bioinformática aplicada al estudio del control de la expresión de genes en el cerebro humano

Juan A. Bo*a Ins-tute of Neurology, University College London, UK Facultad de Informá-ca, Universidad de Murcia, Spain

Algorithmic Approaches for the construc3on of gene co-‐expression networks from control brain 3ssue samples mRNA

RNA-‐seq Substan-a nigra and Putamen brain co-‐expression networks on the UKBEC project to study Parkinson’s Disease

03/04/17 Conferencias de Inves-gación para Posgrado, Fac. Informátca, UCM 2

The central dogma of biology

source Wikipedia

We use pre-‐mRNA


Chapter I. The dataset


Braineacv2, RNA-‐seq based, focused on Parkinson’s Disease

l  Affects 1% to 2% of the population older than 65 years l  Symptons: resting tremor, bradykinesia, rigidity and impairment in ability

to initiate and sustain movements l  The hallmark of this disease is the progressive loss of dopaminergic

neurons, mainly in the substantia nigra


Chapter II. The computa-onal model


Network analysis: aprioris-c versus free approaches


Are networks something more than a fancy graph and nice plots?

Yes they are!! • Can be used to iden-fy the ac-ve pathways in specific samples (cases vs. controls)

• Describe subsystems (i.e. cell types) •  Iden-fy candidate genes (GBA)


To create networks we need to es-mate links between genes


From gene expression to gene co-‐expression networks

TREM2 forms a receptor signaling complex with TYROBP, which triggers the ac-va-on of immune responses in macrophages and dendri-c cells, and the func-onal polymorphism of TREM2 is r e p o r t e d t o b e a s s o c i a t e d w i t h neurodegenera-ve disorders such as Alzheimer’s disease (AD).




TYROBP TREM2 0.76


But before reaching that

•  Scale free topology assump-on – The degree distribu-on p(k) of a network follows a power law so p(k) ~ k-‐ϒ

– Evidence supports this for many organisms (ϒ is approx. 2.2)


But before reaching that (& 2)

•  Modularity assump-on – Varia-on coefficient of organisms, Ci = ︎ 2n/ki(ki – 1) with n number of direct links connec-ng the ki nearest neighbours of i-‐th node, suggests strong modular organiza-on

– Evidence suggests the coefficient of varia-on is higher than expected in SFT networks


But before reaching that (& 3)

•  Hierarchies solve this apparent dilemma


Chapter III. The problem


Our main focus: Parkison's Disease l  Affects 1% to 2% of the population older than 65 years l  Symptons: resting tremor, bradykinesia, rigidity and impairment in ability

to initiate and sustain movements l  The hallmark of this disease is the progressive loss of dopaminergic

neurons, mainly in the substantia nigra excitatory

inhibitory

Substantia Nigra Pars Compacta

Brain regions most typically affected by adult-onset disease

03/04/17

Conferencias de Inves-gación para Posgrado, Fac. Informátca, UCM 18

Step 1: RPKM exonic gene quantification and CQN normalization Step 2: RPKM-CQN > 0.2 & missingness < 70% Step 3: Data correcting for Sex, Age and 7/8 Peer axes Step 4: WGCNA “signed” network construction Step 5: k-Means optimization of module partitions Step 6: Network assessment Step 7: Within tissue and between tissues subsystem characterization

33670 Ensembl genes

Approx. 19K genes, two datasets

Two corrected datasets

SNIG and PUTM networks And gene modules assignment

Modified gene modules assignment for SNIG and

PUTM Quality metrics for networks and

Gene partitions

Functional characterization, correlation with traits, gene

function prediction

Steps on the pipeline Outcomes


Co-expression analysis methodology


A measure of similarity between genes, values in [0,1]

From similarity to adjacency, hard thresholding

From similarity to adjacency, sou thresholding

From adjacency to TOM values

03/04/17


From TOM values to clusters by 1-‐TOM as a distance

complete linkage hierarchical approach for clustering summarisa-on based on eigenvalue

03/04/17


l  Hierarchical clustering's results are highly variable depending on linkage (max/complete, min/single, average linkages)

l  Module membership (MM) of g is the correlation of g and

the 1st PC of gene expression (module eigengene)

l  This doesn't necessarily mean all genes are in the best module according to MM

l  Previous approaches based on reassigning some/all genes l  k-means algorithm helps finding a better partition in which

genes are (hopefully) assigned to a module in a more natural way

Why do we need an optimization process for WGCNA

03/04/17


A k-‐means heuris-c How does it work?


