müßt

Peter Bajcsy, PhDAutomated Learning GroupNational Center for Supercomputing ApplicationsUniversity of [email protected]

February 6 & 13, 2003

Data Mining in Bioinformatics

2

Outline

• Introduction—Interdisciplinary Problem Statement—Microarray Problem Overview

• Microarray Data Processing—Image Analysis and Data Mining—Prior Knowledge—Data Mining Methods—Database and Optimization Techniques—Visualization

• Validation

• Artificial Immune Systems

• Summary

3

Introduction: Recommended Literature

1. Bioinformatics – The Machine Learning Approach by P. Baldi & S. Brunak, 2nd edition, The MIT Press, 2001

2. Data Mining – Concepts and Techniques by J. Han & M. Kamber, Morgan Kaufmann Publishers, 2001

3. Pattern Classification by R. Duda, P. Hart and D. Stork, 2nd edition, John Wiley & Sons, 2001

4

Bioinformatics, Computational Biology, Data Mining

• Bioinformatics is an interdisciplinary field about the information processing problems in computational biology and a unified treatment of the data mining methods for solving these problems.

• Computational Biology is about modeling real data and simulating unknown data of biological entities, e.g.— Genomes (viruses, bacteria, fungi, plants, insects,…)— Proteins and Proteomes— Biological Sequences— Molecular Function and Structure

• Data Mining is searching for knowledge in data— Knowledge mining from databases— Knowledge extraction— Data/pattern analysis— Data dredging— Knowledge Discovery in Databases (KDD)

5

Basic Terms in Biology

Example:• The human body contains ~100 trillion cells• Inside each cell is a nucleus• Inside the nucleus are two complete sets of the

human genome (except in egg, sperm cells and blood cells)

• Each set of genomes includes 30,000-80,000 genes on the same 23 chromosomes

• Gene – A functional hereditary unit that occupies a fixed location on a chromosome, has a specific influence on phenotype, and is capable of mutation.

• Chromosome – A DNA containing linear body of the cell nuclei responsible for determination and transmission of hereditary characteristics

6

Basic Terms in Data Mining

• Data Mining: A step in the knowledge discovery process consisting of particular algorithms (methods) that under some acceptable objective, produces a particular enumeration of patterns (models) over the data.

• Knowledge Discovery Process: The process of using data mining methods (algorithms) to extract (identify) what is deemed knowledge according to the specifications of measures and thresholds, using a database along with any necessary preprocessing or transformations.

• A pattern is a conservative statement about a probability distribution. — Webster: A pattern is (a) a natural or chance configuration,

(b) a reliable sample of traits, acts, tendencies, or other observable characteristics of a person, group, or institution

7

Introduction: Problems in Bioinformatics Domain

• Problems in Bioinformatics Domain—Data production at the levels of molecules,

cells, organs, organisms, populations—Integration of structure and function data,

gene expression data, pathway data, phenotypic and clinical data, …

—Prediction of Molecular Function and Structure

—Computational biology: synthesis (simulations) and analysis (machine learning)

8

MICROARRAY PROBLEM

9

Microarray Problem: Major Objective

• Major Objective: Discover a comprehensive theory of life’s organization at the molecular level—The major actors of molecular biology: the

nucleic acids, DeoxyriboNucleic acid (DNA) and RiboNucleic Acids (RNA)

—The central dogma of molecular biology

Proteins are very complicated molecules with 20 different amino acids.

10

Input and Output of Microarray Data Analysis

• Input: Laser image scans (data) and underlying experiment hypotheses or experiment designs (prior knowledge)

• Output: —Conclusions about the input hypotheses or knowledge

about statistical behavior of measurements—The theory of biological systems learnt automatically

from data (machine learning perspective)– Model fitting, Inference process

11

Overview of Microarray Problem

Data Mining

Microarray Experiment

Image Analysis

Biology Application Domain

Experiment Design and Hypothesis

Data Analysis

Artificial Intelligence (AI)

Knowledge discovery in databases (KDD)

Data Warehouse

Validation

Statistics

12

Statistics Community

• Random Variables

• Statistical Measures

• Probability and Probability Distribution

• Confidence Interval Estimations

• Test of Hypotheses

• Goodness of Fit

• Regression and Correlation Analysis

13

Artificial Intelligence (AI) Community

• Issues:—Prior knowledge

(e.g., invariance)

