BIOMEDICAL DATA MANAGEMENT AND ANALYSIS Emanuel Weitschek

www.dia.uniroma3.it/~emanuel

emanuel@dia.uniroma3.it

2

Outline

• Growth of biological data

• DNA Sequencing

• Growth of clinical data

• Analysis of biomedical data

• Bioinformatics

• GenData 2020

• The Cancer Genome Atlas and GenData

• Data Mining: classification and clustering

• Data mining systems: Weka and DMB

• Application to biological data sets

− Clinical patients

− Gene Expression Profiles

− Sequences: DNA Barcode, Polyomaviruses, Bacteria and CNEs (with alignment free methods)

− EEG signal processing

• Projects

3

Growth of biological data

• Advances in molecular biology lead to an exponential growth of biological data thanks to the support of computer science

‒ originated by the DNA sequencing method invented by Sanger in early eighties

‒ late nineties significant advances in sequence generation, e.g. Human Genome Project

‒ actually the genomic sequences are doubling every 18 months

‒ GenBank: collection of all publicly available nucleotide sequences (160 M seq)

the era of “Big Data”

4

Growth of biological data

• Advances in molecular biology lead to an exponential growth of biological data thanks to the support of computer science

‒ Today next generation high throughput data from modern parallel sequencing machines, are collected and huge amounts of biological data are currently available on public and private sources

‒ 10000 Human Genomes project (3000 Mbp)

‒ The future: 1000$ genome

• Very large data sets, that are generated by several different biological experiments, need to be automatically processed and analyzed with computer science methods

5

DNA Sequencing

• DNA (deoxyribonucleic acid) is the hereditary material in almost all organisms

• DNA sequencing is the process of determining the order of nucleotides within a DNA molecule

• It includes any method or technology that is used to determine the order of the four bases—adenine (A), cytosine (C), guanine (G), and thymine (T)

• Originated by the DNA sequencing method invented by Sanger in early eighties

• In late nineties significant advances in sequence generation techniques, largely inspired by massive projects such as the Human Genome Project

• High costs and time, e.g., for the Human Genome Project 5 billions $ and 13 years

6

Next Generation Sequencing (NGS)

• Today: next generation high throughput data from modern parallel sequencing machines ‒ Roche 454, Illumina, Applied Biosystems SOLiD,

Helicos Heliscope, Complete Genomics, Pacific Biosciences SMRT, ION Torrent

‒ Next generation sequencing (NGS) machines output a large amount of short DNA sequences, called reads (in fastq format)

‒ Cannot read entire genome one nucleotides at a time from beginning to end

‒ shred the genome and generate shorts reads

‒ Low cost per base (5000$ for a whole human genome)

‒ High speed (24h to sequence a whole human genome )

‒ Large number of reads

‒ Problems: data storage and analysis, high costs for IT infrastructure

7

How do we assemble the genome?

8

How do we assemble the genome?

• Take small tissue or blood sample containing millions of cells with identical DNA

• Use biochemical methods to break the DNA into fragments

• Sequence these fragments to produce reads

• Genome assembly: putting a genome back together from its reads (it is just like reassembling a newspaper)

• Biologists can easily generate enough reads to analyze a large genome, but assembling these reads still presents a major computational challenge

• The difficulty is that researchers do not know where in the genome these reads came from, and so they must use overlapping reads to reconstruct the genome

9

NGS experiments

DNA-seq • Sequence the DNA (whole genome) of an organism/sample

• De Novo assembly and Short read mapping

• Align the sequence to a reference genome (mutations analysis)

RNA-seq • Transcriptome sequencing to detect the gene expression

profiles

• Count the reads that map on a particular gene region (quantitative measures)

Chip-seq • Detection of genome-wide protein-DNA interactions

• Determines if a protein is bound to potential target regions

DNA methylation • Determines epigenetic modifications which influence gene

expression and cell phenotypes

• Epigenetics: study of changes in gene expression or cellular phenotype, caused by mechanisms other than changes in the underlying DNA sequence

10

Growth of clinical data

• The wide spread of electronic data collection in medical environments lead to an exponential growth of clinical data from heterogeneous patient samples

‒ Electronic health records

‒ European and national projects

‒ Integration with genomics

• Very large data sets, that are generated by several different clinical experiments, need

to be automatically processed and analyzed with computer science methods

the era of “Big Data”

11

Growth of clinical data

• Collection of different electronic health patient records

• Data set obtained often from different health care facilities and hospitals

• Data collected by medical doctors or assistans: manual insertion of the clinical trials values or copy from non electronic medical records

• Normally no integrated IT system

• Therefore clinical data sets are often noisy, full of missing data and outliers

• To perform a data mining analysis the records have to be integrated in an unique data set

– by joining on common attributes

– by leaving all the non shared attributes as supplementary data

• Privacy issues: every patient must remain anonymous by mapping it with a private id, that has not to be visible to the data analyst

