Division Harvard-MIT of Health Sciences and Technology HST.512: Genomic Medicine Prof. Zoltan Szallasi
Limitations of massively
parallel technologies
Zoltan Szallasi, MD
Children's Hospital Informatics Program
www.chip.org
New technology
All problems will be solved within
a couple of years
Realistic Expectations (limitations)
Limitations: (you want to make predictions) Accuracy noise Sensitivity - completeness Inherent limitations (think about unpredictability > chaos)
NOISE: - what is noise ? (and what is signal ?) - noise as an inherent feature of complex systems - noise in continuous and discrete measurements - noise as the limitation of the technology - what can be done about noise ? Statistics Normalization as a way to deal with systematic errors
c : an unwanted signal or a disturbance (as static or a variation of voltage) in an electronic device or instrument (as radio or television); broadly :
a disturbance interfering with the operation of a usually mechanical device or system d : electromagnetic radiation (as light or radio waves) that is composed of several frequencies and that involves random changes in frequency or amplitude e : irrelevant or meaningless data or output occurring along with desired information
Noise may turn out to be an important signal !!!! -Penzias and Wilson >>> cosmic background radiation - discovery of the chemotherapeutic agent cis-platinum
What we perceive as noise/error might be a key component of biological processes: 1) Mutations in evolution 2) "Junk" DNA 3) Asymmetric cell division may contribute to differentiation
4) Stochastic fluctuations may be important for the
stability of complex physicochemical systems
Genetic networks are stochastic systems: 1) A couple of hundred copies of a given transcription factor/nucleus 2) Intracellular environment is the not a free solution 3) Reaction kinetics is often slow etc.
A B C D E F A B C D E F
A B C D E F
A B C D E F
Please see Science. 2002 Aug 16; 297(5584):1183-6.
Comment in:
Science. 2002 Aug 16; 297(5584):1129-31.
Stochastic gene expression in a single cell.
Elowitz MB, Levine AJ, Siggia ED, Swain PS.
-measuring population averaged data.
That is true even if single cells are quantified due to stochasticity > two cells can get from a given state to another one via different paths
Noise in measurements
There is no measurement without noise - (it is the accuracy/sensitivity of your measurement that is low ) For continuous variables it is expected to obtain data with a certain "spread"
Consequently: Statistics was invented
- 0.5, -0.3, 0.2, 1.4, -1.5.....etc what is the true value
of the observed variable ?
- Did the variable change due to a given treatment? Etc.
Lots of measurements and/or fairly good idea about the nature of the noise (e.g. normal distribution)
Statistical analysis in biology: 1) What is the true value of a given parameter ? 2) the most common analysis Bayesian
3) You don't believe the measurements >> normalization 4) There are too many numbers >> permutation etc.
Biological measurements are often expensive !!!!!!!!!
A large number of papers relating to cancer were published in Nature/Science ..... based on single microarray measurements
STATISTICS
Reliable numbers cannot be produced without replicates
The central problem :
In massively parallel biological measurements quantitative or qualitative calls are supposed to be made on a large number of heterogeneous variables using only a few replicates.
Noise of continuous variables, e.g. microarray measurements
Tissue
or
Tissue
under
influence
.....
RNA cDNA/cRNA Tagged with fluorescent dye
Microarray of genes aka gene chips
Ideally: 1 copy of a given RNA will produce 1 unit of a specific signal !!!!!!!!!!!!!!
1) cDNA produced from RNA (initiation of RT step, RT might drop off etc.) 2) cRNA produced in the presence of fluorescent dyes (cRNA production in not linear, Dye incorporation) 3) Breaking down cRNA into small pieces 4) hybridization/cross hybridization
final signal =
(all of the above)
The situation is further complicated by other experimental issues >>> two-color cDNA microarray
Ratio is influenced on background calculations equal amounts of labelled
cDNA samples
There is no truly blank spot !!!! Background
mRNA reference sequence 5` 3` Spaced DNA probe pairs
Reference sequence
... TGTGATGGTGGGAATGGGTCAGAAGGACTCCTATGTGGGTGACGGAGGCC ...
Fluorescence Intensity Image
Perfect match probe cells
Mismatch probe cells
AAT G G G T C A G AA G G A C T C C TAT G T G G G T G AAT G G G T C A G AA C G A C T C C TAT G T G G G T G
Perfect match Oligo Mismatch Oligo
Data representation
If we express gene expression measurements as "per unit RNA" then decrease in the level of a given message unavoidably leads to a relative increase in the level of other messages.
Distribution of probe intensities of several Affymetrix data sets belonging to the same set of experiment.
Systematic error
Density (x = x[, 1], from = 4, to = 16)
1.0
0.8
Density
0.6
0.4
0.2
0.0 4 6 8 10 12 14 16
N = 131822
Bandwidth = 0.1128
Normalization
Normalization You don't believe the numbers 1) "most or certain things do not change" 2) Error model
Shifting the means or medians and adjusting the distributions by Cubic spline fit/ Lowess etc. (Overfitting !!!)
Density (x = x[, 1], from = 4, to = 16)
1.0 0.8
Density
0.6
0.4
0.2
0.0 4 6 8 10 12 14 16
N = 131822
Bandwidth = 0.1128
Density (x = y[, 1], from = 4, to = 16)
0.6
0.4
0.2 0.0 4 6 8 10 12 14 16
N = 131822 Bandwidth = 0.09808
cDNA microarray: the R/G ratios are intensity dependent
Values should scatter about zero.
Courtesy of Natalie Thorne. Used with permission.
