Mismatch String Kernels for SVM Protein Classification

1
Mismatch String Kernels for SVM Protein
Classification

• By Leslie. et .al
• Presented by Huzefa Rangwala
• (rangwala_at_cs.umn.edu)

2
Outline

• Problem Definition
• Support Vector Machines
• Feature Extraction Scheme
• Evaluations
• Comparison with other Schemes
• Conclusion

3
Problem Definition

• Remote Homology
• Motivation Classify proteins based on sequence
  data into homologous groups (evolutionary
  similar) to understand the structure and
  functions of proteins.
• Previous known approaches pairwise sequence
  alignment, profiles, HMM
• New approaches Discriminative Models

4
Problem Definition

• Protein sequences are seen as a set of labeled
  sequences, i.e positive if they belong to the
  same super-family or family and negative
  otherwise
• Our classifiers will detect as positives those
  test sequences that are remotely related to the
  positive training instances.

5
Classification Definition
Taken from Using fisher kernel method to
detect remote protein homologies by Jaakkola,
Diekhans, et. Al.
6
SVM Linear Separators

• Which of the linear separators is optimal?

7
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.

8
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.

9
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.
• The paper in discussion introduced a new feature
  extraction scheme, which is very simplistic but
  performs very well.
• All future research in this domain is to extract
  the best possible features to capture the right
  signals for fold prediction.

10
Feature Extraction

• Spectrum Kernel (Simpler) This kernel simply
  counts the occurrences of k-length subsequences
  for each of the sequence in consideration.
• This can be done by sliding a window of length k
  across a sequence and updating a hash table which
  has 20k entries.

11
Spectrum Kernel

• Spectrum 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.

12
Mismatch Kernel

• Slight modification to the Spectrum kernel
• Define a parameter m which allows up to m
  mismatches in the counting of occurrences
• This means gt

13
Mismatch Kernel

• 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.

14
Putting all together

• SVM can be thought of as a magic box
• However it computes the dot product between the 2
  vectors to find similarity between two sequences.

Model Learnt
SVM Classifier
SVM Learner
N Vectors, each of dimension 20k
N Training sequences
Test Vector
Classification result
Unknown Test Sequence
15
Evaluation ROC (Explain?)
16
Conclusion

• A simple use of just protein sequence gives
  better performance than blast
• Current state of the art use profiles as an
  addition into the kernels, they seem to do better
  at this prediction problem than the ones that use
  structure directly.
• My research revolves around this idea
• https//wwws.cs.umn.edu/tech_reports/index.cgi?sel
  ectedyear2005modeprintreportreport_id05-007

17
Questions ?

• Huzefa Rangwala
• PhD in Computer Science
• Email rangwala_at_cs.umn.edu