ml-cancer-research
ml-cancer-research
2 0
ml-cancer-research
Talk on how machine learning helps cancer researchOn Github evelinag / ml-cancer-research
How machine learning helps cancer research
Evelina Gabasova
@evelgab
MRC Cancer Unit, University of Cambridge
DNA
DNA path – 2.5 km 10,000 colored stripes representing BRCA2 gene, one of approx. 30,000 genes in human 3 billion BPs Whole DNA – 20 times around the EarthCost of whole genome sequencing
The whole area is 15 years old Human genome project finished in 2003 Joint effort of 20 research centresSequencing data
It is projected that soon just the amount of sequencing information we’ll have will be larger than what data astronomy, particle physics processes Projected in a couple of years, genomics will have more data than Youtube (300 hours of video are uploaded to YouTube every minute) -> challenges – high dimensionality (DNA is HUGE)DNA and genes
DNA is stable storage Contains genes, each gene encodes a protein To synthetize a protein, we need to copy out a piece of DNA Explain what is gene expression – crutial part of analysing cellsCancer
- Genetic mutations
- Oncogenes and tumour suppressors
BRCA1 and BRCA2 are chromosome guardians
- Cancer is not a single disease
Clustering
Clustering is about finding groups in data, that are more similar within themselves and different across themselves Identifying types of cancer, because cancer is not a single disease Using gene expression to find subtypes of cancer, that behave differently in different patientsExample
Clustering wholesale customers
440 wholesale customers
Annual spending on
- Fresh produce
- Milk products
- Grocery products
- Frozen products
- Detergents and paper
- Delicatessen
Methods for clustering
- k-means clustering
- hierarchical clustering
- spectral clustering
- Gaussian mixture model and other probabilistic methods
- ...
Visualisation of high-dimensional data
Principal component analysis
Clustering cancer data
Genes instead of customers
Gene expression instead of spending on products
Why would we even want to cluster cancer data?Conventional medicine
Precision medicine
Clustering in cancer research
TCGA breast cancer Clustering 368 tumour samples based on expression of 648 genes.
Find different sybtypes of cancer [DEMO - cluster gene expression data from TCGA] Survival analysisIntegrative clustering
Integrative clustering
how single data in isolation don’t tell much Example of what I do Identify groups that exist among these different data types. Problems: cannot just put the data together, because they measure something different, have to devise more intelligent methodsCollaborative filtering
Example
The Netflix prize
User
Film 1
Film 2
Film 3
Film 4
...
Film 1000
Alice
5
2
x
x
...
4
Bob
x
1
2
x
...
2
Carol
2
x
3
?
...
3
...
...
...
...
...
...
...
Zoe
x
5
4
5
...
x
Matrix factorization
what methods are used to compute matrix factorisation?Collaborative filtering of cancer data
Patients instead of users
DNA mutations instead of film ratings
Mutational signatures
Patient
C/A
C/G
C/T
T/A
T/C
T/G
Alice
5
2
0
0
3
4
Bob
0
1
2
0
0
2
Carol
2
0
3
0
1
3
...
...
...
...
...
...
...
Zoe
0
5
4
5
2
0
C/T and T/C are associated with environmental damage different cancers have different signaturesMatrix factorization to identify features
Matrix factorization to identify features
Matrix factorization to identify features
Proving system stability
Theorem proving
SAT
(A ∨ ¬B ) ∧ (¬ A ∨ B)
A = true B = true
Theorem proving
Satisfiability Modulo Theories (SMT)
(A ∨ ¬B ) ∧ (¬ A ∨ B)
((a > 3) ∨ (b < 1)) ∧ ((a < 5) ∨ (b = 0))
a = 4 b = 0
Software verification
Z3 theorem prover
Preconditions Postconditions Loop conditions ➜ SMT formulas Assertions ...Software verification
Spec#
Software verification
Spec#
Proving stability of biological processes
Proteins
Genes ➜ Variables Receptors … ➜ v + 1 if v < T(v) v ➜ v if v = T(v) ➜ v - 1 if v > T(v) Traditional way to do this is via systems of differential equations DNA pathways - inhibition and activation Target function T - average effect of inhibitors and activators If we can get to the state where v = T(v) for all variables, the system stabilises Transform this into SMT formula and find a value of the state Iterate until the system stabilises of ends up in a cycleChronic myeloid leukemia
Proving stability of biological systems
Chronic myeloid leukemia
Model constructed from identified interactions from literature Relatively simple system because it’s caused by a single change when two chromosomes overlap and exchange parts The model correctly corresponds to the currently existing drug targets Allows testing what happens when we disable a receptor, or a specific protein Turns out the system is remarkably stable, needs disruption of at least two elements 54 nodes, 104 interactions, sourced from published literature (160 papers)Machine learning is not just
for targeted advertising
or algorithmic trading
Links
How machine learning helps cancer research
Evelina Gabasova
@evelgab
MRC Cancer Unit, University of Cambridge