Systems Biology for Personalised Medicine? Marc Wilkins Topics of - - PDF document

1

Systems Biology for Personalised Medicine?

Marc Wilkins Topics of this lecture:

1) What is systems biology

2) Systems biology and disease 3) Case study – patient classification by „network type‟ 4) Confounding factors for personalised medicine

2

What is Systems Biology?*

The study of an organism, viewed as an integrated and interacting network of genes, proteins and biochemical reactions, which together give rise to life. Systems Biology differs from previous biological approaches as it:

• does not focus on individual components
• focuses on the interactions of all components and

how they work as a system Systems biology has become possible through the rise of the „omics‟.

* summarised from www.systemsbiology.org

Why is Systems Biology Important?

There are properties of biology which cannot be discovered by analysing individual components. These are known as emergent properties. These are said to be irreducible (cannot be broken down).

This engine has many parts. The power and torque produced by the engine are emergent properties. Are there others? These would be difficult to discover without studying it as a system.

3

Systems Biology and Disease

Some diseases are monogenic

• mutations in a single gene lead to disease
• e.g. cystic fibrosis, retinoblastoma, Huntington‟s

BUT… Most human diseases are complex and/or polygenic

• many cancers
• diabetes
• cardiovascular disease

Many also have environmental influences. Study of genes and proteins as part of a system should help understand the basis of complex disease, and to personalise therapies.

Figure: Peltonen and McKusick (2001) Science 291, 1224- 29.

4

Systems Biology for Personalised Medicine

genomics, transcriptomics, proteomics, kinomics,

metabolomics, lipidomics, glycomics

PPI networks metabolic pathways signaling networks

patient classification? therapy?

Patient Classification by „Network Type‟

5

Sarah-Jane Schramm

Cancer cells show dysregulation, as compared to normal cells.

Following Taylor et al. (Nat. Biotech. 2009. 27:199-204) do networks of „bad outcome‟ melanoma (surviving < 1 year) show dysregulation compared to „good outcome‟ melanoma (surviving > 4 years)??

Does this reveal proteins of interest?

Prognostic or possible therapeutic value…

6

Some considerations…

• The correlation of expression of interacting proteins can

be measured.

• This can be done for hubs…. Why hubs…???

But…..

• Every expression set from every melanoma sample is

genetically different …

• Human network data is sparse and of medium to low

quality… there is no gold standard…

• How can we define a hub in a human cell?

Select 4 melanoma gene expression microarray datasets (REMARK compliant) Partition datasets by patient outcome 65 „good‟ outcome 93 „bad‟ outcome Prepare 4 human protein-protein interaction networks iRefWeb, BioGRID, MetaCore, HPRD Undertake hub analysis Identify hubs from networks using 2 approaches >5 interactions or top 15% of hubs

Data Preparation

7 Aim: identify hubs with patterns of gene co-expression that are different between good and bad outcome patients

Hub Analysis

good outcome bad outcome Correlation of expression between hub and partner Is the above hub of interest….. Why? Results Edges show correlation with hub. Experiment done 32 times, concordance measured.

HDAC1

Network topology preserved.

HDAC1

8

• 32 hubs showed dysregulation in bad versus good outcome

patients in 5 or more of the 32 experiments

• 21 of 32 hubs were of interest (REDISCOVERED) here

– 4 hubs are known correlates of melanoma prognosis

(CCNA2), progression (HIF1A), or tumour thickness (TNF and SMAD2)

– 9 hubs are already drug targets

(AKT1, HDAC1, HIF1A, IKBKB, JAK1, PIM1, PTPN11, TNF, and TGM2)

– 8 are causally implicated in other cancers

(AKT1, CIITA, CREBBP, FANCG, JAK1, NF2, PIM1, and PTPN11)

Results

Hanahan-Weinberg „cancer hallmarks‟: functional significance of the 32 hubs

9

Cohort:

Mann Bogunovic Jonsson John Sample size

(ngood outcome; npoor

• utcome)