—Model deviation from true model

—Sampling distributions

—Computational complexity

—Model complexity (overfitting)

Collect Data

Train Classifier

Choose Model

Choose Features

Evaluate Classifier

Design Cycle of Predictive Modeling

14

Knowledge Discovery in Databases (KDD) Community

Database

15

Microarray Data Mining and Image Analysis Steps• Image Analysis

— Normalization— Grid Alignment— Spot Quality Assurance Control— Feature construction (selection and extraction)

• Data Mining— Prior knowledge— Statistics— Machine learning— Pattern recognition— Database techniques— Optimization techniques— Visualization

• Validation— Issues— Cross validation techniques

?

16

MICROARRAY IMAGE ANALYSIS

17

Microarray Image Analysis

18

DATA MINING OF MICROARRAY DATA

19

Why Data Mining ? Sequence Example

• Biology: Language and Goals• A gene can be defined as a region of DNA.• A genome is one haploid set of chromosomes with the

genes they contain.• Perform competent comparison of gene sequences

across species and account for inherently noisy biological sequences due to random variability amplified by evolution

• Assumption: if a gene has high similarity to another gene then they perform the same function

• Analysis: Language and Goals• Feature is an extractable attribute or measurement

(e.g., gene expression, location)• Pattern recognition is trying to characterize data

pattern (e.g., similar gene expressions, equidistant gene locations).

• Data mining is about uncovering patterns, anomalies and statistically significant structures in data (e.g., find two similar gene expressions with confidence > x)

20

Types of Expected Data Mining and Analysis ResultsHypothetical Examples:• Binary answers using tests of hypotheses

—Drug treatment is successful with a confidence level x.

• Statistical behavior (probability distribution functions)—A class of genes with functionality X follows Poisson

distribution.• Expected events

—As the amount of treatment will increase the gene expression level will decrease.

• Relationships—Expression level of gene A is correlated with

expression level of gene B under varying treatment conditions (gene A and B are part of the same pathway).

• Decision trees —Classification of a new gene sequence by a “domain

expert”.

21

PRIOR KNOWLEDGE

22

Prior Knowledge: Experiment Design

• Microarray sources of systematic and random errors

• Feature selection and variability

• Expectations and Hypotheses

• Data cleaning and transformations

• Data mining method selection

• Interpretation

Collect Data

Choose Features

Data Cleaning and Transformations

Choose Model and Data Mining Method

Pri

or K

now

ledg

e

23

Prior Knowledge from Experiment Design

Complexity Levels of Microarray Experiments:1. Compare single gene in a control situation versus a treatment

situation• Example: Is the level of expression (up-regulated or down-regulated)

significantly different in the two situations? (drug design application)• Methods: t-test, Bayesian approach

2. Find multiple genes that share common functionalities• Example: Find related genes that are dependent?• Methods: Clustering (hierarchical, k-means, self-organizing maps,

neural network, support vector machines)3. Infer the underlying gene and protein networks that are

responsible for the patterns and functional pathways observed• Example: What is the gene regulation at system level?• Directions: mining regulatory regions, modeling regulatory networks

on a global scaleGoal of Future Experiment Designs: Understand biology at the system

level, e.g., gene networks, protein networks, signaling networks, metabolic networks, immune system and neuronal networks.

24

Data Mining Techniques

S ta tis t ics M a ch in e lea rn ing

D a ta b ase te chn iqu es P a tte rn re co g n it ion

O p tim iza tio n te ch n iq u es

D a ta m in in g tech n iq u e s d ra w from

Visualization

25

STATISTICS

26

Statistics

Inductive Statistics

Statistics

Descriptive Statistics

Are two sample sets

identically distributed ?

Make forecast and inferences

Describe data

27

•Gene Expression Level in Control and Treatment situations

•Is the behavior of a single gene different in Control situation than in Treatment situation ?

Statistical t-test

• m – sample mean

• s – variance

Normalized distance Normalized distance t follows a Student distributionwith f degrees of freedom.

If t>thresh then the control and treatment data populations are considered to be different.

?