Overview of normalization: - to correct for systematic errors 1) Choose a set of that elements will be used - housekeeping genes - special control genes etc. 2) Determine the normalization function - global mean/median normalization - intensity dependent normalization
Microarray Gene Expression Data Society
www.mged.org
Intensity dependent normalization by error models
Error model:
Low concentrations High concentrations Rocke, Vingron
x=+
x = e
x = e +
( ~ N
0, )
2
~
2 N(0,
)
Noise will limit the useful information content of measurements: A reliable detection of 2-fold differences seems to be the practical limit of massively parallel quantitation. (estimate: optimistic and not cross-platform)
Level of gene expression
Measurements with error bars
Time window
T
me
i
A rational experiment will sample gene-expression according to a time-series in which each consecutive time point is expected to produce at least as large expression level difference as the error of measurement: approximately 5 min intervals in yeast, 15-30 min intervals in mammalian cells.
Limitations: (you want to make predictions) Accuracy noise Sensitivity - completeness
Inherent limitations (think about unpredictability > chaos)
Sensitivity completeness How many parameters are we measuring ?
How many parameters should we measure ?
How many bionodes ?
Cautious estimate: on the order of 1-2x105 10,000-20,000 active genes per cell < 3 posttranslational modifications/protein in yeast 3-6 (?) posttranslational modifications/protein in humans The number of bionodes is probably less than 10 times the number of genes Splice variants < > modules
The coverage of microarray chips and proteomics keeps increasing >>>> complete genome
Holland MJ. Transcript abundance in yeast varies over six orders of magnitude. J Biol Chem. 2002
Sensitivity : 2 copies/cell MOST transcripts are not seen by microarray
Please see J Biol Chem. 2002 Apr 26; 277(17): 14363-6. Epub 2002 Mar 06.
Transcript abundance in yeast varies over six orders of magnitude.
Holland MJ.
The utmost goal of technology :
Single copy/ single cell
BUT even if you measure everything accurately there
might be problems with predictions
Even a relatively simple set of ODEs can produce a
rather strange behavior.
Edward Lorenz 3 linked ODEs produced a behavior
very sensitive to the initial conditions.
(Chaos theory, Bifurcations etc.)
Small changes in the initial conditions can cause huge
changes at later time points
The problem of way too many correlated numbers:
Can this be
due to chance ?
-Analytical solution
- Computational solution: Permutate and look for similar patterns
In some cases analytical solution may exist Six breast cancer cell lines yielded 13 consistently mis-regulated genes (H-cadherin, S1002A, keratin 5 etc.) Can this be due to chance ?
"E" different cell lines "N"-gene microarray Mi genes mis-regulated in the "i"-th cell line, K consistently mis-regulated across all E cell lines. What is the probability that the K genes were mis-regulated
by chance ?
This translates into a simple combinatorics problem
BUT !!! - what if more genes are involved
Distribution of pair-wise correlation coefficients in cancer associated gene expression data
randomized
Real
The problem of way too many correlated numbers is a particularly nasty one.
Significance can be off by orders of magnitude when comparing
completely random permutations with "structural permutations"
Noise in discrete measurements: DNA sequences Measurement error: Sequencing errors (0.1%-1%) Solution: sequence a lot
AAATAACTCGGTGACCAAAAAAGAGTGTGAGGATAGATGTCA GAATGGTTGCTAAGGCACCTATTATTAGGTCGCTTATTAGTTTT CATGCCGTACATTGCACCTGGCAGACCTTGCCTTATTTCTCTGT ACATTTTTATTTTCCCGCGTGCTGCGCGGTGTTACACTGCGTTG TGTATTGCGCTGTGCACGGGGTCTGCGTAAGCGATGTTTTAGG GCACGGTTTGCTTCTAGAGTGGCCTCTCGCTCTTTTATTACCTCG CGCTTGTCAATTAGCTTTTTACCTCGCGCAAGGGATATAAGAA GCTTCGCGCGGCCGTTCCTGAAATAAAACTTGATGGGCACCAG GGTTATACCAGG........................
3 billion
-Find genes, introns, exons, transcription factor
binding sites etc.
Help can be found --- cDNA libraries etc. BUT 1) Yelin et al. Widespread occurrence of antisense transcription in the human genome. Nat Biotechnol. 2003:379-86. ~1600 ACTUALLY transcribed antisense transcriptional Units 2) Kapranov et al. Large-scale transcriptional activity in chromosomes 21 and 22. Science, 2002 As much as one order of magnitude more of the genomic sequence is transcribed than accounted for by the predicted and characterized exons.
TF binding site: TGGACT It can also be: TGCACT TGG/CACT
TCG/CNCT
Try to add constraints
1) Within 500 bp from the ATG
2) Tends to cluster in the same region
Even if you do all this you will find that many "obviously" TF binding site-looking sequences do not function as such. (due to higher level DNA organization etc.) AND You often do not know what sequence to start with.
1. Statistical overrepresentation
You define the rules 2. Cross-species conservation 3. Using artificial intelligence/Machine learning Hidden Markov models for exon/intron/gene identification (GENIE)
Please see Nature. 2003 May 15; 423(6937): 241-54. Sequencing and comparison of yeast species to identify genes and regulatory elements. Kellis M, Patterson N, Endrizzi M, Birren B, Lander ES.
S. cerevisiae S. bayanus S. mikatae S. paradoxus Number of genes ~ 5,500
High level of synteni
Courtesy of Eric Lander. Used with permission.
Slow and rapid evolution:
YBR184W 32% nucleotide and 13% aa identity
MATa2 - 100 % nucleotide and 100 % aa identity !!!!!!!!
Courtesy of Eric Lander. Used with permission.
XYZn(0-21)ABC
Intergenic conservation Intergenic vs. genic conservation Upstream vs. downstream conservation A given motif is also enriched in front of genes with similar function
Courtesy of Eric Lander. Used with permission.