47 (25;22) 33 (23;10) 54 (7;47) 24 (10;14) Outcomes compared

survival >4yr with no sign of relapse

• r <1yr after

surgical resection

• f stage III

disease survival ≥ 1.5yr

• r<1.5yr since

metastasis

• verall

survival time taken to tumor progression from stage III to stage IV disease ≥2yr or <2yr

Good / bad prediction error (LOOCV under KNN) 0.33 0.24 0.20 0.29

Classification of Patients by Hubs

32-hub expression signature, K-nearest neighbour classification

Prediction error by clinical parameters*

0.56

* tumor-positive lymph nodes, tumor burden at the time of staging, presence or absence of primary tumor ulceration, and thickness of the primary melanoma (Balch et al. 2009).

Conclusions…..

Networks showed reproducible, survival-associated differences Hubs with correlative differences were functionally relevant Hub expression signatures could classify patients (why?)

and Implications

Systems-based approaches have been successful in achieving one aspect

• f personalised medicine….

And suggesting a number of novel protein candidates of interest. But….. classification was not by network…. despite efforts to do so…..!!

10

Network Analysis Formalised: VAN: an R-package for identifying biologically perturbed networks via differential variability analysis. Is network dysregulation widespread?

• V. Jayaswal, SJ Schramm,

G Mann, MR Wilkins, J Yang. submitted.

Confounding Factors for Systems-based Personalised Medicine? Inter-cellular noise

11

Single Cell Proteomics: yeast

Newman et al. 2006 Nature 441: 840-6. GFP chromosome-based fusions of 4159 proteins, measured cell by cell with FACS Noise is related to protein function:

• high noise proteins include stress-

response, amino acid biosynthesis, and heat shock

• low noise proteins include translation

initiation, ribosomal and degradation Noise is related to localisation:

• High noise include mitochondria and

peroxisome proteins

• Low noise include Golgi proteins

Noise or Variance in Networks

Gene expression variance calculated from 270 yeast microarray experiments. A: proteasome regulatory lid B: mediator complex C: SAGA complex D: SWR1 complex

Data: Komurov & White 2007 Mol Syst Biol. 3: 110.

static dynamic

12

Single Cell Proteomics: human cell line

Cohen et al. (2008) Science 322, 1511-16. Response of human lung carcinoma cell line to camptothecin:

• Cell line double-transformed with mCherry

(nuclear) then YPF fusions

• 1020 proteins studied
• Video microscopy and image analysis

Proteins showed cell to cell variability:

• standard deviation of 10% to 60% of mean
• 20% of this variability due to cell-cycle
• remainder due to stochastic processes.

13

Relationship between average expression level in single cells (μ, x axis) and standard deviation (σ, y axis) for 6,313 genes.

“We find extensive, and previously unobserved, bimodal variation in messenger RNA abundance and splicing patterns, which we validate by RNA-fluorescence in situ hybridization for select transcripts… Moreover, splicing patterns demonstrate previously unobserved levels of heterogeneity between cells. ”

Confounding Factors for Systems-based Personalised Medicine? Inherited noise?

14

Gene expression data:

• B cells
• 233 individuals, 14 families,

3 generations per family PPA2, CDNK1A, CD44 show small to large family-associated differences in median and variance of expression Little, Williams, Wilkins (2009) Trends Biotech, 27: 5-10.

Cancer Biomarkers have Different „Normal‟ Ranges Approved and In-Clinic Biomarkers have Lower Inter-Individual Variation

1,261 „biomarker‟ genes. 9 are FDA approved. 32 in clinical use. Approved & clinical biomarkers have statistically lower inter-individual variation.

15

Final Comments

• Integration of networks with expression (or variance)

data is a powerful experimental paradigm

• Hubs can be a biologically relevant focus
• although be aware of biases
• Much is still to be learned about networks at different

levels

• per individual
• per tissue
• per cell
• Incompleteness of human networks

remains a significant constraint

Acknowledgements

Graham Mann Jean Yang Marc Wilkins

Sarah-Jane Schramm Anna Campain Vivek Jayaswal Richard Scolyer Apurv Goel Simone Li Chi Nam Ignatius Pang David Fung