In the new version (Version 3, launched on December 7, 2007) of dbs-pssm, we have made following changes:
1. Training data sets have been replaced by a new one with 160 NR Protein chains derived from Protein-DNA complexes.
2. Neural network complexity has been increased to include 8 neighbors or 17 residue window. 
	This was not permissible earlier due to a small data set and large network could have given 
	exaggerated performance. Larger data sets allow us to improve this.
3. Cross validation has been changed from 3-fold proteins-wise data sets to leave-third-pattern-out type.
4. Activation function has been changed from linear to ArcTan function for hidden and Sigmoid for output layer.
5. Overall protein amino acid composition has been added as additional descriptors. 
6. PSSM on the web version is now generated against NCBI NR data instead of smaller data sets used earlier.
7. Sequence size limit that can be submitted on the web is now 2000 residues instead of 200 used earlier.

As a result:
1. New sensitivity and specificity are 76% and 76% respectively. New
correlation value reaches 0.3983 
2. PSSM generation takes longer and a single query may take about 3-5 minutes.