GO Fuzzy cmeans Clustering Application

1
GO Fuzzy c-means Clustering Application

• Presented by
• Neha Somani

2
Outline of the presentation

• Clustering and different algorithms (In general)
• GO Fuzzy c-means algorithm (My understanding)
• Basic details about the algorithm
• Differences from Fuzzy c-means algorithm.
• The assigned project Application for the
  underlying algorithm
• Whats already there.
• What I plan to do.

3
Clustering

• Clustering - Classification of objects into
  different groups such that elements of each group
  share some common features.
• In gene expression data, it is used to observe
  simple gene-to-gene and sample-to-sample
  similarities so that potential functions of genes
  can be found.
• An important criteria for any clustering
  algorithm is the distance measure that helps in
  determining the similarity of two elements. For
  e.g. The Euclidean distance.
• Commonly used Stopping Criterion Distance wise
  or Number wise.

4
Standard Clustering Methodologies
Hierarchical
Partitional

• K-Means its derivatives
• K-means
• Fuzzy c-means
• QT clustering

Locality-sensitive hashing
Graph-theoretic methods
Agglomerative (Bottom-up approach)
Divisive (Top-down approach)

• Other common classifications
• Symmetric/ Asymmetric distances
• Two-way clustering Clusters objects as well as
  the features of the objects.
• Source Wikipedia Cluster Analysis

5
GO Fuzzy c-means algorithm

• Its an extension of Fuzzy c-means with prior
  biological knowledge.
• Allows assignment of multiple gene functions as
  well as incorporate prior biological knowledge
  such that genes can be associated with more than
  one biological function and thus help in
  capturing the actual behavior of genes.
• The prior knowledge used here is obtained from
  Gene Ontology and is used for initial membership
  assigment. (Can support other sources of
  knowledge as well).
• Uses the biological process ontology.
  Specifically GOSlim biological process terms are
  used to interpret the functions of genes at a
  higher level.
• GOSlim terms a set of general GO terms for
  various organisms and generic use. E.g. Cell
  cycle, Meiosis.
• Developed by Luis Tari under guidance of Dr. Kim
  and Dr. Baral.

6
Some important terms

• GOSlim
• Narrows down the clustering to general biological
  processes.
• For experimental purposes , 32 GOSlim biological
  process terms have been used.
• GO Annotations
• Annotations provided by experts.
• Evidence codes determine the degree of belief (0
  lt p_ij lt 1) of these annotations.
• E.g. IEA - 0.5, NAS - 0.6, IC - 0.7, IDA - 0.9
• Two customizable parameters play an important
  role in initialization (range 0 lt alpha, r lt 1)
• Alpha Degree of dependency.
• Alpha 0 implies initialization of membership is
  totally dependent on Gene annotation and less
  dependent when it approaches 1.
• r Degree of belief when a gene is not
  associated with a particular biological process.
• Value of r should be smaller but not too small
  either.
• Validity of Cluster Measure S
• Minimal value of S implies that the cluster is
  most compact and the furthest separation exists
  between clusters. (Inter and Intra cluster
  distances)
• Stopping Criteria

7
Algorithm steps

• Initial membership assignment
• Each GOSlim biological process term (GOBP) is
  considered as a cluster. (Thus, number of
  clusters is automatically determined).
• Suppose as per the annotation, gene g is
  associated with a process bp. Now using GO
  hierarchy the parent (direct or indirect) sbp of
  bp is found such that sbp belongs to GOBP. In
  this case g is assigned to the process sbp and
  thus associated with the corresponding cluster.
• The actual membership value is calculated based
  on evidence codes (degree of belief), alpha
  (degree of dependency) and r (constant that acts
  as degree of belief when g is not associated with
  a process b).
• Repeat the following steps until stopping
  criteria is achieved.
• At each step (k), compute the fuzzy centroid. The
  parameters involved are m (fuzzy parameter),
  expression vector for gene, number of genes, and
  current membership.
• Update the fuzzy membership based on expression
  vector, centroid and initial membership.
• Calculate the validity of cluster measure (S).
• If validity of cluster S lt S, then assign S S
  and update the optimal cluster and memberships
  with the values calculated above (in the kth
  step).