12

Analysis of biomedical data

• Analyzing these enormous amount of data is becoming very important in order to shed light on biological and medical questions

• Challenges:

‒ data collection

‒ data integration

‒ managing this huge amount of data

‒ discovering the interactions

‒ the integration of the biological know-how

13

Analysis of biomedical data

• Major issues:

– incompleteness (missing values)

– different adopted measure scales

– integration of the disparate collection procedures

• Major problems in clinical databases

– Complexity

– Inaccuracy

– Frequent missing values

– Different adopted measure scales

– Mixed (numeric – textual) variables, e.g. BUN = {12, 16, high, 9, 11 …}

• Final goal: extract relevant information from huge amounts of biomedical data

14

Bioinformatics

• New methods are demanded able to extract relevant information from biological data sets

• Effective and efficient computer science methods are needed to support the analysis of complex biological data sets

• Modern biology is frequently combined with computer science, leading to Bioinformatics

• Bioinformatics is a discipline where biology and computer science merge together in order to design and develop efficient methods for analyzing biological data, for supporting in vivo, in vitro and in silicio experiments and for automatically solving complex life science problems

• Bioinformatician: a computer scientist and biology domain expert, who is able to deal with the computer aided resolution of life science problems

15

Analysis of genomic sequences

• Massive sequencing experiments lead to an increasing availability of biological sequences

• Sequence analysis

− Alignment based methods

Exact methods

Heuristics

− Alignment free methods

Substrings frequencies

Sequence compression

16

GenData 2020

• Data standard definition for genomics and biomedical data • http://www.bioinformatics.deib.polimi.it/genomic_computing

• Build the abstractions, models, and protocols for supporting a network of genomic data

• Genome servers located in the major biologist laboratories in the world

• Distributed computing

• Definition of methods for querying, searching, and analyzing genomic data

• Problems: data management

• huge amount of data

• diversity of the platforms

• diversity of the formats

• How to model and store genetic data so as to gain their integrated accessibility

• Data definition based on the standards proposed by the Functional Genomics Data Society

• Use of Distributed Annotation System (DAS) that defines a communication protocol used to exchange data on genomic or protein sequences

17

GenData 2020

• Final aim: Definition of an unique and global platform for effectively storing, searching and retrieving genomic data via distributed computing

• Focus NGS experiments:

– DNA-seq, RNA-seq, ChIP-seq, DNA methylation

• GenData Format:

– Within the same dataset, two kinds of data:

• Region values aligned w.r.t. a given reference, with specific left-right ends within a chromosome

• Metadata, with free-format attributes, storing all the knowledge about the dataset

18

GenData 2020

19

GenData 2020

• Genometric Query Language (GMQL) is defined as a sequence of algebraic operations following the structure: < variable > = < operator > (< parameters >) < variable >

• Every variable is a collection of datasets describing experiments of the same level

• Offers high-level, declarative operations based on:

– regions and their properties

– general-purpose meta-data

• Inspired by Pig Latin and targeted towards cloud computing

• Will be further developed by adding genometric operations and statistical operations (e.g. clustering) in cooperation with biologists.

• The language allows for queries on the genome involving large datasets describing:

– genomic signals

– reference regions

– distance rules

• http://www.bioinformatics.deib.polimi.it/genomic_computing/GMQL/

20

GenData 2020

21

GenData 2020

22

GenData 2020

23

The Cancer Genome Atlas (TCGA)

• Comprehensive genomic characterization and analysis of more than 30 cancer type tissues to accelerate the understanding of the molecular basis of cancer through the application of genome analysis technologies, including large-scale genome sequencing

• (2006) Coordinated joint effort of the National Cancer Institute (NCI) and the National Human Genome Research Institute (NHGRI) both of the National Institute of Health (NIH) - USA

• Aim: improve the ability to diagnose, treat and prevent cancer

• A free-available platform to search, download, and analyze data sets containing clinical information, genomic characterization data, and high level sequence analysis of the tumor genomes

• Open access

24

Data types

• Tissue pathology data

• Images

• Pathology reports

• Copy-number alterations for non-genetic platforms

• Epigenetic data

• Data summaries, such as genotype frequencies

• Clinical data (biotab and xml)

• DNA Sequencing (whole genome, whole exome, mutations)

• DNA Methylation

• Gene expression data (miRNA, mRNA, Total RNA, Micro- and Protein-arrays)

• Copy numbers

• https://tcga-data.nci.nih.gov/tcga/tcgaDataType.jsp

25

Data archive description

• https://wiki.nci.nih.gov/display/TCGA/Data+archive

26

Metadata (Clinical Data)

• Clinical information about the participant

• Information about how participant samples (biospecimens) were processed by the TCGA Biospecimen Core Resource Center (BCR)

• Available in XML and as a flatfile biotab format

• Example of TCGA clinical data:

− vital status at time of report disease-specific diagnostic information, and initial treatment regimens

