1. Computing Co-Expression Relationships Wen-Dar Lin

2. Contents Motivation Basic Idea Case Studies –An Example of Single Experiment –An Example of Time-Course Experiment Potential Applications Availability Future Works

3. Motivation Given a set of differentially displayed genes that are reported by an array experiment. –We would like to know relationships among these genes. –These relationships may recover important modules or motifs with respect to the experiment.

4. Motivation Co-expression relationships are one kind of the most biologically meaningful and easily computable relationships. –Co-expression relationships form modules that may infer important biological information. –They can be computed from a large amount of publicly available array data.

5. Basic Idea Array data can be retrieved from publicly available data repository –like the NASCarrays, NCBI GEO, EMBL-EBI ArrayExpress They should be normalized before computing the co-expression relationships. –e.g. normalized by the RMA method

6. Basic Idea Defining co-expression relationships –We define that a co- expression relationship between two genes exists if the pearson correlation coefficient between their normalized expression levels is greater than or equal to a certain threshold. slide #1234 … gene X12103 … gene Y52124 … X Y

7. Basic Idea Properties of pearson correlation coefficient –Let Correl(A, B) be the pearson correlation coefficient between normalized expression levels of gene A and gene B. –0   Correl(A, B)   1 from http://www.gseis.ucla.edu/courses/ed230bc1/notes1/var1.html negative correlation

8. Basic Idea The computational assistance –Given a set of interested genes –Compute co-expression relationships among them –Identify co-expression clusters

9. Case Studies We have implemented aforementioned ideas into a tool kit and applied it to two case studies. –A single experiment –A time-course experiment

10. A Single Experiment In this example, an array experiment was performed –178 differentially displayed genes were identified. –Based on RMA array data of 300 ATH1 slides downloaded from the NASCarrays sample of each slide was derived nonexclusively from roots Threshold for pearson correlation coefficient = 0.7

11. A Single Experiment Two larger clusters One minor subcluster

12. A Single Experiment We may compute co-expression relationships based on all kinds of array experiment data –Based on RMA array data of 1436 ATH1 slides downloaded from the TAIR, co-expression relationships were identified Threshold for pearson correlation coefficient = 0.7

13. A Single Experiment Two larger clusters

14. A Single Experiment Is there any difference between the graphs based on root-array data and that based on all-array data? –By differentially marking clusters of one graph onto the other graph.

15. A Single Experiment Two clusters mapped by the other graph One cluster that should be root-specific

16. A Single Experiment Cluster sizes: 47 & 14 Cluster size: 9

17. A Single Experiment Some remarks –The number of differentially displayed genes reported by the experiment is 178 –The number of clustered genes is 47+14+9 = 70 Reduced by more than 50% –The co-expression relationships are recovered Each cluster may be a module that usually work together. –Finding tissue-specific co-expression relationships Can be done by mapping the graph based on all-array data onto the graph based on tissue-related-array data.

18. A Single Experiment In addition to cluster genes according to co-expression relationships, we also fished genes that may potentially co-expressed. –These genes may not be identified as differentially displayed in the experiment.

19. A Single Experiment A GO enrichment analysis was also carried out –using the GOBU software (gobu.iis.sinica.edu.tw) –which should give a conceptual view of clustered genes.

20. A time-course experiment In this example, a time-course array experiment was performed –Three time points –About 800 genes differentially displayed at least one time point. –Based on array data of 300 ATH1 slides extracted from RMA array data of about 2600 ATH1 slides downloaded from the NASCarrays Threshold for pearson correlation coefficient = 0.8

21. A time-course experiment Time point 1 About 100 genes

22. A time-course experiment Time point 2 About 100 genes

23. A time-course experiment Time point 3 About 100 genes

24. A time-course experiment Though this clustering and time-course expression data shows some biological meaning, –this size of clustered genes (more than 200) makes the graph too complex and is too large to be realized in a short time.

25. A time-course experiment Reducing the size of clustered genes may help –reducing complexity of the graph and –realizing revealed co-expression module We reduced the graph by removing co-expression relationships that generally exist in the entire plant –based on RMA array data of about 2600 ATH1 slides downloaded from the NASCarrays –Threshold for pearson correlation coefficient = 0.7

26. A time-course experiment Edges (relationships) to be removed Y root-related others X

27. A time-course experiment Edges (relationships) to be retained Y root-related others X

28. A time-course experiment About 20 genes About 60 genes About 50 genes Time point 1

29. A time-course experiment About 20 genes About 60 genes About 50 genes Time point 2

30. A time-course experiment About 20 genes About 60 genes About 50 genes Time point 3

31. A time-course experiment Some remarks –The number of differentially displayed genes at least one time point is about 800. –The number of clustered genes is about 60+50+20 = 130 Reduced by more than 80% –The retained graph contains edges, i.e., gene pairs, that are co-expressed in root but not in the entire plant The recovered clusters should be root specific.

32. Potential Applications We have created a tool kit that –computes co-expression relationships based on array data where probe names can be replaced by aliases made by something like orthologous mapping can be used for studying non-model organism using array data of a model organism.

33. Potential Applications We have created a tool kit that –fills colors according to graphs by intensity fold-changes, or clusters in another graph

34. Potential Applications We have created a tool kit that –removes/retains co-expression relationships in another graph –finds specific or common co-expression relationships 200 genes120 genes

35. Potential Applications We have created a tool kit that –fishes genes that are potentially co-expressed with assigned bait

36. Future Works Incorporate pathway database –like the AraCyc –for finding relationships between co-expression clusters and known pathways A user-friendly interface which would –facilitate using this tool kit and –help manage output data

37. Availability The tool kit is now an open-source project –http://maccu.sourceforge.net –Project name: MACCU Multi-Array Correlation Computation Utility –A detailed description of each program module has been created. –A running script with example is provided.

38. Special Thanks I would like to thank –Drs. Chang (Bill), Schmidt & Wu for raising this idea, the initial implementation, and valuable comments.

39. Thank you!