NetChop 2.0 Prediction Server

The NetChop server produces neural network predictions for cleavage sites of the human proteasome.

NetChop has been trained on human data only, and will therefore presumably have better performance for prediction of the cleavage sites of the human proteasome. However, since the proteasome structure is quite conserved, we believe that the server is able to produce reliable predictions for at least the other mammalian proteasomes.

At the moment there are four different networks that can be used for predictions: C-term 1.0, C-term 2.0, N-term, and 20S. The different networks are trained on different data sets. C-term 1.0 network is trained with 229 MHC class I ligands extracted from public databases (using only C-terminal cleavages). N-term network is trained with the same data set, but by using N-terminal cleavages. C-term 2.0 network is trained with a larger database consisting of 1110 publicly available MHC class I ligands (using only C-terminal cleavage site of the ligands). 20S network is trained with in vitro degradation data published in Toes, et al. and Emmerich et al. We believe C-term 2.0 network performs best in predicting the boundaries of CTL epitopes. It is possible to get the averaged prediction from different networks.

4-3-2004: There had been some discussion about whether or not C-term 1.0 and C-term 2.0 networks can predict proteasome specificity, because they are trained using MHC ligands. In a recent article Saxova et al., we have shown that these networks can predict proteasome cleavage sites identified by in vitro experiments much better than other methods available at the moment. Still, if you feel uncomfortable using C-term networks, you can use 20S network. Our 20S network has a very good performance in predicting in vitro proteasomal cleavage.

Another proteasome prediction server is available in Tubingen University: PAProc