− additional clinical follow up information for some or all participants

• Data elements and fields must follow the rules of XML Schema Document (XSD) http://tcga-data.nci.nih.gov/docs/xsd/BCR/

27

TCGA: some statistics

• 30 different tumor types

• 9404 patients

• 13.45 TB experimental data (clinical and genomic)

• 300 different metadata attributes − clinical reports (pdf), images (dcm), blood tests (biotab)

0

200

400

600

800

1000

1200

LAM

L

BLC

A

BR

CA

CH

OL

ESC

A

HN

SC

KIR

C

LIH

C

LUSC

MES

O

PA

AD

PR

AD

SAR

C

STA

D

THYM UC

S

UV

M

Number of patients

Patients

0

1000

2000

3000

4000

5000

6000

7000

LAM

L

BLC

A

BR

CA

CH

OL

ESC

A

HN

SC

KIR

C

LIH

C

LUSC

MES

O

PA

AD

PR

AD

SAR

C

STA

D

THYM UC

S

UV

M

Data quantity (GB)

Data (GB)

28

RNA-Seq, DNA-Methylation, Mutations in Cancer

GELA (Gene Expression Logic Analyzer)

1. Data extraction from The Cancer Genome

Atlas (TCGA), the major repository for

cancer data

2. Focus on breast cancer (1000 patients)

3. Transformation in GenData format

4. Matrix extraction in GMQL

5. Knowledge extraction with GELA and

supervised machine learning techniques

In collaboration with:

Masseroli M., Fiscon G., Cumbo F., Ceri S.

29

Data Mining

• The interdisciplinary field of data mining, which guides the automated knowledge discovery process, is a natural way to approach the complex task of biological data analysis… with the aid of a good bioinformatician

• Data Mining: extraction of knowledge with computer science algorithms for discovering hidden information in heterogeneous data

• The knowledge discovery process:

• Main Applications:

– Classification

– Clustering

• Classification is the action of assigning an unknown object into a predefined class after examining its characteristics

• Clustering partitions objects into groups, such that similar or each other related objects are in same clusters

30

Classification and supervised machine learning

• Classification is the action of assigning an unknown object into a predefined class after examining its characteristics

• Goal: assign an unknown sample to a known class starting from its attributes

• The classification problem may be formulated in the following way:

− given a reference library (training set) composed of samples of known class and

− a collection of unknown samples (query set or test set)

− recognize the latter into the class that are present in the library

• To obtain reliable results

− the query set has to contain only samples from the same class that are present in the reference library

− the reference set has to contain a sufficient number of samples for each class

31

Classification and supervised machine learning

• The user has to provide as input a training set (reference library) containing samples with a priori known class membership

• Based on this training set, the software computes the classification model

• Subsequently, the classification model can be applied to a test set (query set) which contains samples that require classification

• The test set can contain query samples with unknown class membership or, alternatively, samples that also have a priori known species membership, allowing verification of the classifications

32

• Also called logic data mining or classification with logic formulas

• The classifier uses logic propositional formulas in disjunctive or conjuctive normal form (”if then rules”) for classifying the given records, this classification method is also called ruled based

• A logic classifier is a technique for classifying records using a collection of logic propositional formulas (in CNF or DNF), “if… then rules”:

− Antecedent Consequent

− (Condition1) or (Condition2) or … or (Conditionn) Class

− Conditioni: (A1 op v1) and (A2 op v2) and … and (Am op vm)

− A: attribute

− v: value

− op: operator {=, ≠, <, >, ≤, ≥}

Ruled based classification

If pos13 = A and pos400 = G then the virus is RhinoA

33

• WEKA (Waikato Environment for Knowledge Analysis) machine learning software is widely adopted for classification

• WEKA contains several methods to perform supervised classification of general problems

• Input a reference library in arff format

− Data and sequences have to be converted

− sequences have to be of the same region or pre-aligned to the same region

• Weka computes the classification model

• The classification model can be applied to a query set

• For using Weka in biomedical data classification reference and query set have to be converted in arff format or csv

Supervised machine learning: Weka classification software

34

Supervised machine learning: Weka classification software

Mining Big Data using Weka 3 http://www.cs.waikato.ac.nz/ml/weka/bigdata.html

• command-line interface (CLI) to interact with Weka

• use Weka's Knowledge Flow graphical user interface

• write code directly in Java or a Java-based scripting

• train in an incremental fashion

• MOA data stream software containing state-of-the-art algorithms

• Reservoir sampling: incremental sampling

• new packages for distributed data mining:

− distributedWekaBase (“map" and "reduce" tasks)

− distributedWekaHadoop (Hadoop-specific wrappers)

− in the future, other wrappers - for example Spark

35

DMB: a logic data mining system

DMB is based on several computational steps, which have been integrated in the software:

