Biosystems Data Analysis
Inhoud
Week 1 ............................................................................................................................................................. 2
A ................................................................................................................................................................... 2
Lecture 1 Data pre-treatment – Initial analysis and preparation of ‘omics’ data ............................. 2
R Tutorial and practical........................................................................................................................... 11
B ................................................................................................................................................................. 15
Lecture 2 Principal Component Analysis......................................................................................... 15
R practical ............................................................................................................................................... 20
C ................................................................................................................................................................. 24
Lecture 3 Clustering methods and Self Organising Maps (SOM) .................................................... 24
Week 2 ........................................................................................................................................................... 30
D ................................................................................................................................................................. 30
Lecture 4 BDA classification methods - supervised approach ....................................................... 30
R practical ............................................................................................................................................... 35
1
,Week 1
A
Pre-processing and pre-treatment of data is an important aspect of data analysis to remove instrumental
artefacts and add biological content to the data. One of the problems in Next Generation Sequencing
methods is the nonconstant variability in the data. Besides the variance stabilization approach we will also
discuss the meaning of the p-value and the false discovery rate. Read the Nuzzo paper for preparation and
make the questions in the Discussion_Nuzzo2014 pdf.
Web-lecture link: https://webcolleges.uva.nl/Mediasite/Play/0c3570ffed5b4f1ca3701fbdc2d591191d
Lecture 1 Data pre-treatment – Initial analysis and preparation of ‘omics’ data
Goals of the lecture:
- Learn the role of the chain of experimental techniques that determine data quality (e.g.: RNAseq)
- Learn techniques to explore the variation of omics data (bias and random effects)
- Learn techniques to normalize data (remove bias)
- Learn data transformations to remove heteroscedasticity (unequal random error)
- Know the consequences of random error for subsequent statistical analysis
- Learn the ideas behind Multiple Hypothesis Testing
The techniques mentioned above are part of the computer practicals: i.e. the topics treated in the practical
are subject of the exam.
Multiplex: quantification of a large number of (related) components in a single sample (such as omics).
VS
High throughput technologies: quantification of single component in a large number of samples (in a short
time, so not omics) .
Omics experiment is really low throughput, because lots of data takes lots of processing.
Multiplex technologies in biology:
Genomics reading multiple gene sequences in a single sample.
Transcriptomics: quantification of multiple transcript levels (mRNA) in a sample.
Proteomics: quantification & characterization of multiple proteins in a sample.
Metabolomics: quantification of many metabolites in a sample.
RNA-seq: do transcriptomics but in a way in which you sequence each transcript.
RNA-sequencing experimental procedure:
- Stopping all activity = quenching (because concentrations deviate very quick, otherwise noise)
- Isolation of mRNA (isolate out of the cells)
- Reverse transcription: RNA → DNA (because we cannot sequence RNA, thus use DNA)
- Optional amplification by PCR (polymerase chain reaction, create many DNA sequences)
- Library construction : attaching sequence tags/adaptors (to later trace the sequence)
- Sequencing
The experimental procedures affect the outcome:
Quenching because
- RNA’s have short half-lives in living cells.
- RNAses are abundant and have to be stopped
- Handling living cells cause stress which can change gene expression.
- Breaking cells (or bacteria) can be difficult
- Obtaining sample can be time-consuming.
2
,RNA isolation:
- Most RNA is ribosomal RNA (rRNA)
- Eukaryotic messenger RNA (mRNA) can be enriched by poly-A tail hybridization
Sample storage & quality control →
- Storage of mRNA should be done at – 80 ˚C.
- Quality control: 18S/28S rRNA ratio is
measured:
You see how quick mRNA is degraded in the image:
The long mRNA’s become shorter == degraded.
Sequencing and mapping sequences procedure >>
Results: a table of counts. Counts are number of
sequences mapped to a gene.
5 samples taken (A1 – B2) and x genes measured ^
You see large variation when adding all: total is not
equal for all A or all B conditions. This bias should be
removed.
The goal: detect differences in gene expression
between conditions.
Sources of variation
Technical sources (most can be removed) Biological sources
- Sample preparation (medium, temp) - Variation of interest
- Sample isolation (handling, speed of quenching…) - Variation between similar samples /
- Differences in mRNA quality individuals (can be noise but also interesting)
- cDNA synthesis
- Amount of cDNA added
- Sequence bias
- Random measurement error (only error that can’t be removed)
3
, Bias must be removed before statistical testing. Statistical tests are designed to handle random errors, but
not bias or systematic errors.
Normalization: correction of bias.
