geneCBR a case-base reasoning tool for cancer diagnosis using microarray datasets

1
geneCBRa case-base reasoning tool for cancer
diagnosis using microarray datasets

• dr. florentino fdez-riverola
• university of vigo

Computer System of New Generation
2
Outline

• DNA Microarray Technology
• characteristics and model operation overview
• Bioinformatics and AI
• new challenges and emerging research areas
• CBR systems
• case-based reasoning
• GENE-CBR
• human genome analysis using CBR systems
• Demo
• geneCBR in action cancer diagnosis using
  microarrays

2/26
3
Microarrays characteristics

• silicon chips that can measure the expression
  levels of thousands of genes simultaneously
• microarrays are base on a database over 40000
  fragments of genes called expressed sequence tags
  (ESTs)
• allow us for the first time to obtain a global
  view of the cells belonging to
• different individuals
• different time-intervals for the same individual
• different tissues of the same individual
• gene expression profiles can be used as inputs to
  large-scale data analysis as
• fingerprints to build more accurate molecular
  classification
• discovering hidden taxonomies
• Increasing our understanding of normal and
  disease states

3/26
4
Microarrays model operation overview

• how does the chip work?
• Microarray chips incorporate different dyed genes
  tiled in a grid-like fashion
• The individuals DNA to analyze is dyed with a
  different colour
• Both sets of labelled DNA strands are allowed to
  hybridize or bind
• hybridization events are detected identifying
  fluorescent changes in the strands or DNA
• an scanner and the associated software perform
  various forms of image analysis to measure and
  report raw gene expression values
• the scanned intensities show how active the genes
  represented by the ESTs are in the cell
• strong fluorescence indicates that the gene is
  very active in the cell
• no fluorescence indicates that the gene is
  inactive in the cell

scanner
preprocessing
microarray data file
4/26
5
Available data

• bone narrow samples from 43 adult patients with
  Acute Myeloid Leukemia (AML) plus 6 sane
  individuals
• 10 patients with Acute Promyelocytic Leukemia
  APL
• 4 patients with Acute Myeloid Leukemia with
  inv(16) AML-inv(16)
• 7 patients with Acute Monocytic Leukemia
  AML-mono
• 22 patients with Acute non-Monocytic Leukemia
  AML-other
• 6 samples belonging to sane individuals
  control samples

• volume of information processed
• each microarray contains 22.283 ESTs (? genes)
• 49 microarrays 1.091.867 gene expression values

• today available data
• 150 microarrays (Human Genome 133A) 210
  microarrays (Human Genome - plus)

5/26
6
Challenges for microarray Data Mining

• three main types of data analysis needed for
  biomedical applications
• gene selection (? attribute selection in AI)
• find the genes most strongly related to a
  particular class
• classification (? supervised classification in
  AI)
• classifying diseases or predicting outcomes based
  on gene expression patterns, and perhaps even
  identifying the best treatment for given genetic
  signature
• clustering (? unsupervised classification in AI)
• finding new biological classes or refining
  existing ones

• three parallel research areas
• convenient visualization of experiments and
  results
• discovery of biological knowledge (metabolic
  pathways, etc.)
• low-level analysis providing better readouts
  (preprocessing, normalization, etc.)

6/26
7
Problems with existing data

• analysis of microarrays presents a number of
  unique challenges for Machine Learning and Data
  Mining techniques but
• Its capacity for generating enormous amounts of
  data is, however, also an handicap
• great amount of data belonging to each individual
  (thousands of genes)
• efficiency and memory problems
• lack of initial knowledge
• which is the significance level of each gene?
• given the difficulty of collecting microarray
  samples, the number of samples is likely to
  remain small in many interesting cases
• having so many fields relative to so few samples
  creates a high likelihood of finding false
  positives
• these problems are increased if we consider the
  potential errors that can be present in
  microarray data (symmetric and random errors)
• it is required sophisticated data analysis
  techniques and robust methods capable of
  extracting biologically meaningful knowledge from
  the raw data

7/26
8
CBR systems (Case-Based Reasoning)

• Kolodner (1983a, 1983b). Problem solving paradigm
  in AI. It can be viewed as a methodology for
  reasoning and learning
• reasoning by re-using past cases is a powerful
  and frequently applied way to solve problems for
  humans Joh (1997)
• the memory of the system (case base) stores a
  certain number of previously experienced
  situations
• CASE PROBLEM description applied SOLUTION
  RESULT
• a new problem is solved by finding similar past
  cases and reusing them in the new problem
  situation
• Riesbeck et al., (1989)
• 4 cyclical steps are performed when it is
  necessary to solve a new problem
• Kolodner (1993) Aamodt y Plaza (1994) Watson
  (1997)
• Case-based reasoning is - in effect - a cyclic
  and integrated process of solving a problem,
  learning from this experience, solving a new
  problem, and so on...