1) Discretization

2) Normalization

3) Feature clustering

4) Feature selection

5) Formulas computation

6) Classification

7) Supplementary Analysis

36

DMB: Discretization

Aim of Discretization :

to transform gene expressions into discrete values

• Converts each gene expression into a new discrete variable that is suitable for treatment into a logic framework

–Determines a set of cut points over the range of values that each variable may assume

–Defines a number of cut points over which the original variable may be considered discrete

• DMB supports two types of discretization, that differ on the rule adopted to select the first set of intervals:

– Unsupervised discretization (unsupervised rules)

– Supervised discretization (supervised rules)

0

0.5

1

1.5

2

2.5

A A A A A A B B B B B B B

Values

Cutpoints

37

DMB: Feature Clustering

Aims of feature clustering:

– To group together similar features

– To reduce the data to be analyzed

• Methods to extract a subset of genes able to characterize the classification model

• Discrete Cluster Analysis (DCA) :

– Discretization is applied to numeric features

– An integer mapping is computed for every feature, that represents the interval in which an experiment falls

– This integer mapping can be represented in a binary form

– Two or more features are merged into the same cluster when their binary representation over the intervals is equal

– Features with the same discretized profile over the samples are clustered

– Finally, a feature for each cluster is elected as its representative

– Clusters composed of a single feature may also be present and are considered as non clustered features

38

DMB: Feature selection

• Aim: identify and remove irrelevant and redundant information from the data set

• Feature selection: selection of the relevant attributes (or features) of the data samples; the identification of a small subset of important attributes or features in a large data set

• The size reduction allows the knowledge extraction algorithm to work faster and more effectively

• Feature selection as a combinatorial problem: Variation of the Set Covering

• To solve the feature selection problem DMB uses an efficient

Greedy Randomized Adaptive Search Procedure (GRASP)

}1,0{

,, ,

max

k

k k

k

k

ijk

x

x

jijixd

otherwise 0

)( 1}1,0{

jkikk

ijij

aaifda

− Data Matrix A (exp. X features)

− xi = 1 if fi is chosen and 0 otherwise

− each constraint is associated with a

pair of items belonging to different

classes

− α is a variable that measures the degree of information redundancy

39

DMB: Formula Extraction

Aim of Formula Extraction :

to extract logic relations that explain the data

• After the output of FS, DMB extracts the logic classification formulas

• The Lsquare method is part of DMB for computing the model of the data, (e.g., the separating ”if-then” formulas or rules)

• Lsquare approaches it as a sequence of Minimum Cost Satisfiability Problems (MinSat), a combinatorial NP-Hard optimization problem

• The problem is solved with an algorithm based on decomposition techniques

• DMB computes for every class of the experimental samples the logic classification formulas in Disjunctive Normal Form

• Inequalities in the form of ”IF Aph1b<0.50” are conjucted in ”AND” and ”OR” clauses

40

DMB: Classification

A logic classifier is a technique for classifying records using a collection of “if… then rules”, named logic formulas:

– Antecedent Consequent

– (Condition1) or (Condition2) or … or (Conditionn) Class

– Conditioni: (A1 op v1) and (A2 op v2) and … and (Am op vm)

– A = attribute; v = value; op = operator {=, ≠, <, >, ≤, ≥}

• Example of logic classification formula is

• The evaluation of the logic formulas and the classification of the samples to the right class is performed with the following steps:

– The formulas are firstly weighted with the Laplace Score on the training set and then applied on the test set for performing the classification assignments

– Additional cut offs of logic formulas with sub-optimal coverage are done by considering the false positive and true positive rates

“IF Aph1b<0.507 then the experimental sample is CONTROL”

41

Logic Data Mining software

DMB software tools:

http://dmb.iasi.cnr.it

Train

files

(dmb)

Test file

fasta

Conversion to

dmb file format

<<Fasta2Dmb.java>>

Cutpoints

Definition

<<makeclasses.c>>

Cutpoints

file

Discretization

<<discretize.c>>

Binary

Test

(dmb)

Creation of the

FS Problem

<<creasc.c>>

Lsc

files FS Problem

Resoluton

<<grasp2.c>> Grasp /

features

files

Feature

Selection

<<filterfeatures.c>>

Creation and

Resolution of the

logic Problem

<<crelogic.c>>

Logic

files

Evaluation

<<evaluate.c>>

Output

FIles

Test

files

(dmb)

Train

files

(dmb)

Binary

Train

(dmb)

Filtered

Test

(dmb)

Filtered

Train

(dmb)

Train file

(fasta)

42

Multiple solutions extraction

• Feature elimination techniques: CAMUR – For logic formulas with just one literal,

e.g., if pos122=C then Polyomavirus WU, eliminate feature 122 and run the analysis again, iterate until the accuracy is lower then a given threshold

– For logic formulas with more conjuncts, e.g., if pos266=T OR pos78=T AND pos80=A then RhinovirusA,

eliminate all the possible combinations of features and each time run the analysis again; iterate until the accuracy is lower then a given threshold