Exception: when the statistical model takes bias into account. E.g. for models of count data that
accommodate variations in total count.
Bias in RNA-sequencing can have obvious effects:
Variation in sequence fragments counted can be due to technical effects:
- Variation in amount of isolated mRNA
- Variation in quality of isolated MRNA
- Variation in quenching efficiency
- Variation in cDNA synthesis efficiency
- Variation in sequencing efficiency (number of sequences
read of ‘total count’)
Or it can be due to interesting biological effects.
OR bias in RNA-sequencing can have subtle effects →→
Fragment length optimum for sequencing is present. Position: RNA is more
degraded at the end than at the start. Sequence bias: bias due to GC and AT concentrations.
High GC content results in lower counts.
Normalization / correct bias:
You need to propose a hypothesis about the origin of observed variations in sequencing counts.
Hypothesis 1: approximately equal concentration of mRNA in each sample.
- Implies: variations in total counts per sample are due to technical reasons.
- Solution: divide sequence count for each gene by the total sequence count in the sample.
- This is called RPM: Rate Per Million reads. Allows comparison of the same gene between samples.
Hypothesis 2: approximately constant number of sequences per kilobase of mRNA.
- Implies: variations in counts between genes are due to gene length.
- Solution: divide sequence count for each gene by total sequence count and by length of
gene in kilobases.
- RPKM: Rate Per Kilobase per Million reads. Allows comparison between different genes
and samples. But underlying hypothesis is debatable due to putative subtle bias effects (e.g. due
to GC percentage).
RNAseq has a high dynamic range: expression levels of known transcripts are quantified by the number of
reads per kilobase of transcripts per million mapped read (RPKM).
Random variation in counts.
Technical error: sequencing the same sample multiple times will yield different counts of gene X. The count
X is a random variable, having a certain distribution.
To do statistical testing of hypothesis, you need to know what kind of random error is present in the data.
Distribution properties of random variable X, (expected distribution of number of reads per gene):
- It is a discrete distribution, because we count reads per target gene
- The probability p of sequencing a fragment of gene X by randomly picking one from all fragments
equals its proportion among all fragments
- The probability p is usually very small
- The probability p remains constant after the first, second, etc. fragment of the target gene has
been sequenced (i.e. the number of sequenced fragments << available DNA fragments)
4
Inhoud
Week 1 ............................................................................................................................................................. 2
A ................................................................................................................................................................... 2
Lecture 1 Data pre-treatment – Initial analysis and preparation of ‘omics’ data ............................. 2
R Tutorial and practical........................................................................................................................... 11
B ................................................................................................................................................................. 15
Lecture 2 Principal Component Analysis......................................................................................... 15
R practical ............................................................................................................................................... 20
C ................................................................................................................................................................. 24
Lecture 3 Clustering methods and Self Organising Maps (SOM) .................................................... 24
Week 2 ........................................................................................................................................................... 30
D ................................................................................................................................................................. 30
Lecture 4 BDA classification methods - supervised approach ....................................................... 30
R practical ............................................................................................................................................... 35
1
,Week 1
A
Pre-processing and pre-treatment of data is an important aspect of data analysis to remove instrumental
artefacts and add biological content to the data. One of the problems in Next Generation Sequencing
methods is the nonconstant variability in the data. Besides the variance stabilization approach we will also
discuss the meaning of the p-value and the false discovery rate. Read the Nuzzo paper for preparation and
make the questions in the Discussion_Nuzzo2014 pdf.
Web-lecture link: https://webcolleges.uva.nl/Mediasite/Play/0c3570ffed5b4f1ca3701fbdc2d591191d
Lecture 1 Data pre-treatment – Initial analysis and preparation of ‘omics’ data
Goals of the lecture:
- Learn the role of the chain of experimental techniques that determine data quality (e.g.: RNAseq)
- Learn techniques to explore the variation of omics data (bias and random effects)
- Learn techniques to normalize data (remove bias)
- Learn data transformations to remove heteroscedasticity (unequal random error)
- Know the consequences of random error for subsequent statistical analysis
- Learn the ideas behind Multiple Hypothesis Testing
The techniques mentioned above are part of the computer practicals: i.e. the topics treated in the practical
are subject of the exam.
Multiplex: quantification of a large number of (related) components in a single sample (such as omics).
VS
High throughput technologies: quantification of single component in a large number of samples (in a short
time, so not omics) .
Omics experiment is really low throughput, because lots of data takes lots of processing.
Multiplex technologies in biology:
Genomics reading multiple gene sequences in a single sample.