8/26
9
The CBR cycle

• RETRIEVING
• one or more previously experienced cases

most similar cases
New problem
(1) RETRIEVE
REUSING the case(s) in one way or another
(2) REUSE
REVISING the solution based on reusing a previous
case(s)
(4) RETAIN
(3) REVISE
proposed solution
confirmed solution
RETAINING the new experience by incorporating it
into the existing knowledge-base (case base).
9/26
10
Main characteristics of CBR systems

• adaptive and dynamic systems the number of cases
  stored in the memory of the model changes,
  allowing the system adaptation to new situations
• CBR allow the utilisation of general knowledge in
  the resolution of a particular problem
• CBR facilitate the indexation of the available
  information
• CBR can use uncompleted cases
• CBR are advised about their limitations (perhaps
  a problem has no solution)
• CBR facilitate the utilisation of representative
  and flexible data structures
• case adaptation aids to discover
  inter-connections and hided structures in the
  available data
• CBR can be completely automated

10/26
11
GENE-CBR
11/26
12
Goals
Objectives
GENE-CBR
Develop an effective and reliable system able to
diagnose cancer subtypes based on the analysis of
microarray data
CBR system (Case-Based Reasoning) Solve new
problems (new patient) based on the previous
experience (diagnosed patients)
Implement a flexible tool for designing and
testing new techniques and experiments
AI techniques selection, clustering, inference
BeanShell Programmer interface
Construct an advanced edition module for run-time
modification of coded techniques
12/26
13
Logic architecture
DFP
GCS
DFP
GCS
1 RETRIEVE
2 REUSE
CASE BASE
3 REVISE
4 RETAIN
13/26
14
Model overview
reclassification
Gene Selection
most relevant genes DFP
Clustering
revised prediction and final diagnostic
genetically similar patients
Knowledge Discovery
Initial prediction
Prediction
14/26
15
GENE-CBRi retrieval

• objectives
• perform gene selection without losing information
• extracting simplified fuzzy patterns (FP) for
  each pathology
• possibility of using AI techniques initially
  discarded
• main phases
• supervised fuzzy discretisation of gene
  expression values
• Low, Medium, High and overlapping labels (LM, MH)
• supervised gene selection for each pathology
• advantages
• independence of the ordering existing in data
• takes into account data variability
• allows for discovering new knowledge
• obtained results are interpretable

15/26
16
GENE-CBRi retrieval
16/26
17
GENE-CBRi retrieval
FP_AML-other
FP_healthy
FP_AML-inv()
FP_APL
FP_AML-monocytic
DFP
. . .
17/26
18
GENE-CBRii reuse

• objectives
• unsupervised identification of genetic
  similarities between patients
• taking only into account the previous selected
  genes (DFP)
• main phases
• training a GCS network DFP-dimensional
• Growing Cell Structures. Fritzke, B. (1993)
• presenting the new patient to the network
• classifying using a proportional weighting voting
  schema
• advantages
• clustering without taking into account the
  patient class
• definition of an indexing and similarity
  structure between nodes (? relating patients)
• generation of clusters containing new subtypes of
  unknown cancer (knowledge discovery)

18/26
19
GENE-CBRii reuse
Similarity
- Similarity
AML-inv()
19/26
20
GENE-CBRiii revise

• objectives
• provide doctors with meaningful information about
  the classification carried out by the system
• help in discovering new knowledge
• if-then rules as decision making support
  mechanism
• information supplied
• identification of similar patients (from a
  genetically point of view)
• proportional weighting voting and assigned
  weights
• rules generation using See5. Quinlan, J.R. (2000)
• DFP genes belonging to the set of patients
  retrieved by the GCS network
• advantages
• doctors can supervise the final decision proposed
  by the system
• new knowledge generation in the form of easy
  understandable rules

20/26
21
GENE-CBRiii revise
CARIOTYPE
BIOLOGICAL AND CLINICAL CHARACTERISTICS
Rule 6 (45 / 4, lift 1.1) If
X65962 (AFFX-HSAC07/X00351_5_at) is LOW then
If U96781 (AFFX-BioDn-3_at) is
LOW-MEDIUM then AML-other
Else If D87845 (AFFX-hum_alu_at) is HIGH then
AML-inv() 0.968
21/26
22
GENE-CBRiv retain

• objectives
• feedback the system with new knowledge
• new subclassification of existing cancer
  pathologies
• reclassification of existing patients
• identification of correlated genes
• discovering of new marks able to distinguish new
  pathologies
• Identification of prototypical patients and rare
  cases
• main phases
• update the case base with new a microarray every
  time a new classification is generated
• modification of the parameters of the model
• advantages
• possibility of easily integrating new biological
  knowledge in the hybrid system

22/26
23
Applied technologies

• Design patterns
• Action
• Future
• MVC
• Singleton
• Wizard

• 100 Java
• Swing
• BeanShell
• Log4j
• JFreeChart

• Unified Modeling Language
• Poseidon for UML

23/26
24
Future work

• going through a plug-in architecture
• designing a core where each technique is
  implemented as a plug-in gt aiBENCH
• implementing fold-cross validation
• generation of multiple training and test cases in
  an automatic way
• supporting standard microarray data formats
• MIAME Minimum Information About a Microarray
  Experiment
• deploying of GENE-CBR with JavaWebStart
• remote and automatic access to latest versions of
  GENE-CBR project
• on-line access to genetic sequence databases
• geneBank (http//www.ncbi.nlm.nih.gov/Genbank)

24/26
25
Demo GENE-CBR in action
25/26
26
geneCBRa case-base reasoning tool for cancer
diagnosis using microarray datasets

• dr. florentino fdez-riverola
• university of vigo

Computer System of New Generation