28

MACHINE LEARNINGAND

PATTERN RECOGNITION

29

Machine Learning

Supervised

Machine Learning

Unsupervised

Reinforcement“Natural groupings”

Examples

30

Pattern Recognition

Pattern Recognition

Linear Correlation and Regression

Neural Networks

Statistical Models

Decision Trees

Locally Weighted Learning

NN representation and gradient based optimization

NN representation and genetic algorithm based optimization

k-nearest neighbors, support vectors

31

Unsupervised Learning and Clustering

• A cluster is a collection of data objects that are similar to one another within the same cluster and are dissimilar to the objects in other clusters.

• Examples of data objects:—gene expression levels, sets of co-regulated genes

(pathways), protein structures

• Categories of Clustering Methods—Partitioning Methods—Hierarchical Methods—Density-Based Methods

“Natural groupings”

32

Unsupervised Clustering: Partitioning Methods

• K-means Algorithm partitions a set of n objects into k clusters so that the resulting intra-cluster similarity is high but the inter-cluster similarity is low.

• Input: number of desired cluster k

• Output: k labels assigned to n objects

• Steps:1.Select k initial cluster’s centers 2.Compute similarity as a distance between an object

and each cluster center3.Assign a label to an object based on the minimum

similarity4.Repeat for all objects5.Re-compute the cluster’s centers as a mean of all

objects assign to a given cluster6.Repeat from Step 2 until objects do not change their

labels.

Example: Centroid-Based Technique

33

Unsupervised Clustering: Partitioning Methods

• K-medoids Algorithm partitions a set of n objects into k clusters so that it minimizes the sum of the dissimilarities of all the objects to their nearest medoid.

• Input: number of desired cluster k• Output: k labels assigned to n objects• Steps:1.Select k initial objects as the initial medoids2.Compute similarity as a distance between an

object and each cluster medoid3.Assign a label to an object based on the minimum

similarity4.Repeat for all objects5.Randomly select a non-medoid object and swap

with the current medoid it would decrease intra-cluster square error

6.Repeat from Step 2 until objects do not change their labels.

Example: Representative-Based Technique

34

Unsupervised Clustering: Hierarchical Clustering

• Hierarchical Clustering partitions a set of n objects into a tree of clusters

• Types of Hierarchical Clustering

—Agglomerative hierarchical clustering– Bottom-up strategy of building clusters

—Divisive hierarchical clustering– Top-down strategy of building clusters

35

Unsupervised Agglomerative Hierarchical Clustering• Agglomerative Hierarchical Clustering partitions a set

of n objects into a tree of clusters with a bottom-up strategy.

• Steps:1. Assign a unique label to each data object and form n

clusters2. Find nearest clusters and merge them3. Repeat Step 2 till the number of desired clusters is equal to

the number of merged clusters.

• Types of Agglomerative Hierarchical Clustering— The nearest neighbor algorithms (minimum or single-linkage algorithm, minimal

spanning tree)— The farthest neighbor algorithms (maximum or complete-linkage algorithm)

36

Unsupervised Clustering: Density-Based Clustering

• Density-Based Spatial Clustering with Noise aggregates objects into clusters if the objects are density connected.

• Density connected objects:—Simplified explanation: P and Q are density

connected if there is an object O such that both P and Q are density connected to O.

—Aggregate P and Q if they are density connected with respect to R-radius neighborhood and Minimum Object criteria

37

Supervised Learning or Classification

• Classification is a two-step process consisting of learning classification rules followed by assignment of classification label.

38

Supervised Learning: Decision Tree

• Decision tree algorithm constructs a tree structure in a top-down recursive divide-and-conquer manner

Car Insurance: Risk Assessment

Age < 25 ?

Risk: LowRisk: High

Sports car ?Risk: High

Age Car Type

Risk

23 family High

17 sports High

43 sports High

68 family Low

32 truck Low

20 family High

yes no

noyes

Attributes

Answers

Visualization of Decision Boundaries

39

Supervised Learning: Bayesian Classification

• Bayesian Classification is based on Bayes theorem and it can predict class membership probabilities.

• Bayes Theorem (X-data sample, H-hypothesis of data label)—P(H/X) posterior probability—P(H) prior probability

• Classification-maximum posteriori hypothesis

40

Statistical Models: Linear Discriminant

• Linear Discriminant Functions form boundaries between data classes.