Transcriptomics: quantification of multiple transcript levels (mRNA) in a sample.
Proteomics: quantification & characterization of multiple proteins in a sample.
Metabolomics: quantification of many metabolites in a sample.
RNA-seq: do transcriptomics but in a way in which you sequence each transcript.
RNA-sequencing experimental procedure:
- Stopping all activity = quenching (because concentrations deviate very quick, otherwise noise)
- Isolation of mRNA (isolate out of the cells)
- Reverse transcription: RNA → DNA (because we cannot sequence RNA, thus use DNA)
- Optional amplification by PCR (polymerase chain reaction, create many DNA sequences)
- Library construction : attaching sequence tags/adaptors (to later trace the sequence)
- Sequencing
The experimental procedures affect the outcome:
Quenching because
- RNA’s have short half-lives in living cells.
- RNAses are abundant and have to be stopped
- Handling living cells cause stress which can change gene expression.
- Breaking cells (or bacteria) can be difficult
- Obtaining sample can be time-consuming.
2
,RNA isolation:
- Most RNA is ribosomal RNA (rRNA)
- Eukaryotic messenger RNA (mRNA) can be enriched by poly-A tail hybridization
Sample storage & quality control →
- Storage of mRNA should be done at – 80 ˚C.
- Quality control: 18S/28S rRNA ratio is
measured:
You see how quick mRNA is degraded in the image:
The long mRNA’s become shorter == degraded.
Sequencing and mapping sequences procedure >>
Results: a table of counts. Counts are number of
sequences mapped to a gene.
5 samples taken (A1 – B2) and x genes measured ^
You see large variation when adding all: total is not
equal for all A or all B conditions. This bias should be
removed.
The goal: detect differences in gene expression
between conditions.
Sources of variation
Technical sources (most can be removed) Biological sources
- Sample preparation (medium, temp) - Variation of interest
- Sample isolation (handling, speed of quenching…) - Variation between similar samples /
- Differences in mRNA quality individuals (can be noise but also interesting)
- cDNA synthesis
- Amount of cDNA added
- Sequence bias
- Random measurement error (only error that can’t be removed)
3
, Bias must be removed before statistical testing. Statistical tests are designed to handle random errors, but
not bias or systematic errors.
Normalization: correction of bias.
Exception: when the statistical model takes bias into account. E.g. for models of count data that
accommodate variations in total count.
Bias in RNA-sequencing can have obvious effects:
Variation in sequence fragments counted can be due to technical effects:
- Variation in amount of isolated mRNA
- Variation in quality of isolated MRNA
- Variation in quenching efficiency
- Variation in cDNA synthesis efficiency
- Variation in sequencing efficiency (number of sequences
read of ‘total count’)
Or it can be due to interesting biological effects.
OR bias in RNA-sequencing can have subtle effects →→
Fragment length optimum for sequencing is present. Position: RNA is more
degraded at the end than at the start. Sequence bias: bias due to GC and AT concentrations.
High GC content results in lower counts.
Normalization / correct bias:
You need to propose a hypothesis about the origin of observed variations in sequencing counts.
Hypothesis 1: approximately equal concentration of mRNA in each sample.
- Implies: variations in total counts per sample are due to technical reasons.
- Solution: divide sequence count for each gene by the total sequence count in the sample.
- This is called RPM: Rate Per Million reads. Allows comparison of the same gene between samples.
Hypothesis 2: approximately constant number of sequences per kilobase of mRNA.
- Implies: variations in counts between genes are due to gene length.
- Solution: divide sequence count for each gene by total sequence count and by length of
gene in kilobases.
- RPKM: Rate Per Kilobase per Million reads. Allows comparison between different genes
and samples. But underlying hypothesis is debatable due to putative subtle bias effects (e.g. due
to GC percentage).
RNAseq has a high dynamic range: expression levels of known transcripts are quantified by the number of
reads per kilobase of transcripts per million mapped read (RPKM).
Random variation in counts.
Technical error: sequencing the same sample multiple times will yield different counts of gene X. The count
X is a random variable, having a certain distribution.
To do statistical testing of hypothesis, you need to know what kind of random error is present in the data.
Distribution properties of random variable X, (expected distribution of number of reads per gene):
- It is a discrete distribution, because we count reads per target gene
- The probability p of sequencing a fragment of gene X by randomly picking one from all fragments
equals its proportion among all fragments
- The probability p is usually very small
- The probability p remains constant after the first, second, etc. fragment of the target gene has
been sequenced (i.e. the number of sequenced fragments << available DNA fragments)
4
