From microarray images to Biological knowledge. Junior Barrera BIOINFO-USP DCC/IME-USP

Transcription

1 From microarray images to Biological knowledge Junior Barrera BIOINFO-USP DCC/IME-USP

2 Team Hugo A. Armelin Junior Barrera Helena Brentaini Marcel Brun Y. Chen Edward R. Dougherty Roberto M. Cesar Jr. Daniel Dantas Gustavo Esteves Marco D. Gubitoso Nina S. T. Hirata Roberto Hirata Jr Luiz F. Reis Paulo S. Silva Sandro de Souza Walter Trepode...

3 Layout Introduction Chip design Image analysis Normalization Genes Signature Clustering Genetic networks An environment for knowledge discovery

4 Introduction

5 Knowledge evolution in genetics Gene expression

6 Knowledge evolution in genetics

7 Data acquisition

8 Data acquisition

9 Analysis Phases G. Signature Measure Normal, Clustering Clustering Clustering G. R. Net.

10 Biological questions What genes are enough to separate a cancerous tissue from a normal one? What genes can separate to types of cancer? What genes differentiate two types of chickens or trees? What genes regulate a pathway?

11 Chip design

12 Clone Selection 1. Clustering for the same gene 5 Known gene 3 2. Choice of a clone representing the gene 5 3 Cuidado com sítios de Poliadenilação!!

13 Affect hybridization process Clone size - they should have similar size Clone position in the gene - they should come from similar regions Clone similarity - they should not have large similar regions

14 Hybridization cdnas Captura das Imagens Cy5 Misturas de cdnas marcados CY3 e CY5 Cy3

15 Image Analysis

16 A scanned image of a microarray slide

17 Processo de Segmentação de Imagens de Microarrays Hirata R, Barrera J, Hashimoto R, Dantas D, Esteves G. In press, 2002.

18 Raw data to the gene expression estimation step

19 Available solutions Scanalyze: usually doesn t t find misaligned spots. SpotFinder(TIGR): subarrays must be placed manually. Arrayvision: very good on locating misaligned spots; many options. UCSF Spot: : does everything automatically if the image is perfect. Quantarray,, F-scanF scan, Dapple, Genepix, Imagene etc. All of them require user interaction to some level.

20 Our aim... Is to reduce the user interaction, doing the job automatically and measuring correctly the relative mrna concentrations. This will make the process cheaper and faster. User interaction makes the segmentation subjective. Eliminating that, the results may be more reproducible.

21 Our software

22 Parameter setting In this window the user sets parameters for a whole family of arrays He can save in a file for reusing them

23 A vertical image profile is the sum of the spots values of each image line

24 The subarray gridding Is done by filtering the horizontal and vertical profiles

25 And finally taking the local minima of the filtered profile

26 the same is done with the horizontal profile. Here the result

27 Spots gridding is done separately for each subarray

28 The profile filtering is simpler having just one step, and also uses local minima

29 The spots detection step is basically the application of the Watershed operator

30 To avoid oversegmentation the image must be filtered

31 The filtered image also gives markers that will be used in Watershed

32 We give as input to the Watershed the markers, grid and the filtered image gradient

33 Here the resulting grid in white and spots cortours in light blue

34 Segmentation example

35 Segmentation example

36 Raw data to the gene expression estimation step The raw data of a spot consists on: the pixels values of both channels inside its rectangular region of interest which pixels belong to foreground or backround Foreground is the region with spotted cdna Background is the region without it.

37 Raw data to the gene expression estimation step

38 Gene expression estimation Is to find a value that represents the relative quantity of mrna in the two samples.

39 Some techniques to estimate gene expression Linear regression or least-squares squares fit of the values of pixels in the two channels.

40 Some techniques to estimate gene expression (ch1i-ch1b) ch1b) / (ch2i-ch2b) ch2b) where chxi is the estimated foreground intensity and chxb is the estimated backround intensity of channel X.

41 Some techniques to estimate gene expression To estimate chxi and chxb we can do: mean or median of all pixels in the foreground and background. mean or median of some percentiles in the foreground and background (fixed region method) mean or median of higher percentiles of all the pixels in the rectangle to estimate chxi and of lower percentiles to estimate the chxb.. Foreground and background information is ignored (histogram method)

42 Gene expression generation The program saves the expression data in a tab separated text file The file has the same format of the ones generated by ScanAlyze

43 Validation We made controlled experiments to test the expression estimation techniques. The objective of the experiment was to test how expression was affected by: position in the slide dilution of cdna length of mrna fragments being marked with cy3 or cy5

44 Validation We spotted microarrays with 32 blocks, each block with 6 genes x 5 dilutions x 2 repetitions + 4 landmarks = 64 spots We made six slides like this and, onto them, we poured six different mrna soups: Dilution gene 43A 43B 44A 44B 45A 45B Irf Trp ST IL Q Lys

45 Validation Here each point is the value of a spot obtained by the fixed region method. Spots from different dilutions are grouped. The black ones are from the three bigger mrna fragments, and the red, from the three smaller. sqrt( ( (ch1i A -ch1b A ) x (ch2i B -ch2b B ) ) sqrt( ( (ch2i A -ch2b A ) x (ch1i B -ch1b B ) )

46 Validation And here is the best result, obtained with the histogram method. sqrt( ( (ch1i A -ch1b A ) x (ch2i B -ch2b B ) ) sqrt( ( (ch2i A -ch2b A ) x (ch1i B -ch1b B ) )

47 Normalization

48 Normalization The expected expression of the gene IRF was 1.0 but the expression found was 1.6 This is due to the physical properties of the dyes.