• Finding Linear Discriminant Functions is achieved by minimizing a criterion error function.

Linear discriminant function

Quadratic discriminant function

Finding w coefficients:

-Gradient Descent Procedures

-Newton’s algorithm

41

Artificial Neural Networks

• Artificial Neural Network (ANN) is a computational analogue of neurons.

• Artificial neural network is a set of connected input/output units where each connection has a weight associated with it.

• Phase I: learning – adjust weights such that the network predicts accurately class labels of the input samples

• Phase II: classification- assign labels by passing an unknown sample through the network

Network topology or “Structure”

42

Artificial Neural Networks (cont.)

• Steps:1. Initial weights from [-1,1]

2. Propagate the inputs forward3. Backpropagate the error

4. Terminate learning (training) if (a) delta w < thresh or (b) percentage of misclassified samples < thresh or (c) max number of iterations has been exceeded

• Pros & Cons of ANN: Good performance with noisy data, rule extraction & long training, poor interpretability, trial-and-error network design

Interpretation

Unit or node j

43

Support Vector Machines (SVM)

• SVM algorithm finds a separating hyperplane with the largest margin and uses it for classification of new samples

44

DATABASE TECHNIQUESAND

OPTIMIZATION TECHNIQUES

45

Data Types and Databases

• Relational Databases

• Data Warehouses

• Transactional Databases

• Advanced Database Systems—Object-Relational—Spatial and Temporal—Time-Series—Multimedia—Text—Heterogeneous, Legacy, and Distributed—WWW

Structure - 3D Anatomy

Function – 1D Signal

Metadata – Annotation

46

Database Techniques

• Database Design and Modeling (tables, procedures, functions, constraints)

• Database Interface to Data Mining System

• Efficient Import and Export of Data

• Database Data Visualization

• Database Clustering for Access Efficiency

• Database Performance Tuning (memory usage, query encoding)

• Database Parallel Processing (multiple servers and CPUs)

• Distributed Information Repositories (data warehouse)

MINING

47

Search and Optimization Techniques: Search Types• Types of search methods:

• Calculus-based

•Indirect (solve a nonlinear set of equations)

•Direct (follow local gradient - hill climbing)

• Enumerative (search objective function values at every point – dynamic programming)

• Random (search with random sampling)

• Randomized search methods: guide the search with random processes – simulated annealing, genetic programming

48

Search and Optimization Techniques: Challenges• Search and optimization challenges:

• Global versus local maxima• Existence of derivatives (calculus-based)• High dimensionality• Highly nonlinear search space (global

versus local maxima)• Large search space

• Example: A genome with N genes can encode 2^N states (active or inactive states, regulated is not considered). Human genome ~ 2^30,000; Nematode genome ~ 2^20,000 patterns.

49

Genetic Algorithm

• Genetic Algorithm (GA) based optimization is a computational analogue of Darwin’s evolution theory (survival of the fittest).

• Description of GA based optimization:

• Uses coding of the parameter set (not the parameters themselves)

• Searches from a population of points (not a single point)

• Uses an objective function (not derivatives or other auxiliary knowledge)

• Employs probability transition rules (not deterministic rules)

• Is composed of three operators

•Reproduction (or selection)

•Crossover

•MutationReference: D. Goldberg: Genetic Algorithms in Search, Optimization & Machine

Learning,Addison-Wesley Publishing Co., 1989.

50

Genetic Algorithm: Additional Operators

• Additional operators

• Niching for optimization of multimodal and multiobjective functions

•Fitness sharing: the number of individuals residing near any peak will be proportional to the height of that peak (reduce individual fitness according to their similarity)

•Crowding: spread individuals among the most prominent peaks and do not allocate individuals proportionally to fitness (maintain diversity)

• Speciation for optimization of multimodal functions

•Mating restriction scheme (restrict mating or crossover according to the similarity among individuals)

51

• Steps:• Randomly generate initial population of size n=2; e.g.,

strings 0110 & 1100• Reproduction is a process of copying strings according to

their objective function – “a roulette wheel” • Crossover proceeds in two steps (1) random mating of

strings and (2) selecting random positions of each string for mating; e.g., obtain 1110 & 0100

