Protein Classification Using Proteome Analyst

1
Protein Classification UsingProteome Analyst

• MITACS Bioinformatics Seminar
• Duane Szafron, Paul Lu, Russell Greiner, David
  Wishart, Zhiyong Lu, Brett Poulin, Roman Eisner,
  John Anvik,Cam Macdonell, Gary Van Domselaar,
  Chris Upton, Alona Fyshe, David Meeuwis.

2
Outline

• Mathematical Background
• Naïve Bayes Classification
• Protein Classification
• Proteome Analyst walkthrough
• Proteome Analysts Explain Feature

3
Mathematical Background

• Conditional Probability
• P(AB)P(A,B)/P(B)
• Bayes Rule
• P(AB)P(BA)P(A)/P(B)

4
Mathematical Background

• Conditional Independence
• C Class (Hypothesis), F Feature (Evidence)
• So
• P(C,F1Fn)
• P(C) P(F1C) P(F2C,F1)P(FnC,F1Fn-1) (chain
  rule)
• P(C)p1..n P(FiC) P(AB,C) P(AC)
• (assuming conditional independence of
• Features given Class)

5
Naïve Bayes

• Naïve Bayes classifiers
• Are based on standard Bayesian statistical theory
• Assume conditionally independent evidence
• Are explainable

6
Naïve Bayes

• Classification/Analysis (Common Use)
• Given a set of evidence without a known class
  compute the probability for each class
• P(CFiFn) for each C
• Training (Seldom Done)
• Given a set of training data with known classes
  we estimate the parameters of the classifier
• P(C)p1..n P(FiC)

7
Proteome Analyst Process
Training Process
Sequences
Classes
PA Tools
Features
Classes
Analysis Process
ML
Classes
Sequences
PA Tools
Features
Classifier
Explain
8
Proteome Analyst Process
Training Process
Sequences
Classes
PA Tools
Features
Classes
Analysis Process
ML
Classes
Sequences
PA Tools
Features
Classifier
Explain
9
Proteome Analyst Process
Training Process
Sequences
Classes
PA Tools
Features
Classes
Analysis Process
ML
Classes
Sequences
PA Tools
Features
Classifier
Explain
10
Analysis Process(Classification)

• INPUT
• sequences
• Extract Features
• sequences ? features
• Classifier
• features ? classes, explanation
• OUTPUT
• Classes
• explanation

11
Analysis INPUT

• gtABC1_YEAST
• MVTNMVKLRNLRRLYCSSRLLRTIQNGRSSV
• gtABP1_YEAST
• MALEPIDYTTHSREIDAEYLKIVRGSDPD
• gtACE1_YEAST
• MVVINGVKYACETCIRGHRAAQCTHTDG
• .
• .
• .

12
Analysis Feature Extraction

• sequences ? features (Evidence)

Black Box

• Sequence ? ? Features

13
Feature Extraction

• Black Box

Protein 1 - F1 F3 F5 F9 Protein 2 - F6 F7
Protein 3 - F2 F4 Protein 4 - F8 Protein 5 -
F1 F3 F5 F2 F10 F4 F9 Protein 6 - F3 F5
F6 Protein 7 - F4 F6 F8 F9 Protein 8 - F1 F3 F5
Protein 9 - F1 F3 F5 F9
14
Feature Extraction

• Black Box (significantly similar proteins)

Protein 1 - F1 F3 F5 F9 Protein 2 - F6 F7
Protein 3 - F2 F4 Protein 4 - F8 Protein 5 -
F1 F3 F5 F2 F10 F4 F9 Protein 6 - F3 F5
F6 Protein 7 - F4 F6 F8 F9 Protein 8 - F1 F3 F5
Protein 9 - F1 F3 F5 F9
15
Feature Extraction

• Black Box (significantly similar proteins)

Protein 1 - F1 F3 F5 F9 Protein 2 - F6 F7
Protein 3 - F2 F4 Protein 4 - F8 Protein 5 -
F1 F3 F5 F2 F10 F4 F9 Protein 6 - F3 F5
F6 Protein 7 - F4 F6 F8 F9 Protein 8 - F1 F3 F5
Protein 9 - F1 F3 F5 F9
Features (evidence) for the input sequence F1,
F3, F5, F6, F7
16
Analysis Classification

• features ? classes
• Select the most probable class based on the
  evidence gathered in feature extraction
• Maxj(P(CjFiFn) )

17
Training

• INPUT
• sequences, classes
• Extract Features
• sequences ? features
• ML Algorithm
• features, classes ? Classifier
• OUTPUT
• Classifier

18
Training INPUT

• gtCellular processeslt YCB9_YEASTgi00140351
• MESQQLSNYPHISHGSACASVTSKEVHTNQDPLDVSAS
• gtTranslationltPEPQ_ECOLIgi00417465
• MESLASLYKNHIATLQERTRDALARFKLDALLIHSGELFN
• gtRegulatory functionsltSWI4_YEASTgi00730855
• MPFDVLISNQKDNTNHQNITPISKSVLLAPHSNHPVIEIAT
• .
• .
• .

