Fingerprint Clustering Comparative Study

1
Fingerprint Clustering Comparative Study

• This presentation will probably involve audience
  discussion, which will create action items. Use
  PowerPoint to keep track of these action items
  during your presentation
• In Slide Show, click on the right mouse button
• Select Meeting Minder
• Select the Action Items tab
• Type in action items as they come up
• Click OK to dismiss this box
• This will automatically create an Action Item
  slide at the end of your presentation with your
  points entered.

• 606 Project Presentation by
• Dean Cheng

2
Outline

• A Brief Background Introduction
• Some Definitions
• Clustering Algorithms
• Results
• Conclusion and Future directions

3
Oligonucleotide Fingerprinting (1)

• DNA clone long active linear sequence of
  nucleotides (unknown)
• Probe short linear sequence of 6-8 nucleotides
  (known)

Hybridizations A-T A C A G G-C G
T T G A-T G A G T C-G G-C C
A T-A A-T G G (Total) (Half) (None)
4
Oligonucleotide Fingerprinting (2)

• Fluorescent labeled probes, higher intensity
  better hybridization

5
Oligonucleotide Fingerprinting (3)
Fingerprint vector A vector of intensity values
of all probes for a clone. Example c1 has a
fingerprint of 1 0 1 0 0 .
6
Oligonucleotide Fingerprinting (4)

• Applications
• DNA sequencing
• Gene expression
• DNA clone classification
• Clustering
• Similarity/Distance functions
• Algorithms
• Focus on binary (ternary) domain

7
Project Specifications

• Three algorithms and two fingerprint
  representations.
• UPGMA-0-1, FREQ-0-1, UPGMA-0-1-N, GCP-0-1-N
• 0-1-N 2 thresholds
• Intensity above positive control 1
• Intensity below negative control 0
• Anything in between N (unknown)
• Hamming distance. For 0-1-N fingerprint, ignore
  N.

8
Binary Clustering with Missing Values (BCMV)

• Two 0-1-N vectors are compatible if they do not
  differ at any position or they only differ at
  positions where one or both of them has N.
• A 0-1-N vector is resolved if value of N is
  determined (0 or 1).
• The challenge is to cluster only mutually
  compatible 0-1-N fingerprint vectors such that
  fingerprint vectors within a cluster can be
  resolved in the same way.

9
Graphs and 0-1

• For 0-1 fingerprint vectors, we can try to find
  resolve vector of 1-hamming distance.
• For example, the set 100, 010, 001 has a resolve
  vector of 000.

10
GCP (Greedy Clique Partition)

• The algorithm first finds unique maximal cliques
  and removes them from the graph.
• Then a greedy action is used to find and remove
  maximum cliques from the graph until all vertices
  in the graph belong to some clique.
• Implementation from Zhipeng Cai.

11
UPGMA

• Given a distance matrix, join two nodes most
  similar to each other and update distance matrix
  with average distance to the two nodes. Output
  is a tree.
• PHYLIPs neighbor joining has UPGMA.

12
FREQ (1)

• Frequent itemset mining find itemset that has a
  support above a minimum support.

Support of 1, 2, 3 is 40. Support of 1, 2
is 60.
Apriori if an itemset is frequent then all its
subsets must be frequent.
Eg 1 80, 260, 340, 1-260, 2-3 40
13
FREQ (2)

• Transform fingerprints into transactions use a
  fingerprint itself and its compatible
  fingerprints.
• Eg
• 1. 111001NN T1 1, 2, 3, 4, 5
• 2. 11100N1N T2 1, 2, 3, 4
• 3. 11100N11 T3 1, 2, 3, 4
• 4. 1110N11N T4 1, 2, 3, 4
• 5. 1110N10N T5 1, 5

14
FREQ (3)

• Can do same for 0-1 vectors using 2-hamming
  distance (eg. 010, 100, 001 has a resolve vector
  of 000).
• Experiment preprocess the fingerprints first,
  only uses fingerprints that do not have 0-hamming
  distance identical. An implementation issue.
  Eg 101, 101, 111, 111, 100, 010, 001.
• Take the longest frequent itemset containing
  un-clustered fingerprints as a cluster.

15
Validation methods

• Number of cluster with size gt 2. Number of
  singleton.
• Average homogeneity and average separation.
  Euclidean distance and hamming distance used.
• Minkowski measure and Jaccards coefficient using
  GCPs result as standard.

16
Datasets and thresholds

• Bacteria dataset 1491 clones and 27 probes.
  Fungi dataset 1507 clones and 26 probes.
• Thresholds
• Bacteria 1 0.35-0.50 and 0.425
• Bacteria 2 0.35-0.55 and 0.45
• Fungi 425-775 and 600

17
Results
18
Discussion (1)

• GCP-0-1-N perform the best, guarantee to find
  mutually compatible clusters.
• UPGMA-0-1-N also return good results but fewer
  larger-size cluster.
• Both GCP-0-1-N and UPGMA-0-1-N outperform
  UPGMA-0-1 and FREQ-0-1. UPGMA-0-1-N superior to
  UPGMA-0-1. 0-1-N is a better representation.
• UPGMA find more singleton and fewer larger-size
  clusters.

19
Discussion (2)

• FREQ-0-1 better than UPGMA-0-1 but clustering
  qualities unstable. FREQ-0-1 is comparable to
  GCP-0-1-N and UPGMA-0-1-N in Bacteria 1.
• Must add constraints to FREQ. Incremental
  decrease of support can help efficiency and
  quality.
• Minkowski measure and Jaccards coefficient
  indicate clustering solutions different.
  UPGMA-0-1-N gt FREQ-0-1 gt UPGMA-0-1.

20
Conclusion Future Directions

• Uses cluster size homogeneity, separation,
  Minkowski measure and Jaccards coefficient to
  evaluate cluster qualities.
• 0-1-N is a better representation. Use UPGMA-0-1-N
  instead of UPGMA-0-1.
• GCP perform the best.
• FREQ works in both 0-1 and 0-1-N but need more
  testing. Adding constraints are key. Other
  frequent itemset mining algorithms can be
  explored such as Max-Miner. Assigning a
  fingerprint to multiple clusters a benefit?

21

• Thank You