Protein Fold Classification

1
Protein Fold Classification

• -Huzefa Rangwala
• (rangwala_at_cs.umn.edu)
• http//www.cs.umn.edu/rangwala

Some Slides are from a talk given by Prof. Karypis
2
Need to Know About Proteins

• Primary Structure Sequence
• ARN20 Amino Acids
• Secondary Structure
• Alpha, Beta, Coil
• Tertiary Structure
• Three dimensional atom coordinates
• Fold Classification
• SCOP, CATH, databases classify

3
Protein Family Prediction

• The goal is to assign a newly identified protein
  to an existing family.
• The complexity of this problem depends on how
  similar the new protein is to previously
  identified/classified proteins.
• Popular Approaches
• Blast
• Profile HMMs
• Discriminative models

4
Classification Approaches

• Nearest-neighbor classification
• Local/global sequence alignment, BLAST, etc.
• Markov Model-based approaches
• Markov chains, Hidden Markov Models, Profile
  HMMs, etc.
• Discriminating Models
• Neural networks, Support Vector Machines, etc.
• Arbitrary combination of the above basic
  approaches.

5
Discriminative Models

• Within the machine learning community,
  classification algorithms based on discriminative
  models, especially kernel-based support vector
  machines, have become very popular as they
  produce very accurate classifiers.
• The key idea of these discriminative models is to
  build a model that separates the various classes
• discriminate between classes as opposed to
  describing/generating each class.
• In the last 5 years, similar models have been
  developed for sequence classification.
• The work focused on developing string-based
  kernel functions that can be used within the
  context of support vector machines.

6
Key Difference

• Question
• So what is the key difference between a
  discriminative model like SVM and a generative
  model like HMM ?

7
Linear Separators

• Which of the linear separators is optimal?

8
Classification Margin

• Distance from example to the separator is
• Examples closest to the hyperplane are support
  vectors.
• Margin ? of the separator is the width of
  separation between classes.

9
The Kernel Trick

• The linear classifier relies on inner product
  between vectors K(xi,xj)xiTxj
• If every datapoint is mapped into
  high-dimensional space via some transformation F
  x ? f(x), the inner product becomes
• K(xi,xj) f(xi) Tf(xj)
• A kernel function is some function that
  corresponds to an inner product into some feature
  space.

10
Feature representation

• We are basically having a transformation function
  to convert sequences, profiles to be fed into a
  SVM black box to train and then classify.

11
Pairwise Kernel

• Each dimension corresponds to a sequence in the
  training set
• The value to a particular dimension corresponds
  to the sequence similarity score

Model Learnt
SVM Classifier
SVM Learner
N Vectors, each of dimension N
N Training sequences
Test Vector
Classification result
Unknown Test Sequence
12
How do we classify ?

• We build binary classifiers for each of the
  classes and then get predicted output from each
  of the models.
• Decision can be based on maximum result, or some
  weights,
• We can also build a class against another class
  classifier ? How???

13
k-mer or Spectrum Kernel

• k-mer kernel
• each dimension corresponds to a distinct k-mer
• each sequence is represented as a frequency
  vector of the various k-mers that it contains.
• k 4
• Eg - ACDAAA. gt
• hash ACDA
• hash CDAA
• hash DAAA
• .
• We have our vector of _______ features.

14
Mismatch kernel (k,m)

• Adding more information to the kmer kernel
• extends the k-mer kernel to allow for up to m
  mismatches
• Eg- (4,1)
• ACDAAA. gt
• h ?CDA , h A?DA, hAC?A,hACD?
• h ?DAA ,
• h?AAA
• .
• We have our vector of _______ features.

15
HMM kernel / Fisher kernel

• It is derived from a profile HMM
• We train a HMM and learn the emission and
  transmission probabilities
• Each of these probabilities become a dimension in
  the feature vector.
• How will this method do compare to the others ?
  Predict ?

16
Profile Kernel

• Simple extension of the kmer model but instead of
  using the sequence, use kmer scores from the PSSM
  matrix.
• Motivation ?
• -PSSM captures a proteins signature as it is
  build after multiple alignment and taking into
  consideration what regions are conserved.

17
isites Kernel

• Here a library of motifs, commonly occurring
  three dimensional structures are used to score a
  sequence.
• It basically gives a probability score of how
  similar a segment or kmer of the sequence looks
  to each of the motifs.
• These motifs come from isites library.

18
Comparisons/Performance
19
Questions ?

• Feel free to email me for doubts
• rangwala_at_cs.umn.edu
• Please go through the reading list.