Support vector machines, Kernel methods, and Applications in bioinformatics

Transcription

1 1 Support vector machines, Kernel methods, and Applications in bioinformatics Ecole des Mines de Paris Computational Biology group Machine Learning in Bioinformatics conference, October 17th, 2003, Brussels, Belgium

2 2 Overview 1. Support Vector Machines and kernel methods 2. Application: Protein remote homology detection 3. Application: Extracting pathway activity from gene expression data

3 3 Partie 1 Support Vector Machines (SVM) and Kernel Methods

4 The pattern recognition problem 4

5 5 The pattern recognition problem Learn from labelled examples a discrimination rule

6 6 The pattern recognition problem Learn from labelled examples a discrimination rule Use it to predict the class of new points

7 7 Pattern recognition examples Medical diagnosis (e.g., from microarrays) Drugability/activity of chemical compouds Gene function, structure, localization Protein interactions

8 8 Support Vector Machines for pattern recognition φ Object x represented by the vector Φ(x) (feature space)

9 9 Support Vector Machines for pattern recognition φ Object x represented by the vector Φ(x) (feature space) Linear separation in the feature space

10 10 Support Vector Machines for pattern recognition φ Object x represented by the vector Φ(x) (feature space) Linear separation with large margin in the feature space

11 Large margin separation 11

12 12 Large margin separation H

13 13 Large margin separation m H

14 14 Large margin separation m e1 e3 e2 e4 e5 H

15 15 Large margin separation m e1 e3 e2 e4 e5 H min H,m { 1 m 2 + C i e i }

16 16 Dual formulation The classification of a new point x is the sign of: f(x) = i α i K(x, x i ) + b, where α i solves: n max α i=1 α i 1 n 2 i,j=1 α iα j y i y j K(x i, x j ) i = 1,..., n 0 α i C n i=1 α iy i = 0 with the notation: K(x, x ) = Φ(x). Φ(x )

17 17 The kernel trick for SVM The separation can be found without knowing Φ(x). kernel matters: K(x, y) = Φ(x). Φ(y) Only the Simple kernels K(x, y) can correspond to complex Φ SVM work with any sort of data as soon as a kernel is defined

18 18 Kernel examples Linear : Polynomial : K(x, x ) = x.x K(x, x ) = (x.x + c) d Gaussian RBf : K(x, x ) = exp ( x x 2 ) 2σ 2

19 19 Kernels For any set X, a function K : X X R is a kernel iff: it is symetric : K(x, y) = K(y, x), it is positive semi-definite: a i a j K(x i, x j ) 0 i,j for all a i R and x i X

20 20 Advantages of SVM Works well on real-world applications Large dimensions, noise OK (?) Can be applied to any kind of data as soon as a kernel is available

21 21 Examples: SVM in bioinformatics Gene functional classification from microarry: Brown et al. (2000), Pavlidis et al. (2001) Tissue classification from microarray: Mukherje et al. (1999), Furey et al. (2000), Guyon et al. (2001) Protein family prediction from sequence: Jaakkoola et al. (1998) Protein secondary structure prediction: Hua et al. (2001) Protein subcellular localization prediction from sequence: Hua et al. (2001)

22 22 Kernel methods Let K(x, y) be a given kernel. Then is it possible to perform other linear algorithms implicitly in the feature space such as: Compute the distance between points Principal component analysis (PCA) Canonical correlation analysis (CCA)

23 23 Compute the distance between objects φ( g 1 ) d 0 φ( g 2 ) d(g 1, g 2 ) 2 = Φ(g 1 ) Φ(g 2 ) 2 ( = Φ(g1 ) Φ(g ) ( 2 ). Φ(g1 ) Φ(g ) 2 ) = Φ(g 1 ). Φ(g 1 ) + Φ(g 2 ). Φ(g 2 ) 2Φ(g 1 ). Φ(g 2 ) d(g 1, g 2 ) 2 = K(g 1, g 1 ) + K(g 2, g 2 ) 2K(g 1, g 2 )

24 24 Distance to the center of mass φ( g 1 ) m Center of mass: m = 1 N N i=1 Φ(g i ), hence: Φ(g 1 ) m 2 = Φ(g 1 ). Φ(g 1 ) 2 Φ(g 1 ). m + m. m = K(g 1, g 1 ) 2 N N i=1 K(g 1, g i ) + 1 N 2 N i,j=1 K(g i, g j )

25 25 Principal component analysis PC2 PC1 It is equivalent to find the eigenvectors of ( K = Φ(gi ). Φ(g ) j ) i,j=1...n = ( K(g i, g j ) ) i,j=1...n Useful to project the objects on small-dimensional spaces (feature extraction).

