02.03.2013 Views

Target Discovery and Validation Reviews and Protocols

Target Discovery and Validation Reviews and Protocols

Target Discovery and Validation Reviews and Protocols

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

44 Imoto et al.<br />

β k (k = 1,…,q j ) are parameters <strong>and</strong> ε ij (i = 1,…,n) are independently <strong>and</strong> normally<br />

distributed with mean 0 <strong>and</strong> variance σ j 2 . This model assumes the relationships<br />

between genes are linear, <strong>and</strong> it is unsuitable to extract effective<br />

information from the data with complex structure. To capture even nonlinear<br />

dependencies, Imoto et al. (9) proposed the use of the nonparametric additive<br />

regression model (19) of the form<br />

where m jk ( . ) (k =1,…,q j ) are smooth functions from R to R. We construct m jk ( . )<br />

by the basis function expansion method with B-splines (20,21):<br />

where are parameters, is the prescribed<br />

set of B-splines, <strong>and</strong> Mjk is the number of B-splines. We then have the<br />

Bayesian network <strong>and</strong> nonparametric regression model<br />

f j (xij | pa(X j ) i ,θ j ) =<br />

( j ) ( j ) (j )<br />

1 x − γ<br />

ij ∑ bsk ( pik<br />

{ ∑<br />

)<br />

k s sk } exp −<br />

2<br />

2πσ j<br />

2<br />

( j )<br />

γ sk (s = 1,…,M jk )<br />

⎡<br />

⎤<br />

⎢<br />

⎥<br />

⎢<br />

2<br />

⎥.<br />

⎢<br />

2σ j<br />

⎥<br />

⎣<br />

⎦<br />

Note that the Bayesian network model based on the linear regression is<br />

included in this model as a special case. To extend from the additive regression<br />

model to a general regression, see Note 7.<br />

Let the prior distribution on the parameter be specified by the hyperparameter<br />

<strong>and</strong> let log p( | ) = O(n), then the Laplace approximation for integrals<br />

(22–24) gives an analytical solution for computing p(D|G) as follows:<br />

where<br />

( j) ( j ) ( j) ( j) t<br />

xij =β0 +β1p +⋅⋅⋅+βqj i1 piq +εij , where pa(X<br />

j<br />

j ) i = (pi1 ,…, piqj ) ,<br />

( j ) ( j )<br />

xij = m j1 (pi 1 ) +⋅⋅⋅+m (p jqj iq )+εij ,<br />

j<br />

p(D |G) = (2π / n) r /2 | J λ ( ˆ θ | D)| –1/2 exp{nl λ( ˆ<br />

θ | D)}{1+O p (n −1 )},<br />

l λ (θ | D) = 1<br />

n<br />

M jk<br />

∑<br />

m jk (p) = γ sk<br />

s =1<br />

{log p(D | θ, G)+ log p(θ | λ)},<br />

J λ (θ | D) = –<br />

( j ) ( j )<br />

bsk (p),<br />

∂ 2<br />

∂θ∂θ t l λ<br />

( j) ( j)<br />

{b1k (⋅),⋅⋅⋅⋅⋅bM<br />

jk k (⋅)}<br />

(θ |D),

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!