19
Training INPUT

• gtCellular processesltYCB9_YEASTgi00140351
• MESQQLSNYPHISHGSACASVTSKEVHTNQDPLDVSAS
• gtTranslationltPEPQ_ECOLIgi00417465
• MESLASLYKNHIATLQERTRDALARFKLDALLIHSGELFN
• gtRegulatory functionsltSWI4_YEASTgi00730855
• MPFDVLISNQKDNTNHQNITPISKSVLLAPHSNHPVIEIAT
• .
• .
• .

classes
20
Training Feature Extraction

• Same as analysis

Black Box

• Sequence ? ? Features

21
Training OUTPUT

• Final Output
• Features classes Classifier

22
Results Subcellular localization



23
Proteome Analyst Walkthrough
24
Subcellular Locations Animal
Golgi Nucleus Extracellular Mitochondrion Cytoplas
m Plasma Membrane Lysosome Peroxisome Endoplasmic
Reticulum Secreted
http//www.probes.com/handbook/figures/0908.html
25
Step 1 Obtain the translated sequence (FASTA
format)
gtPEX6_HUMANQ13608Homo sapiens (Human)
cytoplasmic MALAVLRVLEPFPTETPPLAVLLPPGGPWPAAELGLVL
ALRPAGESPAGPALLVAALEGPDAGTEEQGPGPPQLLVSRALLRLLALGS
GAWVRARAVRRPPALGWALLGTSLGPGLGPRVGPLLVRRGETLPVPGPRV
LETRPALQGLLGPGTRLAVTELRGRARLCPESGDSSRPPPPPVVSSFAVS
GTVRRLQGVLGGTGDSLGVSRSCLRGLGLFQGEWVWVAQARESSNTSQPH
LARVQVLEPRWDLSDRLGPGSGPLGEPLADGLALVPATLAFNLGCDPLEM
GELRIQRYLEGSIAPEDKGSCSLLPGPPFARELHIEIVSSPHYSTNGNYD
GVLYRHFQIPRVVQEGDVLCVPTIGQVEILEGSPEKLPRWREMFFKVKKT
VGEAPDGPASAYLADTTHTSLYMVGSTLSPVPWLPSEESTLWSSLSPPGL
EALVSELCAVLKPRLQPGGALLTGTSSVLLRGPPGCGKTTVVAAACSHLG
LHLLKVPCSSLCAESSGAVETKLQAIFSRARRCRPAVLLLTAVDLLGRDR
DGLGEDARVMAVLRHLLLNEDPLNSCPPLMVVATTSRAQDLPADVQTAFP
HELEVPALSEGQRLSILRALTAHLPLG..

• Example
• PEX6, a protein from the Swiss-Prot database
  (accession Q13608)
• 980 Amino Acids long
• Codes for Human Peroxisome Assembly Factor
• Cytoplasmic

26
Step 1 Continued

• Peroxisome
• Cell structure that rids the body of toxic
  substances in the liver, kidneys, and brain
• Zellwegers Syndrome
• Caused by a one amino mutation in PEX6
• Caused by interference in peroxisome
  manufacturing so that few or none are made.
• Symptoms
• Enlarged liver
• High levels of iron and copper in the blood
• Vision disturbances
• Facial abnormalities
• Early death, often within 6 months of onset

27
Step 2 Log in to Proteome Analyst
28
Step 3 Upload the Sequence
29
Step 4 Process the sequence.

• Processed using Proteome Analysts built in
  classifiers.
• Subcellular Classifier
• Trained on PAs Animal dataset
• Different datasets used for different
  phylogenetic families

30
Step 5 View Results
31
Feature Extraction

• Features collected from the Black Box
• repeat, disease mutation, cytoplasmic,
  atp-binding, membrane protein,
• ipr003959, ipr003593, ipr003960, peroxisome
• InterPro is a database of protein families.
• Ipr003593 - AAA ATPase
• Ipr003959 - AAA ATPase, central region
• Ipr003960 - AAA-protein subdomain
• ATPase - an enzyme that uses ATP as a source of
  energy to drive some other
• process e.g. Peroxisome Biogenesis
• AAA 'A'TPases 'A'ssociated with diverse
  cellular 'A'ctivities
• Proposed that these proteins act as ATP-dependent
  protein clamps

32
Classification
33
Explain

• P(C)pi P(EiC)
• Factors P(EiC) are multiplied to get the
  probabilities for each class
• Explain graph is on a log scale so the
  contributions become additive and easily
  visualized
• log(P(C)pi P(EiC) )
• log(P(C) )?i log(P(EiC) )

34
Explain

• PA produces interpretable, transparent results

35
Explain Champion vs. Contender
36
Questions?
http//www-gap.dcs.st-and.ac.uk/history/PictDispl
ay/Bayes.html

• Contactbioinfo_at_cs.ualberta.ca
• http//www.cs.ualberta.ca/bioinfo/PA
• Special Thanks to Thomas Bayes and the PA-Team

37
Classifier Evaluation

• K-fold Cross Validation
• E.g. k 3
• Confusion Matrix
• 2 classes positive negative

38
Information Content and Wrapping

• Compute for each Feature Fj
• I(Fj) ?i?k P(TjVk, CCi)log(p(P(TjVk,
  CCi)/ (P(TjVk)P( CCi))
• Order Features by I(Fj)
• for(int i 0 i lt 100 i i5)
• Create classifier with i features with the lowest
  I(F)removed from the training set
• In practice best classifiers are created with 60
  of features removed