26 26 Canonical correlation analysis CCA2 CCA1 CCA2 CCA1 K 1 and K 2 are two kernels for the same objects. CCA can be performed by solving the following generalized eigenvalue problem: ( ) ( ) 0 K1 K 2 K 2 ξ = ρ 1 0 ξ K 2 K K2 2 Useful to find correlations between different representations of the same objects (ex: genes,...)

27 27 Part 2 Local alignment kernel for strings (with S. Hiroto, N. Ueda, T. Akutsu, preprint 2003)

28 28 Motivations Develop a kernel for strings adapted to protein / DNA sequences Several methods have been adopted in bioinformatics to measure the similarity between sequences... but are not valid kernels How to mimic them?

29 29 Related work Spectrum kernel (Leslie et al.): K(x 1... x m, y 1... y n ) = m k i=1 n k j=1 δ(x i... x i+k, y j... y j+k ).

30 29 Related work Spectrum kernel (Leslie et al.): K(x 1... x m, y 1... y n ) = m k i=1 n k j=1 δ(x i... x i+k, y j... y j+k ). Fisher kernel (Jaakkola et al.): given a statistical model ( pθ, θ Θ R d) : φ(x) = θ log p θ (x) and use the Fisher information matrix.

31 30 Local alignment For two strings x and y, a local alignment π with gaps is: ABCD EF G HI JKL MNO EEPQRGS I TUVWX The score is: s(x, y, π) = s(e, E) + s(f, F ) + s(g, G) + s(i, I) s(gaps)

32 31 Smith-Waterman (SW) score SW (x, y) = max s(x, y, π) π Π(x,y) Computed by dynamic programming Not a kernel in general

33 32 Convolution kernels (Haussler 99) Let K 1 and K 2 be two kernels for strings Their convolution is the following valid kernel: K 1 K 2 (x, y) = K 1 (x 1, y 1 )K 2 (x 2, y 2 ) x 1 x 2 =x,y 1 y 2 =y

34 33 3 basic kernels For the unaligned parts: K 0 (x, y) = 1.

35 33 3 basic kernels For the unaligned parts: K 0 (x, y) = 1. For aligned residues: K (β) a (x, y) = { 0 if x 1 or y 1, exp (βs(x, y)) otherwise

36 33 3 basic kernels For the unaligned parts: K 0 (x, y) = 1. For aligned residues: K (β) a (x, y) = { 0 if x 1 or y 1, exp (βs(x, y)) otherwise For gaps: K g (β) (x, y) = exp [β (g( x ) + g( y ))]

37 34 Combining the kernels Detecting local alignments of exactly n residues: K (β) (n) (x, y) = K 0 ( K (β) a ) (n 1) K g (β) K (β) a K 0.

38 34 Combining the kernels Detecting local alignments of exactly n residues: K (β) (n) (x, y) = K 0 ( K (β) a ) (n 1) K g (β) K (β) a K 0. Considering all possible local alignments: K (β) LA = i=0 K (β) (i).

39 35 Properties K (β) LA (x, y) = π Π(x,y) exp (βs(x, y, π)),

40 35 Properties K (β) LA (x, y) = π Π(x,y) exp (βs(x, y, π)), lim β + 1 β ln K(β) LA (x, y) = SW (x, y).

41 36 Kernel computation e X0 d X d X2 B M E d Y0 Y Y2 e

42 37 Application: remote homology detection Unrelated proteins Twilight zone Close homologs Sequence similarity Same structure/function but sequence diverged Remote homology can not be found by direct sequence similarity

43 38 SCOP database SCOP Fold Superfamily Family Remote homologs Close homologs

44 39 A benchmark experiment Can we predict the superfamily of a domain if we have not seen any member of its family before?

45 39 A benchmark experiment Can we predict the superfamily of a domain if we have not seen any member of its family before? During learning: remove a family and learn the difference between the superfamily and the rest

46 39 A benchmark experiment Can we predict the superfamily of a domain if we have not seen any member of its family before? During learning: remove a family and learn the difference between the superfamily and the rest Then, use the model to test each domain of the family removed

47 40 SCOP superfamily recognition benchmark No. of families with given performance SVM-LA SVM-pairwise SVM-Mismatch SVM-Fisher ROC50

48 41 Part 3 Detecting pathway activity from microarray data (with M. Kanehisa, ECCB 2003)

49 42 Genes encode proteins which can catalyse chemical reations Nicotinamide Mononucleotide Adenylyltransferase With Bound Nad+

50 43 Chemical reactions are often parts of pathways From

51 44 Microarray technology monitors mrna quantity (From Spellman et al., 1998)

52 45 Comparing gene expression and pathway databases VS Detect active pathways? Denoise expression data? Denoise pathway database? Find new pathways? Are there correlations?