Outline of the op-miza-on

Accepted in BCM Systems Biology 03/04/17


Chapter IV. The results


What we get from the optimization •  More accurate par--on construc-on •  Bever func-on annota-on for modules •  Bever cell markers enrichment •  More preserved modules across similar -ssues

03/04/17


How to assess the accuracy of a co-expression network

cluster driven validation

data driven validation by replication

Are the gene groups good according to a

given index

same tissue similar tissue

same network model

diff. network model

Biology: Does my module

make sense?

functional characterization

03/04/17






given index


same network model

diff. network model


make sense?


03/04/17


Replication in GTEx GNAT networks for Substantia Nigra

lightgreenmidnightblue

cyantan

turquoisegrey60

lightyellowgreen

pinkblue

magentapurpleyellow

redblack

lightcyanbrown

salmongreenyellow

Mantel fold SNIG GTEx coexpression within0.

0

0.5

1.0

1.5

2.0

2.5

3.0

*** 340*** 412

*** 449*** 385

*** 574** 295

*** 250*** 427*** 457*** 783*** 505

*** 417*** 579

*** 521 244

* 410 260 88

372

redpurple

magentaturquoise

blueyellow

lightyellowcyan

lightcyantan

greengrey60

midnightbluelightgreen

pinkbrown

greenyellowblack

salmon

Mantel fold SNIG microarray binary between

0.0

0.5

1.0

1.5

2.0

*** 701*** 624

*** 760*** 837

*** 1070*** 837*** 365*** 624*** 653*** 460

*** 658 477 475

417 743 406 579 402

149

03/04/17


Replication in GTEx GNAT networks for Putamen

lightcyan

grey60

yellow

salmon

greenyellow

pink

green

black

tan

brown

purple

magenta

turquoise

lightgreen

midnightblue

blue

cyan

Mantel fold PUTM GTEx coexpression within

0.0

0.5

1.0

1.5

2.0

2.5

3.0

*** 429

*** 275

* 72

*** 372

*** 268

*** 444

*** 486

*** 484

*** 541

*** 611

*** 574

*** 617

*** 546

*** 440

*** 461

** 386

*** 759

greenyellow

salmon

lightcyan

brown

green

pink

grey60

cyan

tan

magenta

black

lightgreen

purple

turquoise

midnightblue

blue

yellow

Mantel fold PUTM GTEx binary between0.

0

0.5

1.0

1.5

2.0

2.5

*** 268

*** 372

*** 429

*** 611

*** 486

*** 444

*** 275

*** 759

*** 541

*** 617

*** 484

*** 440

*** 574

*** 546

*** 461

*** 386

72

03/04/17






given index


same network model

diff. network model


make sense?


03/04/17


Asignment of biological func3on to modules with gProfiler

•  Based on GO (BP, MF, CC) and gProfileR •  Fisher's exact test and Bonferroni corrected p-‐values •  What should we expect?

•  Normal cell processes like respira-on, cell development, immune func-on

•  But also brain related terms (hopefully movement disorders, signalling) in some of the modules

•  What should we consider when looking for enrichment? •  GO is not a closed world ontology •  Something not found doesn't imply it doesn't exist

•  Genes can play new roles •  Groups of genes can have new func-ons •  It is possible to find modules with no GO and s-ll be valid


Significant similarities in practically all modules

This is a tabular View of significant agreements (Fisher's Exact test) on genes between modules from the two tissues


Subsystems


Subsystems cell type & function

Neuron cells, Synapse/NADH

Microglia cells, Immune system

Nucleus, transcription

Neuron, astrocytes & microglia cell types Response to stimulus

Endothelial cell type, Cell division

Oligodendrocytes cell type, synapse & ion transport

Mitochondrion

Cytosolic rybosome

Ubiqutin 03/04/17 Conferencias de Inves-gación para Posgrado, Fac. Informátca, UCM 42

Lessons learned l  The default WGCNA can be improved to get more

coherent gene groups l  Network analysis reveals

l  cell specific subsystems in putamen and substantia nigra

l  Interesting differences between the two tissues at the subsystem level

Ongoing work l  Models to explain the differences between subsystems l  Function prediction for non coding species and intergenic

regions


03/04/17

Acknowledgements

University College London

Jana Vandrovcova Sebastian Guelfi Karishma D'sha John Hardy Mar Matarin Daniah Trabzuni

King's College London

Mike Weale Mina Ryten Paola Forabosco Adai Ramasamy


bioinformática aplicada al estudio del control de la expresión de genes en el cerebro humano

Engineering