• Feature selection techniques: MISSAL – Explore the solution space by computing equivalent solutions:

Modification of the DMB FS algorithm Take into account adjacent features and

prefer solutions with adjacent features

43

Biological applications of supervised learning and rule-based classification:

Aim: To extract relevant features from the ever-increasing amount of

biological data and to apply supervised learning to classify them

Biology Issue Features Software Data source

Clinical patient classification

Clinical variables (blood, imaging, psicosometric tests…)

DMB, Weka Heterogeneous health care facilities

Gene Expression Analysis

Discretize gene expression profiles

Gela, CAMUR TCGA, EBRI

DNA barcoding Nucleotide sequences of DNA-barcode

Blog, Fasta2Weka Barcode of Life Consortium

Polyoma/Rhyno Viruses

Nucleotide sequences of Polyoma/Rhyno viruses

DMB, MISSAL Istituto Superiore di Sanità

EEG signals processing

Fourier Coefficients extracted from EEG recordings

Matlab, Weka, DMB IRCCS Centro di Neurolesi “Bonino-Pulejo” of Messina

44

Clinical Patients

• The aim of the project is to design a new diagnostic model and possibily a work-flow for the early diagnosis of dementias

• Clinical patients trial samples were collected in different Italian health structures

• Collected data included demographic characteristics, medical history, pharmacological treatments, clinical and neurological examination, psychometric tests, laboratory blood tests, imaging… (700 features x 5000 experiments)

• The data set:

45

Preprocessing and missing values

• Fundamental in clinical data analysis

• Produces clean data and is crucial for obtaining reliable knowledge by the data mining algorithms

• Perform common statistical analysis for every trial: mean, standard deviation, minimum, maximum, number of missing values, modal value, etc.

• Trials filtering:

– exclusion of the trials with a high number of missing values, e.g. more than 20%

– filling of missing values. e.g. by the class mean

– outliers detection by mean - standard deviation - minimum - maximum values analysis and records removal

– correction of the mixed attributes types

46

Preprocessing and missing values

• Missing values treatment:

– Most common attribute value

– Assigning all possible values of the attribute

– Ignoring examples with unknown attribute values

– Event-Covering method

– Treating missing attribute values as special values

– Filling of missing values with the class mean (for numeric variables )

– Filling of missing values with the class modal value (for categorical variables)

• Special scripts and programs are necessary

47

Clinical Patients

• Results:

• Classification Results

Clinical data mining: problems, pitfalls and solutions. E. Weitschek, G. Felici and P. Bertolazzi. IEEE DEXA 2013

48

Gene Expression Profiling

• Major advances in microarray and Next Generation Sequencing technologies lead to an ever-increasing amount of gene expression data available to biological scientists and bioinformaticians

• RNA-seq:

− High-throughput Sequencing and large-scale transcriptome analysis

− Family of methods for detecting the transcriptome expression level

− RNA-Seq technology performs mapping and transcriptome quantification relying on NGS

− It can be used for

• Annotation

• Quantification Determination of abundance of transcripts under different conditions

49

GELA: Gene Expression Logic Analyzer

http://dmb.iasi.cnr.it/gela.php

50

GELA: Software Engineering

• GELA relies on the software architectural pattern Pipes and Filters for computing streams of data.

• Component and flow diagrams:

51

GELA: GUI Release

• Graphic user interface [dmb.iasi.cnr.it/gela.php]

• Guides the user in the RNA-seq clustering and classification process

• Java Swing framework for visualizing the data set, running the software, performing the additional analysis and viewing the clusters, the classification results and the logic separating formulas

• A complete user manual is available on the GELA website.

52

GELA: Command line

• Command line version [dmb.iasi.cnr.it/gela.php]

– Dedicated to the users who want to perform long experiments and batch the entire analysis process

– Compiled and tested on Linux and Windows operating systems

– Source code released for compiling on alternative systems

– Complete user guide is published

53

Study N. 1

• To characterize the gene expression profile of CONTROL versus ALZHEIMER DISEASED mice in early stage (1-3 months) and late stage (6-15 months)

• 119 experimental samples and 16,515 gene expression profiles

• To identify a limited set of genes able to discriminate between the neuro-degeneration and the healthy state

• To identify clusters of corr- and contra-regulated genes

Data Analysis steps

1) Discretization

2) Gene clustering

3) Feature selection

4) Formula extraction

5) Post processing and correlation analysis

GELA: Alzheimer Data Analysis Results

Raw data

54

GELA: Alzheimer Data Analysis Results

• Discrete Clustering method DCA allow to reduce the whole gene set to

‐ 3656 for 1-3 months

‐ 3615 for 6-15 months

• 8 genes are highly co-regulated or contra-regulated ,and able to separate exactly all the healthy from the sick mice (iterative application of the method) in 30-fold cross validation [13]

