Real Value Prediction of Solvent Accessibility of Proteins with additional graphical ouput.

On this server, we provide Real predictions of ASA using neural networks. No calssification into ASA states is carried out for these predictions. These predictions are made by a neural network, which has been trained on more than 80,000 residues from proteins with known three-dimensional sturctures. All the residues from the Barton set, mentioned in our algorithm have been included for training this online prediction server. Mean Absolute Error, defined as per residue difference between the predicted and experimental values of ASA for this data is less than 19%. Other important points of these predictions are:
1. Common sequences formats such as FASTA, SwissProt and PIR are accepted. Sequence can also be entered as a single letter code without any header. Spaces and carriage returns will be ignored in such a case.
2. Largest sequence size is 700. Part of the sequence larger than 700 will be ignored during prediction.
3. Linux binaries of this program to run as standalone are available from Shandar Ahmad.

Reference: Shandar Ahmad and M. Michael Gromiha, Akinori Sarai, Bioinformatics 19 (2003) 1849-1851