tides,

ܡොൌ܁ܟ൅܍ൌ۰ሺ܆ሻܟ൅܍

(3.46)

mple implementation of BBFNN is to use a maximum likelihood

of a logistic regression model to estimate w. The sigmoid

is used for squashing ܡො into the interval between zero and one,

ૈൌ

1

1 ൅݁^ିܡො

(3.47)

rnoulli likelihood function is generated based on the logit vector

ࣦൌ෍ߨ௡

௬೙ሺ1 െߨ௡ሻଵି௬೙

ே

௡ୀଵ

(3.48)

mising this likelihood can estimate w. Based on the estimated

rameters ܟෝ, which is the estimated version of w, a constructed

model can be used to scan a long polyprotein sequence to identify

protease cleavage sites for a specific protease or posttranslational

tion sites for a specific chemical.

NN has been implemented in different versions. For instance, a

BBFNN and an orthogonal kernel machine.

he Bayesian BBFNN algorithm

esian BBFNN is an extension to BBFNN. The aim of developing

sian BBFNN is for generating a more robust BBFNN modelling

m for protease cleavage peptide data analysis [Yang, 2005b].

an error vector is denoted by ܍ൌሺ݁ଵ, ݁ଶ, ⋯, ݁ேሻ^௧, a linear

ion of K bio-basis functions is defined as below,

ݕො௡ൌ෍ݓ௞ࣜሺܠ௡, ܛ௞ሻൌݕ௡െ݁௡

௄

௞ୀଵ

(3.49)