• More genes are strongly co-regulated with the 8 genes network

If Aph1b<0.50 then the individual is control

55

GELA vs other methods

• Classification Analysis We tested the following widespread algorithms (WEKA [23] implementations) : K-Nearest Neighbor (KNN), Support Vector Machine (SVM), Random Forest (RF) , and C4.5 Decision Tree (C4.5)

• Results:

• GELA outperforms other methods

• Advantages of GELA:

₋ to extract meaningful and compact models

₋ clustering capabilities

₋ availability as an integrated tool

56

GELA: Data Analysis Results

Study N. 2 • Data sets downloaded from public repositories ArrayExpress and GEO

• Files normalized using the standard Affymetrix Expression Console software (ver 1.2), by the MAS5 algorithm

• To classify into diseased and healthy status samples of:

- 54,613 gene expression profiles of 176 Psoriasis human subjects

(85 control and 91 diseased)

- 22,215 gene expression profiles of 178 Multiple Sclerosis human subject (44 control and 134 diseased)

Results:

• GELA outperforms the other methods

57

GELA: Data Analysis Results

Study N. 3

• Data sets downloaded from the public repository The Cancer Genome Atlas (TCGA)

• To characterize the gene expression profile of CONTROL versus BREAST CANCER human samples

• To classify samples of 20531 gene expression profiles of 783 subjects into the tumor

and the healthy state

Results:

• GELA achieves comparable results of SVM and RF, that unfortunately produce classification models whose interpretation is very difficult for human beings

58

CAMUR: cancer analysis

Classification of RNA-seq cancer data through multiple rule-based models

• RNA-seq gene expression analysis and specifically on case-control studies for cancer with rule-based classification algorithms designed to build models, that discriminate cases from controls

• State of the art algorithms typically extract a single classification model that contains few features (genes)

• Goal: elicit a higher amount of knowledge by computing more alternative classification models and therefore to discover several features that are related to the predicted class (cancer)

Database of rules

59

CAMUR: cancer analysis

CAMUR (Classifier with Alternative and MUltiple Rule-based models): a new method and software package able to extract multiple, alternative, and equivalent classification models

– iteratively computes a rule-based classification model

– calculates the power set (or a partial combination) of the features present in the rules

– iteratively eliminates those combinations from the data set

– performs again the classification procedure until a stopping criterion

60

DNA Barcoding: alignment based analysis

• When analyzing biological sequences with a classical approach an overlapping gene region is necessary, due to the fact that an analysis of the characteristics nucleotides present in a determined position for every class is performed

• This is obtained through an alignment between the sequences to analyze, i.e. align areas of the sequences sharing common properties

IF BASE IN POSITION 466 IS C AND

BASE IN POSITION 595 IS T THEN SPECIES IS …1

IF BASE IN POSITION 340 IS G AND

BASE IN POSITION 451 IS A AND

BASE IN POSITION 493 IS C THEN SPECIES IS … 2

IF BASE IN POSITION 340 IS T AND

BASE IN POSITION 466 IS A AND

BASE IN POSITION 625 IS G THEN SPECIES IS … 3

Species 123456789

Species1 ACGTTAACA

Species1 GTCAGTCCA

Species2 GTCGTATAT

Species2 AGCTACGAT

61

DNA Barcoding

• DNA is used for classifying species

• Barcode: small sequence of DNA containing all the information necessary to identify a species of an individual (classification), or to identify new species:

‒ Gene region COI for animals (approx. 650 bp)

‒ Gene regions rbcl and matk for plants (approx. 1000 bp)

‒ Gene regions ITS for fungi (approx 400 bp)

• The sequences are from the same gene region and aligned

• A positional analysis is performed for distinguishing the different species

IF BASE IN POSITION 466 IS C AND

BASE IN POSITION 595 IS T THEN SPECIES IS …1

IF BASE IN POSITION 340 IS G AND

BASE IN POSITION 451 IS A AND

BASE IN POSITION 493 IS C THEN SPECIES IS … 2

IF BASE IN POSITION 340 IS T AND

BASE IN POSITION 466 IS A AND

BASE IN POSITION 625 IS G THEN SPECIES IS … 3

If pos3 = A and pos458 = C then the specimen is a

62

DNA Barcoding

BLOG 2.0

63

BLOG 2.0 vs other methods

Data set NJ PAR NN BLAST BAR BLOG

1,000 83.69 73.31 86.18 86.18 86.25 85.96

10,000 85.53 79.79 86.11 86.09 86.83 88.15

50,000 84.20 79.53 84.76 84.56 85.24 84.58

Overall 84.47 77.54 85.68 85.61 86.11 86.23

70.00

72.00

74.00

76.00

78.00

80.00

82.00

84.00

86.00

88.00

90.00

Ne = 1,000 Ne = 10,000 Ne = 50,000 overall

NJ

PAR

NN

BLAST

DNA-BAR

BLOG

Data set NJ PAR NN BLAST BAR BLOG

Drosophila 83.90 83.90 82.20 83.90 83.90 96.61

Inga 91.28 81.40 88.37 82.56 94.19 90.12

Cypraeidae 91.53 85.31 91.24 92.66 93.22 92.66

Overall 88.90 83.53 87.27 86.37 90.43 93.13

70.00

75.00

80.00

85.00

90.00

95.00

100.00

Inga Cypraeidae Drosophila Overall

NJ

PAR

NN

BLAST

DNA-BAR

BLOG

64

LAF: Logic Alignment Free

• Alignment free k-mer frequency count analysis: – oligomers frequencies: computation of the substring frequencies of a given length k – the k-mers are computed by counting the occurrences of the substrings with a

