Prediction of siRNA functionality using generalized string kernel and support vector machine

Prediction of siRNA functionality using generalized string kernel and support vector machine

Abstract:

GSK+SVM

GSK:generalized string kernel

result : classify effective and ineffective siRNAs use GSK + SVM

Introduction

useful method before this paper:

(1),siRNA duplexes 21nt have a 2 nt 3’ overhang at each mRNA end

(2),a target sequence should be begining 50-100 downstream of the start codon of the mRNA

(3)，G/C 50%

Materials and methods

DATASET:

Dataset : Khvorova’s dataset , total 94 siRNAs , 53 siRNA effective , 41 siRNA ineffective

effective : with 90% or more gene silencing activity

ineffective : with less than 50% gene silencing activity

FEATURE map for siRNAs

GSK is based on mismatch string kernel(MSK) as well as in the spectrum kernel

$k$:the length of the sub-sequence of one string

$m$:at most m mismatches

MSK:

$$K_{(k,m)}(x,y)=<{\Phi }_{(k,m)}(x),{\Phi }_{(k,m)}(y)>$$

$$K_{(k,m)}(x,y)=\frac{K_{(k,m)}(x,y)}{\sqrt{K_{(k,m)}(x,x)}\sqrt{K_{(k,m)}(y,y)}}$$

GSK is a sum of all the $(k_i,m_i)$-mismatch kernels:

$$K_{k_1,m_1,...,k_s,m_s}=\sum _i<{\Phi }_{(k_i,m_i)}(x),{\Phi }_{(k_i,m_i)}(y)>=\sum _i{K}_{(k_i,m_i)}(x,y)$$

SVM implementation:

linear kernel and soft margin

RESULT

Subsequence->Weight

TP,TN,FP,FN,Acc

LOOCV of the GSK/SVM algorithm

Validation of predictive performance of GSK/SVM algorithm against other genes

Discussion

advantage:without a prior knowledge , we could determine contribution of each parameters to siRNA,and it can be applied to siRNAs shorter or longer than 21-mer in length
disadvantage:we can not deduce the sequence of the other useful siRNA