8
Results Output

• On reaching the stopping criteria, Optimal
  Clusters and memberships are provided as output.
• A cluster is considered optimal if the validity
  measure (S, initialized to infinity) of the
  cluster is minimal among the iterations.
• For mathematical equations, results of
  experimental evaluation, please refer the paper.
  (Its beyond the scope of this presentation.)

9
Major differences from Fuzzy c-means

• In fuzzy c-means, initial assignment of
  memberships is random while GOFuzzy uses gene
  annotations for initialization.
• This ensures that results generated are
  repeatable.
• It uses predefined classes based on GOSlim
  biological processes, so number of clusters is
  not needed as an input from user.
• In recursive steps, the membership is updated
  based on data as well as GO annotations.
• Flexibility of degree of reliability on the GO
  annotations for clustering using alpha (degree of
  dependency) and r (degree of belief).
• GOFuzzy eliminates the extra effort to identify
  the functions associated with the clusters as
  they are pre-defined based on biological
  processes.

10
The assigned project Application for the
underlying algorithm

• Developing a user friendly and robust interface
  for the clustering algorithm developed.
• The functionalities should include
• Visualization of results
• Regular updates of annotation files, data, etc.
  through the online repositories
• Input Parameters for user customization.
• File input/output.
• Report generations based on some input and
  generated parameters.
• The objective is to make a stand-alone
  application for GO Fuzzy c-means clustering in
  order to make it useful for all the users,
  especially, biologists.

11
Whats already present?

• GOFuzzyv1.0 - A command line program for the
  algorithm has been programmed in JAVA (1.5) and
  runs perfectly fine.
• Input file format
• Tab-delimited format .txt file
• Tab-delimited format in Eisen's CDT format - .cdt
  file
• The command line program takes input file name,
  taxonId, number of genes, number of samples and
  type of input file as input.
• TaxonId current version can take Yeast Human
  data-sets
• Uses static files saved in the working directory
  for
• GO data files
• GO Slim yeast file
• GO annotation files
• File containing probabilities for various
  evidence codes
• Output
• Three .txt output files are generated.
• out-geneexp Indicating initial assignment of
  genes according to the GO annotation file
• multiple-out-geneexp All genes assigned to the
  optimal clusters

12
What I plan to do?

• Start with a basic version
• Interface for file input.
• Following input parameters can be set by the
  user (Need more input)
• alpha ( 0 lt alpha lt 1)
• r (0 lt r lt 1)
• Type of organism/species for which the dataset is
  to be clustered
• Output
• Display alpha and r in the output file. Also, as
  the users can change this parameter it will be a
  better idea to name the generated file such that
  it indicates the value of alpha and r.
• Generate output files such that it can be viewed
  in Maple Tree software. (Luis has worked some
  more on this particular function)
• Display compactness and separation in the output
  file.
• Build up-on the basic version to add the required
  features
• Use the Maple Tree source code /develop our own
  method so that the user can view the output of
  clustering directly in our application and need
  not open another application to view the results.
• Allow the user to get periodic/manual updates of
  GO Annotations (based on evidence codes), GO data
  files and GO Slim files.
• Give users the flexibility to choose which
  version of the above mentioned files they want to
  use.
• Additional features (if time permits)
• Some way/tools to be implemented/collaborated to
  match the names of genes for species when they
  are not uniform across all the datasets. (e.g.
  human).
• More input required in this area. GO Browser ?

13
References

• Tari L et al., Fuzzy c-means clustering with
  prior biological knowledge, J Biomed Inform
  (2008),doi10.1016/j.jbi.2008.05.009
• http//sysbio.fulton.asu.edu/gofuzzy/
• http//en.wikipedia.org/wiki/Data_clustering
• http//www.scholarpedia.org/article/Fuzzy_C-means_
  cluster_analysis
• http//www.geneontology.org

14
Any Questions

• Hope there arent any!
• Thanks everyone!

15
Files for references