sliding window of length k, starting at position 1 and ending at position n−k+1 – frequency vector, where each component of the vector is associated to the

frequency of a particular k-mer – coordinate space, that is mathematically tractable

• Combination of alignment free methods and logic data mining:

• http://dmb.iasi.cnr.it/laf.php

65

LAF: Logic Alignment Free

• The LAF analysis is based on the supervised machine learning paradigm: class a priori assigned; training and test set

• Considering each sequence g in the data set, the following steps are performed by LAF

1. The reverse complement g’ of the sequence g is calculated 2. The sequence g is concatenated with its reverse complement g’, obtaining the

sequence G = g + g’ 3. The k-mer counts are computed on G with k = 3 … 6 and

stored in the feature vector F 4. The frequency vector F’ is obtained from F 5. The frequency vectors are collected in a

matrix, whose rows correspond to k-mers, and whose columns correspond to sequences

6. A discretization of the frequencies is performed

7. The data matrix is then processed by a rule-based classifier

66

LAF and Bacteria classification

• Whole genome bacteria classification: – Whole genomes are very difficult to align, because of their length – The method is applied for classifying bacteria in the different levels of the

phylogenetic tree (phylum, class, family, genus, species) – 1964 bacteria whole genome sequences downloaded from GenBank

– Very promising results, especially at the phylum level (98% correct classification rates)

67

Virus genomic sequences analysis

• Viruses are organisms that are hard to identify

• Cause of several diseases which affect human and animal species

• Bioterrorism

68

Rhinoviruses identification

• Rhino from the Greek "nose"

• Most common viral infective agents in humans and are the predominant cause of the common cold

• 1316 Rhinovirus A, B, C sequences of VP42 gene region (~370 bp)

• 116 complete genomes of the three different rhinoviruses (~7200 bp)

69

Rhinoviruses identification

• Positional analysis on VP42 gene region

– High recognition rates (accuracy, recall, precision, f-measure > 90%)

• LAF analysis on complete genomes

– High recognition rates (accuracy, recall, precision, f-measure > 90%)

RHINOVIRUSA RHINOVIRUSB RHINOVIRUSC

pos165=C OR pos165=T pos364=T pos185=C AND pos226=C

pos120=G pos363=T pos70=T

pos190=A pos226=G pos269=T

pos293!=A pos163=G AND pos293=A pos163!=G

pos98=T pos73=C AND pos98=A pos73!=C AND pos98=A

pos50=A AND pos334=G pos334=A pos50=T

pos266=T OR pos78=T AND

pos80=A pos80!=A AND pos 266 =C

pos78!=T AND pos80=A AND

pos266=C

pos53=C AND pos317!=T OR

pos314=C

pos53=T AND pos314=A OR

pos53=A AND pos317=A

pos53=A AND pos317!=A OR

pos53=C AND pos317=T

RHINOVIRUSA RHINOVIRUSB RHINOVIRUSC

GCA<31.92 AND

GCC<16.16

18.71<=CCC<24.18 AND

GCA>=31.92 OR

GCA>=31.92 AND

CC<16.16

18.71<=CCC<24.18 AND

GCA>=31.92

OR

GCA>=31.92 AND GCC<16.16

70

Polyomaviruses identification

• Polyomaviruses are oncogenic (tumor-causing) viruses

• They often persist as latent infections in a host without causing disease, but may produce tumors in a host of a different species

• The name polyoma refers to the viruses' ability to produce multiple (poly-) tumors (-oma)

• Aim of this study :

– find nucleotide positions that allow to distinguish the five human polyomaviruses known to date

– examine and explain them for an effective biological analysis

71

Polyomaviruses identification

• Genome length of each polyomavirus (1200 individuals):

– about 5 Kb in length (in term of nucleotide string from 700 to 2500)

•

• Results: The logic data mining method

– identifies a subset of positions that characterize the polyomaviruses and the gene regions

