
10
SignalP-HMM
Source Nielsen and Krogh
Mature protein
Signal Peptide
11
Contents

• Introduction
• Proteins and Their Subcellular Locations
• Importance of Protein Cleavage-Site Prediction
• Information in Amino Acid Sequences
• Existing Approaches to Cleavage Site Prediction
• Conditional Random Field (CRF)
• CRF for Cleavage Site Prediction
• Experiments and Results
• Effectiveness of Amino Acid Properties
• Effectiveness of Different Feature Functions
• Fusion with SignalP

12
Conditional Random Fields

• Conditional Random Fields (CRFs) were originally
  designed for sequence labeling tasks such as
  Part-of-Speech (POS) tagging

• Given a sequence of observations (e.g., words), a
  CRF attempts to find the most likely label
  sequence, i.e., it gives a label for each of the
  observations.

13
HMM Vs. CRF

• Hidden Markov Models Learn

Label
Observation

• Conditional Random Fields Learn

More direct
y1
y2



yT
Label
Observation
14
Advantages of CRF

• Avoid computing likelihood p(observationlabel).
  Instead, the posterior p(labelobservation) is
  computed directly.
• Able to model long-range dependency without
  making the inference problem intractable.
• Guarantee global optimal.

Depends on
M A R S S L F T F L C L A V F I N G C L S Q I E Q
Q
15
CRF for Cleavage Cite Prediction
Cleavage site
Weights
Length of Sequence
Transition features
n-grams of amino acids
State features
16
CRF for Cleavage Cite Prediction
e.g. bi-gram and query sequence T Q T W A G S H
S . . .
17
CRF for Cleavage Cite Prediction
Position
18
Contents

• Introduction
• Proteins and Their Subcellular Locations
• Importance of Protein Cleavage-Site Prediction
• Information in Amino Acid Sequences
• Existing Approaches to Cleavage Site Prediction
• Conditional Random Field (CRF)
• CRF for Cleavage Site Prediction
• Experiments and Results
• Effectiveness of Different Feature Functions
• Effect of Varying Window Size
• Fusion with SignalP

19
Experiments

• Data 1937 protein sequences extracted from
  Swissprot 56.5. The cleavage sites locations of
  these sequences were biologically determined
• Ten-fold cross validation
• For 1st-order state features, up to 5-grams of
  amino acids
• For 2nd-order state features, up to bi-grams of
  amino acids.
• Use CRF software

20
Results
Effectiveness of using AA Properties
Observations (1) Amino acids provide the most
relevant information (2)
Hydrophobicity and charge/polarity can help
21
Results
Effectiveness of Different Feature Functions

• Observations
• Transition feature by itself is no good.
• But, once combined with state-features,
  performance improves

(Transition only)
(Transition State)
22
Results
Effect of Varying the Window Size
e.g. query sequence T Q T W A G S H S . . .
23
Results
Compared with Other Predictors
Observations (1) CRF is slightly better than
SignalP (2) CRF is complementary to SignalP
24
Web Server
http//158.132.148.858080/CSitePred/faces/Page1.j
sp
25
Web Server
http//158.132.148.858080/CSitePred/faces/Page1.j
sp
Available in May 2009
26
(No Transcript)
27
Conditional Random Fields

• Conditional Random Fields (CRFs) were originally
  designed for sequence labeling tasks such as
  Part-of-Speech (POS) tagging

Observations
x
x
y
Labels

• Given a sequence of observations, A CRF attempts
  to find the most likely label sequence, i.e., it
  gives a label for each of the observations.