We de
scribe improvement
s of the cur
rently mo
st popular method for prediction of cla
ssically
secreted protein
s, SignalP. SignalP con
si
st
s of two diffe
rent predictor
s ba
sed on neural network and hidden Markov model algorithm
s, where both component
s have been updated. Motivated by the idea that the cleavage
site po
sition and the amino acid compo
sition of the
signal peptide are correlated, new feature
s have been included a
s input to the neural network. Thi
s addition, combined with a thorough error-correction of a new data
set, have improved the performance of the predictor
significantly over SignalP ver
sion 2. In ver
sion 3, correctne
ss of the cleavage
site prediction
s ha
s increa
sed notably for all three organi
sm group
s, eukaryote
s, Gram-negative and Gram-po
sitive bacteria. The accuracy of cleavage
site prediction ha
s increa
sed in the range 6&nda
sh;17 % over the previou
s ver
sion, wherea
s the
signal peptide di
scrimination improvement i
s mainly due to the elimination of fal
se-po
sitive prediction
s, a
s well a
s the introduction of a new di
scrimination
score for the neural network. The new method ha
s been benchmarked again
st other available method
s. Prediction
s can be made at the publicly available web
server
s.dtu.dk/services/SignalP/"">http://www.cbs.dtu.dk/services/SignalP/