– exhibits high recognition and classification rates

72

Polyomaviruses identification

73

Polyomaviruses identification

1 ---------- --TTATCT-- -GAAGAG--- AAGCAGCTA- --CAA-----

----->

51 -TGTTTGGAC ---AGT---- ---------- ---------- ---AGACAA-

101 --TTAGTAAA G--------- TAT---CCA- -----GGA-- ----------

151 AGC---AAGC TT-------- ---------- ---------G ACTCA---AG

201 CAGTGTACAT GATTTA---- ---------- -AATATA--- CAA---TCT-

251 -----CCT-- -TAC---ACA CCT------- --GAG----- -TCA---TTC

301 AACACC---- --------GA GTTA------ ---GAATCC- --CCATCT--

351 -CCAAAAAGA AGTGCACCA- --GAGGAGCC TAGCTGTTCT CAGGCAACCC

<----

401 CTCCTAAGAA AAAACATGCA TTTGATGCTT CTTTAGAATT TCCTAAAGAG

----->

451 TTGTTAGAGT TTGTTTCACA TGCTGTATTT AGTAATAAGT GTATAACGTG

<-------

501 C---GTAGTA CAT---ACTA GAGAAAAA-- -GAAGTACTT TAT---AAGT

---->

...

74

Supervised Learning and Logic Data Mining for Virus sequences analysis

Recognize the unknown samples into the virus classes of the training

set of virus sequences with a priori known class (training set)

set of unknown samples (test set)

• Classification

to assign an unknown object to a predefined class after examining its features

• Classification with logic formulas

to classify the unknown object by using logic propositional formulas in disjunctive or conjunctive normal form (”if then rules”)

If pos13 = A and pos400 = G then the virus is RhinoA

Aim: To extract relevant features from the ever-increasing amount of

biological data and applying supervised learning to classify them

75

Multiple solutions extraction

• Feature selection technique

− Explore the solution space by computing equivalent solutions

− Each solution evaluated in term of 𝜶, 𝜷, 𝝈

− Find the smallest set of adjacent features within a segment

that are able to classify the given objects

• Advantages:

− reduce sequencing costs

− point to the most compact meaningful portions of sequences

− more reliable classification models

all the solutions equivalent in term of 𝜶, 𝜷 but

differ for 𝝈 value

𝛽

𝜎

𝛽

Starting point 𝜎 Segment size 𝛽

Aim: to identify set of adjacent equivalent interesting solutions in different biological issues

76

• Results: Compact Adjacent Solutions Extraction

Aim:

- find nucleotide positions that allow to distinguish and characterize gene regions of

Polyoma-, Rhino-, and Influenza viruses

⁻ identify set of multiple compact and locally adjacent interesting solutions

⁻ efficiently extract a large number of highly reliable and compact classification rules

Influenza

Polyoma

Rhino

Viruses sequences identification: MISSAL (MultIple Solutions Sequence AnaLyzer with proximity control)

77

• Results: Solutions Map onto the sequence

Aim:

- find nucleotide positions that allow to distinguish and characterize gene regions of

Polyoma-, Rhino-, and Influenza viruses

⁻ identify set of multiple compact and locally adjacent interesting solutions

⁻ efficiently extract a large number of highly reliable and compact classification rules

Viruses sequences identification: MISSAL (MultIple Solutions Sequence AnaLyzer with proximity control)

78

• Results: Classification results on training set

Aim:

- find nucleotide positions that allow to distinguish and characterize gene regions of

Polyoma-, Rhino-, and Influenza viruses

⁻ identify set of multiple compact and locally adjacent interesting solutions

⁻ efficiently extract a large number of highly reliable and compact classification rules

Viruses sequences identification: MISSAL (MultIple Solutions Sequence AnaLyzer with proximity control)

If pos13 = A and pos400 = G → 𝑉𝑖𝑟𝑢𝑠 = PolyomaBK

79

EEG signal processing and patients classification

• Electro Encefalo Grams (EEG) of 102 patients: – 51 Alzheimer Diseased – 37 Mild Cognitive Impairment – 14 Control

• 19 electrodes • Signal duration of 300 seconds • sampling frequency of 1024 or 256 sp/s • 9 GB of raw signals

• Features extraction: – extraction of 180-seconds signals – transform each one at 256 sp/s – frequency analysis by using Fast Fourier Transform (FFT)

80

EEG signal processing and patients classification

• Analysis: We apply the FFT functions to the collected data

₋ taking into account the 180-seconds signal and extracting N Fourier Coefficients (N equal to 16 and 32);

₋ dividing the signal in 6 epochs of 30 seconds and extracting for each one 16(32) Fourier Coefficients

• Classification:

Classification performances [%] for 180-sec EEG signals

Classification performances [%] for EEG epochs of 30-sec

81

Collaborations

82

Contacts

• Emanuel Weitschek University Roma Tre Department of Computer Science and Automation Rome, Italy www.dia.uniroma3.it/~emanuel emanuel@dia.uniroma3.it

• Giulia Fiscon University Sapienza Department of Computer, Control, and Management Engineering Rome, Italy http://www.iasi.cnr.it/~gfiscon/ fiscon@dis.uniroma1.it