• Mutation is the occasional random alteration of the value of a string position to protect premature loss of information; obtain 0110 & 0100

Genetic Algorithm: Example

Objective FunctionOn Off

OnOff

(on,off,on,off) input sequence is converted to a string (1010)

52

VISUALIZATION

53

Visualization

• Data: 3D cubes,distribution charts, curves, surfaces, link graphs, image frames and movies, parallel coordinates

• Results: pie charts, scatter plots, box plots, association rules, parallel coordinates, dendograms, temporal evolution

Pie chart Parallel coordinates

Temporal evolution

54

Novel Visualization of Features

Feature Selection and Visualization

Feature Selection

Mean Feature Image

55

Novel Visualization of Clustering Results

Isodata (K-means)Clustering

Class Labeling and Visualization

Mean Feature Image Label Image

56

VALIDATION

57

Why Validation?

• Validation type:— Within the existing data— With newly collected data

• Errors and uncertainties:— Systematic or random errors— Unknown variables - number of classes— Noise level - statistical confidence due to noise— Model validity – error measure, model over-fit or under-fit — Number of data points - measurement replicas

• Other issues— Experimental support of general theories— Exhaustive sampling is not permissive

58

Error Detection: Example of Spot Screening

Mask Image – No ScreeningMask Image – Location and Size Screening

Mask Image – SNR Screening

59

Cross Validation: Example

• One-tier cross validation— Train on different data than test data

• Two-tier cross validation— The score from one-tier cross validation is

used by the bias optimizer to select the best learning algorithm parameters (# of control points) . The more you optimize the more you over-fit. The second tier is to measure the level of over-fit (unbiased measure of accuracy).

— Useful for comparing learning algorithms with control parameters that are optimized.

— Number of folds is not optimized.• Computational complexity:

— #folds of top tier X #folds of bottom tier X #control points X CPU of algorithm

60

ARTIFICIAL IMMUNE SYSTEMS

61

Artificial Immune Systems

• Artificial Immune Systems (AIS) are adaptive systems, inspired by theoretical immunology and observed immune functions, principles and models, which are applied to problem solving.

• Other types of AIS are hybrids of ANN, GA and fuzzy systems combined with theoretical immunology models

• Applications of AIS:— Pattern recognition (surveillance of infectious diseases)— Fault and anomaly detection ((image inspection and segmentation)— Data analysis (reinforced, unsupervised learning)— Agent-based systems— Scheduling (adaptive scheduling)— Autonomous navigation and control (walking robots)— Search and optimization methods (constrained, time-dependent

optimization)— Security of information systems (virus detection, network intrusion)

62

Basic Terms Used in Artificial Immune Systems

• Immune system is understood as a complex set of cells and molecules that protect our bodies against infection under constant attack by antigens (foreign or self-antigens)

• Immune system consists of two-tier line of defense: adaptive (lymphocytes: B-cells & T-cells) and innate (granulocytes & macrophages) immune systems. Both systems depend upon the activity of white blood cells (leukocytes).

The organs that make up the immune system (lymphoid organs) are thymus & bone marrow (primary) and tonsils,adenoids, spleen, appendix, lymph nodes, lymphatic vessels, peyer’s patches (secondary).

63

Mechanisms Adapted in Artificial Immune Systems• Pattern recognition: lymphocytes (B-cells & T-

cells) carry surface receptors capable of recognizing antigens— Example: recognition via complementary

regions

• The clonal selection principle: only cells capable of recognizing an antigen stimulus will proliferate and differentiate into effector cells

• Immune learning and memory: reinforced learning strategy

• Self/Nonself discrimination: distinguish between molecules of its own cell (self) and foreign molecules (nonself)- positive and negative selection, clonal expansion and ignorance

64

Why Artificial Immune System?

• Pattern recognition: cells and molecules of the immune system have several ways of recognizing patterns

• Uniqueness: each individual possesses its own immune system• Self identity: other than native “elements” to the body can be

recognized and eliminated by the immune system• Diversity: there exist varying types of elements that together protect

the body• Disposability: no single native element is essential for the functioning

of the immune system• Autonomy: there is no central element controlling the immune system • Multi-layered: multiple layers of different mechanisms provide overall

security• No secure layer: any cell of the organism can be attacked by the IS• Anomaly detection: the IS can recognize and react to pathogens that

the body has never encountered before• Dynamically changing coverage: the IS maintains a circulating

repertoire of lymphocytes constantly being changed through cell death, production and reproduction

65

Why Artificial Immune System? (cont.)

