Exploring gene pathway interactions using SOM Keala Chan SoCalBSI August 20, 2004.

Exploring gene pathway interactions using SOM

Keala Chan

SoCalBSI

August 20, 2004

Microarray data analysis

Idea: Study relationships between functional terms or pathways

Gene expression data

Annotate and partition genes using functional terms

Interacting Gene Pathways

Hypothesis: Some relationship exists between Pathway 1 and Pathway 4

Network of pathways

Pathway 18

Pathway 4

Pathway 3

Pathway 2

Pathway 1Pathway 35

Pathway 12

Pathway 18

Pathway 4

Pathway 3

Pathway 2

Pathway 1Pathway 35

Pathway 12

Why use Self-Organizing Map?

• Serves as a data structure to represent the network

• Maps the network onto a 2-D grid, preserving the topological relationship between input vectors

(SOM)

Pathway 12Pathway 18

Pathway 1,

Pathway 2Pathway 4

Pathway 3Pathway 35

What is SOM?

• Tool for mapping similar input patterns onto contiguous locations in the output space

1. Clustering, or the creation of abstractions of the input space

2. Visualization of high-dimensional data in two-dimensional display

The SOM has two major effects:

Example

Each circle represents a number of input vectors. Hence, the input vectors have been clustered, or abstracted. Also, the topology has been preserved: neighboring representative vectors are similar.

Recall: SOM maps similar input patterns onto contiguous locations in the output space, resulting in clustering of the input space and 2-D visualization of the input space

Representative vectors

x

xx

x

x

x

xx

x

x

x

xxx

xx

x

x

x

x

x

The representative vector comes to represent this group of similar input vectors

The best-matching (closest) representative vector and its neighbors are pulled towards the highlighted input vector

2-D representative vector

Method

Partition genes into GO terms

Apply GSEAAffymetrix data

Recall: The general goal is to train a SOM on a large dataset to form a network of pathways for further study.

Data:

Human healthy tissue from 31 adult sources (brain, kidney, skin, etc…), 108 replicants

Baseline: average

Method (continued)

GSEA scores

Train SOM on the pathway dataset

GSEA scores normalized so

mean=0 and stdev=1

Visualizing first resultsThese terms all map to, or are represented by, the same hexagon.

Biological_Process_glycolysis_(10)

Molecular_Function_3-oxo-5-alpha-steroid_4-dehydrogenase_(4)

Molecular_Function_ATP-binding_cassette_(ABC)_transporter_(65)

Molecular_Function_blood_coagulation_factor_IX_(3)

Molecular_Function_blood_coagulation_factor_VII_(4)

Molecular_Function_blood_coagulation_factor_X_(3)

Molecular_Function_fructose-bisphosphate_aldolase_(9)

Molecular_Function_interleukin_receptor_(6)

Molecular_Function_pyruvate_kinase_(3)

Molecular_Function_sodium:phosphate_symporter_(5)

Molecular_Function_transaminase_(24)

These pathways are most activated in the liver

K-means clusteringk-means (15) clustering of the representative vectors groups pathways that are often activated at the same time

Next: Examine which k-means clusters are activated under each condition.

Projecting a new dataset To test for pathways that interact consistently, I projected GSEA scores for 16 different brain tumor types onto the SOM

Biological_Process_glycolysis_(10)

Molecular_Function_3-oxo-5-alpha-steroid_4-dehydrogenase_(4)

Molecular_Function_ATP-binding_cassette_(ABC)_transporter_(65)

Molecular_Function_blood_coagulation_factor_IX_(3)

Molecular_Function_blood_coagulation_factor_VII_(4)

Molecular_Function_blood_coagulation_factor_X_(3)

Molecular_Function_fructose-bisphosphate_aldolase_(9)

Molecular_Function_interleukin_receptor_(6)

Molecular_Function_pyruvate_kinase_(3)

Molecular_Function_sodium:phosphate_symporter_(5)

Molecular_Function_transaminase_(24)

Mapped pathways and GSEA scores to the same location in the SOM

Brain tumor dataQuestions to ask:

What is the best we can do with respect to the visual smoothness of the projection?

What characterizes a “good” projection?

Next: Plot histogram of distances between any two pathways mapping to the same hexagon.

Calculate activation scores for kmeans clusters trained on healthy data.

Fetal tissue

Next?

• Validation by biologists

• Choose parameters wisely (projection data, normalization, distance metric)

• Study k-means clustering of SOM

• More projections on SOM

Acknowledgments

• SOM Toolbox

• All BioDiscovery software

• Stan Nelson Lab microarray data

• Michael Sneddon

• Dr. Bruce Hoff

• Dr. Soheil Shams

• Everyone at SoCalBSI