I

– 6

8

– 2

T

– 4

– 2

7

he bio-basis function neural network algorithm

basis function neural network (BBFNN) has been developed

n this idea, i.e., mapping a non-numeric peptide space to a

l space which is supported by the homology alignment scores

peptides using an amino acid mutation matrix [Thomson, et al.,

ng and Berry, 2004; Berry, et al., 2004; Yang and Chou, 2004b;

gse, et al., 2005; Yang, et al., 2005; Yang, 2005; Yang, 2005b;

05c; Yang and Thomson, 2005; Sidhu and Yang, 2006, Yang, et

; Yang and Hammer, 2007; Maji and Das, 2010; Maji and Das,

ose a peptide is denoted by ܠ௡ and its functional status is denoted

0,1ሽ, which is either functional (cleaved) or non-functional (non-

Suppose a set of cleaved peptides is denoted by Ω^ା and a set of

ved peptide is denoted by Ω^ି. Such a protease cleavage pattern

y problem is a discrimination problem. It is well-understood that

aved peptides will show the amino acid composition trend.

r, a basis used in a basis function neural network mode is served

pporting coordinate in the kernel space. Such a supporting

e must be composed of rich information. Non-cleaved peptides

inly non-informative. Therefore, only cleaved peptides are

d as bio-bases in BBFNN. Suppose a bio-basis, which is a cleaved

s denoted by ܛ௞. The similarity between ܠ௡ and ܛ௞ is denoted by

ሻ, which stands for the non-gap alignment score between ܠ௡ and

calculated using a mutation matrix.

hus assumed that a cleaved peptide (ܠ௡∈Ω^ା) will show a high

y score (similarity) with some bio-basis peptides because they

the same category and should show a similar biological property,

mino acid composition pattern or trend for a protease to recognise