49 Normalization When we have a single slide, we must eliminate the constant k assuming, when appropriate, that we can normalize all the spots using the expression of a housekeeping gene we can normalize assuming that the mean of normalized expression rates is one x = k (ch1i ch1b) (ch2i ch2b)

50 Normalization by swap Consists on eliminating the influence of the dyes properties by using two slides, and swapping the dye used to label the mrna sample. Use it if you find the single slide normalization hypotheses too strong.

51 Normalization by swap Better results can be achieved by doing swap experiments. x = k (ch1i A ch1b A ) = (ch2i B ch2b B ) 1 (ch2i A ch2b A ) (ch1i B ch1b B ) k

52 Normalization by swap Better results can be achieved by doing swap experiments. x = (ch1i (ch2i A A ch1b ch2b A A ) ) (ch2i (ch1i B B ch2b ch1b B B ) )

53 Genes Signature

54 Cancer tissue characterization Problem: from a small set (60) of microarrays, find a minimum number of genes that are enough to separate cancer A and B.

55 v erb b2 avian erythroblastic leukemia viral oncogene homolog LINEAR CLASSIFIER (DISPERSED GAUSSIAN) w/ σ = the tumor sample from Patient phosphofructokinase, platelet 7 th ( ) BRCA1 BRCA2/sporadic 1 th ( )

56 Approach randomize data compute classifier using genes subsets measure error for different dispersions choose the subset that balance small error and high dispersion. A supercomputer is required.

57 Linear classifier Dispersion centered in the sample Flat round dispersion model Error computed analytically (faster) X i [R n-1 ] 2 R h 2 r k R x k h X j [R 1 ] Misclassification error, ε n (h,r) Hyper-plane section: V n-1

58 Robustness analysis r k r j

59 Error curves under various dispersion levels, σ σ = 0.20 σ = 0.30 σ = 0.40 σ = 0.50 σ = 0.60 σ = 0.70 misclassification erorr, ε # of variables, d

60 Linear programming optimization

61

62

63

64 Steps The best linear classifier uses about genes Genes used are eliminated and the best linear classifier is computed, more genes are separated The procedure is repeated till having about 100 genes The full search is applied in the selected subset of genes

65 Separation

66 Validation Expression of chosen subsets of genes are measured several times in low cost experiments If the experiments reveal compact clusters the subset of genes chosen should be correct.

67 Clustering

68 Time course data 6 time course data C1 C2 C3 4 Clustered by dendrogram 2 Original clusters log2(ratio) 0 Dendrogram time KL plot multidimensional space different views of Microarray data KL plot mis-classified C1 C2 C

69 Example No error! Tighter clusters due to small variance Results from Fuzzy c-means

70 Gene Regulation Networks

71 Cell peptide other signals GENES NETWORK mrna Transcr. Translat. Proteins Pool Metabolic Pathways

72 Cell Cycle Modeling Division Steps x1fp... x1 z u1 y1fp x1f y1f I u2 u5 p w1fp vfp w2fp w1f v w2f y2fp w1 w2 x3f x4f y1f y2f y2f y1 y2 z Forward Signal Feedback to p Feedback to previous layer... x6fp x6f x6

73 Modeling Dynamical Systems

74 Simulator Architecture

75

76

77

78 Knockout x1fp... x1 Division Steps z u1 y1fp x1f y1f I u2 u5 p w1fp vfp w2fp w1f v w2f y2fp w1 w2 x3f x4f y1f y2f y2f y1 y2 z Forward Signal Feedback to p Feedback to previous layer... x6fp x6f x6

79

80

81 An environment for knowledge discovery

82 Objected oriented database P1 P2 Pn Pi : analytical and mining procedures (kernel parallel)

83 Integrated Environment Genbank database Another databases Access control module Clinical data Sequence module P1 microarray module P1 analitical operational P2 P3 analitical operational P2 P3 query Pn Pn query query

84 Users 1 Analysis tools writer SpreadSheet Dataminig 2 M_database MOLAP/ROLAP server 3 Metadata Data Warehouse Data Mart Integrator Wrapper Wrapper Wrapper 4 IMS RDMS Others

85 GRID Computer - DCC-IME-USP Internet and Intranet E4000 E4000:Databases and web application services (by SEFAZ) 10 processors 16 G-bytes main memory 1 T-bytes disk Processing services E3500 V880 V880?... (by CAGE-FAPESP) 4 processors 8 gigabytes (by Malaria-FAPESP) 16 processors 32 G-bytes (by eucalipto-cnpq???) 16 processors 32 G-bytes

86 Σ What genes regulate the pathway A->B->C->D? Wet Lab Proteome Transcriptome Genome Pathways