• Distributivity: the immune elements are distributed all over the body

• Noise tolerance: an absolute recognition of pathogens is not required (tolerance to molecular noise)

• Resilience: the IS is capable of functioning despite disturbances• Fault tolerance: the complementary roles of several immune

components allow the re-allocation of tasks to other elements• Robustness: diversity & number of immune elements• Immune learning and memory: the molecules of the IS can adapt

to themselves, structurally and in number, to the antigenic challenges

• Predator-prey pattern of response:#pathogens goes up =>#immune cells goes up

• Self-organization: clonal selection and affinity maturation are responsible for selecting the most adapted cells to be maintained as long living memory cells

• Integration with other systems: the IS communicates with parts of the body

66

General Framework for Artificial Immune Systems• General Framework for AIS:

—A representation for the components of the system—A set of mechanisms to evaluate the interaction of

individuals with the environment and each other (input stimuli, 1 to N fitness functions or other means) – Affinity measures

—Procedures of adaptation that govern the dynamics of the system (e.g., behavior over time) - Algorithms

Reference:L. N. de Castro and J. Timmis, “Artificial Immune Systems: A New Computational Intelligence Approach,”Springer 2002.

67

Components of Artificial Immune Systems

• Representation: —Generalized shape of any molecule in shape space is

described by an attribute string (set of coordinates) of length L.

—Shape-space describes interactions between molecules of the immune system and antigens.

—Immune system is represented as a pattern (molecular) recognition system that is designed to identify shapes.

• Affinity Measures: —Euclidean, Manhattan and Hamming —Real-valued, integer, symbolic or alphabet sub-string

spaces

68

Components of Artificial Immune Systems

• Immune system algorithms: — Bone marrow model: generate repertoire of cells and

molecules (generate random attribute strings)— Thymus model: generate repertoire of cells and molecules

capable of performing self/non-self discrimination (Positive selection: initialize strings, evaluate affinity and keep strings with affinity < threshold; Negative selection: eliminate strings > threshold)

— Clonal selection algorithms: modeling interaction control of the IS and external environment or antigens (similar to GA without crossover and with affinity proportional to reproduction and mutation)

— Immune network models: simulate immune networks (differential equations describing the dynamics)

69

Examples of Artificial Immune Systems

Example ANN GA AIS

Representation

Artificial neurons & interconnection of neurons (summing junction, connection strength, activation)

Genetic representation (gene = a single bit or a block of bits, chromosome= bitstring)

IS representation of molecules (strings of coordinates), and their interactions (shape-space)

Affinity Measures

Backpropagation measures

Fitness function Affinity defined on a shape-space

Algorithms

Learning algorithms Procedures for reproduction, genetic variation and selection

Procedures for generation, cloning, selection and IS network dynamics

70

SUMMARY

71

Summary: Interdisciplinary Science

• CS and ECE have been used to gain a better understanding of biological processes through modeling and simulation

• CS and ECE have been enriched with the introduction of biological ideas, e.g., ANN, GA, cellular automata, artificial life, artificial immune systems (AIS)

• New fields: bio-informatics, bio-medical engineering

• Bilateral interactions between CS, ECE and Biology:—Biologically motivated computing (ANN, GA, artificial

immune systems) —Computationally motivated biology (cellular automata)—Computing with biological mechanisms (silicon-based

computing => quantum and DNA computing)

72

Summary: Bioinformatics

• Bioinformatics and Microarray problem— Interdisciplinary Challenges: Terminology— Understanding Biology and Computer Science

• Data mining and image analysis steps— Image Analysis— Experiment Design as Prior Knowledge — Expected Results of Data Mining— Which Data Mining Technique to Use?— Data Mining Challenges: Complexity, Data Size, Search Space

• Validation— Confidence in Obtained Results?— Error Screening— Cross validation techniques

• Artificial Systems— Biologically motivated computing

73

Backup