53 46 A useful first step and g1 g8 g5 g6 g2 g7 g3 g4 g6 g1 g3 g8 g2 g7 g5 g4

54 47 Using microarray only PC1 g1 g8 g5 g6 g2 g7 g3 g4 PCA finds the directions (profiles) explaining the largest amount of variations among expression profiles.

55 48 PCA formulation Let f v (i) be the projection of the i-th profile onto v. The amount of variation captured by f v is: h 1 (v) = N i=1 f v (i) 2 PCA finds an orthonormal basis by solving successively: max v h 1 (v)

56 49 Issues with PCA PCA is useful if there is a small number of strong signal In concrete applications, we observe a noisy superposition of many events Using a prior knowledge of metabolic networks can help denoising the information detected by PCA

57 50 The metabolic gene network GAL10 Glucose HKA, HKB, GLK1 Glucose 6P PGT1 Fructose 6P PFK1,PFK2 Fructose 1,6P2 FBA1 Glucose 1P PGM1, PGM2 FBP1 HKA1 PGM1 PFK1 GAL10 HKA2 PGT1 FBP1 FBA1 GLK1 PGM2 PFK2 Link two genes when they can catalyze two successive reactions

58 51 Mapping f v to the metabolic gene network 0.8 g1 0.7g2 0.4 g3 g g8 g7+0.5 g6 g g1 g8 g5 g6 g2 g7 g3 g Does it look interesting or not?

59 52 Important hypothesis If v is related to a metabolic activity, then f v should vary smoothly on the graph Smooth Rugged

60 53 Graph Laplacian L = D A 1 2 L =

61 54 Smoothness quantification h 2 (f) = f exp( βl)f f f is large when f is smooth h(f) = 2.5 h(f) = 34.2

62 55 Motivation For a candidate profile v, h 1 (f v ) is large when v captures a lot of natural variation among profiles h 2 (f v ) is large when f v is smooth on the graph Try to maximize both terms in the same time

63 56 Problem reformulation Find a function f v and a function f 2 such that: h 1 (f v ) be large h 2 (f 2 ) be large corr(f v, f 2 ) be large by solving: max (f v,f 2 ) corr(f v, f 2 ) h 1(f v ) h 1 (f v ) + δ h 2(f 2 ) h 2 (f 2 ) + δ

64 57 Solving the problem This formultation is equivalent to a generalized form of CCA (Kernel-CCA, Bach and Jordan, 2002), which is solved by the following generalized eigenvector problem ( 0 K1 K 2 K 2 K 1 0 ) ( α β ) ( K 2 = ρ 1 + δk K2 2 + δk 2 ) ( α β ) where [K 1 ] i,j = e i e j and K 2 = exp( L). Then, f v = K 1 α and f 2 = K 2 β.

65 58 The kernel point of view... g1 g2 g3 g4 g8 g7 g5 g6 g1 g8 g5 g6 g2 g7 g3 g4 Diffusion kernel Linear kernel g7 g3 g4 g1 g2 g5 g8 g6 Kernel CCA g2 g4 g1 g3 g6 g8 g7 g5

66 59 Data Gene network: two genes are linked if the catalyze successive reactions in the KEGG database (669 yeast genes) Expression profiles: 18 time series measures for the 6,000 genes of yeast, during two cell cycles

67 60 First pattern of expression Expression Time

68 61 Related metabolic pathways 50 genes with highest s 2 s 1 belong to: Oxidative phosphorylation (10 genes) Citrate cycle (7) Purine metabolism (6) Glycerolipid metabolism (6) Sulfur metabolism (5) Selenoaminoacid metabolism (4), etc...

69 Related genes 62

70 Related genes 63

71 Related genes 64

72 65 Opposite pattern Expression Time

73 66 Related genes RNA polymerase (11 genes) Pyrimidine metabolism (10) Aminoacyl-tRNA biosynthesis (7) Urea cycle and metabolism of amino groups (3) Oxidative phosphorlation (3) ATP synthesis(3), etc...

74 Related genes 67

75 Related genes 68

76 Related genes 69

77 70 Second pattern Expression Time

78 71 Extensions Can be used to extract features from expression profiles (preprint 2002) Can be generalized to more than 2 datasets and other kernels Can be used to extract clusters of genes (e.g., operon detection, ISMB 03 with Y. Yamanishi, A. Nakaya and M. Kanehisa)

79 Conclusion 72

80 73 Conclusion Kernels offer a versatile framework to represent biological data SVM and kernel methods work well on real-life problems, in particular in high dimension and with noise Encouraging results on real-world applications Many opportunities in developping kernels for particular applications